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
1501.06009	2	1501	2015-02-06T19:58:17	Are Effective Leaders Creative?	[ "cs.MA" ]	This paper explains in layperson's terms how an agent-based model was used to investigate the widely held belief that creativity is an important component of effective leadership. Creative leadership was found to increase the mean fitness of cultural outputs across an artificial society, but the more creative the followers were, the greater the extent to which the beneficial effect of creative leadership was washed out. Early in a run when the fitness of ideas was low, a form of leadership that entails the highest possible degree of creativity was best for the mean fitness of outputs across the society. As the mean fitness of outputs increased a transition inevitably occurs after which point a less creative style of leadership proved most effective. Implications of these findings are discussed.	cs.MA	cs	Gabora, L. (2010). Are effective leaders creative? Psychology Today (online). https://www.psychologytoday.com/blog/mindbloggling/201102/are-effective-leaders-creative Are Effective Leaders Creative? Okanagan Campus, Arts Building, 333 University Way, Kelowna BC, V1V 1V7, CANADA Department of Psychology, University of British Columbia Liane Gabora Abstract This paper explains in layperson’s terms how an agent-based model was used to investigate the widely held belief that creativity is an important component of effective leadership. Creative leadership was found to increase the mean fitness of cultural outputs across an artificial society, but the more creative the followers were, the greater the extent to which the beneficial effect of creative leadership was washed out. Early in a run when the fitness of ideas was low, a form of leadership that entails the highest possible degree of creativity was best for the mean fitness of outputs across the society. As the mean fitness of outputs increased a transition inevitably occurs after which point a less creative style of leadership proved most effective. Implications of these findings are discussed. It is widely assumed that creativity is an important component of effective leadership. (Bellows, 2004, Basadur, 2004; Puccio, Murdock, & Mance, 2006; Simon, 1986; Sternberg, Kaufman, & Pretz, 2003). This paper explains in layperson’s terms how a computer model was used to examine the relationship between creativity and leadership. Although the results we obtained made sense when we thought about them, there were some interesting surprises. First we studied how leadership affects the effectiveness and diversity of ideas in a society. Second we studied to what extent creativity is desirable in a leader. I'll begin by telling you a bit about the computer model itself, and then explain the experiments. THE COMPUTER MODEL The current model's predecessor was called Meme and Variations or MAV (Gabora, 1995). Its name is a pun on the musical form, ‘theme and variations'. MAV was the earliest computer program to model culture as an evolutionary process in its own right. MAV was inspired by the genetic algorithm (GA), a search technique that finds solutions to complex problems by generating a 'population' of candidate solutions (through processes akin to mutation and recombination), selecting the best, and repeating until a 1 satisfactory solution is found. The computer model is composed of an artificial society of agents in a two-dimensional grid-cell world. Agents consist of (1) a neural network, which encodes ideas for actions and detects trends in what constitutes an effective action, and (2) a body, which implements their ideas as actions. The agents can do two things: (1) invent ideas for new actions, and (2) imitate their neighbors' actions. The computer model enables us to investigate what happens to the diversity and effectiveness of actions in the artificial society over successive rounds (called ‘iterations') of invention and imitation. Since the ideas in the model are ideas for actions, diversity is measured by counting how many different actions are being implemented by the agents. Evolution in the biological sense is not taking place; the agents neither die nor have offspring. But evolution in the cultural sense is taking place through the generating and sharing of ideas for actions amongst agents, which over time leads to more effective actions. In MAV, all agents were equally capable of both inventing and imitating. In the latest version of the computer model called EVOC (for EVOlution of Culture), it is possible to vary how likely an agent is to invent versus imitate. A TYPICAL RUN Each iteration, every agent has the opportunity to (1) acquire an idea for a new action, either by imitation, copying a neighbor, or by invention, creating one anew, (2) update their knowledge about what constitutes an effective action, and (3) implement a new action. Effectiveness of actions starts out low because initially all agents are just standing still doing nothing. Soon some agent invents an action that has a higher effectiveness than doing nothing, and this action gets imitated, so effectiveness increases. Effectiveness increases further as other ideas get invented, assessed, implemented as actions, and spread through imitation. The diversity of actions initially increases due to the proliferation of new ideas, and then decreases as agents hone in on the fittest actions. Thus MAV successfully models how 'descent with modification' can occur in a cultural context. LEADERSHIP EXPERIMENTS These experiments made use of EVOC's ‘broadcasting function'. This enables the actions of a particular agent, the leader or ‘broadcaster' to be imitated by not just its immediate neighbors (as is normally the case) but any other agent. Thus broadcasting enables the action implemented by a leader to be visible to all the other agents in the artificial society, referred to as followers. In these experiments, societies consisted of one leader and ninety-nine followers. The leader was chosen randomly and broadcasted throughout the 100-iteration run. In a first set of simulations, the leader was no more or less creative than the followers. We found that the presence of a leader accelerates convergence on optimal ideas, but does so at the cost of consistently reducing the diversity of ideas (Gabora, 2008a,b). In other words, although they find optimal solutions faster, they end up finding fewer of them. This echoes previous simulation findings that when agents can communicate or exchange ideas, leadership can have adverse effects (Gigliotta, Miglino, & Parisi, 2007). The result suggests that in a fast-changing world where diversity of ideas is beneficial because what is effective today may not be effective tomorrow, it may be particularly 2 important to watch out for situations in which leaders pull individuals off their own creative paths. The goal of the next set of experiments was to investigate how creative versus uncreative leadership affects the effectiveness and diversity of ideas, and how creative leadership is affected by how creative the followers are (Leijnen & Gabora, 2010). There are two ways a leader can be creative in EVOC. The first way has to do with how OFTEN the leader invents; that is, the ratio of iterations it spent inventing versus imitating. When an agent's frequency of invention is at the maximum of 1.0, it invents a new action every iteration. When the frequency of invention is at the minimum of 0.0, the agent never invents new actions; it only imitates its neighbors' actions. We tried many possibilities ranging between these two extremes, for both leader and followers. What we found was that when the followers are uncreative, the degree of creativity of the leader matters a lot; the effectiveness of ideas across the society as a whole is positively correlated with the frequency of invention of the leader. However, the more creative the followers are, the greater the extent to which the beneficial effect of creative leadership is washed out. When the followers themselves are creative, the degree of creativity of the leader has almost no impact; in this case, the ideas generated by the society increase over the duration of a run at more or less the same pace, no matter how creative the leader is. The results suggest that creativity may be an important quality for a manager of a relatively uncreative team, but not such an important quality for a manager of a creative team. We then wanted to know whether the decreased diversity associated with the presence of a leader is still observed when leaders are highly creative or highly uncreative compared to followers. We found that while in the early stages of a run, creative leadership is associated with higher cultural diversity, eventually all agents converge on what the leader is doing no matter how creative the leader (Leijnen & Gabora, 2010). That is, in the long run, leadership diminishes cultural diversity regardless of how creative the leader is. Yet another set of experiments investigated the effect of not how often the leader invents, but how creative the leader's inventions are; that is, the extent to which a newly invented idea differs from its predecessor idea. It turned out that the optimal degree of creative leadership with respect to this second measure of creativity depends on how far along the society is. Early on in a run, when the fitness of ideas is still low, a form of leadership that entails the highest possible degree of creativity (highest rate of change per new idea) is ideal. However, this situation changes as the run progresses, and eventually a transition occurs, after which point a much lower rate of change per idea (approximately 40%) is best. Although once again one must be cautious about extrapolating from the results of such simulations to the real world, this result suggests that a new start-up company benefits most from highly creative leadership, while a more established company, or one that has stabilized on an established product line, may benefit most from a more conservative form of leadership. One must be cautious about extrapolating from a simple simulation such as this to the real world. For example, real-world creativity is correlated with emotional instability, affective disorders, and substance abuse (Andreason, 1987; Flaherty, 2005; Jamieson, 1993) which presumably would interfere with effective leadership, and which were not incorporated in these simulations. Moreover, the agents’ neural networks are so small 3 that creative novelty is generated does not involve noticing and refining new kinds of connections the way it happens in real minds (Gabora, 2000). Finally, it is also worth noting though that in this artificial world, unlike the real world, agents had only one task to accomplish. Further experiments will investigate whether this result hold true when there are multiple tasks to be accomplished. However, the results of these computer simulations are provocative, and inspire new ways of thinking about the relationship between creativity and leadership. They suggest that the relationship between leadership and creativity is more complex than previously thought. REFERENCES Andreason, N. C. (1987). Creativity and mental illness; prevalence rates in writers and their first degree relatives. American Journal of Psychiatry, 144, 1288-1292. Basadur, M. (2004). Leading others to think innovatively together: Creative leadership. The Leadership Quarterly, 15(1), 103-121. Bellows, R. M. (1959). Creative leadership. Upper Saddle River, NJ: Prentice-Hall. Flaherty, A. W. (2005). Frontotemporal and dopaminergic control of idea generation and creative drive. Journal of Comparative Neurology, 493, 147-153. Gabora, L. (2000). Toward a theory of creative inklings. In (R. Ascott, Ed.) Art, Technology, and Consciousness (pp. 159-164). Intellect Press, Bristol, UK. Gabora, L. (1995). Meme and variations: A computer model of cultural evolution. In (L. Nadel & D. Stein, Eds.) 1993 Lectures in Complex Systems, Addison-Wesley, 471-486. http://www.vub.ac.be/CLEA/liane/papers/mav.htm Gabora, L. (2008a). EVOC: A computer model of cultural evolution. In V. Sloutsky, B. Love & K. McRae (Eds.), Proceedings of the 30th Annual Meeting of the Cognitive Science Society. North Salt Lake, UT: Sheridan Publishing. (Held July 23-26, Washington DC) http://www.vub.ac.be/CLEA/liane/papers/evoc.pdf Gabora, L. (2008b). Modeling cultural dynamics. Proceedings of the Association for the Advancement of Artificial Intelligence (AAAI) Fall Symposium 1: Adaptive Agents in a Cultural Context,. Nov 7-9, The Westin Arlington Gateway, Arlington VA, (pp. 18-25). Menlo Park, CA: AAAI Press. http://www.vub.ac.be/CLEA/liane/papers/AAAI08FS01Gabora.pdf Gigliotta, O. Miglino, O. & Parisi, D. (2007). Groups of agents with a leader. Journal of Artificial Societies and Social Simulation, 10(4). http://jasss.soc.surrey.ac.uk/10/4/1.html Jamieson, K. R. (1993). Touched by fire: Manic-depressive illness and the artistic temperament. New York: Free Press. Leijnen, S. & Gabora, L. (2010). An agent-based simulation of the effectiveness of creative leadership. Proceedings of the Annual Meeting of the Cognitive Science Society (pp. 955-960). Austin TX: Cognitive Science Society. (Held August 11-14, 2010, Portland, Oregon.) https://people.ok.ubc.ca/lgaboraa/papers/conf_papers/ 4 Cogsci2010LeijnenGabora.pdf Puccio, G. J., Murdock, M. & Mance, M. (2006). Creative leadership: skills that drive change. Thousand Oaks, CA: Sage. Simon, H. A. (1986). What we know about the creative process. In R. Kuhn (Ed.) Frontiers in creativity and innovative management. Cambridge, MA: Ballinger. Sternberg, R. J., Kaufman, J. C. & Pretz, J. E. (2003). A propulsion model of creative leadership. The Leadership Quarterly, 14(4-5), 455-473. 5
1604.04730	1	1604	2016-04-16T11:53:46	Evolutionary-aided negotiation model for bilateral bargaining in Ambient Intelligence domains with complex utility functions	[ "cs.MA", "cs.AI", "cs.GT" ]	Ambient Intelligence aims to offer personalized services and easier ways of interaction between people and systems. Since several users and systems may coexist in these environments, it is quite possible that entities with opposing preferences need to cooperate to reach their respective goals. Automated negotiation is pointed as one of the mechanisms that may provide a solution to this kind of problems. In this article, a multi-issue bilateral bargaining model for Ambient Intelligence domains is presented where it is assumed that agents have computational bounded resources and do not know their opponents' preferences. The main goal of this work is to provide negotiation models that obtain efficient agreements while maintaining the computational cost low. A niching genetic algorithm is used before the negotiation process to sample one's own utility function (self-sampling). During the negotiation process, genetic operators are applied over the opponent's and one's own offers in order to sample new offers that are interesting for both parties. Results show that the proposed model is capable of outperforming similarity heuristics which only sample before the negotiation process and of obtaining similar results to similarity heuristics which have access to all of the possible offers.	cs.MA	cs	Evolutionary-aided negotiation model for bilateral bargaining in Ambient Intelligence domains with complex utility functions V´ıctor S´anchez-Anguixa, Soledad Valeroa, Vicente Juli´ana, Vicente Bottia, Ana Garc´ıa-Fornesa aUniversidad Polit´ecnica de Valencia, Departamento de Sistemas Inform´aticos y Computaci´on, Cam´ı de Vera s/n, Valencia, Spain, 46022 Abstract Ambient Intelligence aims to offer personalized services and easier ways of in- teraction between people and systems. Since several users and systems may coexist in these environments, it is quite possible that entities with opposing preferences need to cooperate to reach their respective goals. Automated ne- gotiation is pointed as one of the mechanisms that may provide a solution to this kind of problems. In this article, a multi-issue bilateral bargaining model for Ambient Intelligence domains is presented where it is assumed that agents have computational bounded resources and do not know their opponents' pref- erences. The main goal of this work is to provide negotiation models that obtain efficient agreements while maintaining the computational cost low. A niching genetic algorithm is used before the negotiation process to sample one's own utility function (self-sampling). During the negotiation process, genetic oper- ators are applied over the opponent's and one's own offers in order to sample new offers that are interesting for both parties. Results show that the proposed model is capable of outperforming similarity heuristics which only sample before the negotiation process and of obtaining similar results to similarity heuristics which have access to all of the possible offers. Keywords: Automated Negotiation, Bilateral Bargaining, Agreement Technologies, Evolutionary Computation, Multi-agent Systems 1. Introduction Nowadays, the number of computational devices present in our everyday life has grown considerably. The use of technology looks to help us achieve a better quality of life, to make our life easier and more comfortable. However, due to the increasing number of devices, it is necessary that the technology itself adapts to the needs of the user, instead of the human being the one that adapts to technology. In that sense, Ambient Intelligence (AmI) tries to cover that it looks to offer personalized services and provide users with easier necessity: Preprint submitted to Elsevier July 31, 2021 and more efficient ways to communicate and interact with other people and systems [4, 35]. Agent technology has been appointed as a proper technology for the support of AmI solutions [4, 10, 31]. In fact, agents show interesting characteristics for AmI environments since they are reactive, proactive and social [36]. Firstly, reactiveness allows agents to change their behavior according to some new con- ditions in the AmI environment (new users, new services, etc.). Secondly, proac- tiveness makes it possible for agents to act autonomously according to the user's goals, which results in a smooth and non-intrusive interaction with the AmI user. And lastly, the agent's social behavior allows several heterogeneous entities to cooperate and offer new complex services to the AmI user. Over the last few years, researchers in the area of agent technology have shown a growing interest in automated negotiation. Negotiation can be defined as a process in which a joint decision is made by two or more parties. The parties first verbalize contradictory demands and then move towards agreement by a process of concession-making or search for new alternatives [29]. Therefore, au- tomated negotiation consists in such a joint decision being automatically decided by means of autonomous entities (e.g., agents representing different users). The parties participating in a negotiation process have opposing preferences, thus negotiation can be considered as a conflict resolution mechanism. Such conflict circumstances are not alien to AmI applications. For instance, shopping malls may be converted into ubiquitous environments where several vendors offer their products to passing shoppers [16, 2]. In many cases, the shoppers know what they want but do not have time to check every shop that offers such products. A possible way of enhancing the customer experience is to automatically negotiate with all of the vendors. A list with the best agreements may be presented to the user through his mobile device. This way, the user does not have to check every possible shop since his mobile device has negotiated with every shop taking into account the user preferences. Nevertheless, there are also benefits for the vendors since automated negotiation allows a more flexible commerce than classic e-commerce. For instance, they may negotiate issues such as price, payment method, discounts, and dispatch dates, which is what often happens in traditional non-electronic commerce. Flexibility in e-commerce may result in client loyalty since the vendor is able to adapt as much as possible to the client preferences. Therefore, automated negotiation is a proper technology for e-commerce-based AmI applications such as shopping malls. The process of negotiation has been traditionally studied by the field of Game Theory [26, 34], providing solutions that reach optimal results under different criteria (e.g., Pareto efficiency, Nash Product, etc.). However, such solutions require unbounded computational resources that are not available in most real applications. In such cases, the research area of artificial intelligence (AI) has tried to provide a solution by means of heuristics that achieve results that are as close as possible to the optima [18, 14]. Artificial intelligence has tradition- ally studied multi-issue negotiations where utility functions are represented as a linear combination of the issues involved in the negotiation process [5, 6, 9, 15]. In linear utility functions, issue values are usually monotonic, so these functions 2 usually have a single global optimum and consequently, the utility function is easy to optimize. Nevertheless, most real world problems are hardly modeled by linear utility functions since they have a higher degree of complexity than that offered by linear utility functions (e.g. e-commerce [17, 33, 13] ). Some of the issues in the negotiation setting may present interdependence relationships. Thus, the value of the utility function may be drastically changed by the pos- itive/negative synergy of interdependent issues. The result is that the utility function is no longer linear, and there may, therefore, be several local optima. Optimizing non-linear utility functions is hard by itself (e.g. it may require non-linear optimizers such as simulated annealing, genetic algorithms, etc.), as is learning opponent preferences and looking for good agreements. Utility func- tions that have the trait of being non-linear are usually known in the literature as complex utility functions. Over the last few years, there has been an effort to research negotiation strategies that are capable of working with such complex utility functions where issues may have relationships of interdependence. Works in these complex do- mains have focused on negotiation strategies that require a mediator [17, 13, 22, 23], or non-mediated strategies that are devised for very specific utility functions [33]. However, non-mediated strategies are more interesting from the point of view of AmI environments due to the fact that users enter and leave the sys- tem in an extremely dynamic way. Thus, it may be difficult to find a trusted mediator for every possible user. Although non-mediated strategies are more interesting from the point of view of different domains, there has been a lack of work in non-mediated strategies for complex utility functions. The work of Lai et al. [20] presents a non-mediated strategy for general utility functions, which obviously includes complex utility functions. The strategy is based on the cal- culation of current iso-utility curves and a similarity heuristic that sends offers from the current iso-utility curve that are the most similar to the last offers received from the opponent. However, the entire calculation of the iso-utility curve may require an exhaustive exploration of the utility function, which may not be tractable in the case of a large number of issues. Furthermore, if the ex- ploration of one's own utility function is not performed in an intelligent way, the result may be that most of the offers sampled are of no use for the negotiation process since they might not interest the opponent. Mechanisms that sample as few offers as possible are needed, especially for environments where devices may have limited computational resources as is the case with AmI environments. In this work, a non-mediated bilateral multi-issue negotiation model for AmI environments is presented. Its main goal is to optimize the computational re- sources while maintaining a good performance in the negotiation process. The proposed model is inspired by the seminal work of Lai et al. [20]. The three main differences between this present work and the work of Lai et al. are: (i) The present approach assumes that it is not possible to exhaustively search the utility function. Before the negotiation process starts, each agent samples its own utility function by means of a niching genetic algorithm (GA) [12, 21]. The effect of this sampling is that offers obtained are highly fit and significantly dif- ferent;(ii) A few additional samples are obtained during the negotiation process 3 by means of genetic operators that are applied over received offers and one's own offers. The heuristic behind this sampling is that offers obtained by ge- netic operators have genetic material from one's own agent and the opponent's offers. Thus, these new offers may be interesting for both parties. (iii) Genetic operators act as a learning mechanism that implicitly guides the offer sampling and selection of which offers must be sent to the opponent. Results show that the proposed work outperforms similarity heuristics that are able to sample the same number of offers before the negotiation process starts. Additionally, it is also shown how the proposed strategy is capable of achieving similar results to those of similarity heuristics that sample the entire utility function with far fewer samples. This result is accomplished due to the learning mechanism provided by genetic algorithms. This paper is organized as follows: section 2 describes an example of appli- cation where automated negotiation and ambient intelligence can be combined in order to offer a useful service for the user; section 3 describes the negoti- ation model, explaining the chosen protocol and the new negotiation strategy in detail. In Section 4, the experimental setting and the results obtained are discussed. In Section 5 related work is discussed. Finally, the conclusions and future lines of work are explained in Section 6. 2. An Example of Automated Negotiation and Ambient Intelligence Synergy: Product Fairs In this section we introduce an example of application where automated negotiation may be used along with well-known AmI techonologies in order to provide a profitable service for users. The example is focused on product fairs. Fairs are public events where manufacturers/sellers/producers exhibit their products to a wide range of consumers who go from small consumers to big retailers. At this kind of events there are usually a large number of exhibitors and products. Therefore, it is extremely difficult to explore the whole fair or find interesting deals for one's interests. It is also difficult for sellers to attract interesting clients. Thus, both consumers and sellers would be benefited by a tool which allows them to attract/search prospective deals quickly. At this point automated negotiation in an AmI environment may come in handy. Let us suppose the following scenario at a fair: each vendor has been assigned a booth where he attends to clients. As well as setting up the typical equipment, a hardware device with Bluetooth wireless communication is pro- vided (e.g. a PC). An agent, which can be downloaded and configured by the vendor prior to the fair event day, is installed in this hardware device, and it complies with rules and communication protocols established by fair organiz- ers. These agents should be provided with information regarding its owner's preferences by means of user modeling methods such as questionnaires, past experiences, and so forth. Additionally, consumers are allowed to download an agent to their mobile devices prior to the fair event. The only requirement for the mobile device is 4 Bluetooth wireless capabilities. The consumer's agent follows the communica- tion protocols established by the fair organizers and can be configured similarly to the vendor's agent. More specifically, the agent may ask what its owner would be interested in buying and general questions about the preferences regarding the possible negotiation attributes. When consumers and vendors enter the fair, they should start the execution of their respective agents. Each consumer's agent offers a negotiation service which can be invoked by vendors' agents. Whenever this service is invoked by a vendor agent, a negotiation process starts between the vendor agent and the consumer agent. The negotiation process continues until a deal has been found or one of the parties has decided to leave. If the deal is considered as interesting by both parties (i.e. utility of the deal higher than a certain threshold or reservation utility) and the deal is among the best ones for the consumer in that specific area (determined by which vendors can be reached by Bluetooth in that space point), the consumer agent and the vendor agent notify their respective owners regarding the possible deal. However, deals discovered by this automatic process are not to be considered as binding but as recommendations. If the deal is considered as interesting enough by the consumer, it may result in the consumer approaching the vendor's booth. At that point, both parties may decide to renegotiate or polish the deal which has been found by their agents. Since Bluetooth technology has coverage limitations, the service can usually only be discovered by vendor agents that are nearby. Therefore, negotiation processes help consumer and vendor agents to find prospective deals as con- sumers walk around the fair. More specifically, it allows consumers to save physical time by recommending them the vendors that seem more suitable for their needs in the area. That way, they only approach vendors in an area who may have an interesting deal for them. Indirectly, it also helps vendors since it attracts consumers with high probabilities of buying their goods instead of losing time with clients with whom the possibilities of making a deal are very low. 3. Negotiation Model Negotiation models are composed of a negotiation protocol and a negotiation strategy. On the one hand, the negotiation protocol defines the communication rules to be followed by the agents that participate in the negotiation process. More specifically, it states at which moments the different agents are allowed to send messages and which kind of messages the agents are allowed to send. For instance, the Rubinstein alternating protocol specifies [27] that agents are allowed to send one offer in alternating turns. Basically, the negotiation protocol acts as a mechanism for the coordination and regulation of the agents that take part in the negotiation process. On the other hand, the negotiation strategy defines the different decisions that the agent will make at each step of the negotiation process. It includes the opponent's offers acceptance rule, the selection of which offers are to be sent to the opponent, the concession strategy, the decision of whether the agent 5 should continue in the negotiation process or not, and so forth. Therefore, the negotiation strategy includes all the decision-making mechanisms that are involved in the negotiation process. The negotiation protocol used can be categorized as an alternating proto- col for bilateral bargaining [27]. More specifically, the protocol used is the k-alternating protocol proposed by Lai et al. [20]. The proposed negotiation strategy belongs to the family of negotiation strategies that use a similarity heuristic in order to propose new offers to the opponent [6, 20]. 3.1. Negotiation Protocol As mentioned above, the negotiation protocol belongs to the family of alter- nating protocols for bilateral bargaining. In this kind of protocols, two different agents negotiate without the need of a mediator. As previously stated, non- mediated strategies are more adequate for AmI applications since users enter and leave the AmI system in a very dynamic way. Thus, it may not be feasible to find a trusted mediator for every possible pair of agents. Furthermore, in some AmI domains such as shopping malls, where there are different competing vendors and lots of potential users, it is difficult to determine who will mediate the negotiation process. The protocol used is the k-alternating protocol proposed by Lai et al. [20]. This protocol is composed of several rounds where the agents exchange offers in an alternating way. One of the agents, called the initiator, is responsible for starting the current round. He can accept one of the previous offers received from the opponent in the last round, exit from the negotiation process, or send up to k different offers to the opponent agent. Once the initiator has performed one of the possible actions, the opponent agent is able to accept one of the offers he has just received, exit from the negotiation process or propose up to k different offers to the initiator. Then, the round ends and a new round is initiated by the initiator agent. The negotiation process ends when one of the agents accepts an offer (the negotiation succeeded) or one of the agents decides to abandon the negotiation (the negotiation failed). Some of the properties of the k-alternating protocol proposed by Lai et al. are: • The protocol is adequate for situations where both agents are equal in power (e.g. none of them has the monopoly over a resource). • Each agent is capable of sending up to k different offers, making it more probable that one of the proposed offers satisfies the requirements of the opponent agent. • Since k different offers are proposed in each agent's turn more informa- tion about opponent preferences can be inferred, increasing the chances of finding a win-win situation. This may produce faster agreements, which is inherently interesting for every domain but particularly for AmI do- mains since it may reduce the number of messages exchanged and thus the bandwidth consumption. 6 Figure 1: An example of two agents in the k-alternating protocol proposed by Lai et al. [20] An example of two agents negotiating with a 3-alternating protocol (k = 3) can be observed in Figure 1. Agent A is the initiator of the negotiation round, whereas Agent B is the responding agent. The first round starts with 3 offers proposed by the initiator. Once the offers reach Agent B, he decides whether he should accept one of them or not. Since the 3 offers are not interesting for Agent B, he decides to counteroffer 3 different offers. When the 3 first offers from Agent B reach Agent A, the second round starts. Due to the fact that none of the offers proposed by Agent B are of interest to the initiator, he decides to send 2 offers. The 2 offers from the initiator reach Agent B, who analyzes the offers in order to determine whether they are interesting. Since he found OfferA4 to be interesting, he decides to accept it and the protocol thus ends with an agreement. 3.2. Negotiation Strategy The proposed negotiation strategy can be classified within the group of strategies that use similarity heuristics to propose new offers to the opponent [6, 20]. The proposal complements some of the benefits introduced in the inspir- ing work of Lai et al. [20], making it especially interesting for AmI environments. The goal is to optimize the computational resources while maintaining a good performance in the negotiation process. The main traits of the proposed model are twofold. Firstly, it is not necessary to sample the entire utility function. Secondly, the proposed strategy provides an implicit learning mechanism that guides the offer sampling and which of the offers sampled are to be sent to the opponent. The different decision-making mechanisms of the negotiation strategy can be grouped according to the period during which they are applied: pre-negotiation and negotiation. The former group of decision making is applied before the negotiation process starts. Basically, since utility functions are complex and it is not feasible to completely explore them, each agent samples its own utility function by means of a niching GA (self-sampling). The latter group of mechanisms is applied during the negotiation process. It includes the acceptance criteria for opponent offers, the concession strategy, the evolutionary sampling, and the selection of which offers are to be sent. 7 AGENT AAGENT BPropose (OfferA1,OfferA2,OfferA3)Propose (OfferB1,OfferB2,OfferB3)Propose (OfferA4,OfferA5)Accept (OfferA4)The most remarkable part is introduced with evolutionary sampling: genetic operators are carried out over received offers and one's own offers in order to sample new offers that may be of interest to both parties. Evolutionary sampling acts as an implicit learning mechanism of the opponent's preferences. The result of evolutionary sampling may be used afterwards when the offers to be sent to the opponent are selected. A brief outline of the proposed strategy can be observed in Algorithm 1. A more detailed outline of the strategy used before the negotiation process and during the negotiation process can be observed in Algorithms 2 and 3. Algorithm 1 A brief outline of the negotiation strategy Negotiation Strategy Pre-negotiation 1.Self-sampling Negotiation Process 2.Receive opponent offer(s) if there are any offers 3.Acceptance criteria: accept an offer and end the negotiation, or reject all of them and continue the negotiation process 4.Concession strategy 5.Evolutionary sampling 6.Select which offers to send 7.Send offer(s) and go to step 2 3.2.1. Pre-negotiation: Self-sampling When an agent uses complex utility functions to represent its preferences it may be difficult to find own offers with good utility. If the number of issues is not very large the complete sampling of the utility function may be feasible. However, when the number of issues is large, this complete sampling may be an extremely expensive process. For instance, a complete sampling of a negotiation domain formed by 10 integer issues from 0 to 9 requires sampling 1010 offers. The cost associated to this sampling can be exorbitant, especially if agent preferences change with a frequency that is greater than the time invested in the sampling. Furthermore, this sampling is unacceptable for AmI domains. Not only does it take too much computational time and power, but it would also need too much storage for the limited devices usually found in these domains. The sampling process can be reduced by skipping offers that are of very low quality for the agent (i.e., offers with utility equal to zero). A possible solution to this problem is to use mechanisms that enable to sample good offers for the negotiation process and skip those of low quality. Due to the highly non-linear nature of complex utility functions, non-linear optimizers are required for this task. The main goal is to sample a set of 8 different offers that have good utility and are significantly different, because these offers may point to different regions of the negotiation space where a good deal may be found for the agent. In this work, a genetic algorithm (GA) was used to solve this problem. GA's are general search and optimization mechanisms based on the Darwinian selec- tion process for species [11, 12]. Genetic operators such as crossover, mutation, and selection are employed in order to find near-optimal solutions for the re- quired problem. Nevertheless, the problem posed by classic GA's is that the entire population converges to one optimal solution. As already stated, differ- ent interesting offers for the negotiation process need to be explored. Niching methods are introduced to confront problems of this kind [21, 25]. These meth- ods look to converge to multiple, highly fit, and significantly different solutions. A possible family of niching methods for GA's is the crowding approach [25]. Crowding methods achieve the desired result by introducing local competition among similar individuals. One advantage of crowding methods is that they do not require parameters beyond the classic GA's. Euclidean distance is usually used to assess the similarity among individuals. Probabilistic Crowding (PC) and Deterministic Crowding (DC) [25] are two of the most popular crowding methods. They only require a special selection rule with respect to classic GA's. Both rules are employed to select a winner given n different individuals. On the one hand, DC selects the individual that has the highest fitness value, resulting in an elitist selection strategy. On the other hand, PC allows lower fitness value individuals to be selected as winners with a certain probability. This probability is usually proportional to the fitness of each individual. PC behavior is more exploratory than DC. In both cases, the niching effect is achieved by applying either of the two rules to those individuals that are similar. Each parent is usually paired with one of its children in such a way that the sum of the distances between pair elements is minimal. For each pair, one of the two crowding rules is employed to determine which individuals will form the next generation. DC and PC can be observed in more detail in Equations 1 and 2, respectively. Pc(s1, s2) = Dc(s1, s2) =  s1 s2 s2 s1 s1 ∨ s2 with pi = (1) (2)  s1 f (s1) > f (s2) f (s1) < f (s2) other s2 s1 ∨ s2 f (s1) > f (s2) ∧ rand ≤ p1 f (s1) > f (s2) ∧ rand > p2 f (s1) < f (s2) ∧ rand ≤ p2 f (s2) < f (s1) ∧ rand > p1 other f (si) f (si) + f (si(cid:48)) 9 where rand ∈ [0, 1], f (.) is the fitness function, s1 and s2 are two solutions, and p1 and p2 are the probability of acceptance of both solutions given the pair (s1, s2). The designed mechanism uses a GA that employs crowding methods to find significantly different good offers. This GA is individually executed by the agent before the negotiation process begins. The chromosomes of this GA represent possible offers in the negotiation process, whereas the fitness function used is one's own utility function. A portfolio with DC and PC is used. The population has a fixed number of individuals and the whole population is selected to form part of the genetic operator pool. Pairs of parents are selected randomly and multi-point crossover or mutation operators are applied over them. In both cases, the result is two children. Each parent is paired with the child that is more similar to it according to Euclidean distance. PC or DC is applied to each of the pairs according to an established probability pdc and 1 − pdc respectively. Those individuals that are selected as winners by the crowding replace the current generation. The stop criterion was set to a specific number of generations. At the end of the process, the whole population should have converged to different good offers that are to be used by the negotiation process as an approximation to the real utility function of the agent. This population, called P , is used as an input for the negotiation process. A more detailed outline of the proposed GA can be observed in Algorithm 2. 3.2.2. Negotiation: Concession strategy A concession strategy determines which utility the agent will try to achieve at each negotiation step. The agent usually proposes offers that have a utility equal or above the utility level defined by its concession strategy at a specific negotiation round. In this work, we assume a time-dependent strategy, where the utility required by each agent depends on the remaining negotiation time. This kind of concession strategies are adequate for environments such as AmI, where time is a limitation (e.g., limited power devices, goods that loose their value as time passes, real-time environments, etc.). Some examples of concession strategies are sit-and-wait [1] (no concession until the deadline, e.g. one of the agents has monopoly), linear (same concession rate at each step), boulware [5, 29] (no concession until the last rounds, where it quickly concedes to the reservation value), and conceder [5, 30] (at the start, it quickly concedes to the reservation value). One of the traits of similarity-based strategies is that they are usually in- dependent of the underlying concession strategy. However, this work assumes an environment where agents have similar market power (similar concession rate), and similar computational resources (similar deadlines). Thus, a linear concession strategy is assumed. In each negotiation round, the agents concede according to their strategy un- til a private deadline is reached. The minimum utility that an agent a demands for a negotiation round t can be formalized as follows: ) ± δ Ua(t) = 1 − (1 − RUa)( t Ta (3) where Ua(t) is the minimum demanded utility level for agent a at negotiation round t, RUa is the reservation utility, δ is a small value that allows to ac- 10 Algorithm 2 Pre-negotiation: Genetic algorithm with niching mechanism. Its goal is to sample the agent utility function P : Explored preferences, good quality offers Dc : Deterministic crowding rule Pc : Probabilistic crowding rule pcr : Probability of crossover operator pdc : Probability of DC n : Current number of generations nmax : Maximum number of generations pairi : Pair of solutions Initialize P n = 0 Do n = n + 1 shuffle P Paux = ∅ i = 1 While i ≤ P − 1 p1 = Pi p2 = Pi+1 If Random() ≤ pcr (c1, c2) = crossover(p1, p2) pi − ck + pj − cl Else c1 = mutate(p1) c2 = mutate(p2) EndIf (pair1, pair2) = argmin pi(cid:54)= pj ck(cid:54)= cl If Random() ≤ pdc Add(Paux, Dc(pair1)) Add(Paux, Dc(pair2)) Else Add(Paux, Pc(pair1)) Add(Paux, Pc(pair2)) EndIf i = i + 2 EndWhile P = Paux While n ≤ nmax Return P 11 cept/select offers which are relatively close, and Ta is the private deadline of the agent. 3.2.3. Negotiation: Acceptance criteria The acceptance criteria for an agent usually depend on its concession strat- egy. Normally, an opponent offer is accepted if it provides a utility that is equal or greater than the demanded utility for the next negotiation round. Conse- quently, given the set of offers X t b→a received by agent a from agent b at instant t, the acceptance criteria for agent a can be formalized as depicted in the fol- lowing expression:  accept Va(xt,best reject otherwise b→a ) ≥ Ua(t + 1) Acceptt a(X t b→a) = (4) a(X t where Acceptt b→a) is the offer acceptance function, Va(x) valuates the utility of an offer, xt,best b→a is the best offer received from the opponent at negotiation round t, and Ua(t + 1) is the utility demanded for the next negotiation round. 3.2.4. Negotiation: Evolutionary sampling One of the keys of the proposed strategy is evolutionary sampling. This provides an implicit mechanism for learning opponent preferences and making an intelligent sampling. Basically, it is based in the application of some genetic operators to offers received from the opponent in the last negotiation round and one's own good offers from P . The idea behind the evolutionary sampling is that offers generated by this method have genetic material from the opponent and one's own agent. Therefore, these offers may yield a greater probability of being accepted by the opponent that offers that have been sampled in a blind way. The new offers are added to a special population called Pevo which contains offers that have been generated by genetic operators. b→a = [xt,1 Let us consider X t b→a], which is the set of offers sent by agent b to agent a at negotiation round t, and U (t) the current desired utility to generate offers at negotiation round t. For each offer xt,i b→a, a total of M offers are selected from the current iso-utility curve ICP (offers with a utility equal to U (t)) defined in the population P . These M offers minimize the expression: b→a, xt,2 b→a, ..., xt,k M(cid:88) j=1 argmin C ∈ ICP C = M xt,i b→a − cj (5) b→a−cj is the Euclidean distance where C is the set of M different offers, and xt,i between one of the offers in C and the offer received from the opponent. Thus, these M offers are the ones most similar to xt,i b→a from iso-utility curve in P and they will be involved in the evolutionary process. Offers are selected from the current iso-utility curve since offers with much greater utility may generate new offers with a utility that is no longer useful in the negotiation process 12 Figure 2: An example of a crossover operation (e.g. a utility greater than the current utility), and offers with lower utility may produce new offers that are not to be used until the last rounds of the negotiation process. Furthermore, the M selected offers are the most similar since applying crossover operators over offers that are too different may disrupt the quality of the solution for both agents (the resulting offer is too far from both agents' offers). Once the M closest offers have been selected, a total of ncross crossover op- b→a, cj), where cj ∈ C. The crossover b→a is chosen randomly from 1 and N − 1, with erations are performed for each pair (xt,i operator takes two parents and generates one child. More specifically, the num- ber of issues that come from xt,i N being the number of issues. The rest of the issues come from cj. Which par- ticular issues come from each parent is also decided randomly. This way, each agent's preferences are taken into account in a statistically equal manner. Each child is added to a special pool, called Pevo, that contains new offers sampled during the different evolutionary sampling phases. An example of a crossover operation can be observed in Figure 2. A total of nmut mutation operations are carried out for each generated child by crossover operations. The mutation operator changes issue values randomly, according to a certain probability of mutating individual issues (pattr). When pattr is low, mutated offers are close to the original offer, so the effect is the exploration of the neighborhood of the offer. The operator is applied nmut times to each child that is produced by crossover operations and to the original offers from the opponent. Mutation also generates new children that are added to the special pool Pevo 13 2101185811460xt,ib→acjAgent proposals: Each phenotype corresponds to the value of a negotiation issueTotal number of issues from the opponent (agent b): 3Specific issues from the opponent (agent b) proposal: 1, 4, 5Specific issues from agent's a proposal: 2, 3Crossover2114185s1Note that no offer from Pevo is discarded even though their utility may be considered too low for the current negotiation round. The reason for this mechanism is that offers that are not currently acceptable may be interesting in future negotiation rounds due to the concession strategy. Furthermore, since they have genetic material from the opponent's offers, they are more likely to be accepted. As can be observed in Algorithm 3, if the negotiation process lasts nround rounds, the Evolutionary Sampling will have explored a total number of offers that is equal to: Samplesevo = nround ∗ ((k ∗ M ∗ ncross) + (k ∗ M ∗ ncross) ∗ nmut + k ∗ nmut) = nround ∗ k ∗ (M ∗ ncross ∗ (1 + nmut) + nmut) Then, the number of offers sampled during the negotiation process depends on the number of rounds that the negotiation lasts, k, M , and the number of genetic operators that are performed per offer selected from the iso-utility curve. 3.2.5. Negotiation: Select which offers to send The next step in specifying the negotiation strategy consists of defining the mechanism to propose new offers. In this case, it is necessary to devise a mech- anism that is capable of proposing up to k different offers to the opponent and taking into account the preferences of the opponent. The applied heuristic takes into account the k offers received from the opponent and the offers in P and Pevo. In order to select these offers, k offers from the current iso-utility curve are sent. More specifically, two different iso-utility curves are calculated. The first one is the iso-utility curve calculated using offers in P , called ICP . The second one is the iso-utility curve calculated using offers in Pevo, called ICE. Basically, the first iso-utility curve has offers that were generated during the self-sampling (only taking into account one's own preferences), whereas the second iso-utility curve only has offers that were generated in the evolutionary sampling (they may take into account both agents' preferences). The negotiation strategy defines a proportion of ppevo offers to come from ICE. The rest of the offers come from ICP . The offers selected from ICE are those that minimize the distance to any offer received from the opponent in the previous negotiation round. This selection may be formalized as: (6)  C(cid:88) j=1 argmin C⊂ICE C=ppevo∗k  cj − x min x∈X t b→a On the other hand, offers are also selected from ICP . The total number of offers corresponds to a proportion that is equal to 1 − ppevo. In this case, offers that are the closest to any offer received from the opponent in the previous negotiation round are selected. This selection can be formalized as: 14  D(cid:88) j=1 min x∈X t b→a  dj − x argmin D⊂ICP D=(1−ppevo)∗k (7) The parameter ppevo determines the degree of relevance of the new offers sampled during the evolutionary sampling with respect to the offers sampled before the negotiation process. When ppevo = 0, the strategy ignores the re- sults that come from Pevo. Consequently, only offers that were sampled in the pre-negotiation phase (self-sampling) are sent to the opponent. In this particu- lar case, the strategy is equivalent to a negotiation strategy that only samples before the negotiation process and does not take into account the opponent's preferences. In contrast, when ppevo = 1, the offers sampled during the evolu- tionary sampling are the only ones taken into account. In any case, ppevo is a parameter to be adjusted. 3.2.6. Case Study We prepared a very simple case based on the product fair example. To be more specific, it depicts a purchase in a furniture fair where one buyer is interested in buying chairs and tables from a seller. It should be pointed out that the goal of this case study is not to test the performance of the model, which will be thoroughly studied in Section , but to show show a trace of the negotiation model from the point of view of one of the agents. In this case, we will focus on the buyer. We used the weighted constraint model proposed by Ito et al. [13] to repre- sent the utility functions of the buyer and the seller. The weighted constraint model was introduced as a complex utility function to model agent preferences. Let us consider a negotiation model where the number of issues is N , si repre- sents the i-th issue, each issue has a domain si ∈ [0, X] that sets its maximum and minimum value, and (cid:126)s = (s1, s2, ..., sN ) represents a particular offer. These settings make up an N-dimensional space for the utility function. In the weighted constraint model, a constraint cl represents a specific region of the space. Any point of the space enclosed in that region is said to satisfy the constraint cl. Basically, the term constraint represents an interdependence relationship among the negotiation issues. Each constraint cl has a certain value v(cl, (cid:126)s) that is added to the utility of (cid:126)s when the constraint is satisfied by the point (cid:126)s. For instance, a constraint defined as cl = (1 ≤ s1 ≤ 10 ∧ 3 ≤ s2 ≤ 4) and v(cl, (cid:126)s) = 10 would hold a utility of 10 for the point (2,3) of the space. A utility function in the weighted constraint model is formed by l constraints whose values are summed up whenever the constraints are satisfied. The utility of a point (cid:126)s given l constraints can be defined as: (cid:88) U ((cid:126)s) = v(cl, (cid:126)s) (8) where (cid:126)s is the point/offer, cl is a constraint, L is the set of constraints, and v(cl, (cid:126)s) is the value of the constraint if it is satisfied (0 otherwise). cl∈L 15 Algorithm 3 Negotiation strategy during the negotiation process P : Offers from self-sampling k: Number of offers of the protocol M : Number of selected offers ncross: Number of times to crossover nmut: Number of times to mutate ppnew: Proportion of offers from Pnew Pnew: Offers from evolutionary sampling Receive X t b→ a If Va(xt,best b→ a) ≥ Ua(t + 1) then Accept Update current utility t=t+1 /Evolutionary sampling/ For each xt,i b→ a in X t b→ a C = (cid:80)M j=1 xt,i b→ a − cj argmin C⊂ ICP C=M For each cj in C b→ a, cj ) Repeat ncross times s1=Crossover(xt,i If s1 (cid:42) Pnew then Add(Pnew,s1) Repeat nmut times s2=Mutate(s1) If s2 (cid:42) Pnew then Add(Pnew,s2) EndRepeat EndRepeat EndFor Repeat nmut times s1=Mutate(xt,i If s1 (cid:42) Pnew then Add(Pnew,s1) b→ a) EndRepeat EndFor /Select which offers to send/ k1 = ppnew ∗ k X1 = j=1 (cid:80)C cj − x min x ∈ Xt b→ a argmin C⊂ ICE C=k1 (cid:80)D j=1 dj − x min x ∈ Xt b→ a k2 = (1 − ppnew) ∗ k X2 = argmin D⊂ ICP D=k2 a→ b = X1 ∪ X2 X t+1 Send X t+1 a→ b 16 As stated in [13], although the expression seems linear, it produces a non- linear utility space due to the interdependence among the issues represented by the constraints. Furthermore, the utility function may generate spaces with several local maxima, which makes the problem highly non-linear and very diffi- cult to optimize. Additionally, the agents do not have any knowledge about the possible constraints of the opponent, thus making the problem of negotiation still more difficult. This negotiation case consists of 3 different attributes: price (P) [0-9], chair model (CM) [0-9], and table model (TM) [0-9]. Next, we introduce the utility functions we employed to represent the preferences of both consumer and seller: Buyer Utility Function (v1 = 100) (0 ≤ P ≤ 1) (v2 = 50) (2 ≤ P ≤ 4) (v3 = 25) (5 ≤ P ≤ 7) (v4 = 30) (0 ≤ CM ≤ 3) ∧ (0 ≤ T M ≤ 3) (v5 = 10) (0 ≤ CM ≤ 3) ∧ (6 ≤ T M ≤ 9) (v6 = 50) (0 ≤ CM ≤ 3) ∧ (5 ≤ T M ≤ 6) (v7 = 30) (4 ≤ CM ≤ 6) ∧ (0 ≤ T M ≤ 3) (v8 = 20) (4 ≤ CM ≤ 5) ∧ (4 ≤ T M ≤ 5) (v9 = 10) (4 ≤ CM ≤ 5) ∧ (8 ≤ T M ≤ 9) (v10 = 50) (7 ≤ CM ≤ 9) ∧ (2 ≤ T M ≤ 4) (v11 = 20) (7 ≤ CM ≤ 9) ∧ (6 ≤ T M ≤ 8) Seller Utility Function (v1 = 80) (8 ≤ P ≤ 9) (v2 = 60) (6 ≤ P ≤ 7) (v3 = 45) (4 ≤ P ≤ 5) (v4 = 20) (1 ≤ P ≤ 3) (v5 = 15) (1 ≤ CM ≤ 2) (v6 = 10) (0 ≤ CM ≤ 1) (v7 = 10) (2 ≤ CM ≤ 5) (v8 = 5) (5 ≤ CM ≤ 9) (v9 = 20) (8 ≤ CM ≤ 9) (v10 = 60) (0 ≤ T M ≤ 1) (v11 = 30) (1 ≤ T M ≤ 4) (v12 = 5) (4 ≤ T M ≤ 6) (v13 = 20) (6 ≤ T M ≤ 9) (v14 = 10) (8 ≤ T M ≤ 9) The consumer shows attribute interdependences relating the two types of furniture (e.g. some pairs of models fit better than other pairs). In the case of the seller, no interdependences are found but he may present preferences regarding which models to sell (e.g. some of them need to be manufactured; some models only have a few units, etc.). As for the parameters of the self-sampling phase, they were set to P = 16, nmax = 100, pdc = 80% and pcr = 80%. The rest of parameters of the negotiation model were set to δ = 0.05, k = 2, T = 10, ppevo = 100%, ncross = 2, nmut = 2, and M = 2. The next table shows the 16 offers found by the self-sampling process carried out by the buyer. It depicts the value for each attribute and the utility of the offer. In this case the utility has been scaled to [0,1] for the sake of simplicity. (u = 0.81) 1 3 0 (u = 0.81) 1 5 3 (u = 0.62) 0 2 4 (u = 0.62) 1 9 1 P=Self-sampling results for the buyer (u = 1.00) 1 1 6 (u = 1.00) 0 1 6 (u = 0.93) 1 7 3 (u = 0.93) 1 7 4 (u = 0.93) 1 9 2 (u = 0.93) 1 2 5 (u = 0.93) 1 8 3 (u = 0.93) 1 8 4 (u = 0.93) 0 7 3 (u = 0.93) 1 9 3 (u = 0.93) 0 1 5 (u = 0.81) 1 5 0 Round 1 Us(1) = 1 − 0.95 Ub(1) = 1 − 0.95. Once the self-sampling phase has ended, the negotiation process starts with the buyer acting as initiator. Since 17 there are no opponent offers to value, evolutionary sampling is skipped and the agent directly proposes offers to the opponent. Due to the fact that no evolutionary sampling has been carried out, Pevo is empty and only the iso- utility curve which can be calculated comes from P. X=(1 1 6) and Y=(0 1 6) are randomly selected since there is no opponent offer to compare with. The opponent rejects the offers since they yield a utility of 0.35 and 0.25 respectively. The opponent makes a counteroffer which contains W=(8 1 1) and Z=(9 1 1). Both of them are rejected since their utilities (0.18 for both of them) are lower than Us(2) = 0.85. Round 2 Us(2) = 0.95− 0.85 Ub(2) = 0.95− 0.85. Two offers have been received from the opponent. Thus, the evolutionary sampling phase is carried out. The iso-utility curve from P (Us(2) = 0.95 − 0.85) is shown in the following tables. It shows the offers and the euclidean distance to W and Z. For both W and Z, the M = 2 offers which are more similar are selected. These offers selected from the iso-utility curve become one of the parents for the genetic operations, which are also shown in the following tables. For the sake of simplicity, genetic operations which produced children that were already in Pevo are not included (nor are they stored more than once). All of the offers generated during this phase are added to Pevo. Iso-utility curve (P) Offer d(Z) 1.00 1 2 5 1.09 0 1 5 1.13 1 7 3 1.16 1 7 4 1.20 1 8 3 0 7 3 1.22 1.22 1 8 4 1.26 1 9 2 1 9 3 1.27 d(W) 0.90 0.99 1.04 1.07 1.12 1.13 1.14 1.18 1.20 Parent 1 Parent 2 Child Parent 1 Child Crossover Mutation Genetic Operations 8 1 1 8 1 1 8 1 1 9 1 1 9 1 1 9 1 1 9 1 1 1 2 5 1 2 5 0 1 5 1 2 5 1 2 5 0 1 5 0 1 5 (u=0.31) 8 2 5 (u=0.81) 1 1 1 (u=0.81) 0 1 1 (u=0.31) 9 1 5 (u=0.31) 9 2 5 (u=0.31) 9 1 5 (u=0.81) 0 1 1 8 2 5 1 1 1 1 1 1 8 1 1 8 1 1 9 1 5 9 1 5 9 2 5 9 2 5 9 1 1 9 1 1 (u=0.34) 6 2 1 (u=0.68) 1 1 7 (u=0.18) 8 1 1 (u=0.31) 2 1 4 (u=0.15) 5 7 1 (u=0.46) 6 1 5 (u=0.62) 1 8 5 (u=0.81) 1 2 3 (u=0.15) 7 6 5 (u=0.50) 4 0 1 (u=0.37) 9 2 6 Next, it is necessary to select which offers to send to the opponent. Since ppevo = 100%, if possible, all of the offers will come from the iso-utility curve calculated using Pevo. If it is not possible, it will take as many offers as possible from the iso-utility curve from Pevo and take the rest from the iso-utility curve from P . In this case, X=(1 2 5) and Y=(0 1 5) are selected from P since Pevo does not contain elements to form a current iso-utility curve. The opponent receives the offers X and Y. Since they yield a utility of 0.25 and 0.15 respectively, both are rejected. The seller sends W=(6 1 1) and Z=(9 4 1) as counteroffers. Both of them are rejected since their utilities (0.34 and 0.18 respectively) are lower than Us(3) = 0.75. Round 3 Us(2) = 0.85− 0.75 Ub(2) = 0.85− 0.75. Two offers have been received from the opponent. Thus, the evolutionary sampling phase is carried out. The iso-utility curve from P (Us(2) = 0.85 − 0.75) and genetic operations are shown in the following tables. 18 Parent 1 Parent 2 Child Parent 1 Child Crossover Mutation Genetic Operations Iso-utility curve (P) d(Z) Offer 1 3 0 0.90 0.90 1 5 0 1 5 3 0.92 d(W) 0.60 0.72 0.74 6 1 1 6 1 1 6 1 1 9 4 1 9 4 1 1 3 0 1 3 0 1 5 0 1 3 0 1 5 0 (u=0.81) 1 3 1 (u=0.34) 6 1 0 (u=0.81) 1 1 0 (u=0.18) 9 4 0 (u=0.81) 1 5 1 1 3 1 1 3 1 6 1 1 6 1 1 1 1 0 1 8 5 9 4 0 9 4 0 9 4 1 9 4 1 1 5 1 1 5 1 (u=0.00) 8 8 1 (u=0.62) 1 6 7 (u=0.34) 6 2 1 (u=0.21) 6 1 8 (u=1.00) 0 1 6 (u=0.62) 1 8 5 (u=0.18) 9 5 0 (u=0.81) 1 4 1 (u=0.18) 8 4 1 (u=0.18) 9 6 1 (u=0.62) 1 7 0 (u=0.31) 4 7 1 Next, it is necessary to select which offers to send to the opponent. The table below shows the iso-utility curve calculated from Pevo. In this case, X=(1 1 1) and Y=(1 1 0) are selected from Pevo. The opponent receives the offers X and Y. Since they yield a utility of 0.69 and 0.53 respectively, both are rejected. However, in this round, the seller sends W=(4 1 1) as counteroffer. The offer is rejected because its utility is equal to 0.5, and is thus lower than Ub(4) = 0.65. From this point on we will overlook the inner steps of the model due to the fact that the way it works has already been described. Iso-utility curve (Pevo) Offer 1 1 0 0 1 1 1 4 1 1 3 1 1 5 1 1 1 1 1 2 3 d(W) 0.56 0.66 0.64 0.59 0.71 0.55 0.60 d(Z) 0.95 1.05 0.88 0.89 0.89 0.94 0.94 Round 4 Us(2) = 0.75 − 0.65 Ub(2) = 0.75 − 0.65. In this round, the buyer sends X=(1 1 7), which yields a utility of 0.33 for the seller. Therefore, the offer is rejected. Then, the opponent sends W=(1 1 1) and Z=(1 2 1), Z being accepted by the buyer since its utility is equal to 0.81. The negotiation process ends with the deal (Ub = 0.81,Us = 0.69), which is the Nash Bargaining Point for this negotiation case. This section has described the main traits of the proposed negotiation model for AmI environments. More specifically, it has explained the protocol employed, and the negotiation strategy that is adapted to AmI domains thanks to the in- telligent sampling provided by genetic operators during the negotiation process. Additionally, we have also shown how the proposed model works in a small case study. In the next section the proposed model is tested in several scenarios to check its performance. 4. Experiments The performance of the devised strategy is detailed in this section. The proposed negotiation model was tested against the weighted constraint model 19 [13]. This model makes it possible to represent unre- proposed by Ito et al. stricted interdependence relationships among the negotiation issues. Further- more, if the number of constraints is large, it can represent highly non-linear utility functions. Therefore, it represents a proper testbed for the proposed strategy. Nevertheless, as in the work of Lai et al. [20], the proposed nego- tiation model is general and does not depend on a particular utility function. The model of Ito et al. was selected as a testbed because it provides a well studied utility function [13, 22, 23] that holds enough complexity to study the real performance of the negotiation model. Firstly, the negotiation setting employed in the experiments is briefly de- scribed. After this, the different experiments and their results are presented. Finally, a brief discussion summarizing the results of the experiments is included. 4.1. Negotiation setting The aim of these experiments was to evaluate whether or not the proposed model is capable of working in domains where the agents' utility functions are highly non-linear. For that purpose, different negotiation cases where randomly created: • Number of issues N = [4-7]. • Integer issues. si ∈ [0, 9]. • L = N∗5 uniformly distributed constraints per agent. There are con- straints for every possible interdependence cardinality. For instance if N =4, there are 5 unary constraints, 5 binary constraints, 5 trinary con- straints and 5 quaternary constraints. • v(cl, .) for each n-ary constraint drawn randomly from [0, 100 ∗ n]. • For every constraint, the constraint width for each issue si is uniformly drawn from [2, 4]. For instance, if the constraint width for issue s1 is 3, then (0 ≤ s1 ≤ 3), (1 ≤ s1 ≤ 4), (2 ≤ s1 ≤ 5), (3 ≤ s1 ≤ 6), (4 ≤ s1 ≤ 7), (5 ≤ s1 ≤ 8) and (6 ≤ s1 ≤ 9) are all of the possible configurations for issue s1 in the constraint (just one is used in the constraint). • Agent deadline T = 10. Agents do not know their opponent's private deadlines. • Agent reservation utility RU = 0. Agents do not know their opponent's private reservation utilities.It is set to zero in order to find a deal, if possible. Should this be the case, the deal is checked against certain thresholds which will determine whether the application notifies its owner of the possible deal. • Agents do not know their opponent's utility functions 20 For each number of issues, a total of 100 negotiation cases were generated with the above settings. The execution of each case was repeated 30 times in order to allow for the possible differences between different executions of the methods. In order to evaluate the quality of the agreements found by the participant agents, some measures were gathered at the end of each negotiation. • Euclidean distance to the closest Pareto frontier point [28]. This is a measure of economic efficiency for agreements. The closer to the Pareto frontier, the better. • Euclidean distance to the Nash Product [28]. Since both agents have the same concession strategy and the same deadline it is also feasible to study the distance to the Nash Product. This is the point that maximizes the product u1 ∗ u2 in the Pareto Frontier, where u1 is the utility of agent 1, and u2 is the utility of agent 2. • Number of negotiation rounds. Faster agreements are preferred since a lesser number of messages are exchanged, less bandwidth is needed, and limited devices need less power to send messages. Additionally, some experiments were also devised in order to test the com- putational performance of the proposed model in a real environment. Measures such as the time spent in decision making tasks before the negotiation process (self-sampling) and during the negotiation process (opponent offer acceptance phase, evolutionary sampling, and offer proposal) were gathered. For that pur- pose, the proposed model was implemented using a HTC Desire (1 Ghz, 576MB RAM, Android Operating System) as one of the parties and a PC (2 Ghz, 4096MB RAM, Ubuntu Operating System) as the other party. A total number of 30 negotiations were carried out in order to measure the computational cost of the proposed model. 4.2. Results The proposed strategy, which will be named as Evolutionary Sampling or ES, was compared with two different negotiation models. The first strategy is an implementation of the general framework proposed by Lai et al. [20]. This model is provided with the whole sampling of the utility function, so that it can completely calculate iso-utility curves. It is used as a measure of how close the proposed strategy is to the ideal case where all of the offers are available. The second model assumes that it is not possible to completely sample all of the offers. Therefore, it samples before the negotiation process by means of a niching GA (self-sampling) and uses the similarity heuristic (ppevo = 0) during the negotiation process, which will be named as Non Evolutionary Sampling or NES model. The number of samples explored by the NES model before the negotiation process is set equal to the number of samples explored by the ES model (P + Samplesevo). Consequently, both the NES and ES model yield the same computational cost in every experimentation. 21 Four different experiments were carried out in order to test the proposed model. In the first experiment, the three different models are compared as the number of issues is increased. The second experiment, studies the impact of the proportion of offers (ppevo) that are sent from the special pool Pevo in the ES model. Next, the three models are compared as the number of proposals k increases. Finally, the ES and the NES model are compared as the size of the population (P) provided by the self-sampling increases. 4.2.1. Experiment 1: Performance study on the number of issues The goal of this experiment is to study how the proposed strategy behaves for negotiations with a different number of issues N = {4, 5, 6, 7}. It is impor- tant that the proposed model be capable of properly handle negotiations with multiple issues since most real world domains, including AmI domains, need to reach agreements for multiple issues. A negotiation setting where agents are limited to k = 3 proposals per negotiation round is used. The three different models were tested during this experiment. The parameters of the self-sampling were set to nmax = 100, pdc = 80% and pcr = 80%. The number of samples optimized before the negotiation process was set to P = 128 for the ES model and to P = 128 + Samplesevo for the NES model. The parameters of the ES were set to M = 5, ncross = 4, nmut = 4, pattr = 30%, and ppevo = 100%. Therefore, all the offers are sent from the sam- ples generated by the evolutionary sampling carried out during the negotiation process. The distance to the Nash Product, the distance to the closer Pareto Frontier Point and the number of negotiation rounds were measured for the three models. The results for this experiment can be found in Figure 3. Intuitively speaking, since the number of offers sampled remains constant and the number of issues increases, the performance of the NES and the ES model should be worsened with respect to the results achieved by the model of Lai et al. However, the results for the ES do not comply with this intuitive hypothesis. As can be observed, even though the proposed model and the NES model explore the same number of offers, the NES obtains worse results than the other two models. This is particularly true as the number of issues increases, since the performance of this method drastically decreases. On the contrary, the ES model is capable of achieving statistically equal results to the model of Lai et al., which can access the whole iso-utility curve. Nevertheless, the proposed model explores far fewer offers than the complete sampling of the utility function, especially for larger number of issues. For instance, when N = 6, Lai et al. has access to 106 offers, whereas the proposed model has only sampled an average of 1510 samples (128+ average Samplesevo). The ES model has been able to achieve similar results to the case where the full iso-utility curve can be calculated, while maintaining the offers sampled to a small number. This result is particularly interesting for AmI domains where agents may be executed in devices with low computational and storage capabil- ities. Therefore, fewer samples mean less power consumption and less capacity 22 needed to store them. Moreover, it must also be highlighted that the num- ber of rounds was also lower than that obtained by NES, which, consequently means fewer messages sent, less bandwidth needed and, of course, less power consumption by the limited devices. The reason for this improvement is the intelligent sampling achieved by the use of genetic operators during the negotiation process. On the contrary, sampling only before the negotiation process leads to worse results since it is not capable detecting which offers will be interesting for the negotiation. Both, the ES and the NES model, have the same computational cost, but the ES is obviously preferred since it is capable of achieving a better performance. Figure 3: Evolution of the distance to the Nash Product, distance to the closest Pareto Point, and number of negotiation rounds in Experiment 1. The graphic shows the mean and its associated confidence intervals (95%) 4.2.2. Experiment 2: Performance study on ppevo In this case, the experiment's goal is to study how relevant the proportion of offers that are sent from the offers sampled during the negotiation process (gov- 23 0.1 0.15 0.2 0.25 0.3 0.35 4 5 6 7Distance Nash ProductNumber of issuesP=128 k=3Lai et al.No Evo. Samp.Evo. Samp. 0 0.05 0.1 0.15 0.2 4 5 6 7Distance Pareto FrontierNumber of issuesP=128 k=3Lai et al.No Evo. Samp.Evo. Samp. 3.5 4 4.5 5 5.5 4 5 6 7Number of RoundsNumber of issuesP=128 k=3Lai et al.No Evo. Samp.Evo. Samp.erned by the parameter ppevo) in the ES model is. Since all of the configurations sample new offers during the negotiation process, all of them yield a very similar computational cost. In fact, it may only be different if one of the configurations obtains a significantly different number of negotiation rounds. Consequently, the main subject of study in this scenario is the economic efficiency (distance to Nash and Pareto Frontier), although some improvements in the computational cost may be observed due to a lower number of rounds. The same conditions from the previous experiment were set (k = 3 and N = {4, 5, 6, 7}), and the same configuration parameters were set for the ES (M = 5, ncross = 4, nmut = 4, and pattr = 30%). However, in this scenario we compare the ES model results when 1 out of 3 offers (ppevo = 30%), 2 out of 3 offers (ppevo = 50%), and 3 out of 3 offers (ppevo = 100%) come from the offers sampled during the evolutionary sampling phase. The results for this second scenario can be observed in Figure 4. The graphic shows that the three different configurations yield similar results for the distance to the Nash Product, the distance to the closest Pareto Frontier Point, and the number of negotiation rounds. This similarity is explained due to the fact that, on most occasions, the offer accepted by the opponent is the closest one from the evolutionary sampling population (Pevo). Therefore, it is always sent, as long as the results from the evolutionary sampling are not ignored. Nevertheless, it seems that higher values of ppevo have a slightly better economic and computa- tional performance than lower ones. The reason for this slight improvement is that, in some cases, the offer preferred by the opponent may be the second or third closest from Pevo. Due to this small improvement, higher values of ppevo are preferred in practice. 4.2.3. Experiment 3: Performance study on k The next experiment aims to study the performance of the three different models (Lai et al., NES, and ES ) as the limit to the number of offers k sent per agent's round is increased. As mentioned, the number of offers sent may help to reach agreements faster since agents are capable of finding win-win situations. This is very important in AmI environments where devices have limited power [20], demonstrated how and their running time must be optimized. Lai et al. higher values of k helped to reach better agreements. In this scenario, the experiment is repeated in order to evaluate whether the differences between the three models still hold for different values of k. The studied values of k were 1, 3, 5, and 7. The rest of the negotiation setting was configured to use negotiation cases with N = 6 issues. The parameters of the self-sampling were set to the values employed in the previous tests except for P = 256. The parameters of the ES were set to the same conditions described in Experiment 1. As it can be observed in Figure 5, the three models achieve better results as k increases. These results agree with those presented in [20]. Although all of the models improve, the differences observed in Experiment 1 still hold for this scenario. The NES model gets worse results than Lai et al. and the proposed model. On the contrary, the ES obtains results that are statistically equivalent 24 Figure 4: Evolution of the distance to the Nash Product, distance to the closer Pareto Point, and number of negotiation rounds in Experiment 2. The graphic shows the mean and its associated confidence intervals (95%) to the case when the full iso-utility curve can be calculated. As a matter of fact, for higher values of k the proposed model gets slightly better results than Lai et al. Nevertheless, the differences between the two of them are not significant enough to be considered as relevant. It must be noted again that the number of offers sampled for ES and NES is the same and it is much lower than the complete sampling of the utility function. For instance, in this scenario, the complete sampling consists of 106 offers, whereas the other two methods sampled an average of 773 samples for k = 1, 1653 for k = 3, 2497 for k = 5, and 3357 for k = 7. 4.2.4. Experiment 4: Performance study on P and memory performance This last experiment was designed to assess the influence of the population optimized by the self-sampling on the performance of the ES model and the NES model. It is specially relevant to see how many samples the NES model 25 0.14 0.16 0.18 0.2 0.22 0.24 0.26 4 5 6 7Distance Nash ProductNumber of issuesP=128 k=3ppevo=30%ppevo=50%ppevo=100% 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 0.11 0.12 4 5 6 7Distance Pareto FrontierNumber of issuesP=128 k=3ppevo=30%ppevo=50%ppevo=100% 3.8 4 4.2 4.4 4.6 4.8 5 4 5 6 7Number of RoundsNumber of issuesP=128 k=3ppevo=30%ppevo=50%ppevo=100%Figure 5: Evolution of the distance to the Nash Product, distance to the closest Pareto Point, and number of negotiation rounds in Experiment 3. The graphic shows the mean and its associated confidence intervals (95%) needs to achieve similar results to those ones obtained by the model proposed in this article. Obviously, more population means more storage needed and more computational cost since it needs to optimize more samples. The average number of samples explored was analyzed for a negotiation setting where N = 6 and k = 3. The settings used for the self-sampling and the ES in previous experiments were repeated for this scenario. The number of sampled offers was increased by allowing more offers to be optimized in the self-sampling (P = 128, 256, 512, 1024, 2048, 4096). The results for this experiment can be observed in Figure 6. The x axis of the graphics show the average number of offers sampled by both models, thus it shows P + averagerounds ∗ Samplesevo. In the case of the NES model all of the samples were produced before the negotiation process started. Several observations can be made from the data shown in the graphics. On the one 26 0.2 0.25 0.3 0.35 0.4 1 3 5 7Distance Nash ProductKP=256 Issues=6Lai et al.No Evo. Samp.Evo. Samp. 0.05 0.1 0.15 0.2 0.25 1 3 5 7Distance Pareto FrontierKP=256 Issues=6Lai et al.No Evo. Samp.Evo. Samp. 4 4.5 5 5.5 6 1 3 5 7Number of RoundsKP=256 Issues=6Lai et al.No Evo. Samp.Evo. Samp.hand, it seems that the size of P does not have too much of an effect on the performance of the ES model, since it is more dependent on the exploration carried out during the negotiation process and does not need as much sampling to get results similar to the case where the full iso-utility curve can be accessed. Therefore, the behavior of the model remained almost constant for different configurations of P. Again, this behavior is very adequate for AmI environ- ments since the model can properly work with configurations that do not require too many computational resources. On the other hand, the NES model perfor- mance increased along with the number of offers sampled. It must be noted, that when the number of samples for both methods was 5506, the two of them obtained very similar, almost equivalent, results. Therefore, the NES needed 5506 samples to achieve similar results to the same results obtained by the ES model for 1510 samples. It can be concluded that NES needs 5506 1510 = 3.64 times more samples to achieve similar results to ES. Figure 6: Evolution of the distance to the Nash Product, distance to the closest Pareto Point, and number of negotiation rounds in Experiment 4. The graphic shows the mean and its associated confidence intervals (95%) 27 0.18 0.2 0.22 0.24 0.26 0.28 0.3 1510 1912 2431 3461 5506Distance Nash ProductOffers sampledIssues=6 k=3No Evo. Samp.Evo. Samp. 0.06 0.07 0.08 0.09 0.1 0.11 0.12 0.13 0.14 0.15 1510 1912 2431 3461 5506Distance Pareto FrontierOffers sampledIssues=6 k=3No Evo. Samp.Evo. Samp. 4.2 4.4 4.6 4.8 5 5.2 5.4 1510 1912 2431 3461 5506Number of RoundsOffers SampledIssues=6 k=3No Evo. Samp.Evo. Samp.Number negotiations Memory(KB) NES Memory(KB) ES 1 3 5 7 10 129 387 645 903 1290 35 105 175 245 350 Table 1: Approximate ammount of memory needed by the NES model and the ES model when executing several negotiations at the same time It is possible to approximately analyze the total amount of memory employed by both methods when they achieve statistically equivalent results. As has been suggested by the previous experiment (Experiment 4), the NES model needs 5506 samples to achieve statistically equivalent results to those the ES model with 1510 samples. If we assume that the underlying platform is a 32 bit platform, where integers usually need 32 bits to be stored, we can approximately calculate the memory needed by both models as follows: M emory(KB) = Samples ∗ N ∗ 32 ∗ 1 8 (9) where Samples is the number of samples, N is the number of attributes of the negotiation process, 32 is the size of an integer, 1 8 converts from bits to Bytes, and 1 1024 converts from bytes to KBytes. Taking into account the formula above, the NES model would take 129 KB to store the data needed for the previous type of negotiation process (N = 6, Samples = 5506), whereas the ES model would take 35 KB (N = 6, Samples = 5506). Depending on the underlying device, this difference may be quite important (e.g. devices with a few MB of storage available). However, this difference may be still more important if we consider that in some scenarios it may be necessary to perform several negotiations at the same time (e.g. the fair scenario). For instance, Table 1 shows the approximate amount of memory necessary to carry out several negotiations at the same time. 1024 1 ∗ 4.2.5. Experiment 5: Time Performance As introduced earlier, it was also interesting to test the computational per- formance of the model in a real environment. Thus, the proposed model was implemented using a HTC (1 Ghz, 576MB RAM, Android Operating System) as one of the parties and a PC (2 Ghz, 4096MB RAM, Ubuntu Operating Sys- tem) as the other party. The self-sampling parameters were set to nmax = 100, pdc = 80% and pcr = 80%. The number of samples optimized before the nego- tiation process was set to P = 128. As for the parameters employed during the negotiation process, these were set to k = 3, M = 5, ncross = 4, nmut = 4, pattr = 30%, and ppevo = 100%. The number of attributes of the negotiation process was N = 5. The time spent in the whole negotiation process (tt), the time spent in sending/waiting for offers (tm), the time spent in self-sampling (ts), and the time spent in decision-making during the negotiation process (tdm) 28 tt (s) 0.773 ts (s) 0.264 tdm (s) 0.358 tm (s) 0.415 Table 2: Computational performance measures for Experiment 5 were measured. Table 2 shows the mean obtained in seconds for the 30 negoti- ation cases that were studied. As can be observed, the time spent for a negotiation process tt was reason- ably good (less than a second) and it enables negotiations to be carried out in environments where real-time responses are needed (e.g. Ambient Intelligence). Moreover, it can also be observed that the time spent in decision-making tasks tdm does not take as much time as other tasks such as sending/waiting for offers tm. This leaves room for more negotiation processes to be carried out in paral- lel during CPU idle time (e.g. waiting for offers). Again, carrying out multiple negotiation processes simultaneously proves especially interesting again for AmI environments. For instance, in the fair scenario, it makes it possible to negoti- ate simultaneously with those vendors who are available in the area where the consumer is walking at that moment. The time spent in self-sampling is the least problematic since it is a process to be carried out only once until agent preferences change. In some AmI environments, such as the fair, we may con- sider preferences to be static during the fair event. Thus, self-sampling would only be needed once. Despite all those facts, it must be remarked that the time spent in self-sampling is reasonably good (less than a second). 4.3. Discussion Ambient Intelligence domains are characterized as domains where computa- tional resources are of extreme importance. Users interact with its environment through devices with limited capabilities, thus the efficient use of resources is crucial. Furthermore, the environment infrastructures are usually connected by means of a limited bandwidth wireless connection. Thus, network resources must also be optimized. The results obtained by the proposed model, while maintaining fairly good economic performance, cope with the problems found in AmI environments. If we assume that limited devices cannot completely sample the agent's util- ity function and store those samples, some mechanisms are needed to sample as few offers as possible. A straightforward method would be sampling some offers before the negotiation process starts, which is precisely what the NES model does. However, this sampling does not take advantage of the informa- tion revealed by the opponent in the negotiation process. Most of the offers sampled before the negotiation process may be useless since they are of no in- terest to the opponent. However, the proposed model takes advantage of this information and employs it to make a more intelligent sample, optimizing the computational resources. Nevertheless, although computational resources are important, economic efficiency should not be ignored in AmI negotiations. 29 In the previous sections, we were able to observe the behavior of the ES model in different scenarios. Its performance was compared with a method that samples the same number of offers before the negotiation process (NES ), and the ideal case where all of the samples of the utility function are available. The results of the experiments can be summarized as: • The proposed model needs very few computational resources and storage to obtain results statistically equivalent to the ideal case where the all of the offers are available. It obtained similar results in economic per- formance (distance to Nash, distance to Pareto Frontier) and number of negotiation rounds. • When the proposed model and the NES model sample the same number of offers, the first obtains better results. In fact, the NES model needs to sample 3.64 times more offers to obtain similar results. • The proposed model needs less negotiation rounds to achieve better results than the NES model. Therefore, the environment bandwidth is optimized since it needs fewer messages to be sent in order to reach agreements. Consequently, the proposed model fits perfectly for the conditions needed by AmI environments, since it needs less computational resources and it obtains economically efficient results. 5. Related Work Ambient Intelligence looks to offer personalized services and provide users with easier and more efficient ways to communicate and interact with other people and systems [4, 35]. Since several users may coexist in AmI environments, it is quite probable that their preferences conflict and thus mechanisms are needed to allow users to cooperate. For instance, imagine a ubiquitous shopping mall [16, 2] where buying agents have to help users to buy the products, and vendor agents have to maximize their users' profits. Automated negotiation provides mechanisms that solve this particularly interesting problem. Some authors have already claimed that in most real world negotiations such as e- commerce [17, 33, 13], issues present interdependence relationships that make agents' utility functions complex. Therefore, the problem of complex utility functions in automated negotiation is also interesting for AmI applications. Over the last few years, most of the work in automated negotiation has focused on offering solutions for the case of imperfect knowledge and bounded computational resources [18, 14]. The use of heuristics is necessary to provide a solution to problems of this type. The present work can be classified within this same category of solutions. Faratin et al. [5] presented a negotiation model for linear utility functions where a negotiation strategy is composed of different tactics that may be ap- plied depending on the negotiation time, the quantity of the resource and the 30 behavior of the opponent. Nevertheless, the model is only applicable in negoti- ation with linear utility functions, which are easier cases than those presented in this present article. Matos et al. [24] determined the successful strategies for different settings [5]. They employ an evolutionary using the model proposed by Faratin et al. approach in which strategies and tactics correspond to the genetic material in a genetic algorithm. In their experiments, populations of buyers and sellers with different strategies negotiate in a round robin way. After each round robin round, strategies are evaluated by means of a fitness function. Then, strategies are selected to be the parents of the next population according to their fitness function. In the end, a population of strategies implicitly adapted to the en- vironment is obtained. They use genetic algorithms as a learning mechanism of negotiation strategies when placed under certain circumstances. There are two differences between Matos et al. work and the present work. Firstly, the negotiation model of Matos et al. is designed for linear utility functions. Sec- ondly, the genetic algorithm proposed in this present work is an implicit learning mechanism of the opponent's preferences that guides the offer sampling during the negotiation process. Later, Faratin et al. [6] presented a negotiation strategy for bilateral bar- gaining that is focused on achieving win-win situations by means of trade-off. The heuristic applied to perform trade-off is similar to that employed in this present work. Given an agent's current utility, the offer from the iso-utility curve that is most similar to the last offer received from the opponent is sent. The idea behind this heuristic is that, since the proposed offer is the most similar to the last offer received from the opponent, it is more likely to be satisfactory to both participants. A fuzzy similarity criterion is employed to compare offers. Nevertheless, the use of fuzzy similarity requires some knowledge of opponent preferences. The application of criteria of this kind is complicated in complex utility functions due to the interdependencies among the different issues. In this present work, the Euclidean distance is used, as this does not require any knowledge about the opponent, and which is independent of the interdependen- cies among issues. Fatima et al. [7, 9, 8] analyzed the problem of multi-attribute negotiations in an agenda-based framework. Agendas determine in which order the different issues are to be negotiated when negotiations are carried out issue by issue. Once an agreement has been found on a specific issue, it cannot be changed. Thus, the agents face the problem of which issues should be negotiated first and which strategies should be applied. They studied the optimal agendas for different scenarios. Nevertheless, their work focused on linear utility functions, which does not take into account the possible interdependences among the different issues. The work of Krovi et al. [19] opened the path for GA's in automated negoti- ation. Krovi et al. proposed a GA for bilateral negotiations that was performed each time a negotiation round ended. The population of chromosomes was ran- domly initialized with 90 random offers and 10 heuristic offers (the last offer from the opponent and the nine best offers from the previous round). The idea 31 behind using GA's is that the resulting offers have good characteristics for both agents. However, 60 generations were needed during each round in order to obtain the next offer, which may turn out to be computationally expensive in large issue domains. Choi et al. [3] enhanced Krovi's model with more learning capabilities. More specifically, it is capable of learning opponent preferences by means of stochastic approximation and of adapting its mutation rate to oppo- nent behavior. However, these strategies and mechanisms are devised for linear utility functions with few negotiation issues. The performance of these methods is uncertain when a large number of issues or complex utility functions are used. This present work also employs genetic operators to obtain new offers, but it is capable of providing solutions for domains with complex utility functions and domains where the number of issues is large. There have been some works that have studied the problem of negotiation models for complex utility functions. Most of them have focused on mediated negotiation models. The seminal work of Klein et al. [17] proposes a mediated negotiation model where agents' preferences are represented by influence matri- ces. Influence matrices represent binary interdependence relationships between binary issues. Their proposed approach consists of a mediator that generates bids that are voted by the agents participanting in the protocol. Ito et al. [13] propose a mediated negotiation model for multilateral negotiations where agents have their preferences represented by weighted constraints. The agents sample their utility function and carry out a simulated annealing for each point sampled in order to obtain one's own bids. If the utility of such point is above a certain threshold, the constraints that the bid satisfies are sent to the media- tor (constraint bid). After receiving bids from the agents, the mediator tries to look for contracts common to the bids received, while maximizing social welfare. Marsa-Maestre et al. [22, 23] carry out further research in the area of mediated negotiation models for complex utility functions. More specifically, they take advantage of the constraint based model by proposing different bidding mech- anisms that work in the constraint space instead of the bid space. They also allow for a negotiation protocol that may not be one-shot. In fact, the mediator can suggest the relaxation of some constraint bids in order to increase the prob- ability of finding an agreement. Nevertheless, all of these works need a trusted mediator, which may not be available in every domain. Furthermore, their mod- els are highly dependent on the underlying utility function. The present work does not require a mediator and the model is independent of the underlying utility function. Robu et al. [33, 32] presented a non-mediated bilateral negotiation strategy for agents in electronic commerce. Agent utility functions are based on special graphical models called utility graphs. One of the agents, the seller, is responsi- ble for finding agreements that are satisfactory for both parties. In order to do that, the seller models the buyer by means of utility graphs and tries to learn the buyer's preferences. However, utility graphs are only designed for binary issues. Our work differs in that it is capable of working with general complex utility functions and is also capable of working issue domains that are not necessarily binary. 32 In Lai et al. [20], a powerful bilateral bargaining model with general utility functions is presented. The negotiation protocol is based on the Rubinstein alternating protocol [27], but each agent is allowed to send up to k different offers in each round. The offer with highest utility is chosen from the k offers received from the opponent in the last round. The offer from the current iso- utility curve that is the most similar to the one chosen by the agent from the offers made by the opponent is selected. This offer from the iso-utility curve becomes a seed from which k-1 offers in the neighborhood are generated. The selected offer from the curve and the k-1 generated offers are sent back to the opponent. Again, the general ideal behind this heuristic is that, since the offers are similar to one of the last offers received from the opponent, they are more likely to be satisfactory for both parties. The model proposed in this article complements the seminal work of Lai et al. since it adapts similarity models for AmI environments. In the model proposed in this article, only a small number of offers are sampled before the negotiation process, since it is assumed that the utility function cannot be exhaustively explored. This is especially important for scenarios with a large number of issues and scenarios where devices have limited storage and computational resources. Secondly, the proposed model incorporates an implicit learning mechanism that allows, thanks to genetic operators, an intelligent sampling of new offers that may be of interest for both parties. 6. Conclusions and Future Work Ambient Intelligence aims to offer new services and methods of interaction with technology adapted to the users. It has been stated that automated ne- gotiation may provide a conflict-resolution mechanism in Ambient Intelligence applications where several users with opposing preferences need to cooperate (e.g. ubiquitous shopping malls, fairs). In these environments, users' agents may present utility functions that are complex due to the interdependences among the negotiation issues that form the utility function. A multi-issue bilateral bargaining model for Ambient Intelligence domains that deals with complex utility functions has been presented in this article. This work complements the inspiring work of Lai et al. [20] and provides a negotiation model that is adequate for Ambient Intelligence applications. The main goal of this work is to achieve efficient agreements while maintaining the use of computational resources low. The proposed model uses a negotiation protocol where agents are allowed to send up to k different offers in each negotiation round. Before the negotia- tion process starts, each agent samples its own utility function by means of a niching genetic algorithm. This genetic algorithm gets highly interesting and significantly different offers for one's own utility function (self-sampling). After the negotiation process starts, the agents apply genetic operators over the last offers received from the opponent and those offers that are most similar from the current iso-utility curve (evolutionary sampling). The desired effect is to sample new offers that are interesting for both parties. Therefore, the opponent 33 preferences guide the sampling process during the negotiation process. The of- fers that are sent to the opponent are selected from the current iso-utility curve, being those that are the most similar to the last offers received from the oppo- nent. An additional mechanism is introduced that allowing priority to be given to those offers that come from the evolutionary sampling iso-utility curve. Several experimental scenarios have been carried out and studied. In these tests, the proposed model has been compared with a similarity heuristic that has access to all of the possible offers and a similarity heuristic that samples the same number of offers before the negotiation process by means of a niching genetic algorithm (NES ). The results show that the proposed model needs very few computational resources and storage to obtain statistically equivalent results to the ideal case where all of the offers are available. For instance, the full iso-utility curve consists of 106 offers and the proposed model just samples 1510 offers in a negotiation setting where the number of issues is 6, and the number of offers sent per negotiation round is 3. Additionally, although the proposed model and the NES model sample the same number of offers, the first one obtains better results. In fact, the NES model needs to sample 3.64 times more offers to obtain similar results. The low computational cost and the efficient results make the proposed model very adequate for Ambient Intelligence domains. Future work includes studying the effect of changing preferences during the negotiation process, (i.e., when the strategy is integrated with an argumentation mechanism), and introducing different agent behaviors (more self-interested, more cooperative, etc) by means of some modifications to genetic and selection operators. Acknowledgments This work is supported by TIN2008-04446, PROMETEO/2008/051, TIN2009- 13839-C03-01, CSD2007-00022 of the Spanish government, and FPU grant AP2008- 00600 awarded to V.S´anchez-Anguix. [1] B. An, K.M. Sim, C.Y. Miao, Z.Q. Shen, Decision making of negotiation agents using markov chains, Multiagent Grid Syst. 4 (2008) 5 -- 23. [2] J. Bajo, J.M. Corchado, Y. de Paz, J.F. de Paz, S. Rodr´ıguez, Q. Martin, A. Abraham, Shomas: Intelligent guidance and suggestions in shopping centres, Appl. Soft Comput. 9 (2009) 851 -- 862. [3] S.P.M. Choi, J. Liu, S.P. Chan, A genetic agent-based negotiation system, Computer Networks 37 (2001) 195 -- 204. [4] J.M. Corchado, J. Bajo, Y. de Paz, D.I. Tapia, Intelligent environment for monitoring alzheimer patients, agent technology for health care, Decis. Support Syst. 44 (2008) 382 -- 396. [5] P. Faratin, C. Sierra, N.R. Jennings, Negotiation decision functions for autonomous agents, Int. Journal of Robotics and Autonomous Systems 24 (1998) 159 -- 182. 34 [6] P. Faratin, C. Sierra, N.R. Jennings, Using similarity criteria to make nego- tiation trade-offs, in: 4th International Conference on Multi-Agent Systems (ICMAS-2000), pp. 119 -- 126. [7] S. Fatima, M. Wooldridge, N.R. Jennings, Optimal agendas for multi-issue negotiation, in: AAMAS '03: Proceedings of the second international joint conference on Autonomous agents and multiagent systems, ACM, 2003, pp. 129 -- 136. [8] S. Fatima, M. Wooldridge, N.R. Jennings, Optimal negotiation of multiple issues in incomplete information settings, in: AAMAS '04: Proceedings of the Third International Joint Conference on Autonomous Agents and Multiagent Systems, IEEE Computer Society, 2004, pp. 1080 -- 1087. [9] S.S. Fatima, M. Wooldridge, N.R. Jennings, An agenda-based framework for multi-issue negotiation, Artif. Intell. 152 (2004) 1 -- 45. [10] J.A. Fraile, J. Bajo, A. Abraham, J.M. Corchado, Hocama: Home care hybrid multiagent architecture, in: Pervasive Computing, Computer Com- munications and Networks, 2009, pp. 259 -- 285. [11] D.E. Goldberg, Genetic Algorithms in Search, Optimization and Machine Learning, Addison-Wesley Longman Publishing Co., Inc., Boston, MA, USA, 1989. [12] J. Holland, Adaptation in natural and artificial systems, University of Michigan Press, 1975. [13] T. Ito, M. Klein, H. Hattori, A multi-issue negotiation protocol among agents with nonlinear utility functions, Multiagent Grid Syst. 4 (2008) 67 -- 83. [14] N.R. Jennings, P. Faratin, A.R. Lomuscio, S. Parsons, M.J. Wooldridge, C. Sierra, Automated negotiation: Prospects, methods and challenges, Group Decision and Negotiation 10 (Mar 2001) 199 -- 215. [15] C. Jonker, V. Robu, Automated multi-attribute negotiation with efficient use of incomplete preference information, in: AAMAS '04: Proceedings of the Third International Joint Conference on Autonomous Agents and Multiagent Systems, IEEE Computer Society, Washington, DC, USA, 2004, pp. 1054 -- 1061. [16] S. Keegan, G. O'Hare, M. O'Grady, Easishop: Ambient intelligence assists everyday shopping, Information Sciences 178 (2008) 588 -- 611. [17] M. Klein, P. Faratin, H. Sayama, Y. Bar-Yam, Negotiating complex con- tracts, in: AAMAS '02: Proceedings of the first international joint con- ference on Autonomous agents and multiagent systems, ACM, New York, NY, USA, 2002, pp. 753 -- 757. 35 [18] S. Kraus, Negotiation and cooperation in multi-agent environments, Artif. Intell. 94 (1997) 79 -- 97. [19] R. Krovi, A.C. Graesser, W.E. Pracht, Agent behaviors in virtual negotia- tion environments, IEEE Transactions on Systems, Man, and Cybernetics, Part C 29 (1999) 15 -- 25. [20] G. Lai, K. Sycara, C. Li, A decentralized model for automated multi- attribute negotiations with incomplete information and general utility func- tions, Multiagent Grid Syst. 4 (2008) 45 -- 65. [21] S.W. Mahfoud, Niching methods for genetic algorithms, Ph.D. thesis, Champaign, IL, USA, 1995. [22] I. Marsa-Maestre, M.A. Lopez-Carmona, J.R. Velasco, E. de la Hoz, Effec- tive bidding and deal identification for negotiations in highly nonlinear sce- narios, in: AAMAS '09: Proceedings of The 8th International Conference on Autonomous Agents and Multiagent Systems, International Foundation for Autonomous Agents and Multiagent Systems, Richland, SC, 2009, pp. 1057 -- 1064. [23] I. Marsa-Maestre, M.A. L´opez-Carmona, J.R. Velasco, T. Ito, M. Klein, K. Fujita, Balancing utility and deal probability for auction-based negoti- ations in highly nonlinear utility spaces, in: IJCAI, pp. 214 -- 219. [24] N. Matos, C. Sierra, N. Jennings, Determining successful negotiation strate- gies: An evolutionary approach, in: ICMAS '98: Proceedings of the 3rd International Conference on Multi Agent Systems, IEEE Computer Society, Washington, DC, USA, 1998, p. 182. [25] O.J. Mengshoel, D.E. Goldberg, The crowding approach to niching in ge- netic algorithms, Evol. Comput. 16 (2008) 315 -- 354. [26] J. Nash, The bargaining problem, Econometrica 18 (1950) 155 -- 162. [27] M. Osborne, A. Rubinstein, Bargaining and Markets, The Academic Press, 1990. [28] M.J. Osborne, A. Rubinstein, A Course in Game Theory, MIT Press, 1999. [29] D.G. Pruitt, Negotiation Behavior, Academic Press, 1981. [30] H. Raiffa, The Art and Science of Negotiation, Cambridge, USA, 1982. [31] P. Remagnino, G. Foresti, Ambient intelligence: A new multidisciplinary paradigm, Systems, Man and Cybernetics, Part A: Systems and Humans, IEEE Transactions on 35 (2005) 1 -- 6. 36 [32] V. Robu, J.A. La Poutr´e, Retrieving the structure of utility graphs used in multi-item negotiation through collaborative filtering of aggregate buyer preferences, in: Proceedings of the 2nd International Workshop on Ra- tional, Robust and Secure Negotiations in Multi-Agent Systems (RSS'06). Hakodate, Japan. [33] V. Robu, D.J.A. Somefun, J.A. La Poutr´e, Modeling complex multi-issue negotiations using utility graphs, in: AAMAS '05: Proceedings of the fourth international joint conference on Autonomous agents and multia- gent systems, ACM, New York, NY, USA, 2005, pp. 280 -- 287. [34] A. Rubinstein, Perfect equilibrium in a bargaining model, Econometrica 50 (1982) 155 -- 162. [35] W. Weber, J.M. Rabaey, E. Aarts, Ambient Intelligence, Springer-Verlag New York, Inc., Secaucus, NJ, USA, 2005. [36] M.J. Wooldridge, Multi-agent systems : an introduction, Wiley, 2001. 37
1005.4263	1	1005	2010-05-24T07:07:20	Facial Recognition Technology: An analysis with scope in India	[ "cs.MA" ]	A facial recognition system is a computer application for automatically identifying or verifying a person from a digital image or a video frame from a video source. One of the way is to do this is by comparing selected facial features from the image and a facial database.It is typically used in security systems and can be compared to other biometrics such as fingerprint or eye iris recognition systems. In this paper we focus on 3-D facial recognition system and biometric facial recognision system. We do critics on facial recognision system giving effectiveness and weaknesses. This paper also introduces scope of recognision system in India.	cs.MA	cs	(IJCSIS) International Journal of Computer Science and Information Security, Vol. 8, No. 1, 2010 Facial Recognition Technology: An analysis with scope in India Dr.S.B.Thorat Director, Institute of Technology and Mgmt Nanded, Dist. - Nanded. (MS), India [email protected] S.K.Nayak Head, Dept. of Computer Science Bahirji Smarak Mahavidyalaya, Basmathnagar, Dist. - Hingoli. (MS), India [email protected] Miss.Jyoti P Dandale Lecturer Institute of Technology and Mgmt Nanded, Dist. - Nanded. (MS), India [email protected] Abstract— A facial recognition system is a computer application for automatically identifying or verifying a person from a digital image or a video frame from a video source. One of the way is to do this is by comparing selected facial features from the image and a facial database.It is typically used in security systems and can be compared to other biometrics such as fingerprint or eye iris recognition systems. In this paper we focus on 3-D facial recognition system and biometric facial recognision system. We do critics on facial recognision system giving effectiveness and weaknesses. This paper also introduces scope of recognision system in India. Keywords-3-D facial recognition; biometric facial recognition; alignment; matching;FRGC. I. INTRODUCTION Pioneers of Automated Facial Recognition include: Woody Bledsoe, Helen Chan Wolf, and Charles Bisson. During 1964 and 1965, Bledsoe, along with Helen Chan and Charles Bisson, worked on using the computer to recognize human faces (Bledsoe 1966a, 1966b; Bledsoe and Chan 1965). He was proud of this work, but because the funding was provided by an unnamed intelligence agency that did not allow much publicity, so little of the work was published. Given a large database of images (in effect, a book of mug shots) and a photograph, the problem was to select from the database a small set of records such that one of the image records matched the photograph. The success of the method could be measured in terms of the ratio of the answer list to the number of records in the database. Bledsoe (1966a) described the following difficulties: “This recognition problem is made difficult by the great variability in head rotation and tilt, lighting intensity and angle, facial expression, aging, etc. Some other attempts at facial recognition by machine have allowed for little or no variability in these quantities. Yet the method of correlation (or pattern matching) of unprocessed optical data, which is often used by some researchers, is certain to fail in cases where the variability is great. In particular, the correlation is very low between two pictures of the same person with two different head rotations”. This project was labeled man-machine because the human extracted the coordinates of a set of features from the photographs, which were then used by the computer for recognition. Using a graphics tablet (GRAFACON or RAND TABLET), the operator would extract the coordinates of features such as the center of pupils, the inside corner of eyes, the outside corner of eyes, point of widows peak, and so on. From these coordinates, a list of 20 distances, such as width of mouth and width of eyes, pupil to pupil, were computed. These operators could process about 40 pictures an hour. When building the database, the name of the person in the photograph was associated with the list of computed distances and stored in the computer. In the recognition phase, the set of distances was compared with the corresponding distance for each photograph, yielding a distance between the photograph and the database record. The closest records are returned. This brief description is an oversimplification that fails in general because it is unlikely that any two pictures would match in head rotation, lean, tilt, and scale (distance from the camera). Thus, each set of distances is normalized to represent the face this in a frontal orientation. To accomplish normalization, the program first tries to determine the tilt, the lean, and the rotation. Then using these angles, the computer undoes the effect of these transformations on the computed distances. To compute these angles, the computer must know the three-dimensional geometry of the head. Because the actual heads were unavailable Bledsoe (1964) used a standard head derived from measurements on seven heads. After Bledsoe left PRI in 1966, this work was continued at the Stanford Research Institute, primarily by Peter Hart. In experiments performed on a database of over 2000 photographs, the computer consistently outperformed humans when presented with the same recognition tasks (Bledsoe 1968). Peter Hart (1996) enthusiastically recalled the project with the exclamation, "It really worked!" By about 1997, the system developed by Christoph von der Malsburg and graduate students of the University of Bochum in Germany and the University of Southern California in the United States outperformed most systems with those of Massachusetts Institute of Technology and the University of Maryland rated next. The Bochum system was developed through funding by the United States Army Research Laboratory. The software was sold as ZN-Face and used by customers such as Deutsche Bank and operators of airports and other busy locations. The software was "robust enough to make 325http://sites.google.com/site/ijcsis/ ISSN 1947-5500 identifications from less-than-perfect face views. It can also often see through such impediments to identification as mustaches, beards, changed hair styles and glasses—even sunglasses". In about January 2007, image searches were "based on the text surrounding a photo," for example, if text nearby mentions the image content. Polar Rose technology can guess from a photograph, in about 1.5 seconds, why any individual may look like in three dimensions, and thought they "will ask users to input the names of people they recognize in photos online" to help build a database. II. FACIAL TECHNOLOGY AT A GLANCE Identix®, a company based in Minnesota, is one of many developers of facial recognition technology. Its software, FaceIt®, can pick someone's face out of a crowd, extract the face from the rest of the scene and compare it to a database of stored images. In order for this software to work, it has to know how to differentiate between a basic face and the rest of the background. Facial recognition software is based on the ability to recognize a face and then measure the various features of the face. (IJCSIS) International Journal of Computer Science and Information Security, Vol. 8, No. 1, 2010 These nodal points are measured creating a numerical code, called a face print, representing the face in the database. In the past, facial recognition software has relied on a 2D image to compare or identify another 2D image from the database. To be effective and accurate, the image captured needed to be of a face that was looking almost directly at the camera, with little variance of light or facial expression from the image in the database. This created quite a problem. In most instances the images were not taken in a controlled environment. Even the smallest changes in light or orientation could reduce the effectiveness of the system, so they couldn't be matched to any face in the database, leading to a high rate of failure. In the next section, we will look at ways to correct the problem. A. 3D Facial Recognition A newly-emerging trend in facial recognition software uses a 3D model, which claims to provide more accuracy. Capturing a real-time 3-D image of a person's facial surface, 3D facial recognition uses distinctive features of the face -- where rigid tissue and bone is most apparent, such as the curves of the eye socket, nose and chin -- to identify the subject. These areas are all unique and don't change over time. Using depth and an axis of measurement that is not affected by lighting, 3D facial recognition can even be used in darkness and has the ability to recognize a subject at different view angles with the potential to recognize up to 90 degrees (a face in profile). Using the 3D software, the system goes through a series of steps to verify the identity of an individual. a) Detection:- Acquiring an image can be accomplished by digitally scanning an existing photograph (2D) or by using a video image to acquire a live picture of a subject (3D). b) Alignment:- Once it detects a face, the system determines the head's position, size and pose. As stated earlier, the subject has the potential to be recognized up to 90 degrees. While with 2-D the head must be turned at least 35 degrees toward the camera. c) Measurement:- The system then measures the curves of the face on a sub-millimeter (or microwave) scale and creates a template. d) Representation:- The system translates the template into a unique code. This coding gives each template a set of numbers to represent the features on a subject's face. e) Matching :- If the image is 3D and the database contains 3D images, then matching will take place without any changes being made to the image. However, there is a challenge currently facing databases that are still in 2D images. 3D provides a live, moving variable subject being compared to a flat, stable image. New technology is addressing this challenge. When a 3D image is taken, different points (usually three) are identified. For example, the outside of the eye, the inside of the eye and the tip of the nose will be pulled out and measured. Figure 1. Face IT software compares the face print with other images in the database. (Photo Identix Inc.) Every face has numerous, distinguishable landmarks, the different peaks and valleys that make up facial features. FaceIt defines these landmarks as nodal points. Each human face has approximately 80 nodal points. Some of these measured by the software are: • Distance between the eyes • Width of the nose • Depth of the eye sockets • The shape of the cheekbones • The length of the jaw line 326http://sites.google.com/site/ijcsis/ ISSN 1947-5500 Once those measurements are in place, an algorithm (a step-by- step procedure) will be applied to the image to convert it to a 2D image. After conversion, the software will then compare the image with the 2D images in the database to find a potential match. f) Verification or Identification :- In verification, an image is matched to only one image in the database (1:1). For example, an image taken of a subject may be matched to an image in the Department of Motor Vehicles database to verify the subject is who he says he is. If identification is the goal, then the image is compared to all images in the database resulting in a score for each potential match (1:N). In this instance, you may take an image and compare it to a database of mug shots to identify who the subject is. Next, we'll look at how skin biometrics can help verify matches. (IJCSIS) International Journal of Computer Science and Information Security, Vol. 8, No. 1, 2010 The process, called Surface Texture Analysis, works much the same way facial recognition does. A picture is taken of a patch of skin, called a skin print. That patch is then broken up into smaller blocks. Using algorithms to turn the patch into a mathematical, measurable space, the system will then distinguish any lines, pores and the actual skin texture. It can identify differences between identical twins, which is not yet possible using facial recognition software alone. According to Identix, by combining facial recognition with surface texture analysis, accurate identification can increase by 20 to 25 percent. Figure 3. Working of facial recognisation (Surface texture analysis model) FaceIt currently uses three different templates to confirm or identify the subject: vector, local feature analysis and surface texture analysis. • The vector template is very small and is used for rapid searching over the entire database primarily for one- to-many searching. • The Local Feature Analysis (LFA) template performs a secondary search of ordered matches following the vector template. • The Surface Texture Analysis (STA) is the largest of the three. It performs a final pass after the LFA template search, relying on the skin features in the image, which contains the most detailed information. By combining all templates, FaceIt® has an three advantage over other systems. It is relatively insensitive to changes in expression, including blinking, frowning or smiling and has the ability to compensate for mustache or beard growth and the appearance of eyeglasses. The system is also uniform with respect to race and gender. Figure 2. Working of facial recognisation B. Biometric Facial Recognition The image may not always be verified or identified in facial recognition alone. Identix® has created a new product to help with precision. The development of FaceIt®Argus uses skin biometrics, the uniqueness of skin texture, to yield even more accurate results. 327http://sites.google.com/site/ijcsis/ ISSN 1947-5500 (IJCSIS) International Journal of Computer Science and Information Security, Vol. 8, No. 1, 2010 III. CRITICISM A. Weaknesses Face recognition is not perfect and struggles to perform under certain conditions. Ralph Gross, a researcher at the Carnegie Mellon Robotics Institute, describes one obstacle related to the viewing angle of the face: "Face recognition has been getting pretty good at full frontal faces and 20 degrees off, but as soon as you go towards profile, there've been problems." Other conditions where face recognition does not work well include poor lighting, sunglasses, long hair, or other objects partially covering the subject’s face, and low resolution images. Another serious disadvantage is that many systems are less effective if facial expressions vary. Even a big smile can render in the system less effective. For instance: Canada now allows only neutral facial expressions in passport photos. B. Effectiveness Critics of the technology complain that the London Borough of Newham scheme has, as of 2004, never recognized a single criminal, despite several criminals in the system's database living in the Borough and the system having been running for several years. "Not once, as far as the police know, has Newham's automatic facial recognition system spotted a live target." This information seems to conflict with claims that the system was credited with a 34% reduction in crime - which better explains why the system was then rolled out to Birmingham also. An experiment by the local police department in Tampa, Florida, had similarly disappointing results. "Camera technology designed to spot potential terrorists by their facial characteristics at airports failed its first major test at Boston's Logan Airport". Safehouse International Limited, an Australian company, patented software including iMotion and iCount systems. The company claimed this system were able to track moving people and calculate the number of people in a crowd. After 9/11, the software was considered “commercially attractive” by the US administration. It was later revealed by David Mapley, a US shareholder of Safehouse International Limited) that the software actually never worked. C. Privacy concerns Despite the potential benefits of this technology, many citizens are concerned that their privacy will be invaded. Some fear that it could lead to a “total surveillance society,” with the government and other authorities having the ability to know where you are, and what you are doing, at all times. This is not to be an underestimated concept as history has shown that states have typically abused such access before. D. Recent improvements In 2006, the performances of the latest face recognition algorithms were evaluated in the Face Recognition Grand Challenge (FRGC). High-resolution face images, 3-D face scans, and iris images were used in the tests. The results Figure 4. Poor lighting can make it more difficult for facial recognition software to verify or identify someone. However, it is not a perfect system. There are some factors that could get in the way of recognition, including: • Significant glare on eyeglasses or wearing sunglasses. • Long hair obscuring the central part of the face. • Poor lighting that would cause the face to be over- or under-exposed. • Lack of resolution (image was taken too far away). Identix isn't the only company with facial recognition systems available. While most work the same way FaceIt does, there are some variations. For example, a company called Animetrix, Inc. has a product called FACEngine ID® SetLight that can correct lighting conditions that cannot normally be used, reducing the risk of false matches. Sensible Vision, Inc. has a product that can secure a computer using facial recognition. The computer will only power on and stay accessible as long as the correct user is in front of the screen. Once the user moves out of the line of sight, the computer is automatically secured from other users. Due to these strides in technology, facial and skin recognition systems are more widely used than just a few years ago. In the next section, we'll look at where and how they are being used and what's in store for the future. Among the different biometric techniques facial recognition may not be the most reliable and efficient but its great advantage is that it does not require aid from the test subject. Properly designed systems installed in airports, multiplexes, and other public places can identify individuals among the crowd. Other biometrics like fingerprints, iris, and speech recognition cannot perform this kind of mass scanning. However, questions have been raised on the effectiveness of facial recognition software in cases of railway and airport security. 328http://sites.google.com/site/ijcsis/ ISSN 1947-5500 indicated that the new algorithms are 10 times more accurate than the face recognition algorithms of 2002 and 100 times more accurate than those of 1995. Some of the algorithms were able to outperform human participants in recognizing faces and could uniquely identify identical twins. Low-resolution images of faces can be enhanced using face hallucination. Further improvements in high resolution, megapixel cameras in the last few years have helped to resolve the issue of insufficient resolution. E. Future development A possible future application for facial recognition systems lies in retailing. A retail store (for example, a grocery store) may have cash registers equipped with cameras; the cameras would be aimed at the faces of customers, so pictures of customers could be obtained. The camera would be the primary means of identifying the customer, and if visual identification failed, the customer could complete the purchase by using a PIN (personal identification number). After the cash register had calculated the total sale, the face recognition system would verify the identity of the customer and the total amount of the sale would be deducted from the customer's bank account. Hence, face-based retailing would provide convenience for retail customers, since they could go shopping simply by showing their faces, and there would be no need to bring debit cards, or other financial media. Wide-reaching applications of face-based retailing are possible, including retail stores, restaurants, movie theaters, car rental companies, hotels, etc.e.g. Swiss European surveillance: facial recognition and vehicle make, model, color and license plate reader. IV. SCOPE IN INDIA 1. In order to prevent the frauds of ATM in India, it is recommended to prepare the database of all ATM customers with the banks in India & deployment of high resolution camera and face recognition software at all ATMs. So, whenever user will enter in ATM his photograph will be taken to permit the access after it is being matched with stored photo from the database. 2. Duplicate voter are being reported in India. To prevent this, a database of all voters, of course, of all constituencies, is recommended to be prepared. Then at the time of voting the resolution camera and face recognition equipped of voting site will accept a subject face 100% and generates the recognition for voting if match is found. 3. Passport and visa verification can also be done using face recognition technology as explained above. 4. Driving license verification can also be exercised face recognition technology as mentioned earlier. 5. To identify and verify terrorists at airports, railway stations and malls the face recognition technology will be the best choice in India as compared with other biometric technologies since other technologies cannot be helpful in crowdy places. (IJCSIS) International Journal of Computer Science and Information Security, Vol. 8, No. 1, 2010 6. In defense ministry and all other important places the face technology can be deployed for better security. 7. This technology can also be used effectively in various important examinations such as SSC, HSC, Medical, Engineering, MCA, MBA, B- Pharmacy, Nursing courses etc. The examinee can be identified and verified using Face Recognition Technique. 8. In all government and private offices this system can be deployed for identification, verification and attendance. 9. It can also be deployed in police station to identify and verify the criminals. 10. It can also be deployed vaults and lockers in banks for access control verification and identification of authentic users. 11. Present bar code system could be completely replaced with the face recognition technology as it is a better choice for access & security since the barcode could be stolen by anybody else. V. CONCLUSIONS Face recognition technologies have been associated generally with very costly top secure applications. Today the core technologies have evolved and the cost of equipments is going down dramatically due to the integration and the increasing processing power. Certain applications of face recognition technology are now cost effective, reliable and highly accurate. As a result there are no technological or financial barriers for stepping from the pilot project to widespread deployment. Though there are some weaknesses of facial recognition system, there is a tremendous scope in India. This system can be effectively used in ATM’s ,identifying duplicate voters, passport and visa verification, driving license verification, in defense, competitive and other exams, in governments and private sectors. Government and NGOs should concentrate and promote applications of facial recognition system in India in various fields by giving economical support and appreciation. ACKNOWLEDGMENT We are thankful to Hon. Ashokrao Chavan (Chief Minister, Maharashtra) India, Society members of Shri. Sharada Bhawan Education Society, Nanded. Also thankful to Shri. Jaiprakash Dandegaonkar (Ex-State Minister, Maharashtra), Society members of Bahiri Smarak Vidyalya Education Society, Wapti for encouraging our work and giving us support. Also thankful to our family members and our students. REFERENCES [1] R. Brunelli, Template Matching Techniques in Computer Vision: Theory and Practice, Wiley, ISBN 978-0-470-51706-2, 2009 ([1] TM book) [2] Williams, Software". Mark. "Better Face-Recognition http://www.technologyreview.com/Infotech/18796/?a=f. Retrieved 2008-06-02. 329http://sites.google.com/site/ijcsis/ ISSN 1947-5500 [3] Kimmel, recognition". face "Three-dimensional Ron. http://www.cs.technion.ac.il/~ron/PAPERS/BroBroKimIJCV05.pdf. Retrieved 2005-01-01. [4] McNealy, Dead". (Virtually) is "Privacy Scott. http://www.jrnyquist.com/aug20/privacy.htm. Retrieved 2006-12-24. [5] "Mexican Government Adopts FaceIt Face Recognition Technology to Eliminate Duplicate Voter Registrations in Upcoming Presidential Election". Business Wire. http://findarticles.com/p/articles/mi_m0EIN/is_2000_May_11/ai_62019 954. Retrieved 2008-06-02. [6] House, David. "Facial recognition at DMV". Oregon Department of Transportation. http://www.oregon.gov/ODOT/DMV/news/cards_facialrec.shtml. Retrieved 2007-09-17. "Oregon DMV is going to start using “facial recognition” software, a new tool in the prevention of fraud, required by a new state law. The law is designed to prevent someone from obtaining a driver license or ID card under a false name." [7] Schultz, Zac. "Facial Recognition Technology Helps DMV Prevent Identity Theft". WMTV News, Gray Television. http://www.nbc15.com/news/headlines/2684991.html. Retrieved 2007- 09-17. "Madison: ...The Department of Motor Vehicles is using... facial recognition technology [to prevent ID theft]" [8] "Help find Madeleine McCann". Child Exploitation and Online Protection Centre. 2007-05-21. http://www.madeleine.ceopupload.com/. Retrieved 2007-05-21. [9] Brown, David. "We will travel anywhere to find Madeleine, say parents". London: Times Online. http://www.timesonline.co.uk/tol/news/world/europe/article1826735.ece . Retrieved 2008-06-02. [10] http://www.ppt.gc.ca/cdn/photos.aspx?lang=eng[dead link] [11] Meek, James (2002-06-13). "Robo cop". UK Guardian newspaper. http://www.guardian.co.uk/Archive/Article/0,4273,4432506,00.htm. [12] Krause, Mike (2002-01-14). "Is face recognition just high-tech snake oil?". Enter Stage Right. http://www.enterstageright.com/archive/articles/0102/0102facerecog.ht m. [13] "Birmingham City Centre CCTV Installs Visionics' FaceIt". Business Wire. 2008-06-02. http://www.allbusiness.com/government/government- bodies-offices-regional/6111139-1.html. [14] Krause, Mike (2008-06-02). "Is face recognition just high-tech snake oil?". Enter Stage Right. http://www.enterstageright.com/archive/articles/0102/0102facerecog.ht m. [15] Willing, Richard (2003-09-02). "Airport anti-terror systems flub tests; Face-recognition technology fails to flag 'suspects'" (Abstract). USA Today. http://pqasb.pqarchiver.com/USAToday/access/391952771.html?dids=3 91952771:391952771&FMT=ABS&FMTS=ABS:FT&date=Sep+2%2C +2003&author=Richard+Willing&pub=USA+TODAY&edition=&startp age=A.03&desc=Airport+anti-terror+systems+flub+tests+%3B+Face- recognition+technology+fails+to+flag+%27suspects%27. Retrieved 2007-09-17. [16] The Age (Melbourne, Australia) September 28, 2007 “Safehouse finds that America bites” [17] "Civil Liberties & Facial Recognition Software". About.com, The New York Times Company. pp. pp. 2. Archived from the original on 2006- 03-01. http://web.archive.org/web/20060301220151/terrorism.about.com/od/civ illibertiesissues/i/facialrecsoft_2.htm. Retrieved 2007-09-17. "A few examples which have already arisen from surveillance video are: using license plates to blackmail gay married people, stalking women, tracking estranged spouses..." [18] ""Mugspot" Can Find A Face In The Crowd -- Face-Recognition Software Prepares To Go To Work In The Streets". ScienceDaily. 12 November 1997. http://www.sciencedaily.com/releases/1997/11/971112070100.htm. Retrieved 2007-11-06. [19] http://www.military-information- technology.com/article.cfm?DocID=1280 (IJCSIS) International Journal of Computer Science and Information Security, Vol. 8, No. 1, 2010 [20] Electronic Privacy Information Center. "Face Recognition." January 19, 2006. http://www.epic.org/privacy/facerecognition/ [21] Gupta, A. "Biometrics: Eyes Don't Lie." DataQuest. October 14, 2006. http://www.dqindia.com/content/industrymarket/focus /2006/106101402.asp [22] Identix. "FaceIt® G6 Frequently Asked Technical Questions" http://www.identix.com/products/pro_faceit.html AUTHORS PROFILE Dr.S.B.THorat M.E. (Computer Science & Engg.) M.Sc. (ECN), AMIE, LM-ISTE, Ph.D. (Comp.Sc. & Engg.) He is having 24 years teaching experience. From 2001 he is working as a Director, at ITM. He is Dean of faculty of Computer studies at Swami Ramanand Teerth Marathwada University, Nanded (Maharashtra). Recently he is completed his Ph.D. He attended many national and International conferences. He is having 8 international publications. His interested area are AI, Neural network, Data mining, Fuzzy systems, Image processing. S.K.Nayak M.Sc. (Computer Science), D.B.M, B.Ed. He completed M.Sc. (Computer Science) from S.R.T.M.U, Nanded. In 2000 he joined as lecturer in Computer Science at Bahirji Smarak Mahavidyalaya, Basmathnagar. From 2002 he is acting as a Head of Computer Science department. He is doing Ph.D. He attended many national and international conferences, workshops and seminars. He is having 7 international publications. His interested areas are ICT, Rural development, Bioinformatics. Miss.Jyoti P Dandale B.E. (Computer Science & Engg.) She has comleted BE from SSGMCE Shegaon. Since 2 years,She has been working as a lecturer. She has Presented Paper in International conference. She is having 1 international publication. Her interested areas are data and internet security. 330http://sites.google.com/site/ijcsis/ ISSN 1947-5500
1209.0935	1	1209	2012-09-05T11:48:06	Characterizing Successful Formulas: the Multi-agent Case	[ "cs.MA", "cs.LO" ]	Characterization of successful formulas in Public Announcement Logic (PAL) is a well known open problem in Dynamic Epistemic Logic. Recently, Holliday and ICard have given a complete characterization for the single agent case. However, the problem for the multi-agent case is open. This paper gives a partial solution to the problem, characterizing the subclass of the language consisting of unary operators, and discusses methods to give a complete solution.	cs.MA	cs	Characterizing Successful Formulas: the Multi-agent Case Sanchit Saraf and Sumit Sourabh Institute for Logic, Language and Computation, Universiteit van Amsterdam, P.O. Box 94242, 1090 GE Amsterdam, The Netherlands. {s.saraf, s.sourabh}@uva.nl Abstract. Characterization of successful formulas in Public Announcement Logic (PAL) is a well known open problem in Dynamic Epistemic Logic. Recently, Holliday and ICard have given a complete characterization for the single agent case in [6]. However, the problem for the multi-agent case is open. This paper gives a partial solution to the problem, character- izing the subclass of the language consisting of unary operators, and discusses methods to give a complete solution. 1 Introduction The logic of Public Announcements is the simplest form of S5 dynamic epistemic logic, augmenting standard epistemic logic with public announcement operator. It was formulated and axiomatised without the common knowledge operator by Plaza in [8]. The axiomatisation of Public announce- ment logic (PAL) with the common knowledge operator was given by Baltag, Moss and Solecki [3]. In the same paper, the authors show that PAL is not strongly complete because of infinitary nature of the common knowledge operator. For a detailed account on PAL, one can refer to [13]. The notion of a successful update was given by Gebrandy [4] and van Ditmarsch [12]. The formulas which remain true after being announced are called successful formulas. An interesting open problem in PAL is the syntactic characterization of successful formulas [10,11,9,4,13,14,2]. A classic example of a formula which is not successful is the Moore Sentence p∧¬Kp [7], which can be read as “p is true but you do not know p”. The Moore sentences have been analysed by Hintikka in his classical monograph [5]. Their relevance has been extensively studied by van Ditmarsch and Kooi in their paper [14]. Successful formulas have important applications in many security protocols. Together with its practical usefulness, the task of characterizing successful formulas independently presents itself as an interesting mathematical problem. The aim of this paper is to present a partial solution to the problem for the multi-agent case. Other solutions have been proposed, most notably by Holliday and Icard in their recent paper [6], where they completely solve the problem for the single agent case. In [6], it is also shown that for a single agent, the source of failure is Moorean in nature, which implies that unsuccessful formulas contain at least one binary operator. In contrast, for the multi-agent case, formulas with only unary operators can also be unsuccessful. The simplest example of such a formula is KaLbp. The full syntactic characterization of successful formulas in the multi-agent case is a difficult task and not a simple generalization of the single agent case. We give a characterization for the successful formulas in the fragment Lsterm (for multiple agents) which we call single term formulas in our notation. The formulas in Lsterm are terms without any binary connectives and are inductively defined as φ := p ¬φ Kiφ where i ∈ I is the set of agents, p ∈ Prop is the the set of propositional letters and Kiφ is interpreted as agent i knows φ. We further classify single term formulas into simple single term and compound single term formulas to distinguish between single or multiple occurrences of an epistemic operator Ei corresponding to an agent i. We give a few examples to motivate why we need separate analysis for the compound single term formulas. We have also considered the fragment Lmterm, where we allow for binary connectives. We present some preliminary results on characterization of conjunctions of single term formulas. We also have some general results which connect the class of successful formulas with other known classes of formulas, such as self refuting formulas [6]. Our work is relevant as it provide insights to the nature and complexity of the problem in hand. The full characterization for the multi-agent case is still open and we briefly discuss possible ways to go about for solving the problem. The organization of the paper is as follows: In Section 2 we state the preliminaries and previous work. We present our characterization results on single term formulas Lsterm in Section 3. In Section 4, we consider the multiple term language Lmterm and present the characterization results for conjunctions of Lsterm formulas. In Section 5, we present results which shed more light on the properties of successful formulas and their connections with other known classes of formulas. We conclude the paper in Section 5, discussing a possible approach to solve the general problem. 2 Preliminaries and Previous Work In this section, we present the syntax and semantics of Public Announcement logic as given in [13]. We also define successful formulas and list the existing results on characterization of successful formulas. The syntax of PAL (LPAL) is given as follows: φ := p ¬φ φ ∧ φ φ ∨ φ Kiφ Cφ [φ] ψ where p ∈ Prop is the set of propositional letters, Kiφ is interpreted as agent i ‘knows’ φ and Cφ is interpreted as, it is common knowledge that φ. We use the notation [φ]ψ for saying that ψ is true after φ is announced. An epistemic model is given by the triple (W, Ri, V ) where W is the set of worlds and Ri ⊆ W × W is the accessibility relation between the worlds for each agent i ∈ I, with I being the index set of agents. The map V : Prop → P(W ) is the valuation function specifying which propositional letters are true at a world w ∈ W . Since we restrict ourselves to the S5 case, we can assume that the relations Ri, for all agents i ∈ I are equivalence relations. We use RI to denote the reflexive, i∈I Ri)∗. Given a valuation V , we define the truth of a formula φ in a world w, denoted by M, w = φ inductively below, transitive closure of union of all the relations Ri, RI = ((cid:83) iff w ∈ V (p) iff M, w (cid:50) φ M, w = p M, w = ¬φ M, w = φ ∧ ψ iff M, w = φ and M, w = ψ M, w = φ ∨ ψ iff M, w = φ or M, w = ψ M, w = Kiφ iff ∀t ∈ W s.t. wRit ⇒ M, t = φ M, w = Cφ iff ∀t ∈ W s.t. wRI t ⇒ M, t = φ M, w = [φ]ψ iff M, w = φ ⇒ Mφ, w = ψ where Mφ = (W (cid:48), R(cid:48) defined as W (cid:48) = {w ∈ W M, w = φ}, R(cid:48) i, V (cid:48)) is the restriction of the model to the worlds where φ is true, and is We use Liφ to denote the dual of Kiφ, that is, Liφ = ¬Ki¬φ and it is interpreted as agent i i = Ri ∩ (W (cid:48) × W (cid:48)) and V (cid:48)(p) = V (p) ∩ W (cid:48). considers it possible that φ. Definition 2.1 (Successful formulas). A formula φ is successful in PAL in S5 iff [φ]φ is valid. In other words, M, w = φ implies Mφ, w = φ. Example 2.1. The Moore sentence p ∧ ¬Kp is a familiar example of an unsuccessful formula. We have the following model to illustrate why it is unsuccessful. Suppose we have two agents Ann and Bob. There is a butterfly on Bob’s head but he doesn’t know it, although Ann can see the butterfly. Let p denote the sentence “There is a butterfly on Bob’s head” which is true at the actual world w2. Since Bob does not know if there is a butterfly on his head, he cannot distinguish between the worlds w1 and w2. The models below represent the epistemic situation before and after the announcement. Ra, Rb Ra, Rb Rb w1 w2 p Ra, Rb w2 p Fig. 1. Models before and after the announcement of p ∧ ¬Kbp Before the announcement, M, w2 = p ∧ ¬Kbp. After Ann makes the announcement, “There is a butterfly on your head and you don’t know it”, the model changes to the one on the right, where Bob now knows that he has a butterfly on his head. The formula p ∧ Kbp which is announced, is no longer true in the model on the right at w2, and is therefore unsuccessful. The following result by van Benthem et al. [1,15] gives an immediate subclass of formulas which are successful. Theorem 2.1. A formula is preserved under sub-models (of all relational models) iff it is equiv- alent (in K) to a universal formula. A universal formula in S5 is any formula which can be constructed by p,∧,∨ and K. This proves that the following sub-fragment of PAL is successful, which we refer to as Lsuc φ := p ¬φ Kiφ φ ∧ φ φ ∨ φ [¬φ]φ Other than this, [13] also lists individual formulas, for instance ¬Kap, which are successful. The complexity of judging a formula to be successful for multiple agents is shown to be PSPACE- complete in [6]. In the same paper, a complete characterization for successful and self-refuting formulas for S5 dynamic epistemic logic is also proposed for the single agent case. The authors identify the source of all unsuccessful formulas as being a Moorean-sentence, and all self-refuting formulas as being a Moore-sentence. In addition, they define a super-successful formula as below. Definition 2.2 (Super-successful formulas). A formula φ is super-successful iff given any M, for all M (cid:48) ⊆ M if M, w = φ then M such that Mφ ⊆ M (cid:48) (cid:48) , w = φ. It was shown in [6] that super-successful formulas are closed under disjunction, but in general, successful formulas are not, which is an important result. 3 Characterization of Lsterm We consider the subclass Lsterm of single term formulas in PAL. The formulas in Lsterm are inductively defined as φ := p ¬φ Kiφ Our choice of the subclass is appropriate in the sense that it comprises of a basic language which can be used as a building block for the complete LPAL. We use an operator variable Ei to stand for Ki or Li in the description of formula forms. A formula involving operators with numeric subscripts has operators of only one type. For instance, K1 . . . Knφ denotes that there are exactly n, K operators and no L operator. We will work with formulas in negation normal form. We use α, β for denoting propositional formulas (without any epistemic operators) and φ, ψ are used to denote the formulas with epistemic operators. Definition 3.1 (Single term formula). A formula in negation normal form is single term, if it is of the form E1 . . . Enα, where Ei is either Ki or Li and α is a propositional formula. We now present the characterization results after the above mentioned notations. It is easy to see that any single term formula E1 . . . Enα where α is a contradiction or a tautology is a successful formula. 3.1 Simple single term formulas We first give a characterization for a simplified form of the single termed formulas. Definition 3.2 (Simple single term formula). A single term formula E1 . . . Enα is said to be simple if for any agent i ∈ I, where I is the index set for the set of agents, Ei occurs at most once in E1 . . . En. We further classify the simple single term formulas into K-simple single term formulas which have only Ki as the epistemic operators and L-simple single term formulas which have only Li as the epistemic operators. The characterization of successful formulas is easy to see in both these cases and is given by the following proposition. Proposition 3.1. K-simple single term and L-simple single term formulas are successful. Proof. The K-simple single term formulas with only Ki operators can be seen as a subclass of Lsuc formulas, which we know are successful from Theorem 2.1. In case of L-simple single term formulas, suppose M, w1 = L1L2 . . . Lnα. It gives us a chain of related worlds w1, . . . , wn+1 such that w1R1w2R2w3 . . . wnRnwn+1 and M, wn+1 = α. We know that the frame is reflexive so L1L2 . . . Lnα is true at all the worlds in the chain, and no world is deleted after the public announcement of L1L2 . . . Lnα. As a result, ML1L2...Lnα, w1 = L1L2 . . . Lnα showing that L- simple single term formulas are successful. We next define formulas which have both L and K epistemic operators and present characterization results for them. Definition 3.3 (KL-simple single term formula). An KL-simple single term formula is a formula in which there exists at least one L operator in the scope of a K operator. Example 3.1. As a simple example, the formula K1L2L3K4L5α is KL-simple single term, since the operator L2 is in the scope of K1. The formula L1L2K2 is not KL-simple single term, since there is no L operator in the scope of the K2 operator. We would like to stress the fact that the order of K and L operators in the formula does not make a difference as long as there is an L operator in the scope of some K operator. Proposition 3.2. KL-simple single term formulas are unsuccessful. Proof. In order to prove that KL-simple single termed formulas are unsuccessful, we first observe that for the simple case of 2 agents, the formula K1L2α is unsuccessful. Consider the following counter-model (for the sake of clarity, we omit the reflexive arrows for each of the agents at every world ) R1 R2 w2 R1 w4 w3 p w1 p R1 w2 w1 p Fig. 2. Models before and after the announcement of K1L2p We need to show that M, w1 = K1L2p and MK1L2p, w1 = ¬K1L2p. It can be easily checked that the formula K1L2p is true at the worlds w1 and w2 but false at w3 and w4. Therefore after the public announcement of the formula K1L2p, the worlds where the formula is not true get deleted and we get the model on the right. In the updated model MK1L2p, the formula does not hold at w1, that is, MK1L2p, w1 = ¬K1L2p which shows that K1L2p is unsuccessful. The above argument for the simple case easily generalizes to any KL-simple single term for- mula. Consider the KL-simple single term formula K1 . . . L2 . . . α where K1 is the first K operator and L2 is the first L operator in the formula. We use the notation K1XL2Y α for the formula K1 . . . L2 . . . α, where X and Y are series of K and K, L epistemic operators respectively, in any arbitrary order. In order to show that K1 . . . L2 . . . α is unsuccessful, it suffices to use the same counter-model that we have above for the formula K1L2α. The reflexivity of the frame makes the formula K1XL2Xα true at the world w1 and w2 but false at w3 and w4, irrespective of the form of X and Y . As a result after the announcement of the formula K1XL2Y α, the model reduces to the one on the right where we have MK1XL2Y α, w1 = ¬K1XL2Y α, proving that K1XL2Y α is unsuccessful . Definition 3.4 (LK-simple single term formula). An LK-simple single term formula is one which begins with L operators and does not have any L operator in the scope of a K operator. Example 3.2. The simplest LK simple single term formula would be L1K2p with two epistemic operators. A more extensive example would be the formula L1 . . . LmKm+1 . . . Knα. It is easy to observe that in general, any LK simple single term formula will have a series of L operators followed by a series of K operators, because of the restriction that we cannot have an L operator in the scope of a K operator. Proposition 3.3. LK-simple single term formulas are successful. Proof. We first show that the LK-simple single term formula L1K2α is successful. Suppose, M, w1 = L1K2p which implies ∃ w2, such that w1R1w2 and M, w2 = K2p. But since the frame is reflexive, we have w2R1w2 and therefore M, w2 = L1K2p. As a result, w2 ∈ ML1K2p which would make L1K2p true at w1 in ML1K2p since the relations are preserved under sub-models, thus proving that L1K2p is successful. In order to prove that any LK-simple single term formula is successful, we use a similar argument as above. Consider the formula L1 . . . LmKm+1 . . . Knα which is true at a world w1 in a model M . Since M, w1 = L1 . . . LmKm+1 . . . Knα, ∃w2 such that w1R1w2 and M, w2 = L2 . . . LmKm+1 . . . Knα. We can repeat the same argument to get a chain of related worlds, w1R1w2R2w3 . . . wmRmwm+1 such that M, wm+1 = Km+1 . . . Knα. Using reflexivity of Ri for all i ∈ I , we can show that all the worlds in the chain w1R1w2R2w3 . . . wmRmwm+1, will be present in the model after the announcement of the formula L1 . . . LmKm+1 . . . Knα, since the for- mula is true in all the worlds connected to w1 in the chain. Therefore, ML1...LmKm+1...Knα, w1 = L1 . . . LmKm+1 . . . Knα, proving that LK-simple single term formulas are successful. A nice property of this class is that all formulas which are successful are also super-successful. Thus, they will be closed under disjunction. 3.2 Compound single term formulas In this section we present characterization results for compound single term formulas where we allow multiple occurrences of an epistemic operator Ei within a formula. Definition 3.5 (Compound single term formulas). A single termed formula E1 . . . Enα is said to be compound, if there is at least an agent i ∈ I, where I is the index set for the set of agents, such that Ei occurs more than once in the formula. We generalize the definition of K and L-simple single term formulas to the compound case, as K and L-compound simple term formulas, by allowing multiple occurrences of Ei for agents i ∈ I. Proposition 3.4. K-compound single term and L-compound single term formulas are successful. Proof. The proof for the K-compound single term easily follows from the fact that they form a subclass of Lsuc which are successful. For L-compound single term formulas, the proof is identical as in the case of simple formulas. Since the frame is reflexive, M, w = L1L1α implies M, w = L1α, so any multiple occurrences of epistemic operators occurring together can be reduced to a single occurrence. In case of multiple occurrence of epistemic operators not occurring together, we can use the same argument as in the proof of proposition 3.1. The definition of the KL-simple single term formulas can be generalized to the setting of compound formulas by allowing multiple occurrences of epistemic operators corresponding to an agent. Unlike the simple formula case, where we have a single characterization result for all the KL-simple single term formulas, Proposition 3.2 does not hold for KL-compound simple term formulas. While we don’t have a complete characterization of the KL and LK-compound single term formulas, we present a few examples to show that some of the results for the simple formulas do not generalize to the compound formulas, which motivates separate and more general characterization results. The following proposition shows that the formula K1L2K1p is successful, which would otherwise have been classified as unsuccessful in the simple single term case. Proposition 3.5. The compound single term formula K1L2K1α is successful. Proof. Suppose M, w = K1L2K1α which implies that ∀w2 such that w1R1w2, M, w2 = L2K1p. We want to show that M, w2 = L2K1α. Consider an arbitrary w(cid:48) such that w2R1w(cid:48). Since w1R1w2, by transitivity we have w1R1w(cid:48) which makes L2K1α true at w(cid:48), and therefore M, w2 = K1L2K1α. At w2, we have M, w2 = L2K1α, which implies ∃w3 s.t. w2R2w3 and M, w3 = K1α. Now using similar reasoning as above for w2, using transitivity of R1 we can show that M, w3 = K1L2K1α. Therefore, both w2 and w3 belong to the model MK1L2K1α, after the announcement of K1L2K1α which proves MK1L2K1α, w1 = K1L2K1α. The generalization of above example to the case where we can have any number of epistemic operators and a characterization result for a sub-class of KL- compound formulas is quite involved and beyond the scope of this paper. Next, we have an example of the formula K1K2L1p, which is unsuccessful as it would have been in the simple formula case, but the counter-model which we used earlier, doesn’t suffice for this formula. This shows another deviation from the characterization in case of simple formulas. Proposition 3.6. The compound single term formula K1K2L1α is unsuccessful. Proof. It is easy to check that the counter-model in Figure 1 does not work for the formula K1K2L1α, since it is true at all the worlds in the model and therefore no world is deleted from the model after the announcement of the formula. We extend the counter-model presented earlier so that it makes K1K2L1α unsuccessful. R1 R2 w2 w3 R1 R1 w5 R2 w6 w4 α w1 α R1 w2 w1 α Fig. 3. Models before and after the announcement of K1K2L1α We give an example of a formula L1K2K3L1α, beginning with an L operator, which would have been classified as a KL-simple single term formula earlier and therefore unsuccessful, but in the compound case, it is successful. This further motivates the need for a separate characterization result for the compound case. Proposition 3.7. The formula L1K2K3L1α is successful. Proof. The proof of this proposition is similar to proposition 3.5 and uses the idea that all the worlds make the formula true and are therefore contained in the sub-model. Suppose M, w1 = L1K2K3L1α, ⇒ ∃w2, s.t w1R1w2, M, w2 = K2K3L1α. But the frame is reflexive, so M, w2 = L1K2K3L1α, and therefore w2 ∈ ML1K2K3L1α. Now, ∀w3 s.t. w2R2w3, M, w3 = K3L1α. We can use the same argument as in the proof of proposition 3.5 to show that M, w3 = L1K2K3L1α, so w3 ∈ ML1K2K3L1α. Since, M, w3 = L1K2K3L1α, ⇒ ∃w4, s.t w3R1w4, M, w4 = K2K3L1α. We leave it to the reader to check that, M, w3 = L1K2K3L1α iff M, w4 = L1K2K3L1α. Once we have shown that, w4 ∈ ML1K2K3L1α which implies ML1K2K3L1α, w1 = L1K2K3L1α, thus proving L1K2K3L1α is successful. 4 Characterization of Lmterm In this section we present characterization results for the formulas in Lmterm which includes Boolean combinations of simple single term formulas. We know from Section 3.1 that the KL- simple single term formulas are unsuccessful formulas. The proposition below generalizes the result to any number of Boolean conjunctions of unsuccessful formulas. Proposition 4.1. If φ and ψ are unsuccessful simple single term formulas, their conjunction φ∧ψ is also unsuccessful. Proof. We showed in proposition 3.2 that KL-simple single term formulas are unsuccessful. It is easy to see that a conjunction of two simple single term KL formulas will be unsuccessful. The counter-model for the conjunction will be the model which consists of the counter-models for each of the individual unsuccessful formulas sharing the real world as the common world. Since we are in the simple single term case, this counter-model is sufficient since φ and ψ share no common epistemic operators and therefore the two counter-models corresponding to them will have no interaction. The conjunction of a successful and an unsuccessful formula is unsuccessful as expected. Proposition 4.2. If φ is successful and ψ is unsuccessful, their conjunction φ∧ ψ is unsuccessful. Proof. In order to show that φ ∧ ψ is unsuccessful at a world w ∈ W , we use the counter-model starting at w for proving that ψ is unsuccessful and make φ true w. This is possible as long as the non epistemic parts of φ and ψ don’t depend on one another. The analysis of conjunction of two successful formulas is more complicated and involves a number of cases and their success or failure depends on the non-epistemic parts of the formula. Recalling from Section 3.1, any simple single term successful formula is of the form α, Kα, Lα, LKα, where K and L are series of K and L epistemic operators corresponding to different agents having at least 2 epistemic operators 1, and α is a propositional formula. without epistemic operators and K and K(cid:48) are series of K epistemic operators. Proposition 4.3. The conjunction of simple single term successful formulas is successful or un- successful subject to the following conditions 1. The conjunctions α ∧ β, α ∧ Kβ and Kα ∧ K(cid:48)β are successful where, α and β are formulas 2. The conjunction α ∧ Lβ is successful iff α → β. 3. The conjunction α ∧ LKβ is unsuccessful. 4. The conjunction Kα ∧ Lβ is successful iff α → β. 5. The conjunctions Kα ∧ LK(cid:48)β, Lα ∧ L(cid:48)β, Lα ∧ L(cid:48)Kβ and LKα ∧ L(cid:48)K(cid:48)β are unsuccessful. Proof. The proof for 1 is trivial since we know from Theorem 2.1 that universal formulas are preserved under sub-models. For 2, if we assume α → β, then proving α ∧ Lβ is successful is easy. To see why assume M, w = α ∧ Lβ. Since, α is a propositional formula it will be preserved in every sub-model M(cid:48) ⊆ M and therefore in particular, Mα∧Lβ, w = α. But, α → β and the frame is reflexive so we have Mα∧Lβ, w = α∧ Lβ. The converse direction can be proved using a contrapositive argument. Suppose α (cid:57) β, we can construct a counter-model to show α ∧ Lβ, where L = L1L2 . . . Ln, is 1 for formulas with a single L and K, the characterization may differ for some cases, for eg. α ∧ L1Kβ is successful iff β → α, L1α ∧ (cid:32)L2β is successful iff α ↔ β and α ∧ L1β is successful iff α → β or β → α unsuccessful in the following way. Let w1R1w2R2 . . . wnRnwn+1 be a set of related worlds. We make α true at only w1 and false at all other worlds and β true only at wn+1. After the announcement of α∧ Lβ, only the world w1 will remain in the sub-model making the formula α∧ Lβ false at w1. For 3, we can construct a counter-model in the same way as above. Assume, M1 = α∧ LKβ. We have LK = L1 . . . LmKm+1 . . . Kn. So, there exists a chain of related worlds, w1R1w2R2 . . . wnRnwn+1 such that M, wn+1 = Km+1 . . . Knα. In order to have a counter-model, we make α true only at w1 and false at all other worlds and β false at a w(cid:48) related to w. One can check that irrespective of α ↔ β, α ∧ LKβ is unsuccessful. We leave the proof of 4, which is similar to 2, and of 5 which is similar to 2 to the reader. 5 Other results The following section is a mixed bag of auxiliary results relating to successful formulas which might come in handy for further analysis of different classes. The following theorem relates to the class of successful and super-successful formulas. Theorem 5.1. The following class of S5- PAL formulas are truth-preserved under super-models: φ := p ¬p φ ∧ ψ φ ∨ ψ Laφ ¬[φ]¬ψ (cid:48) , w = φ and M (cid:48) (cid:48) (cid:48) , w = Laφ. , w = ψ which shows M Proof. We prove the above result by induction on the complexity of the formula φ. Consider the case, φ = p and assume M, w = φ . We know that the truth of a propositional formula is local, that is, depends only on the current state, so any super-model of M will contain w and hence the statement is true for any propositional formula. If M, w = φ ∧ ψ then M, w = φ and M, w = ψ. By induction hypothesis, both φ and ψ are true , w = φ∧ ψ. The in any super-model M(cid:48) of M . Hence, M proof is similar as above for φ ∨ ψ. If M, w = Laφ, then M, w(cid:48) = φ where wRaw(cid:48). Thus M(cid:48), w(cid:48) = φ for a super-model M(cid:48) of M by induction hypothesis which shows M If M, w = ¬[φ]¬ψ, then M, w = φ and Mφ, w (cid:54)= ¬ψ holds, i.e. Mφ, w = ψ. Now, by induction hypothesis, M(cid:48), w = φ. Consider, Mφ, M (cid:48)φ, and a world s ∈ Mφ. Assume M, s = φ, which gives us M(cid:48), s = φ by induction hypothesis, and therefore s ∈ M(cid:48)φ. This shows that Mφ ⊆ M(cid:48)φ, that is, M(cid:48)φ is a super-model of Mφ. Hence, by induction hypothesis, M(cid:48)φ, w = ψ which finally proves M(cid:48), w = ¬[φ]¬ψ. The above result implies that if any successful formula belongs to this class, it must be super- successful, as M, w = φ ⇒ Mφ, w = φ and by the above formula any M(cid:48) such that Mφ ⊆ M(cid:48), and M(cid:48), w = φ. We have seen that the class of self-refuting formulas is another class of formulas other than successful formulas which are interesting. Definition 5.1 (Self-refuting formulas). A formula is self-refuting iff [φ]¬φ is valid. The following theorem links the two classes of formulas. Theorem 5.2. A formula in S5-PAL is a contradiction iff it is both successful and self-refuting. Proof. It trivially follows from the definition, that a contradiction is both successful and self refuting. For the converse, suppose φ is both successful and self-refuting. Then [φ]φ and [φ]¬φ are valid. Suppose for a given pointed model (M, w), if we have M, w = φ then Mφ, w = φ and Mφ, w = φ, which is a contradiction to our initial assumption. Therefore, M, w (cid:50) φ which shows φ is a contradiction. In [6], it has been shown that successful formulas are not closed under disjunction for the single agent case. We have a result along similar lines for the closure under L operator. Theorem 5.3. The class of successful formulas is not preserved under L operator in the multi- agent case. Proof. L1KaKbL1p is successful, while L2L1KaKbL1p is not, the proof of which is given by the counter-model in the appendix. 6 Discussion for the General Case and Conclusion One can see that the work we have presented in this paper opens up new directions to be explored. We list a few questions answering which, may help to give a complete characterization. We have seen in Section 3.2 that the characterization for the compound single term formulas is quite involved and does not follow as a generalization of the simple single term formulas. We have some preliminary results regarding their characterization which we have not presented in this paper. The idea is to have additional conditions on KL and LK simple single term formulas which allows us to have their complete characterization. We don’t have any results on the compound multiple term formulas involving boolean connectives, which would be interesting to look into. We have seen that the multiple agent scenario is complicated even for single terms as opposed to single agent case, where single terms are always successful. Recursively combining the single terms using conjunction or disjunction and then binding the whole formula within an epistemic operator may result in formulas of increasing complexity. The way out may be finding a “normal form” in which the formula can be expressed in an equivalent conjunctive normal form (c.n.f.) or disjunctive normal form (d.n.f.). Alternatively, as a weaker attempt, we may be able to find a class of formulas which in spite of not being logically equivalent, can only be successful iff the original formula is successful. We believe that such a reduction algorithm would be of great help in avoiding the complex cases arising out of Boolean combinations of formulas. In a nutshell, a possible way of approaching the task of syntactic characterization could be: 1. Finding a normal form of the formulas which preferably are in c.n.f or d.n.f of single-term formulas 2. Propose a way to classify the formulas thus obtained from 1. Our classification above proceeds in direction of achieving 2. Combining the ideas and results, and those in [6] for single-agent classification, we might be able to achieve 2. But whether 1 holds or not is something which is unknown to us at this stage and may be very important with respect to the difficulty of solving the problem of characterizing successful formulas in PAL. References 1. H. Andr´eka, I. N´emeti, and J. van Benthem. Modal languages and bounded fragments of predicate logic. Journal of Philosophical Logic, 27:217–224, 1998. 2. A. Baltag, H. van Ditmarsch, and L. Moss. Epistemic logic and information update. In P. Adriaans and J. van Benthem, editors, Handbook on the Philosophy of Information. Elsevier, 2008. 3. A. Batlag, L. S. Moss, and S. Solecki. The logic of public announcements and common knowledge and private suspicions. In I. Gilboa, editor, TARK, pages 43–56. Morgan Kaufmann, 1998. 4. J. Gebrandy. Bisimulations on Planet Kripke. PhD thesis, University of Amsterdam, 1998. 5. J. Hintikka. Knowledge and Belief: An Introduction to the Logic of the Two Notions. Cornell University Press, Ithaca, N. Y., 1962. 6. W. H. Holliday and T. F. I. III. Moorean phenomena in epistemic logic. In Advances in Modal Logic’10, pages 178–199, 2010. 7. G. E. Moore. A reply to my critics. The Philosophy of G.E. Moore, The Library of Living Philosophers, 4:535–677, 1942. 8. J. Plaza. Logics of public communications. Synthese, 158:165–179, 2007. 9. J. van Benthem. One is a lonely number : On the logic of communication. Logic Colloquium 02, (December):1–37, 2002. 10. J. van Benthem. Open problems in logical dynamics. In D. M. Gabbay, S. S. Goncharov, and M. Zakharyaschev, editors, Mathematical Problems from Applied Logic I, volume 4 of International Mathematical Series, pages 137–192. Springer New York, 2006. 11. J. van Benthem. Open problems in logical dynamics. In D. Gabbay, S. Goncharov, and M. Zakharya- shev, editors, Mathematical Problems from Applied Logic I, pages 137–192. Springer, 2006. 12. H. van Ditmarsch. Knowledge games. PhD thesis, University of Groningen, 2000. 13. H. van Ditmarsch, W. van der Hoek, and B. Kooi. Dynamic Epistemic Logic. Springer, 2008. 14. H. P. van Ditmarsch and B. Kooi. The secret of my success. Synthese, 2006:2006, 2004. 15. A. Visser, J. van Benthem, D. de Jongh, and G. R. R. de Lavalette. Nnil, a study in intuitionistic propositional logic. Logic Group Preprint Series, 111:535–677, 1994. 7 Appendix In the counter-model below, p is true only the worlds x, y and z and false in rest of the worlds. Clearly M, w = L2L1KaKbL1p. Also, L2L1KaKbL1p is true in the worlds s, t, u, y, z and v and false in the world x, x(cid:48) and v . Note that if v was combined to s by a 1-edge, then both v and x would have satisfied L1KaKbL1p. Hence, this construction cannot be used as a counterexample of L1KaKbL1p (which in fact is successful). Thus, the restricted model has only v and not x, x(cid:48) and v(cid:48) . So, in the restricted model Mφ, t = KaKbp. Thus, Mφ, w = L2L1KaKbL1p. Thus φ = L2L1KaKbL1p is not successful. Fig. 4. Counter-model for L2L1KaKbL1p
1503.08880	1	1503	2015-03-31T00:30:02	A composite constraints approach to declarative agent-based modeling	[ "cs.MA" ]	Agent-based models (ABMs) are ubiquitous in research and industry. Currently, simulating ABMs involves at least some imperative (step-by-step) computer instructions. An alternative approach is declarative programming, in which a set of requirements is described at a high level of abstraction. Here we describe a fully declarative approach to the automated construction of simulations for ABMs. In this framework, logic for ABM simulations is encapsulated into predefined components. The user specifies a set of requirements describing the desired functionality. Additionally, each component has a set of consistency requirements. The framework iteratively seeks a simulation design that satisfies both user and system requirements. This approach allows the user to omit most details from the simulation specification, simplifying simulation design.	cs.MA	cs	A COMPOSITE CONSTRAINTS APPROACH TO DECLARATIVE AGENT-BASED MODELING David Bruce Borenstein Lewis-Sigler Institute for Integrative Genomics Princeton University Carl Icahn Laboratory Princeton, NJ 08544, USA October 20, 2018 ABSTRACT Agent-based models (ABMs) are ubiquitous in research and industry. Currently, simulating ABMs involves at least some imperative (step-by-step) computer instructions. An alternative approach is declarative programming, in which a set of requirements is described at a high level of abstraction. Here we describe a fully declarative approach to the automated construction of simulations for ABMs. In this framework, logic for ABM simulations is encapsulated into predefined components. The user specifies a set of requirements describing the desired functionality. Additionally, each component has a set of consistency requirements. The framework iteratively seeks a simulation design that satisfies both user and system requirements. This approach allows the user to omit most details from the simulation specification, simplifying simulation design. 1 INTRODUCTION The need for descriptive modeling Agent-based models (ABMs) constitute one of the most widely used categories of simulation technology. ABMs (also known as "individual-based models" or IBMs) represent a system as an ensemble of autonomous actors (or "agents"). These agents interact with one another according to predefined behaviors. The set of defined behaviors may be unique to each individual agent, or common to a class of agents. ABMs are widely used for academic research in the fields of ecology, epidemiology and social science [1 -- 3]. Commercial and governmental uses include business analytics, supply chain management, and civil and military planning [4 -- 7]. ABMs provide a link between local and global dynamics. The modeler defines local interactions. As these interac- tions play out, global patterns often become evident. By choosing local rules based on observed real-world processes, ABMs enable the modeler to predict large-scale consequences. Conversely, by selecting rules that recapitulate ob- served large-scale processes, modelers can make predictions about the underlying local interactions. In either case, effective use of ABMs requires deep insight into the process at hand -- a body of knowledge wholly disjoint from the computer expertise required to actually build models. These challenges are all exacerbated by the introduction of spatial structure, where topological details can have major implications for emergent behavior [8, 9]. The predominant representation of ABM outside of software is a standardized rubric of model features, called the 'Overview, Design concepts, and Details" (ODD) approach [10]. In an ODD specification, agent-based models are described in terms of their structure and temporal dynamics. Notably, computer logic is minimized in the ODD 1 the rubric focuses on what the model does, rather than how a programmer chose to accomplish it. specification: Simultaneously, there have been efforts to develop a logic-based descriptive framework for the general description of simulations [11]. Here, we present a declarative agent-based simulation framework, Nanoverse, that is based on the principle of description. In existing approaches, model design is closely linked to simulation design. The model deals with the properties of the agents and the world they occupy. When do they act? What information can they incorporate into their choices? The simulation, on the other hand, is a computer program capable of actualizing the model and integrating it over time [12]. For the most part, existing tools are simulation frameworks: it is up to the user to first envision a model and then articulate a process for simulating it. One exception is the commercial tool Anylogic, discussed below, which uses flowcharts to focus attention on the model. In any case, the user must still define a sequence of computer instructions. The key innovation behind Nanoverse is that it structures ABM implementation as a configuration problem. Rather than specify step-by-step rules, the user imposes constraints (requirements) on the model. The platform then uses this specification to find a configuration of predefined components that satisfies these constraints. By replacing step-by- step (imperative) computer instructions with a (declarative) description of a model's properties, simulation design can be brought closer in line with the ways in which ABMs are discussed. Imperative approaches to ABM design Simplifying ABM development has been the focus of much research and development. Much of this research has focused on general-purpose software tools for spatially structured ABMs. Most ABM software tools have introduced expressive computer languages (or language extensions) created for the specific purpose of ABM simulation [13]. By far the most successful project has been NetLogo, which extends the educational programming language LOGO [14] to a large library of agent-specific structures and actions. NetLogo [15] has been widely adopted in academic research, and remains popular after nearly two decades of continuous use. An extreme form of this approach is purely visual programming, as in StarLogo TNG [16] and Scratch [17]. These K-12 educational tools make modeling easier by representing imperative statements as visual blocks. Paradoxically, the simplicity of these LOGO-derived tools means that complex models can be challenging to express. Another tool, GAMA [18] seeks to address some of these limitations by providing straightforward facilities for GIS and multi-level models. GAMA utilizes a fluent, object-oriented language called GAML. GAML automates many aspects of model design, but still requires the user to specify and manipulate data structures. Other imperative tools for complex simulations include Java libraries such as MASON [19] and Repast [20]. These tools each provide a powerful, object-oriented framework within which to build and simulate ABMs. However, these tools require proficiency with general-purpose programming languages such as Java. Commerical package AnyLogic uses a variety of UML-like charts to represent the states and actions of agents, which it then translates to Java code [21]. While highly accessible, this approach is essentially analogous to imperative programming since it requires the user to define a sequence of logical actions. A component-based architecture for ABMs Many agent-based models can be simulated using a common set of strategies. NetLogo, GAMA, Repast and StarLogo all provide an extensive library of common logical pieces; often the only programming task is to unite these pieces in a manner appropriate to the model. Nanoverse extends this concept further, by hierarchically building up components from a pool of subcomponents. By repeatedly applying this idea, it is possible to define agent-based models from a relatively small body of simple units. Since all imperative logic would be encapsulated in these units, the user's task becomes one of describing conceptual relationships, rather than computational tasks. This is the principle behind component-based (or "modular") software engineering [22, 23]. There is precedent for a component-based approach to simulation: SimKit provides a structure for building and distributin reusable imperative blocks, which can then be composed programatically or visually [24 -- 26].To the author's knowledge, there has been no effort to leverage the declarative nature of component-based software in order to present simulation design as a configuration task. This approach opens up a wealth of existing strategies for simulation implementation, as configuration problems are a cornerstone of knowledge engineering [27]. 2 Configuration problems can be solved using constraint satisfaction approaches. In a constraint satisfaction prob- lem, a "solution" is any value which satisfies every specified constraint. In configuration problems, constraints are either directly specified by the user, or are a consequence of a partial solution. For example, suppose that the user specifies a spatially explicit, two-dimensional model with periodic boundary conditions. It follows that the arena must have four sides. The goal of a constraint-based configuration scheme is to satisfy both the requirements imposed by the user and the requirements imposed by the selected sub-components, given a set of available options. A straightforward approach to constraint satisfaction is backtracking. In a backtracking algorithm, solutions are tested sequentially against the first constraint, being globally eliminated if they violate it. Once a solution is found, the algorithm recurs on the next constraint. If no solution satisfies a constraint, it "backtracks" to the previous constraint, which resumes its search [27]. By specifying a sequence of default subcomponents, the backtracking strategy is sufficient to configure a single component of a simulation, such as a spatial structure. An entire simulation can be specified by nesting constraint satisfaction problems together, in a strategy known as composite constraint satisfaction (CCS) [28]. For example, suppose an agent may move or replicate into a nearby location. Are collisions with other agents allowed? If so, how are they resolved? If not, what happens if no legal move is possible? CCS configures these elements sequentially, either as nested sub-problems or as co-constraints within the same problem. This approach has previously been used to automate other software configuration tasks, such as the deployment of complex software systems [29]. This paper describes part of an ongoing effort to create a constraint-driven, spatially explicit agent-based modeling framework. This framework, called Nanoverse, is being prototyped in stages. The first stage, a working mock-up of which is available online [30], is a component-based simulation environment that is functionally similar to GAMA or MASON. This paper concerns the second stage of the prototype, currently under development, consisting of a multi- stage compiler. The paper begins with a brief synopsis of the runtime environment into which the compiler instantiates simulations. The second part describes the architecture of the Nanoverse compilation pipeline. 2 RUNTIME SCHEME The Nanoverse runtime consists of a network of loosely coupled components. The primary subsystems of the runtime are a collection of topologies called "layers" and a discrete event scheduler [31] The layer encapsulates all topological information, and the schedule encapsulates all scheduling information. Mu- tation of the simulation state is accomplished through scheduling events with a relative waiting time, which is subse- quently resolved by the schedule. Likewise, specific changes to the environment are specified relative to a particular agent. As such, agents remain completely agnostic to the global state of the simulation. In order to accomplish this, events have callbacks that request specific changes to their locale. Agents have a symbol table associating specific names with event triggers, as well as a rule table mapping specific conditions to the triggering of certain events. When a simulation event runs, it notifies the layer, which notifies all affected agents to consult their rule table. Forward integration of the simulation is achieved by repeatedly polling the schedule for the next scheduled event. The system time is advanced to that of the next event, and the event is allowed to run. If running the event triggers immediate events or alters the progression of future events, the schedule is updated. The simulation ends when the event queue is depleted or another terminal condition is met. The loose coupling of simulation components allows for the use of a constraints-based compiler system, which in turn allows us to move away from imperative programming. 3 COMPILE SCHEME 3.1 Overview The Nanoverse compiler prototype consists of a four-stage compilation pipeline, ultimately leading to a discrete-event runtime for spatially explicit agent-based models (Fig. 1). The first stage of the pipeline is a parser that interprets a hierarchical source code into an abstract syntax tree (AST). This abstract syntax tree has no semantic information about the structure of an agent-based model; it reflects only the grammar of the user's specifications. The second 3 stage uses a hierarchy of symbol tables to convert the abstract symbol tree into a semantically rich hierarchy of "build nodes." These build nodes roughly correspond to the Java objects that will represent the simulation in memory. The third stage of the pipeline is the backtracking constraint solver, which is used to interpolate unspecified properties of the simulation. This is done by treating the user's specifications as additional constraints on an ordered sequence of defaults, with over- or underdefined specifications leading to an error. Finally, the completed build tree is visited breadth-first in order to instantiate all nodes into Java objects. The top-level object then triggers the execution of the simulation. Nano- syntax Interpret to AST Translate to build hierarchy Interpolate unspecified properties Build and link objects Runtime Figure 1: Nanoverse compilation pipeline 3.2 The Nanosyntax environment The user writes Nanoverse model descriptions using a hierarchical grammar called "Nanosyntax." Nanosyntax was influenced by the JSON object specification, which is used to serialize data for transmission between internet servers and clients [32]. All statement blocks are terminated by a semicolon. Statements consist of an outer node, and, in the case of an assignment, one or more value nodes. Single value assignments are designated by a colon, and block assignments are specified by curly braces. The top level of a Nanoverse project specification is internally represented as a block assignment to a hidden reference. Nanosyntax consists of three types of nodes: "primitives," "references" and "assignments." Primitives are basic data types, such as numbers and strings. References specify an identifier or a block of identifiers. Assigmments map an identifier to a reference. Additional structures are allowed for mathematical operations, but these are converted internally to the other node types. These three elements are sufficient to specify an arbitrary hierarchy of members in a concise and readily intelligible way. Nanoverse project specifications consist of a nested ensemble of constraints, or explicit requirements concerning the properties of the simulation. Any requirements left unspecified are subsequently interpolated from a set of defaults to match the constraints that were specified. The properties to be specified correspond directly to encapsulated operational units, or components, of the simu- lation's business logic. Thus, Nanosyntax is fully declarative: the only purpose of the source code is to describe what should be done, rather than how it should be implemented. As a result of interpolation, Nanosyntax is also minimal: the source code contains only the requirements of interest to the user. The user can therefore begin running simulations with very little code. By iteratively overriding defaults, the user can then build up the behavior of the simulation until it reflects all desired functionality. As an example, consider the "StupidModel" reference model developed by [13]. The first of 16 instances of the model consists of a population of 100 agents that diffuse around a 100x100 rectangular lattice. Here we present the anticipated Nanosyntax for an even simpler model, which consists of a single agent diffusing around a 32x32 rectangular lattice. (Time is specified in arbitrary units.) initially: scatter: description: Agent: do: Behavior { action: wander; every: 1.0; until: time >= 100.0; }; 4 The Nanosyntax representation of the model describes the entire system in a single statement block. Absent a specific geometry requirement, the system defaults to a 32x32 rectangular lattice with absorbing boundary conditions. An initially assignment specifies one or more events that must occur to set up an initial condition. Since the only action -- diffusion, or wandering -- is encoded in the definition of the agent itself, there is no need to define a main loop. The wander operation itself does have subcomponents dealing with destination selection and collision resolution, but these are also handled with interpolated defaults. At its construction time, the agent will schedule a wander action that, upon firing, will schedule itself for one time unit later, until the time exceeds 100. By default, an image sequence of the process will be generated, with frames recorded at the start of each time unit. Specifying an alternative boundary condition, lattice geometry or arena shape would take one additional line apiece. Existing frameworks require far more code to accomplish the same goal. MASON and Repast both require the user to define diffusion from first principles, instantiate a 2D arena, and place the agents using a random number generator; moreover, all of this must be done in Java. NetLogo eliminates the need for low-level programming, but still requires explicit instructions for each operation involved [33]. GAMA requires that the user first define a geometry and a neighborhood structure, then the conditions for an ongoing behavior (or "reflex") representing movement. The user then defines a visual representation of the agent, and a display mode for the visual representation [34]. The GAMA approach is similar to that of Nanoverse, except that Nanoverse is designed to resolve many of the specified details that are required in GAMA. The Nanoverse compiler parses Nanosyntax using the parser generator ANTLR4 [35]. ANTLR4 generates a parse tree. The Nanoverse compiler then translates the Nanosyntax parse tree into an abstract syntax tree (AST), an example of which is shown in Fig. 2. Nanoverse employs a heterogeneous AST: different nodes are used for each of the three basic data types [36]. By distinguishing between data types, the AST provides structural information that simplifies the next process in the pipeline: semantic analysis. root assignment ref: initially assignment ref: scatter assignment assignment ref:count integer:1 ref: description ... Figure 2: Top portion of an abstract syntax tree. 3.3 Adding semantic information After parsing user syntax, Nanoverse constructs a partial representation of model semantics (Fig. 3). This partial se- mantic model, known as the "object node hierarchy" or "build hierarchy," encodes all requirements explicitly specified by the user. Nanoverse constructs this hierarchy through the use of a graph of symbol tables. At their most basic, symbol tables are mapping functions from a text symbol to some other value [37]. A compiler for an imperative language will typically create a single symbol table at each level of contextual scope to associate identifiers with values [38]. Nanoverse symbols, on the other hand, represent system components: i.e., loosely coupled 5 AST node, table List? No Append to list Recur Yes Children? Yes No Return node No Map? Yes Children? Yes Recur on value No Primitive? Yes No Error Map value to ID Figure 3: A flowchart representing the translation of an abstract syntax tree node to an object node. subsystems that supply specific functionality for the containing system, and which themselves depend on further subsystems [22]. As such, Nanoverse uses a symbol table for every component and component class. To encode its rule base, the Nanoverse compiler constructs two classes of symbol tables: resolving symbol tables (RSTs) and instantiable symbol tables (ISTs). RSTs narrow a particular identifier to a specific subclass of an expected class. ISTs resolve the names of specific subsystems required to instantiate an object of a specific class. Object translation proceeds by alternating between these two symbol table classes. To begin translation, a root IST is supplied, along with the root node of the AST, to the translation visitor, or "translator." The translator visits each child of the root node. If the identifier of the node does not exist in the IST, translation halts with an error. Assuming the symbol is recognized, the translator passes the AST node to the IST for resolution. The IST resolves the node against an RST, which provides another IST for translating the child. The child's IST and the child are then passed back to the translator, resulting in a depth-first traversal. ISTs come in three varieties: list ISTs (LISTs), map ISTs (MISTs), and primitive instantiators. MISTs expect key-value pairs and correspond to single objects; LISTs expect multiple anonymous values and represent collections or predicate blocks. Primitive instantiators redesignate primitive AST nodes as object nodes. Translating a MIST 6 involves resolving its identifier against the parent IST to retrieve an appropraite RST, then resolving its value against the RST to retrieve a child IST and calling back. Every element in a LIST has the same RST, so translating the LIST just involves resolving each LIST child node against the RST and calling back. Primitives are terminals; no callback is required. After translation, the user's requirements have been translated into a hierarchy of constraints. The compiler must now determine whether and how a simulation can be instantiated from the user's specifications. The user's constraints may be expressly incompatible, or they may imply further requirements that are incompatible. In these cases, the model is overdetermined. On the other hand, the user may have omitted required fields (i.e., fields with no default values). If this happens, the model is underdetermined. Assuming neither an overdetermined nor underdetermined model, the compiler's next task is to interpolate sufficient constraints to fully determine the model's configuration. Interpolation and construction 3.4 Nanoverse organizes a simulation into a hierarchy of components. "Components" in Nanoverse are equivalent to "primitives" in NetLogo or GAMA [15, 18], except that most components are not "primitive" in the sense of being discrete, atomic wholes. Rather, a Nanoverse component may have an arbitrary number of subcomponents, which may likewise have subcomponents of their own. Components are only loosely coupled to their subcomponents -- often by a single method -- facilitating interchange. Interchangeable components are at the heart of the configuration-based approach. The configuration of Nanoverse components is accomplished through the hierarchical solution of local constraints. Each subcomponent has its own constraints. These constraints determine whether the component is compatible with the existing partial configuration, and which subcomponents can be supplied to it. Additionally, the subcomponents for a given component may depend on one another, and are thus supplied as additional constraints on the subcomponent. Associated with each subcomponent is one or more defaults, which are given in order of preference. Each default may imply its own set of constraints. If the user has specified a particular value for a subcomponent, the specified subcomponent (and its implied constraints) replaces the default list. Component configurations are then solved depth first until a total solution has been found, or it is determined that no solution exists (Fig. 4). In a constraint satisfaction problem, solutions are often obtained through a backtracking scheme. A backtracking scheme consists of a recursive algorithm. Let v0, ...vi represent the values that must be specified, and let Di represent the domain of solutions for vi. In addition, there exists a set of constraints on the solution set. The backtracker begins by seeking a value v0 = d0d0 ∈ D0 that satisfies the relation C∩ v0 = d0. If such a value is found, the algorithm recurs on v1,D1. For the nth recursion, the algorithm seeks a value of vn = dndn ∈ Dn that satisfies C∩ (vk = dk∀k ≤ n). If this relation is not satisfied, the algorithm returns failure [39]. In Nanoverse, the constraints represent the specific requirements of particular subcomponents. The constraint set C is therefore not globally constant. However, for any given component, the only constraint is that all of its subcompo- nents are legal, given their dependencies. Thus, a subcomponent can verify constraint satisfaction by verifying that all of its subcomponents can find legal instance values. For simple subcomponents with only one possible default value, such a check is relatively straightfoward. More complex subcomponents must perform their own interpolation step. This component-dependent interpolation step is encapsulated in the "Valid?" decision node in Fig. 4. Instantiation proceeds like interpolation. Each component has its own instantiation method, which builds any helper objects as necessary. For the most part, these helper objects are themselves components, albeit not user- specified ones. That is, they are only loosely coupled to the parent component, and they are automatically configured based on the properties of the parent component. Helper components include getters and setters from other runtime objects, which serve the same role as public method calls in traditional APIs. 4 CHALLENGES AND FUTURE DIRECTIONS 4.1 Engineering considerations Component-based software design has been widely incorporated into many software platforms, most recently in the form of Java's "Project Jigsaw" [22, 23]. Difficulty of maintenance is a challenge that is common to all of these 7 approaches [40]. Component-based software essentially shifts some of the user's responsibilities onto the developer. Rather than define logic, the user builds pre-fabricated logical components into the desired configuration. The power and utility of a component-based platform is therefore limited by the breadth and quality of these pre-fabricated components. The developer must provide both the runtime logic and any steps required for the component to compile. In the case of Nanoverse, these steps include specifying acceptable sub-components and the order of preference for those components. This implies a larger codebase than analogous, imperative systems. Nanoverse is also fully declarative, unlike hybrid declarative-imperative languages like GAMA. As a fully declarative language, the only possible business logic is that which is defined in an existing component. Many agent-based models share some similar ideas, especially in Nanoverse's primary domain of spatially explicit ABMs. Special cases abound, however, and new questions lead constantly to new model designs. How can Nanoverse accommodate advanced use cases without complicating the simple ones? Perhaps the most straightforward solution is to incorporate an imperative sub-language, which would be backed by a simple interpreter. This would have the added benefit of allowing model changes on the fly, and even self-modifying code, which can be used for evolutionary simulations. Testing is another major challenge. It is often desirable to test a component both in isolation and in its intended context. However, the number of possible contexts for any given component is limitless: as user models (and the component library) grow, the same component can be nested deep in a hierarchy of other parts. When the assumptions of components are in contradiction, unexpected behavior can result. These risks can be managed through the judicious use of consistency checks, strong interface contracts, and exhaustive unit testing [41]. 4.2 Default generality In the Nanoverse prototype, the only planned constraints have to do with logical compatibility. For example, a spatially explicit model taking place in a hexagonal arena cannot employ a periodic boundary condition, because two of the six sides would remain unmatched. Likewise, a rectangular lattice cannot employ a hexagonal arena. This lowers the skill threshold required for use, but it does not handle another important class of constraint: preventing the selection of many parameters that, while technically not in conflict, may produce unexpected behavior. There are many situations in which the user would expect different defaults based on his or her selections, even if one default could technically satisfy all cases. This is particularly true for spatially structured systems. Consider, for example, the resolution of collisions. What should happen if an agent that is scheduled to move has no vacant space into which it can go? The user must specify how to choose a destination, and, if collisions are possible, how to resolve them. If the user specifies that destinations must include occupied locations, there must be a rule for resolving a collision. Conversely, if the user specifies a rule for resolving collisions, it is likely that the user wishes to allow them. That said, "ignore occupied spaces" remains a valid default behavior for the rule. Likewise, permitting occupied spaces is compatible with a resolution strategy of "throw an error on collisions," though this is unlikely to be desired. Within a specific application area, there may also be practical constraints on default behavior that relate to the meaning of the model. For example, the preferred default resolution of collisions in a forest fire model (intensify the fire) is different from that of a microbial model (push the existing occupant away). One solution is the ability to specify custom constraints and default sets, and to inherit these elements as domain-specific libraries. For the previous example, one might create a default component hierarchy for "bacteria" agents, and another for "flame" agents. 5 CONCLUSION The Nanoverse compiler has the potential to simplify the process of building agent-based models. This greater ease can benefit both novice and experienced users, as fewer parameters need be considered or specified. With the introduction of component libraries, Nanoverse can also function as a medium for the transmission of expert knowledge concerning model design. As with many agent-based modeling platforms, the same approach can be used to simplify the design of interactive systems, such as games. The strict hierarchical structure of the Nanoverse language provides several benefits. The Nanoverse compiler already exploits the most important of these: the availability of algorithms to interpolate missing nodes. Hierarchies are also scale-free, which makes them easy to visualize, e.g. using a zooming user interface [42]. The strict separation of concerns required for hierarchical design also simplifies compiler design, which facilitates optimization of program 8 flow. Finally, a component-based approach facilitates the automatic generation of documentation via traversal of the symbol table hierarchy. The nanoverse compiler is under active development. Our first goal is to port all runtime functionality from the interpreted Nanoverse prototype [30], including modular topology and continuum-valued fields, to the compiler-based edition. Following that, we plan to provide an automatic documentation system and publish it to the Nanoverse web- site. Beyond that, we will focus on addressing the limitations of the language by introducing user-defined variables, user-defined constraints and defaults, object orientation, and code importation. ACKNOWLEDGMENTS The author gratefully acknowledges Ned S. Wingreen for his support and guidance throughout the development of the Nanoverse framework, as well as Anne Maslan for her helpful feedback and testing. References [1] Stephen Eubank, Hasan Guclu, V S Anil Kumar, Madhav V Marathe, Aravind Srinivasan, Zolt´an Toroczkai, and Nan Wang. Modelling disease outbreaks in realistic urban social networks. Nature, 429(May):180 -- 184, 2004. doi:10.1038/nature02541. [2] Volker Grimm, Eloy Revilla, Uta Berger, Florian Jeltsch, Wolf M Mooij, Steven F Railsback, Hans-Hermann Thulke, Jacob Weiner, Thorsten Wiegand, and Donald L DeAngelis. Pattern-oriented modeling of agent-based complex systems: lessons from ecology. Science (New York, N.Y.), 310(5750):987 -- 91, November 2005. URL: http://www.ncbi.nlm.nih.gov/pubmed/16284171, doi:10.1126/science.1116681. [3] Nigel Gilbert. Agent Based Models. Sage, Thousand Oaks, CA, 2008. [4] Mark S Fox, Mihai Barbuceanu, and Rune Teigen. Agent-Oriented Supply-Chain Management. 12:165 -- 188, 2000. [5] Thomas W Lucas and Susan M Sanchez. Proceedings of the 2004 Winter Simulation Conference R .G. Ingalls, M. D. Rossetti, J. S. Smith, and B. A. Peters, eds. Simulation, 2004. [6] S. a. Delre, W. Jager, T. H a Bijmolt, and M. a. Janssen. Targeting and timing promotional activities: An agent-based model for the takeoff of new products. Journal of Business Research, 60:826 -- 835, 2007. doi: 10.1016/j.jbusres.2007.02.002. [7] Xiaoping Zheng, Tingkuan Zhong, and Mengting Liu. Modeling crowd evacuation of a building based on seven methodological approaches. Building and Environment, 44:437 -- 445, 2009. doi:10.1016/j.buildenv.2008. 04.002. [8] R Durrett and S A Levin. The importance of being discrete (and spatial). Theoretical Population Bi- ology, 46(3):363 -- 394, 1994. URL: http://www.cs.unm.edu/~forrest/cas-class/readings-2006/ Durrent-Levin-1994.pdf. [9] David Bruce Borenstein, Yigal Meir, Joshua W. Shaevitz, and Ned S. Wingreen. Non-Local Interaction via Diffusible Resource Prevents Coexistence of Cooperators and Cheaters in a Lattice Model. PLoS ONE, 8(5):e63304, May 2013. URL: http://dx.plos.org/10.1371/journal.pone.0063304, doi:10.1371/ journal.pone.0063304. [10] Volker Grimm, Uta Berger, Finn Bastiansen, Sigrunn Eliassen, Vincent Ginot, Jarl Giske, John Goss-Custard, Tamara Grand, Simone K. Heinz, Geir Huse, Andreas Huth, Jane U. Jepsen, Christian Jø rgensen, Wolf M. Mooij, Birgit Muller, Guy Pe'er, Cyril Piou, Steven F. Railsback, Andrew M. Robbins, Martha M. Robbins, Eva Rossmanith, Nadja Ruger, Espen Strand, Sami Souissi, Richard a. Stillman, Rune Vabø, Ute Visser, and 9 Donald L. DeAngelis. A standard protocol for describing individual-based and agent-based models. Ecological Modelling, 198:115 -- 126, 2006. doi:10.1016/j.ecolmodel.2006.04.023. [11] Giancarlo Guizzardi and Gerd Wagner. Towards an ontological foundation of discrete event simulation. Proceedings - Winter Simulation Conference, pages 652 -- 664, 2010. doi:10.1109/WSC.2010.5679121. In [12] John H. Miller and Scott E. Page. Complex Adaptive Systems: An Introduction to Computational Models of Social Life. Princeton University Press, Princeton, NJ, 2007. [13] S. F. Railsback, S. L. Lytinen, and S. K. Jackson. Agent-based Simulation Platforms: Review and Development Recommendations. Simulation, 82(9):609 -- 623, 2006. doi:10.1177/0037549706073695. [14] Wallace Feurzeig, Seymour Papert, and Bob Lawler. Programming-languages as a conceptual framework for teaching mathematics. Interactive Learning . . . , pages 487 -- -501, 1969. URL: http://www.tandfonline. com/doi/abs/10.1080/10494820903520040. [15] Uri Wilensky. Seth Tisue Center for Connected Learning and Computer-Based Modeling Slightly updated from : environment . Proceedings of the Agent 2004 Conference on Social Dynamics : Interaction , This version has been slightly updated from the original to reflect a fe. 2004. [16] Eric Klopfer and Andrew Begel. StarLogo TNG: An Introduction to Game Development. pages 1 -- 15, 2007. URL: http://langwidge.com/starlogo.pdf. [17] John Maloney, Mitchel Resnick, Natalie Rusk, Brian Silverman, and Evelyn Eastmond. The Scratch Program- ming Language and Environment. 10(4):1 -- 15, 2010. doi:10.1145/1868358.1868363.http. [18] Arnaud Grignard, Patrick Taillandier, Benoit Gaudou, Duc An Vo, Nghi Quang Huynh, and Alexis Drogoul. GAMA 1.6: Advancing the art of complex agent-based modeling and simulation. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), volume 8291 LNAI, pages 117 -- 131, 2013. doi:10.1007/978-3-642-44927-7\_9. [19] S. Luke. MASON: A Multiagent Simulation Environment. Simulation, 81:517 -- 527, 2005. doi:10.1177/ 0037549705058073. [20] Michael J North, Nicholson T Collier, Jonathan Ozik, Eric R Tatara, Charles M Macal, Mark Bragen, and Pam Sydelko. Complex adaptive systems modeling with Repast Simphony. Complex Adaptive Systems Modeling, 1(1):3, 2013. URL: http://www.casmodeling.com/content/1/1/3, doi:10.1186/2194-3206-1-3. [21] Andrei Borshchev, Yuri Karpov, and Vladimir Kharitonov. Distributed Simulation Of Hybrid Systems With Anylogic And Hla. Elsevier Future Generation Computer Systems, 18:829 -- 839, 2002. [22] Felix Bachmann, Len Bass, Charles Buhman, Santiago Comella-Dorda, Fred Long, John Robert, Robert Seacord, and Kurt Wallnau. Technical Concepts of Component-Based Software Engineering, vol. 2. SEI Joint Program Office, Bedford, MA, 2000. [23] Mark Reinhold. JSR 376: JavaTM Platform Module System. Technical report, Oracle Corporation, 2014. URL: https://www.jcp.org/en/jsr/detail?id=376. [24] Arnold Buss. Component Based Simulation Modeling with Simkit. . . . of the 34th conference on Winter simula- tion: exploring . . . , 2002. URL: http://dl.acm.org/citation.cfm?id=1030489. [25] A.H. Buss and P.J. Sanchez. Building complex models with LEGOs (Listener Event Graph Objects). Proceedings of the Winter Simulation Conference, 1:732 -- 737, 2002. doi:10.1109/WSC.2002.1172954. [26] Arnold Buss and Curtis Blais. Composability and component-based discrete event simulation. In Proceedings - Winter Simulation Conference, pages 694 -- 702, 2007. doi:10.1109/WSC.2007.4419663. 10 [27] Bob Wielinga and Guus Schreiber. Configuration-design problem solving. IEEE Expert-Intelligent Systems and their Applications, 12:49 -- 56, 1997. doi:10.1109/64.585104. [28] Daniel Sabin and Eugene C. Freuder. Configuration as Composite Constraint Satisfaction. In Proceedings of the AI and Manufacturing Research Planning Workshop, pages 153 -- 161, 1996. [29] Jules White, Doulas C. Schmidt, Krzyszlof Czarnecki, Christoph Wienands, Gunther Lenz, Egon Wuchner, and Ludger Fiege. Automated model-based configuration of enterprise java applications. Proceedings - IEEE International Enterprise Distributed Object Computing Workshop, EDOC, (1):301 -- 312, 2007. doi:10.1109/ EDOC.2007.4384002. [30] Borenstein. Nanoverse Simulation Framework, 2015. URL: http://nanover.se/. [31] George Fishman. Discrete-Event Simulation: Modeling, Programming, and Analysis. Springer, New York, NY, 2001. [32] Ecma International. The JSON Data Interchange Format. (October):1 -- 14, 2013. URL: http://www.json.org. [33] S. F. Railsback, S. L. Lytinen, and S. K. Jackson. Supporting Files for: Agent-based Simulation Platforms: Review and Development Recommendations, 2006. URL: http://condor.depaul.edu/slytinen/abm/. [34] Edouard Amouroux. GAMA tutorial: StupidModel, 2014. eclipselabs.org/p/gama/wiki/StupidTutorialModel1v14. URL: https://code.google.com/a/ [35] Terrence Parr. The Definitive ANTLR4 Reference. Pragmatic Bookshelf, 2nd ed. edition, 2014. [36] Terrence Parr. Language implementation patterns. Pragmatic Bookshelf, 1st ed edition, 2010. [37] Dick Grune, Henri E. Bal, Ceriel J.H. Jacobs, and Koen G. Langendoen. Modern Compiler Design. Wiley, 2000. [38] Keith Cooper and Linda Torczon. Engineering a compiler. Morgan Kauffmann, 2nd ed. edition, 2011. [39] Stuart Russell and Peter Norvig. Artificial Intelligence: A Modern Approach. Prentice Hall, Upper Saddle River, NJ, 3rd ed. edition, 2009. [40] Elaine J. Weyuker. Testing component-based software: A cautionary tale. IEEE Software, 15(October 1998):54 -- 59, 1998. doi:10.1109/52.714817. [41] Ivica Crnkovic and Magnus Larsson. Building Reliable Component-Based Software Systems. Artech House, London, 2002. [42] Ben Bederson and Jon Meyer. experience build- ing Pad++. URL: http://dx. doi.org/10.1002/(SICI)1097-024X(199808)28:10<1101::AID-SPE190>3.0.CO;2-V, doi:10.1002/ (SICI)1097-024X(199808)28:10<1101::AID-SPE190>3.0.CO;2-V. Practice and Experience, 28:1101 -- 1135, 1998. Software: Implementing a zooming User Interface: 11 Yes Next OK? No Recur on next Yes No Default valid? Yes Untried defaults? No Succeed Obj node, index Yes Has req'd args? No Error Next OK? No Yes Index at end? Yes No No User- specified? Yes Recur on next Yes Valid? No Fail Figure 4: Flowchart representing the constraint satisfaction process used to interpolate unspecified user parameters. 12
cs/0604078	1	0604	2006-04-20T11:20:16	The emergence of knowledge exchange: an agent-based model of a software market	[ "cs.MA", "cs.CE" ]	We investigate knowledge exchange among commercial organisations, the rationale behind it and its effects on the market. Knowledge exchange is known to be beneficial for industry, but in order to explain it, authors have used high level concepts like network effects, reputation and trust. We attempt to formalise a plausible and elegant explanation of how and why companies adopt information exchange and why it benefits the market as a whole when this happens. This explanation is based on a multi-agent model that simulates a market of software providers. Even though the model does not include any high-level concepts, information exchange naturally emerges during simulations as a successful profitable behaviour. The conclusions reached by this agent-based analysis are twofold: (1) A straightforward set of assumptions is enough to give rise to exchange in a software market. (2) Knowledge exchange is shown to increase the efficiency of the market.	cs.MA	cs	The emergence of knowledge exchange: an agent-based model of a software market Maria Chli and Philippe De Wilde Abstract We investigate knowledge exchange among commercial organisations, the rationale behind it and its effects on the market. Knowledge exchange is known to be beneficial for industry, but in order to explain it, authors have used high level concepts like network effects, reputation and trust. We attempt to formalise a plausible and elegant explanation of how and why companies adopt information exchange and why it benefits the market as a whole when this happens. This explanation is based on a multi- agent model that simulates a market of software providers. Even though the model does not include any high-level concepts, information exchange naturally emerges during simulations as a successful profitable behaviour. The conclusions reached by this agent-based analysis are twofold: (1) A straightforward set of assumptions is enough to give rise to exchange in a software market. (2) Knowledge exchange is shown to increase the efficiency of the market. Intelligent agents, Multi-agent systems, Economics, Adaptive behaviour, Agent-based modelling Index Terms 6 0 0 2 r p A 0 2 ] A M . s c [ 1 v 8 7 0 4 0 6 0 / s c : v i X r a This work has been carried out as part of the project Digital Business Ecosystem, funded by the 6th Framework Programme of the European Commission. Contract Number: IST - 2002 - 507953. This work was carried out when M. Chli and P. De Wilde were at Imperial College London The emergence of knowledge exchange: an agent-based model of a software market 1 I. INTRODUCTION The growth of the Internet as a medium of knowledge exchange has stimulated a lot of scientific interest originat- ing from various disciplines. The willingness of individuals, organisations as well as commercial firms to share informa- tion via the Internet has been remarkable. In some sectors like scientific research, the communication of newly acquired knowledge and expertise in a field is considered vital for their advancement. On the other hand, in other sectors, the benefits of such exchanges may not be obvious. For instance, it might even be considered damaging for pharmaceutical companies to make public any innovations generated by their Research and Development (R&D) process. In spite of this view, exchange of intellectual property in some industries occurs quite frequently and in various different ways. These include the forming of strategic partnerships, the participation in open source software projects and the publication of scientific papers by research labs that are part of commercial companies. We study the knowledge exchange that occurs in the software industry. In particular, we focus on analysing the rationale behind this exchange as well as its effect on the industry. The complexity of software requirements is a char- acteristic that distinguishes the software market from others. However, the findings of this work might be relevant to other industries as well. This effort fits within the framework of the Digital Business Ecosystem (DBE) project. The DBE project is an attempt to develop a distributed environment which will interlink European Small and Medium Enterprises (SMEs) that are software providers and foster collaboration between them. Our broader interest lies in understanding the dynamics of ecosystems [11], [15], [43]. Furthermore, we are interested in analysing the global system properties which emerge from the interactions that occur in a market ecosystem. We have been using techniques from agent based modelling to simulate the DBE environment. The main aspects of the DBE market are captured in a model where the SMEs are agents with bounded rationality. This model is then studied using simulations of various settings, and a number of observations are made. One of the most interesting observations is that exchanges between the agents similar to the ones that happen in real-life arise in the system. This behaviour emerges in the market even though the model does not explicitly account for social issues of trust, network effects or managerial strategies. The paper is organised as follows. The following section gives an insight to the Digital Business Ecosystem project and the characteristic of the market that will be developed. In section III we sketch the background of this work, namely we review the types of exchanges that occur in markets, giving particular attention to the software market. Section IV details the model used for the investigation carried out. Section V analyses the experiments performed and the results produced and section VI concludes. II. DIGITAL BUSINESS ECOSYSTEM In this section we give a brief overview of the Digital Business Ecosystem project, highlighting its aims and motiva- tion. The characteristics of the end-product are identified and special attention is given to the efficiency of the market that will be formed. A. A DBE Economy It is stated in [35] that virtual organisations make dynamic coalitions of small groups possible. In this way the companies involved can provide more services and make more profits. Moreover, such coalitions can disband when they are no longer effective. At present, coalition formation for virtual organisations is limited, with such organisations largely static. The overall goal of the DBE project 1 [13] is to launch a new technology paradigm for the creation of a digital business ecosystem that will interlink SMEs and especially software providers. The project is encompassed by the Eu- ropean Union's initiative to become a leader in the field of software application development and to strengthen its SME industry. An open source distributed environment will support the spontaneous evolution, adaptation and composition of software components and services, allowing SMEs that are solution and e-business service providers to cooperate in the production of components and applications adapted to local business needs. This will allow small software providers in Europe to leverage new distribution channels providing niche services in local ecosystems and extending their market reach through the DBE framework. Easy access and large availability of applications, adapted to local SMEs, will foster adoption of technology and local economic growth. It will change the way SMEs and EU software providers use and distribute their products and services. The main objective of this work, which was carried out as part of the DBE project, was to study the properties of this new type of market. It is clear that the interactions and exchanges between the SMEs within the Digital Business Ecosystem environment will have an effect on the dissemination of information and subsequently to the efficiency of the market. B. Market Efficiency Within the environment of the DBE, business alliances, networks and supply chains require much less effort to be 1The web page of the project can be found at www.digital-ecosystem.org formed. This will promote cooperation and easier dissemina- tion of information between the member SMEs. On the other hand, competition for a share of the market between SMEs will become more direct. It is to be hoped, that these factors will raise the levels of efficiency in the DBE market in comparison to a traditional market. While these aspects of the DBE are very interesting and the subject of future research, this work studies how market efficiency is affected by the exchange of information between SMEs. The experiments carried out on our model, confirm that as the agents engage in more information exchanges between them, with time the market efficiency of the system rises. Efficient Markets Theory, as proposed by [19], is a field of economics which seeks to explain the operation of an asset market. Specifically, it states that at any given time, the price of an asset reflects all available information [3], [12]. The efficient market hypothesis implies that it is not generally possible to make above-average returns in the stock market over the long term by trading lawfully, except through luck or by obtaining and trading on inside information. The DBE environment is different from an asset market, so the definition of efficiency needs to be modified, retaining the spirit of the efficient market hypothesis. In the model of the DBE used in this work, the market is driven by demand which is fixed and unaffected by the supplied DBE services. In this case the market is efficient if, at any given time, the supply of a service reflects all available information. This means that, the services supplied are such that they satisfy the underlying market needs optimally. In other words, the SMEs are not concentrating on catering for some needs while others are left unsatisfied. In an efficient DBE market, all the needs will be satisfied evenly, assuming that there is equal demand for each of them. To draw a parallel between the traditional definition of an efficient asset market and the proposed definition for the efficiency of the DBE market consider the following. In an inefficient asset market, a trading agent can earn excessive returns by buying a particular stock which she believes to be undervalued. Similarly, in an inefficient DBE market a company might make excessive profits by satisfying a need which it knows is not sufficiently satisfied. To invert the argument, in an efficient asset market, asset prices adjust in- stantaneously and in an unbiased fashion to publicly available new information, so that no excess returns can be earned by trading on that information. Similarly, in an efficient DBE market, the supply of services will adjust immediately to any arising information about the underlying needs. Cooperation, symbiosis [16], [27] as well as the efficiency [37], [40] of adaptive multi-agent systems has been studied in the context of the simple games. In [40] no verifiable definition of efficiency is given, whereas in [37] the system is considered to be in an efficient market phase when all information that can be used by the agents' strategies is traded away, and no agent can accumulate more points than an agent making random guesses would. In the work presented in this paper, market efficiency, cooperation and competition are studied in the context of a more realistic economic market. 2 III. BACKGROUND In this section we list a number of ways in which exchange of knowledge between companies happens in a market and the rationale for each of them is briefly reviewed. As this work focuses on SMEs that are software providers, we survey the key characteristics of the software industry and the exchanges in this particular market. A. Exchange in economic markets In an economic market there are many ways in which the firms engage in exchanges between them. These include the forming of strategic partnerships, the participation in open source software projects and the publication of scientific papers by research companies like HP Labs and Microsoft Re- search. In the paragraphs that follow we will briefly examine the rationale behind these different forms of exchange. For a strategic partnership to be formed, the partners must mutually benefit from the experience, expertise and talent that all the parties bring to the partnership. There usually is an immediate worthy goal or objective that the partners concerned wish to achieve. For instance, they may wish to operate in a new market, or to bring about a change of leadership in the industry they operate in. Hagedoorn in [24] reports a dramatic rise especially in R&D partnerships, over the past 40 years. These partnerships are mostly limited-time project based collaborations as opposed to long-term alliances. The main motives behind them are reported to be related to cost- cutting as well as risk minimisation whilst the partners attempt to enter new technological areas. Recent economics and management research has studied the phenomenon of commercial firms contributing to open source projects. The main motive indicated by these analyses is strategic [22], as set out in more detail in section III-B where the specifics of the software industry are analysed. This seems to be consistent with the fact that it is not the leaders in the industry who engage in open source development, but the followers. Another form of exchange, which at first might seem counter-intuitive, is the publication of scientific papers con- taining the findings of the research commercial companies perform. It may be argued that it would be in the interest of those companies, to keep their innovative work to themselves. Another argument, however, is that by publicising their re- search they invite others to endorse it, add to it and in effect advance it further. Then, they can use the knowledge acquired by this process to better their products. The model of a software market that we propose as part of this work is simple in the sense that the agents/firms do not have the ability to reason about complex situations. They cannot make decisions to operate in new markets, or form partnerships in order to change the leadership in the industry. They cannot devise strategies to undercut their competitors. However, they operate in a capitalistic economy where the best of them succeed whilst the worst perish. They are thus equipped with a simplistic mechanism of reinforcement learning, i.e. being rewarded or punished for choices that prove to be good or bad respectively. When given the opportunity to engage in exchange of services between them, they learn with time under which circumstances this is beneficial to them and they proceed with it without ever being biased by external factors towards exchanging. B. The software industry Complexity is a key characteristic of software which dis- tinguishes the software industry from others. Typical software products carry a large number of features, with innumerable [2] interactions between them. For a program to be successful in the market, it is necessary that it has the right set of features to satisfy the customer base and that these features operate successfully together. The market of proprietary software providers/publishers is dominated by large companies, not SMEs. Microsoft Corpo- ration holds the lion's share in the software market with com- panies like Oracle, IBM, Hewlett-Packard and Sun following with smaller shares2. At the same time, the open source3 movement has been quite successful in developing relatively complex software products like Linux, Apache or sendmail that are serious competitors of well established proprietary software [38]. Networks of thou- sands of volunteers have contributed to these highly complex products. This appears, as it is pointed out in [2], to counter the economic intuition that private agents, without property rights, will not invest sufficient effort in the development of public goods because of free-rider externalities. Lerner and Tirole in [33] justify the volunteers' motivation for contribution to the open source movement as an oppor- tunity to 'signal their quality'. In other words, the volunteers believe it will enhance their career prospects, as the names of the contributors are always listed in open source projects. Other individual motivations, like altruism or opportunity to express creativity are also mentioned. It is important to point out that in recent years, open source projects have not only received contributions by individuals. There have been organised efforts by firms like Sun, IBM and others that have endorsed such projects. The survey [6] conducted among firms, as well as the account of [20] of Sun Microsystems and [22] list strategic reasons behind the motivation of firms to contribute to open source projects. These reasons include efforts to undercut rival products, gaining a wider tester base for their own products, initiating a gift econ- omy culture between the firm and the open source developer community (where the firm provides the software for free and the community provides debugging or more source code in return) and giving out the software to clients in order to charge for its maintenance and support. 2The information reflects the year 2002-2003 and was obtained from IBIS World, a strategic business information provider. http://www.ibisworld.com/snapshot/industry/default.asp?page=industry&industry id=1239 accessed on 27/05/2005. 3In open source software, the source code for a program is made open and available for anyone to screen. There are different open source licenses which prescribe what one is allowed to do with the source code e.g. screen it, interpret it, make changes etc. This is in contrast to proprietary software licenses where the source code is protected by property rights against modification. 3 Previous work in this area includes that of Johnson in [28] and Bessen in [2] who have used mathematical models to explain the emergence of the open source initiative. John- son focuses more on analysing the individual motives and establishing the relationship between the size of the developer base and whether the development goes on. On the other hand, Bessen concentrates on the firm motives for participation in open source initiatives. Bessen, models software as a bit string, each bit being a certain feature of the software. In this way the notion that the number of combinations of features grows exponentially with the number of features is captured, depicting the complexity the software can have. In his work, he compares open source development with proprietary, pre-packaged provision of software and concludes that the two complement each other, recognising that they serve different groups of customers. The latter suits customers with standard, non-complex software needs, while the former serves customers who have software development capabilities and who need more complex software products. Bonaccorsi and Rossi in [5] have designed a multi-agent system simulation with which they explore the circumstances for adoption of open source software. They also conclude that proprietary and open source software will coexist in the future. Their model of the diffusion of the two competing streams of software production takes into account issues like the effect of advertising, network externalities and achievement of critical mass as in [34]. The stylised model presented in this work simulates a market in which the companies try to satisfy a set of un- derlying software needs with the services that they develop. The companies follow simple, high-level rules imposed by a capitalistic economy. Interestingly, exchanges between the agents similar to the ones that happen in real software markets, arise in the system. This behaviour emerges in the system even though we have avoided modelling issues like social or strategic motives of the contributors or network effects. IV. AN AGENT-BASED MODEL OF THE DBE A. Agent-based Modelling Agent-based modelling has been recently used in Eco- nomics research work to study models of markets, e.g. the Santa Fe artificial stock market [4], [32], and their character- istics [31], in Computing-Economics interdisciplinary work to study information economies of autonomous agents [14], [23], [29], [30], [39] and business processes [26], in Social Sciences to study emergent behaviour [17], issues of trust [18] and to perform syndromic behaviour surveillance [10] and in other disciplines. Much research in multi-agent systems explores how refine- ments to one agent's reasoning can affect the performance of the system [8]. Significant effort has been directed towards formally defining emergence in agent-based systems. A strong emergent property is a property of the system that cannot be found in the properties of the system's parts or in the interactions between the parts [1]. Additionally, in [42] the notion of universality is studied: systems whose elements differ widely may have common emergent features. Agent-based modelling according to [41] "is a method for studying systems exhibiting the following two properties: 1) the system is composed of interacting agents; and 2) the system exhibits emergent properties, that is, prop- erties arising from the interactions of the agents that cannot be deduced simply by aggregating the properties of the agents." In models like the one proposed below, where the interaction of the agents is determined by past experience and the agents continually adapt to that experience, mathematical analysis is typically very limited in its ability to derive the dynamic consequences. In this case, agent-based modelling might be the only practical method of analysis. We follow a 'bottom-up' approach, after a brief overview of the methods used in section VII which follows, in sections IV-B and IV-C we describe the first principles of agent behaviour and in section V we analyse the macro-properties emerging from the agent interactions. B. The setting In this section, the model used for the simulation of the DBE environment is set out. SMEs are modelled as agents in a multi-agent system. The services the SMEs provide are modelled as bit strings in the same manner software services are modelled in [2], each bit symbolising a feature of the service. Finally, the underlying market is modelled by a set of requests (market needs) which are exogenous and are generated randomly. A request is a bit string of the same size as a service bit string. Each SME has a population (or portfolio) of services. This population is not static throughout the lifetime of the SME. If a service is successful, the SME tends to add similar services to the portfolio while an unsuccessful service is usually discarded. The whole process is modelled quite elegantly by a genetic algorithm (GA) within the portfolio which involves mutation and crossover with survival of the fittest. Through this population each SME can choose which request it will try to satisfy. The genetic algorithm represents the R&D businesses perform in order to improve their services. An overview of genetic algorithms is given in appendix VII. The use of genetic algorithms is a natural and simple way to model R&D, with minimal assumptions. The GA captures the following characteristics: 1) trying to find a solution to a particular problem, 2) using a population of possible solutions. Any other method that can capture the above two characteris- tics may be used in place of the GA. The objective of an SME is to increase its fitness. Each SME maintains a portfolio of candidate services, only one of which will be submitted to the market. Each candidate service receives a rating according to how profitable it would be for the SME if it was submitted to the market. This calculation is performed using the services submitted by all other SMEs in the previous round. The rating of each candidate service within the SME portfolio is used to: a) decide on which service to submit to the market and b) evolve the best services in the 4 portfolio (with mutation and crossover) and eliminate the worst services. The fitness of a service measures how profitable it is to its owner. The profitability of a service depends on: 1) how close the service is to the market needs (service- request similarity) and 2) how many other services satisfy those needs (limited demand). The fitness of an SME equals the fitness of the service it offers. In the section that follows we discuss the factors that affect the fitness (or profitability) of a service. 1) Service-Request Similarity and Limited Demand: As- sume there are m SMEs in the market, each one offering a single service. Consider a service S and a request R, each represented by a bit string of fixed length. Similarity is measured by the percentage of shared bit values between S and R, denoted by d(Ri, Sj), 0 ≤ d ≤ 1. If the market requests are R1, R2, ..., Rn, services in the market are S1(t), . . . , Sm(t), the fitness of a service Sj(t) is Uj(t) = n Xi=0 (φ(Ri, Sj(t)) × ρi(t)), φ(Ri, Sj(t)) = e− 1−d(Ri,Sj (t)) α2 . (1) (2) where The variable φ is used to parametrise the fitness landscape (make maxima more or less pronounced), α being a shape parameter. Figure 1 shows the relationship of φ with with the similarity d. The weight/discounting factor ρ is given by ρi(t) = min(1, 1 Pj=1 φ(Ri, Sj(t))) . (3) The variable ρ models the fact that the demand in the market is limited. When a request is saturated (i.e. too many services try to satisfy it) then ρ < 1. Subsequently, the fitness of the service is discounted. Otherwise, when ρ = 1 the fitness of the service equals φ. The fitness of an SME is equal to the fitness of the service it submits to the market. 2) Satisfaction of Requests and Market Efficiency: An additional useful measure is the degree to which a request is satisfied. This is a metric of how saturated it is, in terms of how many services try to satisfy it and how similar their features are to those of the request. The degree of satisfaction Qi(t) of a request Ri at round t is given by: Qi(t) = m Xj=1 φ(Ri, Sj(t)). (4) This measure is necessary for assessing the efficiency of the DBE market. As discussed in section II-B, in an efficient DBE market all the market requests will be equally saturated, assuming there is the same demand for all of them. Thus, we calculate the standard deviation σ(t) of the satisfaction values of all the requests in the market at round t. The smaller it φ 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.2 0.4 0.6 0.8 1 d (percentage similarity between service and request) Fig. 1. The relationship of φ with the service-request similarity d for a = 0.2. The variable φ is used to parametrise the fitness landscape (make maxima more or less pronounced). is, the more similar to each other the saturation levels of the requests are. σ(t) = stdev{Q1(t), . . . , Qn(t)} (5) The mean of the saturation values will be constant due to the demand in the model being fixed. C. Exchange of Services As outlined in III-A exchange of services may encompass many real-life situations that occur in a market. These include the forming of strategic partnerships of companies, partici- pation in free/open source projects and others. The setting described here is a loose model of such situations which aims to identify the basic factors that lead to this general behaviour of exchanging. In our model, the exchange involves selecting a set of services from one SME's portfolio and swapping them with the corresponding set of services of the other SME's portfolio. When a company chooses to swap a set of services, this means that after the exchange has taken place it won't have these services in its portfolio any more. The services in a portfolio of a company are sorted according to their fitness (i.e. how profitable they are to the SME that owns them). The model in its current state supports exchange of services that are in the same rank, in the two portfolios, e.g. the 5th service in the portfolio of one SME with the 5th service in the portfolio of the other4. At each time tick, the SMEs need to decide whether they want to exchange some of their services with one of the other SMEs. A statistical classification algorithm is used to model the decision problems an individual agent faces. An overview of statistical classification is given in Appendix VII. 4Experiments have shown that the rank of the services being exchanged is not of much significance, assuming that services of the same rank are being exchanged, but we plan to investigate this further in the future. 5 1) Exchange decisions: Every SME has a classifier system which it uses to decide on whether they want to exchange some of their services with one of the other SMEs. The rules of the classifier are shown in table I below. The objective of an SME at all times is to increase its fitness. The rules' condition part refers to the rank of the SME in the market with respect to the rank of its colleagues. The action part examines the potential partner's rank and prompts the SMEs either to engage in an exchange with a specific type of partner or abstain from exchanging. For simplicity, the SMEs are clustered in three5 groups according to their rank. Therefore we have upper, middle and lower ranked SMEs. For an exchange to take place both parties need to agree. We experiment both with settings in which the rank is based on the fitness of each company and others where the rank is not linked to SME performance in any way. For example, in experiments where rank is based on SME performance, the SME with the highest fitness will have rank = 1, whilst the SME with the lowest fitness will have rank = number of SM Es. On the other hand, in experiments where rank is unrelated to performance in the market the rank of an SME may be its id number. In section V we analyse these experiments and present the effect the different meanings rank may take have on the learning that occurs. if my rank = lower if my rank = lower if my rank = lower if my rank = lower if my rank = middle if my rank = middle if my rank = middle if my rank = middle if my rank = upper ... then then then then then then then then then ... TABLE I exchangewithlowercluster, exchangewithmiddlecluster, exchangewithuppercluster, donotexchange, exchangewithlowercluster, exchangewithmiddlecluster, exchangewithuppercluster, donotexchange, exchangewithlowercluster, s1 s2 s3 s4 s5 s6 s7 s8 s9 ... A FEW EXAMPLE RULES OF THE CLASSIFIER WHICH AN SME USES TO DECIDE ON WHAT TYPE OF PARTNER TO CHOOSE FOR AN EXCHANGE. The classifier system operates as follows [31]. First, it examines the if part of each rule to determine and shortlist the rules whose conditions are satisfied at a given time t. It then assigns a score b to the shortlisted rules, sk being the strength of the kth rule: bk(t) = sk(t) + ε, where ε ≃ N (0, σ). (6) The rule with the highest score b becomes the activerule. After the active rule has been executed and has generated payoff ω during the previous round t − 1, the classifier system updates its strength s: sk = sk(t − 1) − csk(t − 1) + cω(t − 1), where c ∈ [0, 1]. (7) In other words, ∆sk(t) = c[ω(t− 1)− sk(t− 1)]. Therefore, as long as the payoff in round t− 1 is greater than the strength 5Experiments have been carried out which showed that model behaviour doesn't vary significantly with cluster size. Three is the optimal number of clusters with respect to having a model which is realistic enough while taking a reasonable amount of time to execute and giving us the ability to present the results in an efficient and clear way. of the rule on that round, the strength will increase. If the selection of the rule led to a small payoff being generated, the strength of the rule will decrease, making it less likely to be activated in the future. The strength of each rule converges to some weighted average of the rewards ω generated by the environment in response to that specific rule. In our implementation of the model all the rules have initial strength 0. The rule strengths are adjusted as the simulation goes on. The strength of each rule that is activated is updated at every round using the following payoff from the external environment: ω(t) = Uj(t) − Uj(t − 1). In other words, the payoff is the difference in the fitness of the company between the current and the previous round. The payoff may be negative, zero, or positive according to the change in fitness. 2) Exchange decisions resolution: Once the companies that have decided to participate in an exchange have selected the type of partner they prefer, they are teamed up accordingly. For instance, an SME in the cluster of middle ranked SMEs, who has decided to exchange with a high fitness company will be coupled with a high ranked company who wants to exchange with a middle ranked one. If a suitable partner is not found the exchange does not happen. The strength of the rule that was activated in that case will still be updated even if the transaction was not carried out. This reflects the effect choosing a partner who is unwilling to collaborate has on the fitness of the company. D. Discussion The model outlined above is simple in that it has captured the main aspects of a digital business ecosystem. It is the model of a market in which the companies try to satisfy a set of underlying requests. They do so by producing and making available services that are as close as possible to the specified requests. Each company has its own R&D portfolio of services that it evolves. At each round the companies go to the market with what they believe is the best service in their portfolio. In addition, the companies have an option to exchange services with partners that they select themselves. The simplicity of the model is also inherent in the behaviour of the agents. The agents have to find which is the best service to make available, based on the services that were submitted to the market during the previous round. Also, they need to decide whether and with whom to exchange their services based on their rank in the market. These are all abstractions from reality. We do not assume any network effects in the market. Also, there are no indicators about value of the brand of a company. V. ANALYSIS OF THE MODEL In this section the experiments carried out using the model of the DBE are described. The analysis focuses on two main findings: 1) The companies discover themselves that under certain circumstances it is beneficial to them to exchange ser- vices between them. 2) Allowing exchange to take place in the market, makes for greater market efficiency levels. 6 this point to stress that is important at It the choice to exchange services is not a practice that is imposed by the model mechanism. Instead, it is a feature that emerges from the classifiers as it is a gainful practice for the companies under certain circumstances. The model behaviour is quite general and has been observed for a very wide range of parameters and initial conditions. The graphs and figures shown below come from randomly selected runs of the simulation, unless it is stated otherwise. A. Service Exchange 1) Exchange Decision: As described in section IV-C each agent/company uses a classifier to decide whether or not to exchange some of its services. The decision is based on the company's rank in the market. Figures 2(a) and 2(b) show the average strength of the rules of all the companies' classifiers at the end of a simulation which lasted for 10 000 iterations. The companies are ranked according to their fitness. The fittest company will have rank 1 whilst the least fit company will have rank equal to the number of companies in the market. To make for less time consuming simulations and more readable graphs the companies are grouped into three clusters according to their rank; so they are divided into lower, mid and upper ranked SMEs. Figure 2(a) was generated from a run of the simulation where the DBE market consisted of 21 SMEs, each having 20 services in its portfolio. Each service had 10 features. There were 4 software requests in the market, generated randomly. The run of the simulation which produced figure 2(b) had largely similar parameters, the difference being that there were 30 services in the SMEs' portfolios and there were 5 requests in the market. The strongest of the rules at each situation is the one which is more likely to be activated. In other words, it is shown in figures 2(a) and 2(b) that if a company belongs to the mid or lower cluster it is likely that it will choose to participate in an exchange (preferably with a upper ranked company) while if it belongs to the upper ranked cluster it will avoid engaging in any exchange activities. The graphs show that in the less successful, lower ranked SMEs the classifier rules that correspond to exchange actions have higher strengths than the rule that leads SMEs not to exchange. The opposite holds for higher ranked SMEs, i.e. the rule that corresponds to a not exchange action has higher strength than the exchange rules. For mid-ranked SMEs, a rule prompting the firm to exchange is the stronger of all, but exchanging is not always a profitable practice; the rule that leads the SME to avoid exchanging is often stronger than some exchange rules. The generality in the behaviour of the model is confirmed by figure 2(c). A wide range of parameters and initial conditions were varied in a total of 200 experiments, keeping the number of SMEs in the market constant (21). Figure 2(c) shows the average values of the SME classifiers' strengths over those 200 experiments. The general trend which emerges is that the average performing (mid cluster) and worst performing (lower cluster) SMEs learn that it is to their advantage to exchange services with others while the top performers (upper cluster) learn to avoid exchanging . t h g n e r t l S e u R e g a r e v A 10 8 6 4 2 0 −2 −4 −6 −8 −10 Run 1 Action: Exchange with upper cluster Action: Exchange with mid cluster Action: Exchange with lower cluster Action: Do not Exchange t h g n e r t l S e u R e g a r e v A if my rank is upper if my rank is mid if my rank is lower Rule 15 10 5 0 −5 −10 Run 2 Action: Exchange with upper cluster Action: Exchange with mid cluster Action: Exchange with lower cluster Action: Do not Exchange t h g n e r t l S e u R e g a r e v A if my rank is upper if my rank is mid if my rank is lower Rule 7 Average Rule Strenghts over 200 experiments Action: Exchange with upper cluster Action: Exchange with mid cluster Action: Exchange with lower cluster Action: Do not Exchange if my rank is upper if my rank is mid if my rank is lower Rule 10 8 6 4 2 0 −2 −4 −6 −8 −10 (a) Run 1 (b) Run 2 (c) Average over 200 Experiments Run with Rank based on SME Fitness Growth Rates Run with Rank based on the 20−moving average of the SME fitness t h g n e r t l S e u R e g a r e v A 10 8 6 4 2 0 −2 −4 −6 −8 −10 Action: Exchange Action: Do not Exchange if my rank is upper if my rank is mid if my rank is lower Rule t h g n e r t l S e u R e g a r e v A 10 8 6 4 2 0 −2 −4 −6 −8 −10 Action: Exchange Action: Do not Exchange if my rank is upper if my rank is mid if my rank is lower Rule (d) Rank based on Fitness Growth Rate (e) Rank based on Fitness Moving Average Run with Rank based on the SME Id which is static Run with Rank based on the SME Id which is changing h t g n e r t l S e u R e g a r e v A 10 8 6 4 2 0 −2 −4 −6 −8 −10 Action: Exchange Action: Do not Exchange if my rank is upper if my rank is mid if my rank is lower Rule h t g n e r t l S e u R e g a r e v A 10 8 6 4 2 0 −2 −4 −6 −8 −10 Action: Exchange Action: Do not Exchange if my rank is upper if my rank is mid if my rank is lower Rule (f) Rank based on SME Id which is random and static throughout the simulation (g) Rank based on SME Id which is random and constantly changing throughout the sim- ulation Fig. 2. Average Exchange Rule Strength The graphs show the strength values of each rule at the end of a simulation averaged out over all SMEs' classifiers. The SMEs decide whether to participate in an exchange of services according to their rank. The classifier each SME has is as follows: For figures 2(a)-2(e) the rank of the SMEs is based on measures related to their fitness, while figures 2(f) and 2(g) were created for settings in which the SME rank was unrelated to fitness. The graphs show in settings where the rank is associated with some fitness measure the SMEs that are further down in the rank learn that is beneficial to them to participate in an exchange. 2(a) Run 1 parameters: 21 SMEs, each having 20 services in its portfolio. Each service had 10 features. There were 4 software requests in the market. The rank was based on the fitness value of the SME. 2(b) Run 2 parameters: 21 SMEs, each having 30 services in its portfolio. Each service had 10 features. There were 5 software requests in the market. The rank was based on the fitness value of the SME. 2(c) Average values over 200 experiments This figure confirms the generality of the behaviour of the model. A wide range of parameters and initial conditions were varied in a total of 200 experiments, keeping the number of SMEs in the market constant (21). The rank was based on the fitness value of the SME. 2(d) and 2(e) Average Exchange Rule Strength based on SME performance measures. The SMEs decide whether to participate in an exchange of services according to their performance. In 2(d) the performance measure deciding the rank of the SMEs is their fitness growth rate, while in 2(e) it is the 20-moving average of the SME fitness. When the ranking of the SMEs is performance related information exchange emerges as a gainful strategy. 2(f) and 2(g) Average Exchange Rule Strength not based on SME performance measures. In 2(f) the SMEs decide whether to participate in an exchange of services according to their unique id. In 2(g) the ranking of the SMEs is random and constantly changes. In both cases, the ranking is unrelated to SME fitness or any other performance measure. The rule strengths indicate that no rule is significantly more important than any other one implying that the rules if my rank = lower if my rank = lower if my rank = lower if my rank = lower if my rank = middle ... then then then then then ... exchangewithlowercluster, exchangewithmiddlecluster, exchangewithuppercluster, donotexchange, exchangewithlowercluster, s1 s2 s3 s4 s5 ... To understand better the behaviour of the system we per- formed experiments with different rankings of the SMEs. Amongst the ranking methods we tested were variants of the fitness ranking, as well as rankings unrelated to SME performance altogether. The results seem to indicate that information exchange emerges as long as the ranking is in some way related to SME performance. We show in figure 2(d) the rule strengths in the case the SMEs were ranked according to fitness growth rates ∆Uj(t) = Uj(t) − Uj(t − 1), (8) rather than fitness itself. The graphs produced are similar in pattern to those in figure 2(c). These strengths imply that the rules are significant and learning has taken place in the system. Similar results, shown in figure 2(e), were produced when SMEs were ranked according to the N-moving average of their fitness, given by µ = 1 N t XT =t−N Uj(T ). (9) On the other hand, in figure 2(f) a typical case of a ranking that is unrelated to SME fitness is shown. In that particular case we gave the SMEs an arbitrary ranking that remained fixed throughout the simulation. The rule strengths indicate that no rule is significantly more important than any other one implying that the rules are not relevant and no learning has occurred. We also tried a completely random and constantly changing SME ranking which produced similar results, shown in figure 2(g). 2) Choice of Exchange Partner: An interesting result which arose from the experiments is the choice of potential partners for the companies who decide to exchange. In all three situ- ations (if my rank is upper, if my rank is mid and if my rank is lower) the strength of the rules that prompt SMEs to exchange reveal a decreasing preference from left to right between upper, mid and lower ranked partners. That result is entirely intuitive and confirms the validity of the model. A result that might not be so obvious is the fact that the lower ranked SMEs benefit from exchanging even between themselves. This is reflected in the fairly high strength of the relevant rule and it is better illustrated in figure 3. The experiment that yielded figure 3 is as follows. To make for a more intelligible graph, there are only six SMEs in the market and two distinct requests. Every 400 rounds the underlying requests in the market change. Every 200 rounds (but not when the requests change), the lower ranked SMEs exchanged services between them. As the purpose of this experiment was to verify the finding that exchange among lower ranked SMEs is beneficial, the exchange was done deliberately and not using the classifier. As shown in figure 3, in round 200 the exchange does not upset the equilibrium too much as the SMEs have more or less the same fitness. In round 600 the exchange drives the lower ranked SMEs up, whilst damaging the fitness of the others in the market. In round 1000 the exchange not only drives the under-performers up but also causes one of them, SM E1 to join the upper cluster. Fitness of SMEs(cid:13) s(cid:13) s e n t i F 6(cid:13) 5(cid:13) 4(cid:13) 3(cid:13) 2(cid:13) 1(cid:13) 0(cid:13) 8 SME0(cid:13) SME1(cid:13) SME2(cid:13) SME3(cid:13) SME4(cid:13) SME5(cid:13) 1(cid:13) 201(cid:13) 401(cid:13) 601(cid:13) 801(cid:13) 1001(cid:13) 1201(cid:13) 1401(cid:13) 1601(cid:13) 1801(cid:13) 2001(cid:13) time(cid:13) Fig. 3. This is an experiment that illustrates that exchange among lower ranked SMEs is beneficial to them. Every 400 rounds the underlying requests in the market change. Every 200 rounds (but not when the requests change), the lower ranked SMEs exchanged services between them. In most instances the exchange drives the under-performers up, in terms of fitness. The experiment described above illustrated that exchanges between low-ranked SMEs can be highly beneficial. This is because the fusion of their portfolios might yield services that enable them to operate in a new market segment, in other words it may lead them to satisfy another request which was previously not catered for. This can cause their rank in the market to improve and even bring about a change of leadership in the industry. B. Market Efficiency As discussed in section II-B, the increased flow of informa- tion within the DBE, will make it easier for the participating companies to find the right trading partners. Consequently, it will make for greater market efficiency levels in comparison to a conventional market (e.g. the software industry). An interesting observation which emerged from the analysis of the simulations carried out is that allowing the SMEs to exchange services between them, increases the efficiency further. A DBE market is considered efficient when all the requests are equally saturated. In an efficient DBE market, the supply of services will adjust immediately to any arising information about the underlying requests. In other words, there is no excess profit to be gained by an SME choosing to satisfy another request than the ones it currently does. As mentioned in section IV-B.2, the degree of satisfaction of a request R is given by equation 4. In order to assess the level of efficiency in the market we need to calculate the standard deviation σ(t) of the satisfaction values of all the requests in the market, as given by equation 5. The smaller it is, the more similar to each other the saturation levels of the requests are. It is important to mention at this point that the mean of the saturation levels remains constant, because in the model we assume equal demand for all of them, and it is equal to number of services in the DBE . number of requests Figure 4 shows the standard deviation σ(t) of the saturation values Qi(t) of all the requests {R1, . . . , R4} in the market, for two different runs of the DBE simulation. Both runs had been initialised with the same parameters, for one of them exchange between the SMEs was not permitted, whereas for the other one the SMEs were free to exchange services with each other according to the procedure detailed in section IV-C. In order to train the classifiers used for the exchange decisions, every 500 rounds all SMEs' portfolios were reset to the services they had at round 0. To make comparison easier, the resetting of the portfolios was also done during the run where exchange was not allowed. In effect, in this experiment, 'history' repeats itself every 500 rounds. This is the reason spikes occur in the graph every 500 rounds. When exchange is permitted, the SMEs are given the chance to exchange services with each other at rounds 250, 750, 1250, 1750, etc. The graph shows a period of 5000 rounds, when the classifiers have been sufficiently trained. Market efficiency: with and without exchange(cid:13) With Exchange(cid:13) Without Exchange(cid:13) s(cid:13) l e v e L 3(cid:13) 2.5(cid:13) 2(cid:13) 1.5(cid:13) 1(cid:13) 0.5(cid:13) 0(cid:13) (cid:13) n o i t a r u t a S ' s t s e u q e R e h t f o . v e D . t S 0(cid:13) 500(cid:13) 1000(cid:13) 1500(cid:13) 2000(cid:13) 2500(cid:13) 3000(cid:13) 3500(cid:13) 4000(cid:13) 4500(cid:13) Time(cid:13) Fig. 4. Market Efficiency: We assess the level of market efficiency by plotting the standard deviation of the saturation degrees of the requests in the DBE Market. The smaller the standard deviation, the greater the market efficiency. The graph contrasts these data for a situation in which the SMEs are allowed to exchange services with each other and for a situation where exchange is not allowed. The standard deviation of the saturation degrees of the requests is significantly smaller when exchange is allowed, indicating a more efficient market. For classifier training purposes every 500 rounds all SMEs' portfolios were reset to the services they had at round 0. In the case where service exchanges are allowed, these happen in the middle of each cycle, i.e. at rounds 250, 750, 1250, 1750, etc. It is evident from the graph, that when exchange of services between SMEs is allowed, the standard deviation of the requests saturation values is considerably smaller. In other words, the requests in the market are more evenly satisfied. This result is quite invariant to initial conditions and pa- rameters of the simulation. So in the system described, not only will SMEs adopt information exchange as beneficial to their individual progress, but it will also result in a global improvement to the efficiency of the market. Again this is in agreement with what is observed in real economies where open standards, publication of innovations and dissemination of ideas lead to highly efficient markets. VI. CONCLUDING REMARKS The aim of this work has been to study the rationale as well as the effect of knowledge exchange in economic markets. We focus especially on the software industry, our findings, however, to some extent apply to other industries as well. Sharing of information between commercial firms is considered controversial. Although it is acknowledged that when two companies join forces to develop an innovative product they can both benefit, sharing trade secrets is not undertaken lightly. Our main aim has been to formalise a plausible and elegant explanation of how and why companies adopt information exchange and why it benefits the market as a whole when this happens. 9 An agent based model of a Digital Business Ecosystem market has been implemented to assist us in understanding the dynamics of the market mechanisms. Firms are modelled as agents with minimal reasoning capabilities. We investigated the properties that emerge from the agent interactions that occur in the market. Specifically, we examined two key characteristics that we observed in the simulations carried out. Namely, the fact that the agents discover themselves that under certain circumstances it is beneficial for them to exchange services and that allowing exchange to take place in the market, makes for greater market efficiency levels. The technologic infrastructure of the DBE will facilitate the dissemination of knowledge among the member SMEs, increasing the volume and the speed of the information flowing in the market. As a result, it is expected that it will allow for greater market efficiency levels in comparison to a conven- tional market. Admittedly, it is difficult to compare the market efficiency of two different markets. However, an interesting result arose when we performed simulations of the DBE contrasting settings in which exchanges among SMEs were permitted with settings where exchanges were not permitted. Exchanges among SMEs within the DBE further increase the efficiency of the market, which is in agreement with the common intuition that exchanging information is ultimately beneficial for the entire market. The second and most important conclusion that emerged from the DBE simulation is that exchanges between the agents similar to the ones that happen in real-life arise naturally in our system. At regular time intervals, the SMEs were given the chance to decide whether they wanted to choose a partner and swap some of their services. The decision was taken using classifiers, which were separate for each agent. The agents were not pre-programmed or biased in any way to engage in exchanges. The SMEs, on their own, discovered in which cases exchanging is beneficial for them and what type of partner is the best. Exchange is a practice that emerges, and is not forced upon the agents. This work does not directly advocate knowledge exchange as a means of increasing profitability of software companies. Knowledge exchange, is indeed an already existing phe- nomenon in industry as explained in section III-A. The results presented merely serve as a demonstration of a parsimonious set of assumptions that give rise to exchange in a software market. In other words, we identify the substance of this phenomenon, ridding it from unnecessary assumptions, like network effects, social issues of trust, or managerial strategies and show the minimal set of assumptions that allow it to emerge. VII. METHODOLOGY: EVOLUTIONARY ALGORITHMS In order to model evolution in populations as well as learning we have used several evolutionary algorithms in our model. In this section we give a brief overview of these algorithms. Evolutionary algorithms [7] 'is an umbrella term employed to describe computer-based problem solving systems which use computational models of some of the known mechanisms of evolution as key elements in their design and implementa- tion.' A variety of evolutionary algorithms have been proposed by several researchers. The major ones are: genetic algorithms, evolutionary programming, evolution strategies, classifier sys- tems and genetic programming. They all share a common concept of simulating the evolution of objects/structures using the processes of selection, mutation and reproduction. The processes depend on the performance/fitness of the individuals under consideration as defined by their environment and quantified by a fitness function. More precisely, evolutionary algorithms maintain a popula- tion of structures, that evolve according to rules of selection and other operators, that are referred to as "search operators" (or genetic operators), such as recombination and mutation. Each individual in the population receives a measure of its fitness in the environment. Reproduction focuses attention on high fitness individuals, thus exploiting the available fitness information. Recombination and mutation perturb those indi- viduals, providing general heuristics for exploration. Although simplistic, these algorithms are sufficiently complex to provide robust and powerful adaptive search mechanisms. A genetic algorithm (GA) [21] is a model of machine learning inspired by the mechanisms of genetics, which has been applied to optimisation. It operates with an initial pop- ulation containing a number of trial solutions. Each member of the population is evaluated (to yield a fitness) and a new generation is created from the better of them. The process is continued through a number of generations with the aim that the population should evolve to contain an acceptable solution. In [36] it is stated that GAs are particularly suitable for solving complex optimization problems and hence for applications that require adaptive problem-solving strategies. In order to make genetic algorithms reach an optimal solution faster, parallel implementations of GAs are often used [9]. Genetic algorithms are used for a number of different appli- cation areas. An example of this would be multidimensional optimisation problems in which the character string of the chromosome can be used to encode the values for the different parameters being optimised. In practice, therefore, we can implement this genetic model of computation by having arrays of bits or characters to rep- resent the chromosomes. Simple bit manipulation operations allow the implementation of crossover, mutation as well as other operations. Crossover involves combining strings to swap values, e.g. 101001 + 111111 → 101111. Mutation involves spontaneous alteration of characters in a string, e.g. 000101 → 100101. Although a substantial amount of research has been performed on variable-length strings and other structures, the majority of work with genetic algorithms is focused on fixed- length character strings. Statistical classification is a type of supervised learning algorithm which takes a feature representation of an object or concept and maps it to a classification label. A classification algorithm is designed to learn, or in other words, to approx- imate the behaviour of a function which maps a vector of features [X1, X2, ..., Xn] into one of several classes by looking at several input-output examples of the function. An instance of a classification algorithm is called a classi- 10 fier. Learning Classifier Systems [25] are a machine learn- ing technique which combines evolutionary computing and reinforcement learning to produce adaptive systems. It is a minimal form of modelling learning in the sense that it is not necessary to make assumptions about the way the agents perform their reasoning. In addition to that, the absence of any assumptions or biases in the learning process leads to results that can be generalised. A classifier consists of a set of rules, which have a condition C (if part) an action A (then part) and a strength measure s. An example of a classifier system is shown in table II. if C1 if C2 if C3 if . . . ... then A1, then A2, then A3, then . . . ... s1 s2 s3 . . . ... TABLE II AN EXAMPLE OF A CLASSIFIER SYSTEM. In the model described in detail in section IV-B, genetic algorithms and classification algorithms have been used to model evolution of populations of solutions and learning. REFERENCES [1] Yaneer Bar-Yam. A mathematical theory of strong emergence using multiscale variety. Complexity, 9(6):15 -- 24, 2004. [2] James Bessen. Open source software: Free provision of complex public goods, 2002. Unpublished working paper, Research on Innovation. [3] Z. Bodie, A. Kane, and A. Marcus. Investments, 5th Edition, chapter 12: Market Efficiency. McGraw-Hill and Irwin, 2002. [4] Eric Bonabeau, Marco Dorigo, and Guy Theraulaz. Swarm intelligence: from natural to artificial systems. Oxford University Press, Inc., New York, NY, USA, 1999. [5] Andrea Bonaccorsi and Cristina Rossi. Why Open Source software can succeed. Research Policy, 32(7):1243 -- 1258, 2003. [6] Andrea Bonaccorsi and Cristina Rossi. Altruistic individuals, selfish firms? The structure of motivation in Open Source software. First Monday, 9(1), 2004. [7] P.B. (ed.) Brazdil. Editorial, Machine Learning: Proceedings of the European Conference on Machine Learning. Springer, New York, NY, USA, 1993. [8] T. Brenner. Local Industrial Clusters, Existence, Emergence and Evolution. Studies in Global Competition. Routledge, London, 2004. [9] E. Cantu-Paz. A survey of parallel genetic algorithms. Calculateurs Paralleles, Reseaux et Systems Repartis, 10(2):141 -- 147, 1998. [10] K.M. Carley, D.B. Fridsma, E. Casman, A. Yahja, N. Altman, L.-C. Chen, B. Kaminsky, and D. Nave. Biowar: Scalable agent-based model of bioattacks. IEEE Transactions on Systems, Man and Cybernetics, Part A, 36(2):252 -- 265, 2006. [11] M. Chli, P. De Wilde, J. Goossenaerts, V. Abramov, N. Szirbik, L. Cor- reia, P. Mariano, and R. Ribeiro. Stability of multi-agent systems. In Proceedings of the 2003 IEEE International Conference on Systems, Man, and Cybernetics, pages 551 -- 556, 2003. [12] A. Damodaran. Investment Valuation, 2nd Edition, chapter 6: Market Efficiency - Theory and Models. Wiley, 2001. [13] DBE. Annex I - Description of Work, Digital Business Ecosystem. Technical report, 2002. [14] P. De Wilde. Fuzzy utility and equilibria. IEEE Transactions on Systems, Man and Cybernetics, Part B, 34(4):1774 -- 1785, 2004. [15] P. De Wilde, M. Chli, L. Correia, R. Ribeiro, P. Mariano, V. Abramov, and J. Goossenaerts. Adapting populations of agents. Lecture Notes in Artificial Intelligence, 2636:110 -- 124, 2003. [16] T. Eguchi, K. Hirasawa, J. Hu, and N. Ota. Aa study of evolutionary multiagent models based on symbiosis. IEEE Transactions on Systems, Man and Cybernetics, Part B, 36(1):179 -- 193, 2006. [43] T. Yamasaki and T. Ushio. An application of a computational ecology model to a routing method in computer networks. IEEE Transactions on Systems, Man and Cybernetics, Part B, 32(1):99 -- 106, 2002. 11 [17] Joshua M. Epstein. Modeling civil violence: An agent-based compu- tational approach. Proceedings of the National Academy of Sciences, U.S.A., 99(10, Supplement 3):7243 -- 7250, 2002. [18] R. Falcone and C. Castelfranchi. The human in the loop of a delegated agent: the theory of adjustable social autonomy. IEEE Transactions on Systems, Man and Cybernetics, Part A, 31(5):406 -- 418, 2001. [19] E. F. Fama. Efficient capital markets: A review of theory and empirical work. Journal of finance, 25:383 -- 417, 1970. [20] Richard P. Gabriel and Ron Goldman. Open source: Beyond the fairytales, 2002. [Online; accessed 25-May-2005]. [21] David E. Goldberg. Genetic Algorithms in Search, Optimization and Machine Learning. Addison-Wesley Longman Publishing Co., Inc., Boston, MA, USA, 1989. [22] Simon Grand, Georg von Krogh, Dorothy Leonard, and Walter Swap. Resource allocation beyond firm boundaries: A multi-level model for open source innovation. Long Range Planning, 37(6):591 -- 610, 2004. [23] Nathan Griffiths and Michael Luck. Coalition formation through motivation and trust. joint conference on Autonomous agents and multiagent systems, pages 17 -- 24, New York, NY, USA, 2003. ACM Press. In Proceedings of the 2nd international [24] John Hagedoorn. Inter-firm R&D partnerships: an overview of major trends and patterns since 1960. Research Policy, 31(4):477 -- 492, 2002. [25] J. H. Holland. Adaptation. Progress in Theoretical Biology, 4:263 -- 293, 1976. [26] Chun-Che Huang. Using intelligent agents to manage fuzzy business processes. IEEE Transactions on Systems, Man and Cybernetics, Part A, 31(6):508 -- 523, 2001. [27] Z. Jing, E. Billard, and S. Lakshmivarahan. Learning in multilevel games IEEE Transactions on Systems, Man with incomplete information. ii. and Cybernetics, Part B, 29(3):340 -- 349, 1999. [28] Justin Pappas Johnson. Open source software: Private provision of Journal of Economics and Management Strategy, a public good. 11(4):637 -- 662, 2002. [29] J. O. Kephart. Software agents and the route to information economy. Proceedings of the National Academy of Sciences, U.S.A., 99(10, Sup- plement 3):7207 -- 7213, 2002. [30] Jeffrey O. Kephart, James E. Hanson, and Jakka Sairamesh. Price and niche wars in a free-market economy of software agents. Artificial. Life, 4(1):1 -- 23, 1997. [31] Alan P. Kirman and Nicolaas J. Vriend. Evolving market structure: An ACE model of price dispersion and loyalty. Journal of Economic Dynamics and Control, 25(3):459 -- 502, 2001. [32] Blake LeBaron. Market. http://www.econ.iastate.edu/tesfatsi/blake.sfisum.pdf. Working Building Paper, the Santa Fe Artificial June 2002. Available Stock at [33] Josh Lerner and Jean Tirole. Some simple economics of open source. Journal of Industrial Economics, 50:197 -- 234, June 2002. [34] Christoph H. Loch and Bernardo A. Huberman. A punctuated- equilibrium model of technology diffusion. Management Science, 45(2):160 -- 177, 1999. [35] M. Luck, P. McBurney, and C. Preist. Agent Technology: Enabling Next Generation Computing (A Roadmap for Agent Based Computing). AgentLink, 2003. [36] J. L. Ribeiro Filho, P. C. Treleaven, and C. Alippi. Genetic-algorithm programming environments. Computer, 27(6):28 -- 43, 1994. [37] R. Savit, R. Manuca, and R. Riolo. Adaptive competition, market efficiency and phase transitions. Physical Review Letters, 82(10):2203 -- 2206, 1999. [38] Klaus Schmidt and Monika Schnitzer. Public Subsidies for Open Source? Some Economic Policy Issues of the Software Market. Techni- cal Report 3793, C.E.P.R. Discussion Papers, February 2003. Available at http://ideas.repec.org/p/cpr/ceprdp/3793.html. [39] Kwang Mong Sim and Eric Wong. Toward market-driven agents for electronic auction. IEEE Transactions on Systems, Man and Cybernetics, Part A, 31(6):474 -- 484, 2001. [40] M. Sysi-Aho, A. Chakraborti, and K. Kaski. Searching for good strategies in adaptive minority games. Physical Review E, 69(3):36125 -- 1 -- 36125 -- 7, 2004. [41] Leigh Tesfatsion. Handbook of Computational Economics, Vol. 2: Agent- Based Computational Economics, chapter 1, Agent-Based Computa- tional Economics: A Constructive Approach to Economic Theory. North- Holland, 2005. To appear. [42] H. Van Dyke Parunak, Sven Brueckner, and Robert Savit. Universality in multi-agent systems. the 3rd International Joint Conference on Autonomous Agents and Multiagent Systems, pages 930 -- 937, Washington, DC, USA, 2004. IEEE Computer Society. In AAMAS '04: Proceedings of
1707.01546	1	1707	2017-07-05T19:19:55	Agent based simulation of the evolution of society as an alternate maximization problem	[ "cs.MA", "stat.ML" ]	Understanding the evolution of human society, as a complex adaptive system, is a task that has been looked upon from various angles. In this paper, we simulate an agent-based model with a high enough population tractably. To do this, we characterize an entity called \textit{society}, which helps us reduce the complexity of each step from $\mathcal{O}(n^2)$ to $\mathcal{O}(n)$. We propose a very realistic setting, where we design a joint alternate maximization step algorithm to maximize a certain \textit{fitness} function, which we believe simulates the way societies develop. Our key contributions include (i) proposing a novel protocol for simulating the evolution of a society with cheap, non-optimal joint alternate maximization steps (ii) providing a framework for carrying out experiments that adhere to this joint-optimization simulation framework (iii) carrying out experiments to show that it makes sense empirically (iv) providing an alternate justification for the use of \textit{society} in the simulations.	cs.MA	cs	Agent based simulation of the evolution of society as an alternate maximzation problem Amartya Sanyal‡,Sanjana Garg∗,Asim Unmesh† Dept. of Computer Science And Engineering Email: ‡[email protected] ∗[email protected], †[email protected] Indian Institute of Technology, Kanpur (All the authors have equal contributions and the order is random) 7 1 0 2 l u J 5 ] A M . s c [ 1 v 6 4 5 1 0 . 7 0 7 1 : v i X r a Abstract-Understanding the evolution of human society, as a complex adaptive system, is a task that has been looked upon from various angles. In this paper, we simulate an agent-based model with a high enough population tractably. To do this, we characterize an entity called society, which helps us reduce the complexity of each step from O(n2) to O(n). We propose a very realistic setting, where we design a joint alternate maximization step algorithm to maximize a certain fitness function, which we believe simulates the way societies develop. Our key contributions include (i) proposing a novel protocol for simulating the evolution of a society with cheap, non-optimal joint alternate maximization steps (ii) providing a framework for carrying out experiments that adhere to this joint-optimization simulation framework (iii) carrying out experiments to show that it makes sense empirically (iv) providing an alternate justification for the use of society in the simulations. INTRODUCTION The purpose of this work is to develop a framework for modeling the evolutionary dynamics of a society populated with heterogeneous agents. Though history has often tried to portray society as a coherent organization trying to optimize a global objective, society has often yielded agents bound together by constraints(laws) trying to reach their local opti- mum. What is worthwhile to notice is that, just like language, societies that have developed independently in different parts of the world have developed similar characteristics in terms of organization. This might have been because societies that worked in a different way failed to survive the sands of time or because societies, somehow, ended up developing in a similar way. Either way, this suggests that successful societies were bound with certain constraints just like the theory of Universal Grammar and one of them, as have been proposed by Darwin[1], is the survival of the fittest. In other words, all individuals try to maximize their fitness lest they should perish and all societies try to maximize their happiness lest they should perish. We propose that these objectives are related to each other but not in a trivial sense. In a hugely complex interactive heterogeneous system such as a society, this cannot be tractably reduced in complexity than a complex adaptive system[2]. In this system, all agents are motivated to maximize their well being as much as possible. For this text, we will refer to this well being as happiness. It is to be noted that the people are not given a free hand in this optimization step but are often bounded by constraints imposed by society and the natural environment. Constraints imposed by societies include the presence of norms and those induced by environment include climates and geographical features. In this paper, we model these two optimization steps, we develop the characteristics of the agents of this CAS and we also model the society. Interaction in a society intuitively refers to the O(n2) pairs of interactions possible. An alternative way of looking at the characteristics of the society is to look at it as a facilitator for a mathematical simplification of a more expensive computa- tional problem. The happiness of a society is an aggregation of the O(n2) interactions possible in a society. We present the society as a latent variable modeling these O(n2) in just O(n) interactions here the society is a constant participant of all these two-person interactions. Though it has become common to assume rationality of people in the study of social sciences[3], [4], we align with the view that human decision making is constrained when it comes to rationality. This is not only due to the fact than an average human is not always able to figure out the most rational action due to limits in knowledge and cognitive ability but also because, as we have emphasized before, his objective and the society's objective may not be aligned. In other words, the agents follow the idea of bounded rationality[5] and our model tries to ensure this when it looks upon decision making as an optimization step. Instead of optimizing the happiness function, we rather create an augmented augmented fitness function, which inherits this idea of bounded rationality. In Section 2, we introduce the problem setting, where define the terms and notations. In section 3, we look into the protocol for the simulation itself. In section 4, we look at our experimental simulations and the results and then in section 5, we finally conclude with our observations and our ideas regarding future work. PROBLEM SETTING In our model, we have two agents population and society. • A - {Population,Society} • Sp - Set of pure strategies of population which are essentially the characteristics of an individual that we have considered in our model. a 0.9 0.7 -0.1 -0.9 0.7 -0.5 0.6 0 -0.5 0 0 -0.4 0.2 b -0.5 0.2 0.8 0.9 0.7 0 0.2 0.3 0.5 0.8 -0.8 0 0.2 c 0.5 0 -0.5 0 0 0.8 1 0 0.5 -0.2 0.2 -0.5 0.5 d 0.3 0 -0.5 0 0.4 -0.9 0 0.2 -0.8 0 0.5 1 -1 e 0.3 0.4 0 0.5 0.5 0 0 0 0 0.2 0 0 0 f 0.7 0.7 0 0.6 0.6 0 0.5 0.5 0.5 0 0 0 0 g 0.5 0 -1 0 0 0.4 0.8 0.3 0.4 0.3 0.4 -0.3 -0.6 h -0.2 0 0 0 0 0.8 0.5 0 1 0 0 0.6 -0.5 1 2 3 4 5 6 7 8 9 10 11 12 13 • Ss - Set of pure strategies of society which represent characteristics of a society. Let us denote each of the strategies as follows: • Sp = {a, b, c, d, e, f, g, h} • Ss = {1, 2, 3, 4, 5, 6, 7, , 8, 9, 10, 11, 12, 13} : agents{a, b, c, d, e, f, g, h} We look at the following individual traits while modelling Intellectual/Education the level, Physical Strength, Obedience, Flexibility towards change,Health/Immunity, Sincerity towards work, Family- oriented, Religious with The traits{1, 2, 3, 4, 5, 6, 7, , 8, 9, 10, 11, 12, 13}: rate/Education rate, Agrarian, Industrial, Conservati, Communist following Literacy standards/income, Crime is modellied society the levelL, iving Problem modeling σ • σs - mixed strategy of society • π(σ, σs) - payoff for individual adopting mixed strategy • ISp×Ss - Interaction between society and population p Iσs π(σ, σs) denotes the happiness of an • π(σ, σs) = σT • Payoff Matrix Payoff matrix for both the players is the individual given a mixed strategy chosen by society interaction matrix I. Interaction matrix This is an interaction matrix that we have used in our simulation where each cell represents the correlation between the column and the row property. It also represents the payoff matrix for both the population and the society as it weighs the positive and negative relationships between a strategy of and individual and society. The above payoff matrix has 36 Nash equilibria in total. We have used an online calculator for Nash equilibria. This given the payoff matrix for both the agents calculates the Nash equilibria. Out of these 4 are pure strategy equilibria while the rest are mixed-strategy equilibria. PROTOCOL Simulation setting 2) Happiness =(cid:80)N 1) We simulate each person of the population as having certain characteristics(strategies) and we impart a starting population to the city. i IC where xi is the characteristics i=0 X T of the ith person, I is the interaction matrix and C is the city characteristics. Assumptions • The happiness(fitness) of a person is determined at his birth and remains constant during his lifetime. • The lifespan of a person is determined at his birth and is a function of his happiness. • The mating frequency of a person as well as the success of a mating is a function of his happiness. Algorithm (cid:80)k Algorithm 1 Pseudocode for simulation of city 1: P ← [X1 ··· Xn] ∼ N (µ, Σ) 2: [(Y1 ··· Yk), (Z1 ··· Zk)] ← Available({Xi}i=1···n) 3: while T rue do 4: p∈Π i=0 f (θt, born(Y1, Zpi)) {(Y1, Zpi)}i=1···k argmax { Xi}i=1···k ← born({(Y1, Zpi )})i=1···k 5: P ← Update pop(X1 ··· Xn, X1 ··· Xk) 6: θt+1 ← θt + ∆f (θt, P ) 7: t ← t + 1 8: 9: end while ← Timeline events Available: Every person gets available for mating after a fixed time period called the Mating Gap. After a few years, the males and females available among {Xi}i=1···n form [(Y1 ··· Yk), (Z1 ··· Zk)] Born: • Uniform Selection A child gets its characteristics from either its parents with uniform probability. • Mutation A child mutates its genes to a uniform number with probability p = 0.1 Figures[ 1, 2,3] show below our formulation of the three term along with how they vary with happiness of the individual. happier individuals should be able to survive for a longer time and also mate more frequently. These individuals would also then look for individuals to mate who can increase their happiness quotient. To solve this problem, we perform a bipartite matching with an added gaussian noise that accounts for the mismatches and makes the scenario more realistic. For city, the characteristics should update so that they support the population and hence increase their happiness. Therefore, we do a gradient ascent for updating the city characteristics. The detailed updates are mentioned below. Population characteristics updation Given Y and Z, which represent the available males and females and θt, which represent the city characteristics, we need to find a permutation p such that it solves the following optimization problem. k(cid:88) i=0 f (θt, born(Yi,Zpi)) argmax p∈Π where f (θt, X) = X T (Iθt) This is a maximum weight bipartite matching problem also known as the assignment problem. (cid:80)N Society characteristics updation Given a certain population characteristics x = 1 N i=0 Xi and a current city characteristics, we need to update the city characteristics such that the following two properties are satisfied: • The change is local i.e. a society doesn't undergo drastic changes in its characteristics overnight. • The society must change in order to make its population happier. Fig. 1. Lifespan vs happiness where L = a∗ (1− be−h)) and L is lifespan, a = 150, b = 10 and h is defined as happiness Fig. 2. Mating Gap vs Happiness, g = a/(clip(h, 0, ∞) + ) where g is the mating gap, a = 0.8 and h is defined as happiness. Note that a simple way to perform this update is to perform gradient ascent. θt+1 ← θt + λxTI EXPERIMENTS AND RESULTS Population Survival In these experiments we have tried to look at how population of a city and it's net happiness evolves, depending upon the characteristic initialisation of the city and the characteristic initialisation of it's people. We have also obtained the plots for Population for the city over various years. We have defined the characteristic vectors to represent the various kinds of citirs and populations but we do not report them here for lack of space and will be included in an extended version of the paper. The plots below(Figure[4]) indicate how the various popu- lations survived in a criminal city. Fig. 3. Mating success : This shows the min happiness of the male and the female required to succeed with respect to the current population of the city. m = a · p + max(1 − σ(scale ∗ ph), 1 − σ(scale ∗ qh)) where m is the min happiness, scale= 20 and a = 0.002 OPTIMIZATION STEPS As the city's population evolves, both the population and city characteristics need to be updated accordingly. Intuitively, (a) a (c) c (b) b (d) d Fig. 4. Population and happines plots for a high intellect and a low intellect poulation in a criminal city. Figure(a) shows the population plot for an initial high intellect population in a criminal city. Figuer(b) shows the happines of the same population. Figure(c) and (d) shows the same thing for an initial criminal population in the same city This is the same set of experiments for a intellectual city(Figure[6]). Fig. 5. High Intellect Population Fig. 6. Low Intellect Population A very interesting thing to note in this plots is that the population seems to go on the verge of extinction and then recovers dramatically from it. We visualize this as a test for the population where only the fit people are able to survive. One can also notice that this is the point of time, from when the happiness starts to rise again. This refers to the initial difficult times a society has to suffer before it can actually flourish. Non Optimal mating There are a series of more realistic experiments below which were done after feedback from our presentation. The most important change is to make the mating process more realisitic by removing the optimal matching algorithm we had used previously. Here we use the same matching algorithm i.e. maximal weight bipartite matching algorithm on the bipartite graph with the weight of edges modified by adding a gaussian noise with zero mean and unit variance. In order to simulate locality based maring, we take a more pessimistic view by randomly partitioning the people and then finding the optimal (a) a (b) b (c) b (d) d Fig. 7. These experiments deal with a setting where there are 80% Farmers and 20% Intellectuals in Agrarian city (a): Initial Population Character- istics,(b): Final Population Cluster,(c): Final Population Characteristics,(d): Final Population Cluster matching(with noise) in that partition. The new setting is as follows. Given Y(cid:105) and Z(cid:105), which represent the available males and females of the ith partition and θt, which represent the city characteristics, we need to find a permutation p such that it solves the following optimization problem. k(cid:88) j=0 argmax p∈Π f (θt, born(Y(cid:105)j,Z(cid:105)pj )) where and f (θt, X) = X T (Iθt) + z z ∼ N (0, 1) Surprisingly, the results we obtained were better than the previous case in the following respects. • The drop in population, which is observed initially is not that drastic as observed initially. • The convergent happiness is higher than the previous case. a) Analysis: With this we conducted a new set of exper- iments, where we have a mixed initial population and after running the simulation for 10000 unit times, we again get the final population. We then apply MDS or Multidimensional Scaling to compress the data to two dimensions and then apply K-Means clustering to cluster the people to two clusters. After this, we analyze the characteristics of the initial cluster and the final clusters. We carry out three different experiments in three setting is Figure [7, 8, 9]. Below, we give our analysis in two of these settings. In Figure[7], it is noticeable how the population is no more highly distinguishable after the simulation runs. This is evident from the K-Means clustering shown. One can also notice how the agrarian city forces the people to have uniformly high physically strength and not so high intellect. Health, which was another point of difference between the two populations have also grown to a somewhat more uniform distribution after the simulation. In Figure[8], again the most noticeable feature is how the population grows to be unimodal in its characteristics after the simulation is run. Being an intellectual city, the most important characteristics i.e. intelligence grows after the simulation to a remarkable extent while the other characteristics remains more or less uniform. The other important thing is dedication which remains at a somewhat higher level than the other characteristics. In Figure[9], the experiment suggests that the final pop- ulation need not follow one of its founding sub categories but may indeed develop new characteristics to survive. Here, health and obedience were initially present in different clusters at a high level. However, in the final phase, it is present equally in both the clusters. On the other hand , religiousness, which is a quality favoured by both the dominant seed population as well as the city has seemed to grow the other way i.e. positive. Family oriented and obedience though present in high quantity initially and favoured by the city has grown to stabilize at a more moderate level finally. (a) a (b) b (c) c (d) d Fig. 9. Here there are 75% Criminal 25% High Intellect in a criminal city. Figure(a) shows the Initial Population Characteristics, Figure (b) shows the Initial Population Clusters, Figure (c) shows the Final Population Character- istics and Figure (d) shows the Final Population Cluster Locality based mating In an attempt to make our simulation more realisitic, we also include the factor of locality into the mating selection procedure. This means that while mating, one not only looks at the expected happiness of their child but also at the possibility of the matching itself depending upon the how far the two people are situated. This means that added to the already existing personal traits, we now have (x, y) coordinate of each person/ city block to which the person belongs. After the birth of the child, the new child belongs to either one of the city blocks(father's or mothers) randomly. Hence, the objective function of the mating process is as follows k(cid:88) f (θt, born(Yi,Zpi )) − γd(Yi,Zpi) argmax p∈Π where i=0 f (θt, X) = X T (Iθt) (a) a (b) b (c) c (d) d Fig. 8. Here there are 75% Farmers 25% High Intellect in a high intellect city. Figure(a) shows the Initial Population Characteristics, Figure (b) shows the Initial Population Clusters, Figure (c) shows the Final Population Char- acteristics and Figure (d) shows the Final Population Cluster and γ is a scaling factor and d is the distance function. For simplicity, we use the hamming distance. In Figure[10], we plot the distribution of the population and average happiness of the population initially and finally(i.e. after the simulation). We find that multiple communities evolve in the space and some of these communities are disconnected from each other completely. However, all of these communities maintain a viable population. Another possible extension to this is reducing the dependence on the total population(in mating success) to the total population in a particular grid as opposed to the total population of the entire society. Learning Rate Variations of Learning Rate: Learning rate controls the rate of change of societal characteristics which happens through the (a) a (b) b (c) c Fig. 10. Locality based mating. Figure(a) represents the initial distribution and Figure(b) represents the final distribution. In each plot, the figure on the left hand size represents the average happiness and the figure on the right hand side represents the population of that grid block. Here the grid contains 100 blocks(10x10) gradient ascent Step. We denote Learning rate by λ . We can adjust λ and see the changes it causes.We initialise λ at 10−4 and then vary it in multiples of 1,3,10 and 30 of the original value. Here we report the evolution plots(Figure:[11, 12]) of population which are obtained by varying λ respectively across multiples of 1,3,10 and 30. We report the population and happiness plots seperately. (a) 1X (b) 3X (c) 10X (d) 30X Fig. 11. Population (a) 1X (b) 3X (c) 10X (d) 30X Fig. 12. Happiness In the above plots(Figure:[11, 12) we see that varying lambda affects the time which takes in evolution of happiness levels in society. Higher lambda changes societal characteris- tics fast and achieves the stable happiness values earlier. The happiness value eventually plateau. 1) Feature Sets for these experiments: For these sets of experiments we experimented with a varied set of features than mentioned above, for both individuals as well as the Society. For individuals we took 3 features from the Big Five Personality traits model. Namely Openness to Experience, Conscientiousness and Extraversion. We combined them with IQ and Physical Strength to describe an agent in our model. For societal characteristics we have 5 features, namely, Intel- lectuality in the society, Living Standard, Crime Rate, Indus- trialisation and Cultural Richness of society. We heuristically filled an interaction matrix using corelation values between the Individual and Societal Characteristics to complete our model. A. Dynamic Learning Rate We also propose Dynamic Learning Rate to more real- istically model the evolution procedure. Rate of evolution depends on the flexibility of people towards change. Thus there are various Personality Traits, for example Openness to Experience, which encodes how flexible a person is towards new values and experiences. Thus we propose to use average of this trait(or combination of such traits) over the whole population to dynamically determine the value of λ at any point of time. This understanding of learning rate is meant to model the dependence of rate of evolution which depends upon the liberality of people. CONTRIBUTION AND FUTURE WORK We have built a framework to carry out simulations on this protocol in a very efficient way using the simpy library, which offers the capacity of simulation of multiple events in different timescales in an event based callback type of framework. We also offer real time logging capacity of all the characteristics of the population as well as the population strength and the happiness. Our main contribution is that we propose an algorithm to explain the development of social characteristics of a population, which uses two different kind of optimization algorithms, which are well suited to the particular cause as needed here. It must also be noted that both of these steps are cheap and hence easy to simulate. By viewing it as a two player evolutionary game, further work could include drawing similarity to the fictitious play protocol. We also offer an alternate explanation to the existence of this second player called society in our game. The happiness of a society can be described as the average happiness of all(cid:0)n (cid:1) However, this does not necessarily mean that 2 pairs of people. This requires O(n2) computation. However, by assuming the existence of a latent variable called Society to describe the net effect of the other people in the interaction , this can be done in O(n) time. Alternate maximization technique can here lead to optimizing the net happiness as well as bring the city more close to this proposed latent variable. the city the characteristic serves no other purpose than to model O(n2) interaction in an O(n) representation. It also seves to model the environmental and social constraints faced by the population e.g. existing farm lands in huge quantity, an economy dependant on a certain trade, abscence of fertile lands, rule of a certain kind of government/prevalence of certain norms or laws. This does not necessarily mean that they cannot change. It only means that the rate of change is relatively less and it might actually force the model to converge in a certain local minima, which might respect these constraints. REFERENCES [1] C. Darwin and G. Beer, The origin of species. Dent, 1951. [2] J. H. Miller and S. E. Page, Complex adaptive systems: An introduction to computational models of social life. Princeton university press, 2009. [3] J. Scott, "Rational choice theory," Understanding contemporary society: Theories of the present, vol. 129, 2000. [4] G. W. Downs and D. M. Rocke, "Conflict, agency, and gambling for resurrection: The principal-agent problem goes to war," American Journal of Political Science, pp. 362–380, 1994. [5] H. A. Simon, "Models of man; social and rational." 1957.
1902.01131	4	1902	2019-02-13T14:24:31	On the Enactability of Agent Interaction Protocols: Toward a Unified Approach	[ "cs.MA" ]	Interactions between agents are usually designed from a global viewpoint. However, the implementation of a multi-agent interaction is distributed. This difference can introduce issues. For instance, it is possible to specify protocols from a global viewpoint that cannot be implemented as a collection of individual agents. This leads naturally to the question of whether a given (global) protocol is enactable. We consider this question in a powerful setting (trace expression), considering a range of message ordering interpretations (what does it mean to say that an interaction step occurs before another), and a range of possible constraints on the semantics of message delivery, corresponding to different properties of underlying communication middleware.	cs.MA	cs	ON THE ENACTABILITY OF AGENT INTERACTION PROTOCOLS: TOWARD A UNIFIED APPROACH A PREPRINT Angelo Ferrando∗ Liverpool University United Kingdom Michael Winikoff University of Otago New Zealand [email protected] [email protected] Stephen Cranefield University of Otago New Zealand Frank Dignum Utrecht University Netherlands [email protected] [email protected] Viviana Mascardi University of Genova Italy [email protected] February 14, 2019 ABSTRACT Interactions between agents are usually designed from a global viewpoint. However, the implemen- tation of a multi-agent interaction is distributed. This difference can introduce issues. For instance, it is possible to specify protocols from a global viewpoint that cannot be implemented as a collec- tion of individual agents. This leads naturally to the question of whether a given (global) protocol is enactable. We consider this question in a powerful setting (trace expression), considering a range of message ordering interpretations (what does it mean to say that an interaction step occurs before another), and a range of possible constraints on the semantics of message delivery, corresponding to different properties of underlying communication middleware. Keywords Agent Interaction Protocols · Enactability · Enforceability · Implementability · Realizability · Projectabil- ity · Trace Expressions 1 Introduction In order to organise her staying in Montreal, Alice books an apartment from Bob via the online platform AIPbnb. AIPbnb policy states that owners cannot interact with each other, users can interact with owners only via the platform, and if a user finds a better solution for her accommodation, she must cancel the previous one before she makes a new reservation for the same dates, otherwise she will be charged for one night there. When Alice discovers that Carol rents a cheaper and larger apartment, she decides to cancel the reservation of Bob's apartment and book Carol's one. This sit- Res=⇒ Carol uation can be represented by the global Agent Interaction Protocol modifyRes = Alice where a1 M=⇒ a2 models the interaction between a1 and a2 for exchanging message M , "·" models interaction con- catenation, and Canc and Res are sent to the recipients by using the AIPbnb platform as required. Alice believes that the above protocol correctly meets AIPbnb policy, but she is charged for one night in Bob's apartment by AIPbnb: Canc=⇒ Bob · Alice ∗Work supported by EPSRC as part of the ORCA [EP/R026173] and RAIN [EP/R026084] Robotics and AI Hubs. A PREPRINT - FEBRUARY 14, 2019 Carol received Alice's request before Bob received the cancellation, and this violates the policy. What went wrong is the interpretation of "before". To Alice, it meant that she should send Canc before she sent Res, while for AIPbnb it (also) meant that Bob should receive Canc before Carol received Res. This ambiguity would have had no impact on Alice if the physical communication model underlying AIPbnb guaranteed that between the sending and receiv- ing stages of an interaction, nothing could happen. However, if the communication model provides weaker or no guarantees, it may happen that a message sent before another, is delivered after. This simple example shows that enacting the respect of a global protocol without a clear semantics of the "before" meaning, without guarantees from the platform implementation on message delivery order, and without hidden com- munications between the participants ("covert channels"), may not be possible. Many real situations can be resorted to this one: for example, a citizen must wait for the bank to have received (and processed) the request for adding some money to a new, empty account, before sending a request to move that amount to another account, otherwise he can go in debt. Global protocols are modelled using many different formalisms including global types [1], Petri Nets [2], WS-CDL [3], AUML [4], Statecharts [5], and causal logic [6]. In each of these formalisms the enactability problem, that we define as "by executing the localised versions of the protocol implemented by each participant, the global protocol behaviour is obtained, with no additional communication", has been addressed in some form. Despite their diversity, however, most of these formalisms do not support protocol concatenation and recursion, which are needed to achieve a high expressivity: their expressive power is limited to regular languages. Moreover, although -- from an operational point of view -- these approaches agree on the intuition that a global protocol is enactable if the composition of the local protocols, obtained by projecting the global one onto each participant, behaves exactly in the same way as the global protocol, the semantic definition of enactability is far from being standard and sometimes is also more restrictive than necessary: some protocols will be classified as not enactable, while (under suitable conditions) they could be enacted. The intended message ordering and the communication model of the infrastructure in which the agents will be imple- mented and run are never taken into consideration together. As shown in the example above these two elements are effectively two sides of the same coin which must be both modeled for providing a precise and generally applicable definition of enactability. In a similar way, the need to associate the protocol with a decision structure to enforce consistent choices, is recognised as a necessity and suitably addressed by [7] only, and not in conjunction with the other issues that affect enactability. Finally, the availability of a working prototype to check the enactability of global protocols under message ordering and communication models is usually disregarded in the literature. In this paper we provide a semantic characterisation of enactability which integrates message ordering and communica- tion model in a unified framework, along with decision structures. This combination prevents unnecessary restrictions from the definition, which is as general as possible and suitable for highly expressive protocol representation languages like Trace Expressions [8]. We also developed a working prototype in Haskell for enactability checks, which is one key benefit of out approach. 2 Background Trace Expressions. Trace expressions [8] are a compact and expressive formalism inspired by global types [9] and then extended and exploited in different application domains [10, 11, 12, 13, 14]. Trace Expressions, initially devised for runtime verification of multiagent systems, are able to define languages that are more than context free. A trace expression τ denotes a set of possibly infinite event traces, and is defined on top of the following operators:2 • ǫ (empty trace), denoting the singleton set {hi} containing the empty event trace hi. • M (event), denoting a singleton set {hM i} containing the event trace hM i. • τ1·τ2 (concatenation), denoting the set of all traces obtained by concatenating the traces of τ1 with those of τ2. • τ1∧τ2 (intersection), denoting the intersection of the traces of τ1 and τ2. • τ1∨τ2 (union), denoting the union of the traces of τ1 and τ2. 2Binary operators associate from left, and are listed in decreasing order of precedence, that is, the first operator has the highest precedence. 2 A PREPRINT - FEBRUARY 14, 2019 • τ1τ2 (shuffle), denoting the union of the sets obtained by shuffling each trace of τ1 with each trace of τ2 (see [15] for a more precise definition). Trace expressions are cyclic terms, thus they can support recursion without introducing an explicit construct. As customary, the operational semantics of trace expressions, defined in [16], is specified by a transition relation δ ⊆ T × E × T , where T and E denote the set of trace expressions and of events, respectively. We do not present all the transition rules for space constraints. They are standard ones which state, for example, that δ(ev · τ, ev, τ ) (the protocol whose state is modelled by ev · τ can move to state τ if ev occurs), and that δ(τ1 ∨ τ2, ev, τ ) if δ(τ1, ev, τ ) (if the protocol whose state is modelled by τ1 can move to state τ if ev occurs, then also the protocol whose state is modelled by τ1 ∨ τ2 can). The denotational semantics is defined as follows: JǫK = {hi} JM K = {hM i} Jτ1 · τ2K = {t1 ◦ t2t1 ∈ Jτ1K ∧ t2 ∈ Jτ2K} Jτ1 ∧ τ2K = Jτ1K ∩ Jτ2K Jτ1 ∨ τ2K = Jτ1K ∪ Jτ2K Jτ1τ2K = {z t1 ∈ Jτ1K ∧ t2 ∈ Jτ2K ∧ z ∈ t1 ⊲⊳ t2} Where t1 ⊲⊳ t2 is the set of all interleavings of t1 and t2, and ◦ is concatenation over sequences. Events can be in principle of any kind. In this paper, we will limit ourselves to consider interaction and message events. M=⇒ b and gives information on the protocol from the global perspective, collapsing An interaction has the form a sending and receiving. We say that τ is an interaction protocol if all the events therein are interactions. Interaction protocols take other names in other communities, such as Interaction Oriented Choreography [17] in the Service Oriented Community, and global type in the community working on process calculi and types [1]. Message events have the form aM ! (a sends M ) and bM ? (b receives M ). They model actions that one agent can execute, hence taking a local perspective. A trace expression where all events are messages will be named a message protocol throughout the paper. Message protocols have different names in different communities, such as Process Oriented Choreography [17] and "local type" or "session type" in the global type community [18, 19]. Communication Models. Given that in our proposal we explicitly take the communication model supported by the MAS infrastructure into account, we provide a summary of communication models based on [20]. We use CM0 to CM6 to identify them in a compact way. CM0: Synchronous Communication. Sending and receiving are synchronised: the sender cannot send if the receiver is not ready to receive. CM1: Realisable with Synchronous Communication (RSC). After a communication transition consisting of a send event of a message, the only possible communication transition is the receive event of this message. This asynchronous model is the closest one to synchronous communication and can be implemented with a 1-slot unique buffer shared by all agents. CM2: FIFO n-n communication. Messages are globally ordered and are delivered in their emission order: if sending of M1 takes place before sending of M2, then reception of M1 must take place before reception of M2. This model can be implemented by means of a shared centralised object, such as unique queue. CM3: FIFO 1-n communication. Messages from the same sender are delivered in the order in which they were sent. It can be implemented by giving each agent a unique queue where it puts its outgoing messages. Destination peers fetch messages from this queue. CM4: FIFO n-1 communication. A send event is implicitly and globally ordered with regard to all other sending actions toward the same agent. This means that if agent b receives M1 (sent by agent a) and later it receives M2 (sent by agent c), b knows that the sending of M1 occurred before the sending of M2 in the global execution order, even if there is no causal path between the two sending actions. The implementation of this model can, similarly to FIFO 1-n, be done by providing each agent with a queue: messages are sent by putting them into the queue of the recipient agent. 3 A PREPRINT - FEBRUARY 14, 2019 CM5: Causal. Messages are delivered according to the causality of their emissions [21]: if a message M1 is causally sent before a message M2 then an agent cannot get M2 before M1. An implementation of this model requires the sharing of the causality relation. CM6: Fully Asynchronous. No order on message delivery is imposed. Messages can overtake others or be arbitrarily delayed. The implementation is usually modelled by a bag. Message Ordering. The statement "one interaction comes before another" is ambiguous, as exemplified in Section 1. This ambiguity has been recognised by some authors who suggested how to interpret message ordering, when moving from the interaction (global) level to the message (local) level. In this section we summarise and compare the proposals by Lanese, Guidi, Montesi and Zavattaro [17] and that by Desai and Singh [22]. To identify the interpretations, we will use the acronyms used in [22] when available, and our own acronyms otherwise. M2=⇒ d. For the sake The starting point for interpreting message ordering is the interaction protocol τ = a of clarity, we denote aM1! with s1, bM1? with r1, cM2! with s2, and dM2? with r2; we characterise the message ordering interpretations by the traces of messages that respect them. M1=⇒ b·c RS: a message send must be followed immediately by the corresponding receive, so w.r.t. τ , M1 must be received before M2 is sent. The set of traces that respect this model is {s1 r1 s2 r2}. This interpretation is named RS (receive before send) in [22] and disjoint semantics in [17]. SS: M1 is sent before M2 is, and there are no constraints on the delivery order. The set of traces that respect this model is {s1 r1 s2 r2, s1 s2 r1 r2, s1 s2 r2 r1}. This interpretation is named SS (send before send) in [22] and sender semantics in [17]. RR: M1 is received before M2 is, and there are no constraints on the sending order. The set of traces that respect this model is {s1 r1 s2 r2, s1 s2 r1 r2, s2 s1 r1 r2}. This interpretation is named RR (receive before receive) in [22] and receiver semantics in [17]. RR & SS: this combines the requirements of RR and of SS: M1 is sent before M2 is sent and also M1 is received before M2 is received. The set of traces that respect this model is {s1 r1 s2 r2, s1 s2 r1 r2}: both s1 comes before s2 ("coming before" according to the senders), and r1 comes before r2 ("coming before" according to the receivers). This interpretation is named sender-receiver semantics in [17]. SR: M1 is sent before M2 is received. The set of traces that respect this model is {s1 r1 s2 r2, s1 s2 r1 r2, s1 s2 r2 r1, s2 s1 r1 r2, s2 s1 r2 r1}. This interpretation is named SR (send before receive) in [22]. It is easy to see that the following inclusions among asynchronous models hold: RS ⊂ RR & SS ⊂ SS ⊂ SR and RS ⊂ RR & SS ⊂ RR ⊂ SR. The SS and RR interpretations are not comparable. In the remainder of this paper we consider only the four interpretations defined by Desai & Singh, i.e. we do not consider "RR & SS". 3 Defining Enactability using a Semantic Approach In the following let ComModel = {CM 1, CM 2, CM 3, CM 4, CM 5, CM 6} be the set of pos- Basic Notation. sible (asynchronous) communication models, and MOISet = {SS, SR, RS, RR } the set of possible message order interpretations that can be imposed. We also define A = {a, b, c, d, a1, a2, . . . , an} to be the set of agents involved in the interaction protocol. Recall that we consider both interaction and message protocols. When we say that τ is an interaction protocol, we mean that the protocol represents sequences of interactions. The set of traces recognized is obtained following the semantics defined in Section 2, and for an interaction protocol τ we have that3 I ∈ Jτ K =⇒ ∀i∈I .i ∈ I(τ ), where we define I(τ ) to be the set of interactions involved in the interaction protocol τ . We also define I to be the set of all possible interactions events. Similarly, when τ is a message protocol (rather than an interaction protocol), it represents sequences of send and receive events of the form aM ! (send event) and bM ? (receive event), and given a particular set of possible interactions I, we define EI to be the corresponding set of events: EI = {aM !∃b∈A.a M=⇒ b ∈ I} ∪ {bM ?∃a∈A.a M=⇒ b ∈ I} In a message protocol τ we have that E ∈ Jτ K =⇒ ∀e∈E .e ∈ EI(τ ). Given a message protocol τ we also define E(τ ) to be the set of events that occur in the protocol. 3We use "∈" to also denote membership of an item in a sequence. 4 A PREPRINT - FEBRUARY 14, 2019 Next, we define the language of traces for interaction protocols and message protocols. For interaction protocols, the set of all possible traces is defined to be: LI = I ∗ ∪ I ω. For message protocols the definition is somewhat more complex, since there is a relationship between a send and a receive event. Specifically, the set of all possible traces of events is constrained so that a message being received must be preceded by that message having been sent. We also constrain the set so that each message can be sent at most once, and received at most once (i.e. message names are unique). The assumption is made by most authors, see [20] for example, and it is considered as a harmless one; we can integrate many elements to the notion of "message name", such as content, protocol id, conversation id, etc, to discriminate between messages at design time. Formally: LEI = {E ∈ E ∗ I ∪ E ω I (∀i,j∈dom(E).E[i] = aM ! ∧ E[j] = aM ! =⇒ i = j) ∧ (∀i,j∈dom(E).E[i] = bM ? ∧ E[j] = bM ? =⇒ i = j) ∧ (∀i∈dom(E).E[i] = bM ? =⇒ (∃j∈dom(E).E[j] = aM ! ∧ j < i)) Message Order Interpretation (MOI). An interaction protocol τ defines orderings between messages Mi, whereas a message protocol deals in events (sending and receiving). If a protocol says that M1 comes before M2, how should we interpret this in terms of events? Should sending M1 come before sending M2, or does it mean that receiving M1 should occur before receiving M2? The message ordering interpretation (MOI) specifies this. As discussed earlier, we follow prior work in considering four (natural) interpretations (SS, SR, RS, and RR). We formalise this by defining a variant semantics that takes an interaction protocol τ and returns its semantics in terms of events rather than interactions. The possible sequences of events are constrained: given a situation where τ specifies that M1 must occur before M2, we constrain the possible sequence of events with the appropriate constraint on events corresponding to the selected MOI. Def. 3.1 (Order on interactions in a trace) Let I ∈ LI be a trace of interaction events, E ∈ LEI be a trace of send M2=⇒ d ∈ I two interactions. and receive events, moi ∈ MOISet a message ordering interpretation, and a M1=⇒ b ∈ I, c M2=⇒ d as I2, we define an order on M1 and M2 for moi in E as follows: M1=⇒ b as I1 and c Abbreviating a I1 ≺E I1 ≺E I1 ≺E I1 ≺E where e1 ≺E e2 , ∃i,j∈dom(E).E[i] = e1 ∧ E[j] = e2 ∧ i ≤ j SS I2 , aM1! ≺E bM2! SR I2 , aM1! ≺E dM2? RS I2 , bM1? ≺E bM2! RR I2 , bM1? ≺E dM2? traces. Formalising the MOI is not as simple as it might seem. An obvious approach that does not work is to compute the semantics of the interaction protocol τ , and then map each sequence I ∈ Jτ K to a This does not work because the trace is linear, and therefore a total or- set of message event der, whereas a protocol can specify a partial order. An illustrative example is τ = (M1 · M2) M3. This simple protocol has three sequences of interactions: {hM1, M2, M3i, hM1, M3, M2i, hM3, M1, M2i}. As- sume an RS message ordering interpretation, then each of the message sequences corresponds to exactly one sequence of events, giving4 {hs(M1), r(M1), s(M2), r(M2), s(M3), r(M3)i, hs(M1), r(M1), s(M3), r(M3), s(M2), r(M2)i, hs(M3), r(M3), s(M1), r(M1), s(M2), r(M2)i}. However, the protocol does not specify any con- straint on M3, so should also allow other interpretations where the occurrences of s(M3) and r(M3) are not con- strained relative to the other events, for example hs(M1), r(M1), s(M3), s(M2), r(M2), r(M3)i. Instead, we define a variant semantics, which is compositional. The semantics follow the standard semantics (Sec- tion 2) with a few exceptions. Firstly, the semantics of an interaction I is given as the sequence of sending the message, followed by receiving it (denoted respectively s(I) and r(I)). Secondly, the semantics for a sequence τ1 · τ2 is given by taking the semantics of τ1 and of τ2. These are then combined by interleaving them (rather than simply concatenating them), but with the constraint that the result must satisfy the appropriate MOI constraint (I1 ≺E SS I2) for all possible final messages of τ1 (I1) and all possible initial messages of τ2 (I2). Determining initial and final messages is itself somewhat complex, and is done using partially ordered sets. A partially ordered set (poset) is a pair (E, <) where E is the set of elements (in this case send and receive events) and < is a binary relation on E. We define the union operator to act piecewise on posets, and to take the transitive closure of the resulting relation, i.e. (E1, <1) ∪ (E2, <2) = (E1 ∪ E2, (<1 ∪ <2)∗). We can then define the poset of 4For readability we use s(M ) and r(M ) to abbreviate sending and receiving message M , eliding the identity of the agents involved. 5 A PREPRINT - FEBRUARY 14, 2019 an interaction protocol as follows poset(ǫ) = (∅, ∅) poset(I) = ({I}, ∅) poset(τ1 ∧ τ2) = poset(τ1) ∪ poset(τ2) poset(τ1 τ2) = poset(τ1) ∪ poset(τ2) poset(τ1 ∨ τ2) = poset(τ1) ∪ poset(τ2) poset(τ1 · τ2) = poset(τ1) · poset(τ2) (E1, <1) · (E2, <2) = (E1 ∪ E2, <1 ∪ <2 ∪{(x, y) x ∈ max(E1, <1) ∧ y ∈ min(E2, <2)}) Where we define a sequence of two posets (E1, <1) · (E2, <2) by collecting the orderings of each of E1 and E2, and adding additional ordering constraints between the maximal elements of E1 and the minimal elements of E2. We can now proceed to define Jτ Kmoi. JǫKmoi = {ǫ} JIKmoi = {hs(I), r(I)i} Jτ1 ∧ τ2Kmoi = Jτ1Kmoi ∩ Jτ1Kmoi Jτ1 · τ2Kmoi = {t t1 ∈ Jτ1Kmoi ∧ t2 ∈ Jτ2Kmoi ∧ t ∈ t1 ⊲⊳ t2 ∧ ∀I1 ∈ max(poset(τ1)), ∀I2 ∈ min(poset(τ2)) : I1 ≺t moi I2} Jτ1 ∨ τ2Kmoi = Jτ1Kmoi ∪ Jτ1Kmoi Jτ1τ2Kmoi = {z t1 ∈ Jτ1Kmoi ∧ t2 ∈ Jτ2Kmoi ∧ z ∈ t1 ⊲⊳ t2} Where t1 ⊲⊳ t2 is the set of all interleavings of t1 and t2. Communication Model Semantics. We formalise the defined communication model semantics by defining for each communication model CM i a corresponding language of event traces that incorporates the appropriate restriction, ruling out event sequences that violate the communication model. The definitions below are those already provided in Section 2. For example, for CM 1 the constraint is that immediately after each sending event in u we have its corresponding receiving event, with nothing in the middle; etc. LEI CM1 = {E ∈ LEI ∀ a M1=⇒b∈I .∀k∈dom(E).aM1! = E[k − 1] =⇒ bM1? = E[k]} LEI CM2 = {E ∈ LEI ∀ a M1=⇒b∈I .∀ c M2=⇒d∈I .∀i,j,k,l∈dom(E). bM1? = E[i] ∧ dM2? = E[j] ∧ aM1! = E[k] ∧ cM2! = E[l] ∧ k < l =⇒ i < j} LEI CM3 = {E ∈ LEI ∀ a M1=⇒b∈I .∀ a M2=⇒d∈I .∀i,j,k,l∈dom(E). bM1? = E[i] ∧ dM2? = E[j] ∧ aM1! = E[k] ∧ aM2! = E[l] ∧ k < l =⇒ i < j} LEI CM4 = {E ∈ LEI ∀ a M1=⇒b∈I .∀ c M2=⇒b∈I .∀i,j,k,l∈dom(E). bM1? = E[i] ∧ bM2? = E[j] ∧ aM1! = E[k] ∧ cM2! = E[l] ∧ k < l =⇒ i < j} LEI CM5 = {E ∈ LEI ∀ a M1=⇒b∈I .∀ a M2=⇒b∈I .∀i,j,k,l∈dom(E). bM1? = E[i] ∧ bM2? = E[j] ∧ aM1! ≺E =⇒ i < j} where aM1! ≺u Causal bM2! ⇐⇒ ((a = b ∨ M1 = M2) ∧ Causal aM2! 6 A PREPRINT - FEBRUARY 14, 2019 ∃i,j∈dom(u).(u[i] = aM1! ∧ bM2! = u[j] ∧ i < j)) ∨ (∃ev∈E .aM1! ≺u Causal ev ∧ ev ≺u Causal bM2!) LEI CM6 = LEI We can then apply a particular communication model to an interaction protocol τi using JτiKCM protocol τm using JτmKCM, which are defined as follows: moi, and to a message moi = JτiKmoi ∩ L JτiKCM JτmKCM = JτmK ∩ LE(τ ) CM EI(τ ) CM Projection. Projection is defined, intuitively, as focussing on the aspects of the protocol that are relevant for a given role. It is defined as follows, where we write τ A to denote projecting trace τ for role A. (ǫ)A = ǫ (a M =⇒ b)A = aM !, if a = A = bM ?, if b = A = ǫ, otherwise (aM !)A = if a = A then aM ! else ǫ (aM ?)A = if a = A then aM ? else ǫ (τ1 ⊗ τ2)A = (τ1)A ⊗ (τ2)A Where ⊗ is any operator. We then define the distribution of τ , denoted pτ q, where τ involves roles a1 . . . an as5: pτ q = τ a1 k . . . kτ an To make an example, let us consider again the scenario proposed in Section 1. Alice decided to book Carol's apartment and now Carol needs some pieces of information from Alice in order to complete the reservation. This information can be wrong or incomplete, and Carol might need to ask Alice twice or more times. This can be represented using a cyclic specification reqInfo = Alice Inf o =⇒ Carol · (Carol W rong =⇒ Alice · reqInfo ∨ Carol Booked=⇒ Alice) where if the information provided by Alice is not satisfactory, Carol tells Alice and asks for new one (recursion on reqInfo). Once Carol will be satisfied with Alice' answer, she will confirm the booking. Thanks to cyclic specifications, we can represent protocols with infinite behaviours. Let us consider main as the combination of the two protocols: main = modifyRes · reqInfo. The projection of main on each single agent would generate pmainq = main Alice k main Bob k main Carol main Alice = modifyRes Alice · reqInfo Alice modifyRes Alice = AliceCanc! · AliceRes! reqInfo Alice = AliceInf o! · (AliceW rong? · reqInfo Alice ∨ AliceBooked?) main Bob = modifyRes Bob · reqInfo Bob modifyRes Bob = BobCanc? reqInfo Bob = ǫ 5We use k to distinguish between parallel composition of different agents, and parallel composition within a protocol. This distinction is used later in this section. 7 A PREPRINT - FEBRUARY 14, 2019 main Carol = modifyRes Carol · reqInfo Carol modifyRes Carol = CarolRes? reqInfo Carol = CarolInf o? · (CarolW rong! · reqInfo Carol ∨ CarolBooked!) In order to define the semantics of a projected protocol we need to first define what we term a decision structure. This is needed in the semantics in order to deal correctly with projected protocols. Specifically, the intuition for enactability (see Section 3) is that an interaction protocol τ involving, say, three roles a, b and c is enactable iff there exist three protocols τ a, τ b and τ c such that their concurrent interleaving results in the same behaviour as the original protocol. However, when a protocol contains choices (∨) we need to ensure that the occurrences of ∨ in each of τ a, τ b and τ c M2=⇒ c. This arising from the same ∨ in τ are treated consistently. For example, consider the protocol τ = a protocol is simple: it specifies that agent a can either send a message ("M1") to b, or it can send a different message ("M2") to agent c. When we distribute the protocol by projecting it (see Section 3) and forming τ akτ bkτ c we obtain the distributed protocol (aM1! ∨ aM2!)k(bM1? ∨ ε)k(ε ∨ cM2?). However, if we interpret each ∨ independently (as the semantics would naturally do) then we can have inconsistent choices. For example, we could have (aM1!)k(ε)k(ε) where the message is sent by a, but b does not elect to receive it. So what we need to do is ensure that each of the three occurrences of "∨" represent the same choice, and that the choice should be made consistently. M1=⇒ b ∨ a The heart of the issue is that the trace expression notation offers a choice operator (∨), which is adequate for global protocols. However, for local protocols it is important to be able to distinguish between a choice that represents a free (local) choice, and a choice that is forced by earlier choices. In this example, a can freely choose whether to send M1 or M2. However, the choice of b whether to receive M1 or not is not a free choice, but is forced by a's earlier choice. Our semantics handles this by defining a decision structure which is used to enforce consistent choices. Formally, given a protocol τ we define d(τ ) as a set of decision structures (formal definition below). A decision structure is a syntactic structure that mirrors the structure of τ , except that each ∨ is annotated with a decision (e.g. L or R). We define three operations defined on a decision structure: to get the sub-decision structure corresponding to the left part (denoted d.L), to get the right part (d.R) and to get the decision (L or R) associated with the current ∨ node (denoted d.D). We define d(τ ) to create a set of decision structures, each of which corresponds to the structure of τ , but where all possible assignments of decisions are made. Observe that If τ contains N occurrences of ∨ then the set d(τ ) contains 2N elements. For example, given τ = a M1=⇒ b ∨ a M2=⇒ b we have that d(τ ) = {_ L ∨ _, _ R ∨ _} where we use _ to indicate an irrelevant part of a decision structure, and L ∨ to denote a node tagged with a decision L. LR ∨ t2 . . .). The reason is In addition to decisions of L and R, the definition of d(τ1 ∨ τ2) has a second case (. . . ∪ {t1 that it is only possible to enforce consistent choice if the choice is made by a single agent. If this is not the case, then M2=⇒ b we we annotate with "LR" to indicate that a mixed choice is possible. For example, given τ = b M1=⇒ a ∨ a have that d(τ ) = {_ LR ∨ _} because ag(τ1) = {b} 6= ag(τ2) = {a}. d(ε) = {ε} d(I) = {I} d(τ1 ∨ τ2) = {t1 x ∨ t2 t1 ∈ d(τ1) ∧ t2 ∈ d(τ2) ∧ x ∈ {R, L} ∧ ag(τ1) = ag(τ2) ∧ ag(τ1) = 1} ∪ {t1 LR ∨ t2 t1 ∈ d(τ1) ∧ t2 ∈ d(τ2) ∧ ((ag(τ1) 6= ag(τ2)) ∨ (ag(τ1) 6= 1))} where ag(τ ) = {p p M=⇒ r ∈ min(poset(τ ))} d(τ1 ⊕ τ2) = {t1 ⊕ t2 t1 ∈ d(τ1) ∧ t2 ∈ d(τ2)} (τL ⊗ τR).L = τL (τL ⊗ τR).R = τR (τL X ∨ τR).D = X Where ⊗ is any operator, and ⊕ is any operator other than ∨. We now specify the semantics of a distributed protocol, denoted Jτ Kdist. The semantics is defined in terms of a union over possible decision structures (first line). The remaining of the equations for the semantics carry along the decision structure, and follow it in recursive calls, and for the semantics of ∨ it enacts the decision specified in the structure, 8 A PREPRINT - FEBRUARY 14, 2019 SR ✔ (✔) (✔) (✔) (✔) (✔) SR ✔ (✔) (✔) (✔) (✔) (✔) CM CM1 CM2 CM3 CM4 CM5 CM6 CM CM1 CM2 CM3 CM4 CM5 CM6 a M1=⇒ b · a RR RS ✔ ✔ M2=⇒ c SS ✔ ✘ ✘ ✘ ✘ ✘ ✔ ✔ ✘ ✘ ✘ ✔ ✔ ✔ ✔ ✔ SR ✔ (✔) (✔) (✔) (✔) (✔) a M1=⇒ b · c RR RS ✔ ✔ M6=⇒ b SS ✔ ✘ ✘ ✘ ✘ ✘ ✔ ✔ ✔ ✔ ✔ ✔ ✘ ✔ ✘ ✘ CM CM1 CM2 CM3 CM4 CM5 CM6 a ✔ M1=⇒ b · b RR RS ✔ ✔ (✔) (✔) (✔) (✔) (✔) ✔ ✔ ✔ ✔ M3=⇒ a SS ✔ (✔) (✔) (✔) (✔) (✔) SR ✔ (✔) (✔) (✔) (✔) (✔) a M1=⇒ b ∨ a RR RS ✔ ✔ M2=⇒ c SS ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ CM CM1 CM2 CM3 CM4 CM5 CM6 a M1=⇒ b · c RR RS ✔ ✔ M4=⇒ a SS ✔ ✘ ✘ ✘ ✘ ✘ ✘ ✘ ✘ ✘ ✘ ✘ ✘ ✘ ✘ ✘ a M1=⇒ b ∨ b RR RS ✔ ✔ M3=⇒ a SS ✔ ✘ ✘ ✘ ✘ ✘ ✘ ✘ ✘ ✘ ✘ ✘ ✘ ✘ ✘ ✘ CM CM1 CM2 CM3 CM4 CM5 CM6 CM CM1 CM2 CM3 CM4 CM5 CM6 SR ✔ ✔ ✔ ✔ ✔ ✔ SR ✔ ✘ ✘ ✘ ✘ ✘ SR ✔ (✔) (✔) (✔) (✔) (✔) SR ✔ ✔ ✔ ✔ ✔ ✔ a ✔ M1=⇒ b · b RR RS ✔ ✔ (✔) (✔) (✔) (✔) (✔) ✔ ✔ ✔ ✔ M5=⇒ c SS ✔ (✔) (✔) (✔) (✔) (✔) a M1=⇒ b · a RR RS ✔ ✔ M2=⇒ b SS ✔ ✘ ✘ ✘ ✘ ✘ ✔ ✔ ✔ ✔ (✔) ✔ ✔ ✔ ✔ (✔) CM CM1 CM2 CM3 CM4 CM5 CM6 CM CM1 CM2 CM3 CM4 CM5 CM6 Figure 1: Automatically generated analyses of enactability rather than considering both sub-protocols. Note that projection is defined using k rather than the usual - this differs in the semantics below, in that k passes the same decision structure to both arguments. This ensures consistency between agents, but not within agents. Jτ Kdist = [ Jτ a1 k . . . kτ an Kdt dt∈d(τ ) JM Kdt = {hM i} JεKdt = {hi} Jτ1 · τ2Kdt = {t1 ◦ t2t1 ∈ Jτ1Kdt.L ∧ t2 ∈ Jτ2Kdt.R} Jτ1 ∧ τ2Kdt = Jτ1Kdt.L ∩ Jτ2Kdt.R Jτ1 ∨ τ2Kdt = if dt.D = R then Jτ2Kdt.R elseif dt.D = L then Jτ1Kdt.L else Jτ2Kdt.R ∪ Jτ1Kdt.L Jτ1τ2Kdt = {zt1 ∈ Jτ1Kdt.L ∧ t2 ∈ Jτ2Kdt.R ∧ z ∈ t1 ⊲⊳ t2} Jτ1kτ2Kdt = {zt1 ∈ Jτ1Kdt ∧ t2 ∈ Jτ2Kdt ∧ z ∈ t1 ⊲⊳ t2} Where t1 ⊲⊳ t2 is the set of all interleavings of t1 and t2, and ◦ is concatenation over sequences. Note that if τ does not contain any occurrences of ∨ then the semantics above reduce to the standard semantics. Finally, we define JτiKCM applies a particular communication model CM. dist, which computes the semantics of an interaction protocol τi by distributing it, and also JτiKCM dist = JτiKdist ∩ L EI(τ ) CM Enactability. We are now finally in a position to define enactability. The intuition is that an interaction protocol τ is enactable iff the semantics of τ , with respect to a selected message ordering interpretation and communication model, can be realised by a distributed version of the protocol. In other words, if there exists for each role r a corresponding message protocol τr such that the combination of these protocols realises the same behaviour as τ . However, instead of considering whether there exists some τr, we let τr = τ r, i.e. we take for each role the projected protocol as its protocol. We also consider a notion of weak enactability. This applies in a situation where the a distributed enactment is able to avoid violating the behaviour specified by τ , but is not able to recreate all of the behaviours that τ specifies. This situation can arise with weaker message ordering interpretations (see below for examples). Weak enactability can also M2=⇒ a). In this arise in situations where two ordered messages have two overlapping roles (e.g. τ = a situation the projection operator is too strict: it has τ b = r(M1) · s(M2), but if we adopt an SR message ordering interpretation, then we do not need to ensure that M2 is sent after M1 is received, only that M1 is sent before M2 is received, which role a can ensure on its own. M1=⇒ b · b Def. 3.2 (Strongly/Weakly Enactable) Let τ be an interaction protocol, {a1, a2, ..., an} the set of agents involved in τ , moi ∈ MOISet a message order interpretation and CM ∈ ComModel a communication model. We say that, τ 9 is strongly (weakly) enactable, for moi semantics in CM model iff the decomposition of τ through projection on its agents {a1, a2, ..., an} recognizes the same (a subset of) traces recognized by τ . Formally: A PREPRINT - FEBRUARY 14, 2019 enact (τ )CM moi weak _enact (τ )CM moi iff iff Jτ KCM Jτ KCM dist = Jτ KCM dist ⊆ Jτ KCM moi moi If a protocol is weak enactable, the interleaving of the corresponding local protocols generates a subset of its traces (with a fixed moi and communication model). In practice, this means that our implementation is sound (generates only valid traces), but it is not complete (not all the traces are generated). Consequently, our system will be more restrictive than we wanted. Figure 1 show the results of applying this definition to a number of cases, with different message ordering interpre- tation, and different communication models. These tables were all generated by the Haskell implementation of the definitions in this paper, in which ✔and (✔) denote strongly and weakly enactable, respectively. The prototype counts ~300 LOC. It implements the trace expression standard semantics, message order interpretation, communication model semantics and enactability check6. Looking at the tables in Figure 1, we make the following observations. Firstly, CM1 is quite strict: all the cases considered are enactable under CM1, regardless of the selected message ordering interpretation. This is expected: we know that CM1 is quite strong. Secondly, for many examples there is not a difference in enactability with the different communication models (other than CM1), except where the communication model corresponds to the combination of MOI and the pattern in the protocol. For example, in the top row, second table from the right, the simple protocol is enactable given SS message ordering interpretation only with CM2 and CM4 (and, of course, CM1). This is because for this protocol both messages are received by the same agent but sent by different agents, and, given an RR MOI, the desired constraint that agent B receives the first message before the second, can only be enforced using a communication model that guarantees delivery of messages to the same recipient in the order in which messages were sent. Both CM2 and CM4 provide this guarantee (in fact CM4 provides exactly this, and CM2 is stronger). Thirdly, RS appears to be a good choice for message ordering interpretation, since it is the only MOI where protocols are never weakly enactable. For the other message ordering interpretations, there are protocols that are only weakly enactable (for communication models other than CM1). A protocol being weakly enactable indicates that the desired behaviour specified by the MOI is too loose: it permits behaviours that the distributed realisation cannot realise. On M2=⇒ b), the protocol is not the other hand, in the case of the left-most table on the bottom row (protocol a enactable under RS (except for CM1), but is enactable under SS and under RR. Turning to SR, we observe that it seems to be too weak: almost all the protocols in the figure are enactable (although in most cases only weakly enactable). M1=⇒ b · a Returning to the example from the introduction: modifyRes = Alice Canc=⇒ Bob · Alice Res=⇒ Carol where a1 M=⇒ a2 this example corresponds to the second table from the left in the top row of Figure 1. This shows that, if one desires an RR MOI, i.e. that what is meant by Canc coming before Res is that Bob receives the Canc message before Carol receives the Res message, then the underlying message communication must be CM 1, CM 2 or CM 3, in order for the protocol to be enactable. 4 Discussion Despite the large amount of work on enactability, very few approaches consider how message ordering and decision structures affect its definition, very few come with an implemented prototype, and none considers the issues raised by the communication model. Although one motivation might be that it is generally desirable to have robust protocol specifications that are inde- pendent of the underlying platform implementation, also ensuring separation of concerns, we observe that robustness could make the protocol too complex, or harder to maintain. Considering what the underlying implementation guar- antees w.r.t. communication model, we can relax our specifications, and above all, a protocol that is not enactable in some platform, can be in some other. This makes our work relevant to platform designers, and protocol designers. 6The code is available on the web at: http://enactability.altervista.org/ 10 A PREPRINT - FEBRUARY 14, 2019 Taking all these features into account in a unified semantic-driven way, and demonstrating the potential of the approach on a highly expressive protocol language, are the innovative and original features of this contribution. Desai and Singh [22] limit their investigation to the RS message ordering interpretation, that they consider the standard of correctness. Hence, despite the nice introduction they provide to other message orderings and to the problems they might raise, the definition of enactability they provide is not parametric in the MOI. Lanese et al. [17] move a step further, but the generality of their approach is still limited. They define three different notions of enactability, that they name conformance: sender conformance, receiver conformance, and disjoint confor- mance. That approach is more flexible that the one by Desai and Singh, but less general than ours, where the definition of enactability is parametric in the MOI and does not require different cases. Also, they only consider how sequence and choice are affected by MOIs, leaving the study of other operators for the future. Moreover, when discussing in- teraction protocols whose most external operator is a choice, they put a very strong constraint for enactability, namely that the agents involved in the two branches of the choice (excluding the agents involved in the choice itself) are the same. We added decision structures to overcome this restriction, and provide a notion of enactability that can succeed even when that constraint is not met. Neither Desai and Singh, nor Lanese et al., use formalisms for protocol representation as expressive as trace expres- sions, and neither of them presents experiments obtained from a working prototype, as we do. With respect to the introduction of decision structures to remove unnecessary restrictions on enactability of protocols when choice is involved, our proposal is similar to that by Qiu et al., [7], as for the other works we have discussed in this section, we implemented our enactability checker, whereas their work only provides definitions. Additionally, our approach is simpler in that we do not need to label the choice operator with agents as they do. In the future, we will address both theoretical and practical issues. On the theoretical side, we will carry out a sys- tematic analysis of the relationships between Communication Model and Message Ordering Interpretation, to identify those combinations which provide some guarantees by design. We will also consider the relationships between en- actability and distributed monitorability [11], as they might turn out to resort to the same definition. On the practical part, we plan to improve our working prototype to provide a useful tool to assess protocols for enactability. Apart from providing a user-friendly interface, a key issue to address will be to provide a way to isolate the part of a non-enactable protocol that makes it non-enactable. Also, trace expressions are interpreted in a coinductive way [23] to represent infinite traces of events. Since Haskell does not support coinduction, the existing prototype can be only used on acyclic message and interactions protocols. Haskell has been chosen because the implementation mimics the semantics requiring next to no effort. In order to fully implement the proposed features we are planning to develop the enactability check using SWI-Prolog7, which natively supports coinduction. To stress-test the prototype and assess its performance from a qualitative and quantitative viewpoint we plan to create a library of interaction protocols known to be "problematic" w.r.t. enactability, and perform systematic experiments. Finally, this work highlighted the need of characterising the existing agent infrastructures like Jade [24], Jason [25], Jadex [26], etc, in terms of the communication model they support. This would allow us to state if a protocol is enactable on a given infrastructure, strengthening the potential of our proposal to be exploited in real applications. References [1] Giuseppe Castagna, Mariangiola Dezani-Ciancaglini, and Luca Padovani. On global types and multi-party ses- sions. In Roberto Bruni and Jürgen Dingel, editors, Formal Techniques for Distributed Systems - Joint 13th IFIP WG 6.1 International Conference, FMOODS 2011, and 31st IFIP WG 6.1 International Conference, FORTE 2011, Reykjavik, Iceland, June 6-9, 2011. Proceedings, volume 6722 of Lecture Notes in Computer Science, pages 1 -- 28. Springer, 2011. [2] James L. Peterson. Petri nets. ACM Compututing Surveys, 9(3):223 -- 252, September 1977. [3] W3C. Web Services Choreography Description Language Version 1.0. https://www.w3.org/TR/ws-cdl-10/, 2005. [4] Marc-Philippe Huget and James Odell. Representing Agent Interaction Protocols with Agent UML. In James Odell, Paolo Giorgini, and Jörg P. Müller, editors, Agent-Oriented Software Engineering V: 5th International Workshop, AOSE 2004, Revised Selected Papers, pages 16 -- 30. Springer Berlin Heidelberg, Berlin, Heidelberg, 2005. 7http://www.swi-prolog.org 11 A PREPRINT - FEBRUARY 14, 2019 [5] David Harel. Statecharts: a visual formalism for complex systems. Science of Computer Programming, 8(3):231 -- 274, 1987. [6] Enrico Giunchiglia, Joohyung Lee, Vladimir Lifschitz, Norman McCain, and Hudson Turner. Nonmonotonic causal theories. Artificial Intelligence, 153(1-2):49 -- 104, 2004. [7] Zongyan Qiu, Xiangpeng Zhao, Chao Cai, and Hongli Yang. Towards the theoretical foundation of choreography. In Carey L. Williamson, Mary Ellen Zurko, Peter F. Patel-Schneider, and Prashant J. Shenoy, editors, Proceedings of the 16th International Conference on World Wide Web, WWW 2007, Banff, Alberta, Canada, May 8-12, 2007, pages 973 -- 982. ACM, 2007. [8] Davide Ancona, Angelo Ferrando, and Viviana Mascardi. Comparing trace expressions and linear temporal logic for runtime verification. In TPFM, volume 9660 of LNCS, pages 47 -- 64, 2016. [9] Davide Ancona, Sophia Drossopoulou, and Viviana Mascardi. Automatic generation of self-monitoring MASs from multiparty global session types in Jason. In DALT, volume 7784 of LNCS, pages 76 -- 95. Springer, 2012. [10] Davide Ancona, Angelo Ferrando, and Viviana Mascardi. Parametric runtime verification of multiagent systems. In AAMAS, pages 1457 -- 1459. ACM, 2017. [11] Angelo Ferrando, Davide Ancona, and Viviana Mascardi. Decentralizing MAS monitoring with decamon. In Kate Larson, Michael Winikoff, Sanmay Das, and Edmund H. Durfee, editors, Proceedings of the 16th Confer- ence on Autonomous Agents and MultiAgent Systems, AAMAS 2017, São Paulo, Brazil, May 8-12, 2017, pages 239 -- 248. ACM, 2017. [12] Angelo Ferrando, Louise A. Dennis, Davide Ancona, Michael Fisher, and Viviana Mascardi. Recognising as- sumption violations in autonomous systems verification. In Elisabeth André, Sven Koenig, Mehdi Dastani, and Gita Sukthankar, editors, Proceedings of the 17th International Conference on Autonomous Agents and MultiA- gent Systems, AAMAS 2018, Stockholm, Sweden, July 10-15, 2018, pages 1933 -- 1935. International Foundation for Autonomous Agents and Multiagent Systems Richland, SC, USA / ACM, 2018. [13] Davide Ancona, Angelo Ferrando, Luca Franceschini, and Viviana Mascardi. Parametric trace expressions for runtime verification of Java-like programs. In FTfJP@ECOOP, pages 10:1 -- 10:6. ACM, 2017. [14] Angelo Ferrando, Davide Ancona, and Viviana Mascardi. Monitoring patients with hypoglycemia using self- adaptive protocol-driven agents: A case study. In Matteo Baldoni, Jörg P. Müller, Ingrid Nunes, and Rym Zalila- Wenkstern, editors, Engineering Multi-Agent Systems - 4th International Workshop, EMAS 2016, Singapore, Singapore, May 9-10, 2016, Revised, Selected, and Invited Papers, volume 10093 of Lecture Notes in Computer Science, pages 39 -- 58. Springer, 2016. [15] Sabine Broda, António Machiavelo, Nelma Moreira, and Rogério Reis. Automata for regular expressions with shuffle. Inf. Comput., 259(2):162 -- 173, 2018. [16] Davide Ancona, Angelo Ferrando, and Viviana Mascardi. Comparing trace expressions and linear temporal logic for runtime verification. In Theory and Practice of Formal Methods, volume 9660 of LNCS, pages 47 -- 64, 2016. [17] Ivan Lanese, Claudio Guidi, Fabrizio Montesi, and Gianluigi Zavattaro. Bridging the gap between interaction- and process-oriented choreographies. In Antonio Cerone and Stefan Gruner, editors, Sixth IEEE International Conference on Software Engineering and Formal Methods, SEFM 2008, Cape Town, South Africa, 10-14 Novem- ber 2008, pages 323 -- 332, CA, USA, 2008. IEEE Computer Society. [18] Kohei Honda, Vasco Thudichum Vasconcelos, and Makoto Kubo. Language primitives and type discipline for structured communication-based programming. In Chris Hankin, editor, Programming Languages and Systems - ESOP'98, 7th European Symposium on Programming, Held as Part of the European Joint Conferences on the Theory and Practice of Software, ETAPS'98, Lisbon, Portugal, March 28 - April 4, 1998, Proceedings, volume 1381 of Lecture Notes in Computer Science, pages 122 -- 138. Springer, 1998. [19] Kaku Takeuchi, Kohei Honda, and Makoto Kubo. An interaction-based language and its typing system. In Constantine Halatsis, Dimitris G. Maritsas, George Philokyprou, and Sergios Theodoridis, editors, PARLE '94: Parallel Architectures and Languages Europe, 6th International PARLE Conference, Athens, Greece, July 4-8, 1994, Proceedings, volume 817 of Lecture Notes in Computer Science, pages 398 -- 413. Springer, 1994. [20] Florent Chevrou, Aurélie Hurault, and Philippe Quéinnec. On the diversity of asynchronous communication. Formal Aspects of Computing, 28(5):847 -- 879, 2016. [21] Leslie Lamport. Time, clocks, and the ordering of events in a distributed system. Commun. ACM, 21(7):558 -- 565, July 1978. [22] Nirmit Desai and Munindar P. Singh. On the enactability of business protocols. In Dieter Fox and Carla P. Gomes, editors, Proceedings of the Twenty-Third AAAI Conference on Artificial Intelligence, AAAI 2008, Chicago, Illi- nois, USA, July 13-17, 2008, pages 1126 -- 1131, CA, USA, 2008. AAAI Press. 12 A PREPRINT - FEBRUARY 14, 2019 [23] Davide Sangiorgi. On the origins of bisimulation and coinduction. ACM Trans. Program. Lang. Syst., 31(4):15:1 -- 15:41, May 2009. [24] Fabio Luigi Bellifemine, Giovanni Caire, and Dominic Greenwood. Developing Multi-Agent Systems with JADE. Wiley, 2007. [25] Rafael H. Bordini, Jomi Fred Hübner, and Michael Wooldridge. Programming Multi-Agent Systems in AgentS- peak Using Jason (Wiley Series in Agent Technology). John Wiley & Sons, 2007. [26] Alexander Pokahr, Lars Braubach, and Winfried Lamersdorf. Jadex: A BDI reasoning engine. In Rafael H. Bordini, Mehdi Dastani, Jürgen Dix, and Amal El Fallah Seghrouchni, editors, Multi-Agent Programming: Lan- guages, Platforms and Applications, pages 149 -- 174. Springer US, Boston, MA, 2005. 13
1908.08288	1	1908	2019-08-22T10:03:15	Dealing with uncertainty in agent-based models for short-term predictions	[ "cs.MA" ]	Agent-based models (ABM) are gaining traction as one of the most powerful modelling tools within the social sciences. They are particularly suited to simulating complex systems. Despite many methodological advances within ABM, one of the major drawbacks is their inability to incorporate real-time data to make accurate short-term predictions. This paper presents an approach that allows ABMs to be dynamically optimised. Through a combination of parameter calibration and data assimilation (DA), the accuracy of model-based predictions using ABM in real time is increased. We use the exemplar of a bus route system to explore these methods. The bus route ABMs developed in this research are examples of ABMs that can be dynamically optimised by a combination of parameter calibration and DA. The proposed model and framework can also be used in an passenger information system, or in an Intelligent Transport Systems to provide forecasts of bus locations and arrival times.	cs.MA	cs	DEALING WITH UNCERTAINTY IN AGENT-BASED MODELS FOR SHORT-TERM PREDICTIONS A PREPRINT School of Geography & Leeds Institute of Data Analytics Le-Minh Kieu∗ University of Leeds United Kingdom [email protected] Nicolas Malleson School of Geography University of Leeds and Alan Turing Institute United Kingdom [email protected] Alison Heppenstall School of Geography University of Leeds and Alan Turing Institute United Kingdom [email protected] August 23, 2019 ABSTRACT Agent-based models (ABM) are gaining traction as one of the most powerful modelling tools within the social sciences. They are particularly suited to simulating complex systems. Despite many methodological advances within ABM, one of the major drawbacks is their inability to incorporate real-time data to make accurate short-term predictions. This paper presents an approach that allows ABMs to be dynamically optimised. Through a combination of parameter calibration and data assimilation (DA), the accuracy of model-based predictions using ABM in real time is increased. We use the exemplar of a bus route system to explore these methods. The bus route ABMs developed in this research are examples of ABMs that can be dynamically optimised by a combination of parameter calibration and DA. The proposed model and framework can also be used in an passenger information system, or in an Intelligent Transport Systems to provide forecasts of bus locations and arrival times. Keywords First keyword · Second keyword · More 1 Introduction Agent-based modelling (ABM) [4] is a field that excels in its ability to simulate complex systems. Instead of deriving aggregated equations of system dynamics, ABM encapsulates system-wide characteristics from the behaviours and interactions of individual agents e.g. human, animals or vehicles. ABM has emerged as an important tool for many applications ranging from urban traffic simulation [1], humanitarian assistance [10] to emergency evacuations [39]. Despite the many advances and applications of ABM, the field suffers from a serious drawback: models are currently unable to incorporate up-to-date data to make accurate real-time predictions [28, 46, 47]. Models are typically calibrated once, using historical data, then projected forward in time to make a prediction. Here, calibration is ideal for one point in time, but as the simulation progresses, the prediction rapidly diverges from reality due to underlying uncertainties [47]. These uncertainties come from dynamic (changing over space and time), stochastic (containing inherent randomness) and unobserved (unseen from the data) conditions of the real system under study. An example of such a system can be found in bus routes. Each time a bus reaches a bus stop, the number of alighting passengers is uncertain and the number of waiting passengers downstream is unobserved. The bus route's conditions also change over time, e.g. traffic ∗Corresponding author A PREPRINT - AUGUST 23, 2019 varies over the route and with at off-peak to peak periods. There are methods to incorporate streaming data into models, such as data assimilation (DA) routines [26, 45]. Broadly, DA refers to a suite of techniques that allow observational data to be incorporated into models [45] to provide an optimal estimate of the evolving state of the system. Performing DA increases the probability of having an accurate representation of the current state of the system, thereby reducing the uncertainty of future predictions. This is a technique that has been widely applied in fields such as meteorology, hydrology and oceanography [21]. There are, however, two methodological challenges that must be overcome to apply DA in ABM. First, DA methods are often intrinsic to their underlying models which are typically systems of partial differential equations with functions linearised mathematically. Hence DA methods typically rely on linearising the underlying model [16]. One of the most appealing aspects of agent-based models is that they are inherently non-linear, so it is not clear whether the assumptions of traditional DA methods will hold. Second, it is still unknown how much uncertainty DA can effectively deal with when implemented within ABM. Assimilation of real-time data into ABMs has only been attempted a few times and these examples are limited by their simplicity [28, 46, 47]. This paper is part of a wider programme of work2 that is focused on developing DA methods to be readily used in ABM. This paper focuses on one particular model that aims to make predictions of bus locations in real time. Bus route operation has been chosen due to its inherent uncertainties -- for example a model will need to account for uncertain factors affecting how buses travel on the roads [22] -- but also for its tractability -- there are many fewer interactions than present in, say, a model of a crowd. We also focus on one particular DA algorithm -- the Particle Filter (PF). This method is chosen due to its ability to incorporate data into non-linear models such as ABMs [5]. The objectives of this paper are to: (1) perform dynamic state estimation to reduce the uncertainty in the model's estimate of the current system state; (2) improve the accuracy of short term forecasts. All the numerical experiments in this paper will be tightly controlled, following an 'identical twin' experimental framework [for example see 46]. We will first develop a complex ABM of a bus route to generate fine-grained synthetic GPS data of buses, that are reasonably similar to real GPS data, for use as synthetic 'ground truth' data. We call this model the 'BusSim-truth' model. The next step is to develop companion ABMs that are of simpler nature than BusSim-truth that will not know the parameters of BusSim-truth and will not have the dynamic and stochastic features of BusSim-truth. We will calibrate and evaluate these companion ABMs against the data generated from BusSim-truth. This experiment is designed to be similar to the real-time monitoring and predictions of bus locations, where models are often a simpler version of reality, that are calibrated to be as close as possible to reality. The prediction of bus location and arrival times are essential for bus operators and a topical research challenge [3]. The methods developed here can easily be applied to simulation and forecasting for real bus systems and could, therefore, offer considerable potential impact. This is particularly pertinent in rapidly developing cities where daily bus schedules can be extremely erratic. In these cases accurate, up-to-date estimates of current waiting times will be highly beneficial to citizens who use (or would like to use) public transport. The contributions of this paper are threefold. First, several ABMs of bus routes are constructed that account for the interactions between the bus and passengers, the bus and the surrounding traffic, and between multiple buses are considered. While model development is not the sole focus of this paper, these bus route ABMs are novel and have utility for other studies. Second, this paper introduces a combination of parameter calibration and DA techniques that can dynamically optimise an ABM to enable accurate estimation of the bus system in real time. Third, this paper shows and quantifies the impacts of calibration and DA in dealing the with stochastic and dynamic nature of the system under study. This paper is structured as follows. Section 2 describes the research problem and the related works in the literature. Section 3 outlines the methodology. Section 4 describes the numerical experiments that are conducted and discusses these results. Finally, Section 6 concludes the study and considers the opportunities for future work. 2 Research problem and related works Historical and real-time bus GPS data is often used by operators to locate buses and predict their locations and arrival times. For instance, Public Transport Information and Priority System (PTIPS) is a state-of-the-art system in Australia to give priority to public transport vehicles on the roads and provide information on predicted bus arrival to passengers. The prediction of bus locations and arrival times in real time is a challenging problem [8]. Ideally, perfect knowledge of the current state of the system and any underlying processes is required. However, obtaining this level of knowledge is impossible due to sources of uncertainty and the complex interactions in bus operations. The majority of research within this area has focused on machine learning methods to find a direct mapping between input data and bus arrival 2http://dust.leeds.ac.uk/ 2 A PREPRINT - AUGUST 23, 2019 time. Examples of these methods include Artificial Neural Networks [8], Support Vector Machines [3], and Bayesian techniques [22]. While machine learning methods are generally efficient in real time, they are solely reliant on the quality of available data. Even with high-resolution datasets that record accurate spatio-temporal bus locations, the full complexity of the system will never be captured. There are analytical and simulation models of bus routes that aim to reproduce the underlying processes in bus operations, and shed some light on the associated uncertainties. One of the earliest successes in simulating a simple bus systems was from Cellular Automata modelling [9, 20, 29, 35]. Whilst the dynamical foundations of these models are well understood, they are outperformed by more sophisticated models such as bus-following models [18, 19, 32, 33, 41]; and traffic-following models [7, 15, 42]. Bus-following models aim to model the fundamental dynamics of a bus route by modelling individual buses that follow each other (for example speeding up if the bus ahead is far away). Traffic-following models, on the other hand, aim to model buses as a component of a transport system with private and public transport, where their speeds are affected by the traffic flow, traffic signals [15] or traffic density [42]. The majority of these models are static, i.e. they only have parameters that are fixed over time. We can represent these static simulators with the equation Y = f (X), where f represents the simulator. A run of a simulation is defined as the process of producing one set of data Y for a single set of model parameters X. One way for these models to reduce their uncertainty and fit more closely to the observed data is to adjust the model parameters until the model satisfies some predetermined criteria. This parameter adjustment process is often referred to as parameter calibration. Popular optimisation techniques include simulated annealing [37], genetic algorithms, [17, 30], and approximate Bayesian computation [14]. Parameter calibration, especially with ABMs, is often only implemented once, and therefore cannot account for any changes that may take place within the system. In the traffic context these might include accidents, traffic signal failures, vehicle faults, etc. Static models are simple to implement, but struggle to model dynamic systems. In real-time applications, e.g bus location or bus arrival time prediction in real time, prediction models often have to deal with the fact that there are so much uncertainty in bus operations. Real bus operation is dynamic (changing over time) and also stochastic (contains inherent randomness). In real time, there are also many unobservable information of bus operation, such as the number of passengers who are waiting at downstream stops or the number who plan to get off the bus, and the surrounding traffic conditions. The lack of information about these factors means that any model of bus operation in real time will have to make assumptions thereby introducing uncertainties. Therefore, this research will explore a combination of parameter calibration and a data assimilation (DA) technique to calibrate a dynamic ABM bus route simulator using historical data, and then dynamically optimised it on-the-fly using real-time data. This, in itself, is a novel and important contribution. Few previous efforts have attempted to incorporate data assimilation with agent-based models [for example see 46, 47], and it is unclear how DA methods, that have typically been created for linear models [16], can be adapted for non-linear ABMs. DA methods assume that observational data are sparse and only describe the target system in limited detail. Therefore a model is essential as a means of filling in the gaps in space and time left by the observations through the generation of additional data. In effect, the model propagates data from observed to unobserved areas [6]. Although techniques can be used to perform parameter estimation, they are most often framed as a state estimation problem. The aim is to calculate a posterior probability for the state vector Xt, given prior distributions from a model (in this case, a bus route operation model) and data from observations. It is this marriage of a model and real-time observational data (and the associated uncertainties) that offers the means of allowing all the available information to be used to determine the true state of the system as accurately as possible [40]. Models where the system state at time t are only dependent on the state at time t − 1 are termed Markovian. We are particularly interested in ABMs that can be written in a Markovian nature because DA algorithms require knowledge to the full model state in the form of the state vector Xt. While some ABMs in the literature track agent histories and use this information to decide future states, these can be recast as Markovian ABMs by expanding the state vector to include these histories. Implementing a bus route system as a Markovian model requires variables such as vehicle locations, speeds, occupancies etc. It is reasonable to assume that the system state at the next time step only depends on the value of these variables at the current time step. For simplicity, we assume that the state vector used here has a fixed size. The unused variables (i.e. those for buses that have yet to enter the system) can be set to zero, enabling the state vector to be treated as sparse and passed efficiently between iterations. If the state vector has a fixed size, then all possible states of the system belongs to a state-space X ∈ Rn. The system state evolves in some fixed interval {0,..,K}. We denote the state of the bus route at time t by Xt ∈ X . This paper follows an 'identical twin' experiment framework [46], where experiment data to be used will be generated from simulation, instead of using real data. The reason is that real data often comes with noise that hides the true state of the bus route (e.g. noise from GPS data). A simulated synthetic data would enable us to control the level of noise 3 in the data, and to evaluate the modelling results against the ground truth rather than noisy data. Figure 1 shows the workflow of this study. A PREPRINT - AUGUST 23, 2019 Figure 1: Study workflow. The study workflow generally consists of 2 major steps. It starts with the development of a Markovian ABM of bus route operation that will be referred as BusSim-truth. BusSim-truth is a hypothetical version of reality and will be used to generate synthetic GPS data of bus locations with timestamps. Two sets of data will be generated. The first represents 'historical' GPS data, which are essentially the outputs of multiple runs of the same BusSim-truth model with the same predefined set of parameters. The GPS data will be slightly different each time the model is run because BusSim-truth is stochastic (its outputs vary slightly from one run to another) and dynamic (the parameters that control factors such as the amount of traffic vary during a single model run). The second set of data represent a single run of BusSim-truth, also using the same set of parameters. These data will represent synthetic 'real-time' GPS data and will be used to conduct data assimilation. This situation is similar to the reality, where 'historical' data across multiple days are used to calibrate models and 'real-time' data represent the current state of the world. BusSim-truth will be reasonably realistic and will replicate popular phenomenon in bus operations such as bus bunching (two buses of the same line arrive at the same bus stop at the same time). In reality, any simulation model is a simplification of the actual dynamics. Taking this into consideration, we develop two simpler variations of BusSim-truth, knowing that they would not be able to perfectly represent the dynamics in BusSim-truth. The two variations are: • BusSim-deterministic. This model evolves exactly the same way in each model run; • BusSim-stochastic. This model is stochastic, e.g. the numbers of people waiting at bus stops is drawn from a random distribution As would be necessary in reality, BusSim-deterministic and BusSim-stochastic will first be calibrated against the synthetic 'historical' GPS data. In the second step of the study workflow, DA will be used in an attempt to update the states of the models to the 'real-time' GPS observations in order to produce more accurate short-term forecasts of the system behaviour. 3 Methodology 3.1 A hypothetical version of reality: BusSim-truth and its two simpler variations The first step in the proposed workflow is to develop an agent-based bus route model that will be used to generate synthetic GPS data for each bus on the route (BusSim-truth). BusSim-truth is a stochastic and dynamic model with two 4 A PREPRINT - AUGUST 23, 2019 classes of agents (bus and bus stop) and predefined parameters (see Table 1 ). It is stochastic because the number of boarding passengers is drawn from a random distribution, and dynamic because it parameters gradually change over time. The level of stochasticity and dynamicity in BusSim-truth can also be adjusted to represent bus route systems where conditions are largely stable or volatile over time. Figure 2 illustrates the workflow for BusSim-truth. Only a brief explanation of the model is included here, as more information on how the BusSim-truth model works can be found in the Appendix A. At each current time step, each Bus agent checks whether the next time step would be larger than the vehicle's scheduled dispatch time. If it is, we then check whether the bus is on the road (Status equals M OV IN G), or at a stop for passenger dwelling (Status equals DW ELLIN G), or has finished its service (Status equals F IN ISHED), otherwise the bus remains IDLE. If the status is M OV IN G, we first check whether the bus is at a bus stop, by comparing the GeoF ence area of each bus stop agent with the bus' location. If the bus is not approaching a bus stop, its current speed will be compared with the surrounding traffic speed. In the case it is slower, we assume that the bus will speed up. If the speed already matches the traffic speed, the bus will maintain the same speed. Currently the traffic volume on the whole network is represented as a single dynamic parameter, although in practice it would be relatively trivial to make the traffic volume heterogeneous across the network. The system will first check if the stop is at the last stop when the bus is approaching a bus stop, where the bus' status will be changed to F IN ISHED and the bus speed changed to zero. If it is not the last stop, the system will change the status of agent Bus to DW ELLIN G and its speed to zero. Table 1: Type of agents and their parameters in BusSim-truth Description Unique ID of the bus agent The acceleration value in m/s2 if the bus needs to accelerate Parameter BusID Acceleration StoppingTime Deadtime due to door opening and closing if the bus has to stop Visited States Trajectory bus stopID Position Arrm Depm Arrival_time GeoFence List of visited bus stops Whether the bus is idle, moving, dwelling or finished GPS coordinates of bus locations Unique ID of the bus stop Distance from the first stop Passengers arrived to the stop per second Percentage of onboard passengers alight at the stop Store actual arrival time of buses at the stop A circle area to identify whether the bus is at the bus stop As described in Section 2, we use BusSim-truth to generate two sets of synthetic data: (1) 'historical' GPS data that simulate normal bus route operation over a number of days and are used for calibration; and (2) 'real-time' GPS data that represents a single run of the model and are used to represent the bus system today. These are visualised in Figure 3. Each record in the synthetic data is called an observation vector. The vector contains all of the observations made from the 'real world' (in this case the BusSim-truth). 3.2 Optimising the parameters of BusSim-deterministic and BusSim-stochastic Most agent-based models have a large number of parameters. For the BusSim models, the model parameter vector St at time t contains the arrival rate Arrt m at each stop m, and the traffic speed V t. (1) St =(cid:2) Arrt m, departure rate Dept m Dept m V t (cid:3) m = 1..M The two simpler models (BusSim-deterministic and BusSim-stochastic) first need to be calibrated to the observations (the historical data). Here an automatic parameter calibration process, based on the Cross-Entropy Method (CEM) [38] is used. CEM is a population-based Monte Carlo learning algorithm to combinatorial multi-extremal optimisation and importance sampling. It originated from the field of rare event simulation, where even small probabilities need to be estimated [38]. In principle, CEM develops a probability distribution over possible solutions for the optimal parameters 5 A PREPRINT - AUGUST 23, 2019 Figure 2: Flowchart of BusSim-truth. of the model. New solution candidates are drawn from this distribution and are evaluated. The best candidates are then selected to form a new improved probability distribution of the optimal parameters, until certain criteria are met. CEM is chosen over some other popular optimisation methods in parameter calibration of ABMs, such as Genetic Algorithm [17] and simulated annealing [37], because of its probabilistic nature that facilitates the calibration of stochastic models [34]. The interested reader may refer to [38], and various applications of CEM, such as [34], for a more detailed account. A pseudo-code of the CEM algorithm that we adopted for this paper has also been described in Appendix B. Formally, the parameter calibration is an optimisation problem to minimise some performance index P I(π) over all π ∈ Rk. Here a solution π = (π1, π2, ..., πk) denotes a set of parameters of the model under consideration and k denotes the number of dimension in this set. Let π∗ denote the optimal solution, or the best set of model parameters that we want to find, that is: π ∈ Rn π∗ = argmin P I(π), (2) The above objective function is equivalent to finding π∗ such that P I(π∗) ≤ P I(π) ∀X ∈ Π, where Π is a constrained parameter space such that Π ∈ Rk. The performance index P I(π) is generally the difference between model output and observed data. The complexity of this problem comes from the stochasticity of BusSim, where the same solution π may yield a different realisation P I(π). To reduce this stochastic effect, it is necessary to run the (stochastic) model multiple times, and to evaluate the simulation outputs against a compilation of observed data from multiple days or instances. Let KI be the number of replications required for each model evaluation and KO be the number of instances in the observed data, we can derive a more detailed objective function of the parameter calibration problem: (3) min P I(π) = 1 N · T N(cid:88) T(cid:88) (cid:115)(cid:80)KI(cid:0)sSIM j,i,t − sSIM KI − 1 n=1 j,i,t t=1 (cid:12)(cid:12)(cid:12)(cid:12) (cid:34)(cid:12)(cid:12)(cid:12)(cid:12) 1 (cid:1)2 KI − KI(cid:88) (cid:115)(cid:80)KO(cid:0)sOBS KO(cid:88) j,i,t − 1 sSIM KO j,o,t − sOBS KO − 1 j,o,t sOBS j,o,t (cid:12)(cid:12)(cid:12)(cid:12)+ (cid:35) (cid:1)2 (cid:12)(cid:12)(cid:12)(cid:12) 6 BusSim flowchartEach bus agent at each time step tt > dispatch_time - 1 ?Set Status = IDLEStatus =MOVINGAt a bus stop?Check StatusCheck LocationCheck SpeedSpeed <Traffic_Speed?Maintain thespeedAccelerateBoarding orAlighting > 0?Maintain thespeedSpeed = 0SetStatus=DWELLINGRegisterLeave_stop_timeStatus =DWELLINGt =Leave_stop_time-1 ?AccelerateSpeed = 0Last stop?Set Status =FINISHEDStatus =FINISHEDYESYESYESYESYESNONONONONONOYESSet Status =MOVINGCheck number ofBoarding andAligtingpassengersExported from Pencil - Sun Feb 24 2019 11:31:34 GMT+0000 (GMT) - Page 1 of 1A PREPRINT - AUGUST 23, 2019 Figure 3: Synthetic 'historical' versus 'real-time' GPS bus location data. Each coloured line shows the trajectory of one bus in the 'historical' GPS data. As the BusSim-truth model is stochastic, there are differences between the trajectories. This is similar to the reality where buses operate slightly differently on multiple days. The bold black lines are another instance of bus trajectory that we consider as the 'real-time' GPS data. is the location of simulated bus agent j at time t Where N is the number of buses, T is the number of time steps, sSIM j,i,t for the replication i, and similarly sOBS is the synthetic observed location of bus j at time t for the instance o. The j,o,t objective function in Equation 3 can be seen as the sum of the difference in mean location and standard deviation of locations at each time step for each bus and each replication/instance between simulated outputs and synthetic observed data. We want to evaluate the difference in not just the mean but also the standard deviation of bus locations because the system under study is stochastic, so it is not just the mean but also the spread of bus locations over multiple instances are important. 3.3 Data Assimilation using a Particle Filter (PF) Xt = f (Xt) + t and use data assimilation (DA) to dynamically We can formulate an ABM as a state-space model optimise the model variables with up-to-date data to reduce uncertainty. The state-space model is represented by a state-space vector Xt at time t, which contains all information of the current state of each agent in the model: =(cid:2) ct j Xt = [Ot St] m V t (cid:3) st j vt j Occt j Arrt m Dept (4) The state-space vector Xt must contain all of the information that identifies the current state of the modelled system, allowing it to be projected forward to the next time step. Thus vector Xt usually contains both the observation vector Ot and the model parameters vector St (Equation 1) at time t. Note that St has been calibrated in the previous section, but is still included in the state space vector Xt to allow the model to be dynamically optimised with new data -- this is essential in dynamic situations where parameter values change over time. This approach is often referred to as dynamic calibration [12]. Data Assimilation (DA) is a suite of methods to adjust the state of a running model using new data to better represent the current state of the system under study [47]. DA was born out of data scarcity, where observation data are sparse and insufficient to describe the system. Notwithstanding the proliferation of new data sources, insufficient data is still a 7 0100020003000400050006000Time (s)0500010000150002000025000300003500040000Distance (m)HistoricalReal-timeA PREPRINT - AUGUST 23, 2019 major problem in research. The prediction of bus locations is a clear example where the number of future boarding and alighting passengers are unknown in real time. DA algorithms fill in the spatio-temporal gaps in the observed data by running a model forward in time until new observed data are available. This is typically called the predict step in DA algorithms. After the predict step, DA has an estimate of the current system state and its uncertainty (which is often referred as the 'prior' in Bayesian literature). The next step is typically called the update step, where new observations and uncertainty are used to update the current state estimates. The result is referred to as the 'posterior' in Bayesian literature, and should be the best guest of the system state from both the observations and model. There are several DA algorithms in the literature, ranging from the simple Kalman Filter [31] to more advanced extensions, including extended, ensemble and unscented Kalman Filter [47]. These algorithms generally aim to extend the original Kalman Filter by relaxing the assumption of linearity and introducing methods to work with non-linear models. However, they may not be the most suitable candidate to incorporate data into ABMs for two reasons. First, ABMs are driven by a large number of interacting agents with goals, history and behavioural rules. As a result, they lack an analytic structure, such as differential or difference equations, to facilitate the implementation of the Kalman Filter and its extensions where often the model Jacobian and covariance matrices need to be formulated [46]. Second, although the assumption of linearity has been relaxed, these extensions assume that the noise in the model estimation is Gaussian. There is a flexible Bayesian filtering method that has been designed to work with non-linear, non-Gaussian models without analytical structure; this is the Particle Filter (PF). The key idea is to approximate a posterior distribution by a set of samples or particles, drawn from this distribution. Each particle is a concrete hypothesis of the true system state. The set of particles approximates the posterior distribution. PF is best described as a nonparametric Bayes filter because it develops the belief using a finite number of samples. Hypotheses of the system state at time t is represented by a set Pt of NP weighted random particles: Pt = {(cid:104)X (cid:98)i(cid:101) t (cid:98)i(cid:101) , w t (cid:105) i = 1, ..., NP} (5) is the state vector of the i-th particle and w (cid:98)i(cid:101) where X is the corresponding weight. Weights are non-zero, and sum t over all weights is 1. The core idea of the PF is to update and maintain this set of particles given model outputs and observations. A PF recursively estimates the particle set Pt based on the estimate Pt−1 at the previous time step, and the observation. The PF algorithm can be briefly described in three steps: (cid:98)i(cid:101) t 1. Predict: Generate the next set of particles Pt from the previous set Pt−1. This represents the prior distribution to describe how the system state evolves. 2. Importance Weighting: Compute the importance weight w for each particle in Pt. This is equivalent to the 'Update' step in Kalman Filter, and will give us the posterior distribution (cid:98)i(cid:101) t 3. Resampling: This step has no analogous step in Kalman Filter and its extensions. The resampling step creates a new set of particles from the current set. The likelihood to draw a particle is proportional to its weight. We adopt Sample Importance Resampling (SIR), a popular bootstrap systematic resampling in the PF literature [6, 46]. SIR has been developed to deal with particle deprivation, which is the problem when particles converge to a single particle after several iterations due to one particle outperforming all others [24]. This problem significantly reduces the area of state space covered by the particles in later iterations. Since resampling will generate particles using the existing pool of particles, it will not be able to produce particles where the prediction accuracy is better than the existing particle pool. This means that in classical PF, the model parameter St of both BusSim-deterministic and BusSim-stochastic will be unchanged over time. Because the parameters change over time, we need to dynamically optimise St. This problem is solved in this paper by a simple and generic solution. We improve the quality of the particles by diversification similar to [44], in a process also known as roughening, jittering, and diffusing [36]. This is achieved by adding a random Gaussian white noise σ with mean 0 and a predefined standard deviation, not to the whole state vector Xt, but to the model parameter St, to increase the probability of having particles that represent the current state of the underlying model. The PF is applied to BusSim-deterministic and BusSim-stochastic using up-to-date data from the synthetic 'real-time' GPS data. 8 A PREPRINT - AUGUST 23, 2019 4 Numerical experiment 4.1 Experiment set up To generate the synthetic 'historical' and 'real-time' GPS data used in BusSim-deterministic and BusSim-stochastic, we predetermine a set of model parameters to generate realistic GPS data. Table 2 lists the fixed parameters being used in this experiment. Class Bus Table 2: Fixed parameters in BusSim-truth Parameter FleetSize Acceleration [θ1, θ2, θ3] Value Unique ID of the bus agent 3 m/s2 [3,1,0.85] s 20 2000m 50m BusStop Number of Stops Length between stops GeoFence m V t] (Equation 1) are time-varying and therefore, randomly Second, the dynamic parameter set St = [Arrt generated using fixed rules. We first generate an initial arrival rate Arr0 m at stop m at time 0 by a random generation from an uniform distribution between the minimum and maximum passenger arrival rate [minDemand, maxDemand]. m Dept Arrm = U(minDemand, maxDemand) m = 1, ..., M (6) The departure rate is also generated from an uniform distribution, but also ordered non-decreasingly to represent the fact that more passengers alight at the end of the route than at the beginning. The departure rate at the last stop (stop M) is set as 1 to let every remaining passengers to alight the bus at the last stop. Depm = ordered (U(0.05, 0.5)), DepM = 1 & m = 1, ..., M (7) 4.2 The stochastic and dynamic nature of BusSim-truth Figures 4 and 5 provide a simple verification that demonstrates BusSim-truth generates realistic synthetic GPS data under different sets of parameters. Other variables have also been verified and will be used in the sensitivity analysis. This section outlines how this validation was achieved. We aim to control the stochastic and dynamic level in BusSim-truth using only a single parameter for each. Equation 6 controls the level of stochasticity in BusSim-truth. For instance, a pair of values [minDemand, maxDemand]=[0.5,1] means 0.5 to 1 passenger arriving at the bus stop each minute. By fixing the minDemand to be a small number (e.g. equals 0.5), we can control the stochasticity of BusSim-truth by a single parameter maxDemand, with a larger maxDemand meaning more stochasticity and vice versa. We control the level of dynamicity in BusSim-truth by a dynamic change rate parameter ξ, which gradually changes the arrival rate and surrounding traffic speed over the simulation period. To implement an inner verification of the BusSim-Truth model and to investigate the impacts of the stochastic and dynamic natures of the system under study, we evaluate the outputs from BusSim-truth under different values of stochasticity and dynamicity. Figure 4 gives an insight into the differences in bus trajectories when maxDemand equals 0.5 and 2. Note that when maxDemand equals 0.5, BusSim-truth reduces to a deterministic model (similar to BusSim-deterministic) because maxDemand would then be equal to minDemand. Each line in Figure 4 shows the GPS trajectory of bus location, as generated by BusSim-truth. The solid lines show the trajectory of buses at high and stochastic demand (maxDemand equals 3), whereas the dashed lines are for low and deterministic demand (maxDemand equals minDemand). The trajectories in Figure 4 show that as the maxDemand increases, there are more delays for each individual buses and less likely that buses are able to keep stable headway from each other. The dynamic nature of BusSim-truth is illustrated in Figure 5 when the dynamic change rate parameter ξ is equal to 1% and 10%. Because the arrival rate and traffic speed gradually change, there is little change in the bus trajectories of BusSim-truth with ξ equals 1% and 10%. As time passes, there are more delays for BusSim-truth with ξ equals 10%. This is because there are more passengers (higher arrival rate) and the buses are also travelling slower (lower traffic speed). 9 A PREPRINT - AUGUST 23, 2019 Figure 4: Synthetic bus GPS trajectory at low and high passenger demand. Red, dashed lines are bus trajectories when maxDemand equals 0.5, while black, solid lines are bus trajectories when maxDemand equals 2. Figure 5: Synthetic bus GPS trajectory with two different value of ξ. 10 0100020003000400050006000Time (s)0500010000150002000025000300003500040000Distance (m)Low passenger demandHigh passenger demand0100020003000400050006000Time (s)0500010000150002000025000300003500040000Distance (m)Dynamic change = 1%Dynamic change = 10%A PREPRINT - AUGUST 23, 2019 4.3 Scenario 1: no calibration (benchmark) This scenario aims to evaluate the prediction results from BusSim-deterministic and BusSim-stochastic without calibrating their parameters or performing data assimilation. This is necessary so that later we can evidence the additional predictive performance of the model after calibration and data assimilation. The two models are implemented using random parameters generated from Equation 6 and 7. The outputs from these models are bus locations at each time step t, which can be compiled to space-time trajectories and compared to the synthetic 'real-time' bus trajectories, as illustrated in Figure 6. Figure 6: Prediction results from Scenario 1: no calibration Figure 6 shows one particular case where maxDemand equals 2, and ξ equals 7%, as an example of the prediction results. Both models poorly predict the trajectories of the 'real' buses. This is expected because the models do not have the optimal parameters to capture the bus route operations. These models are therefore not useful for real-time prediction without parameter calibration or data assimilation. 4.4 Scenario 2: Parameter calibration In this scenario, BusSim-deterministic and BusSim-stochastic are calibrated using the Cross-Entropy Method, as described in Section 3.2. The two calibrated models are used to predict the bus locations at each time step t, which can be compiled to trajectories. Figure 7 shows an example of the comparison between the predictions from BusSim- deterministic and BusSim-stochastic versus the synthetic 'real-time' GPS data, where the maxDemand equals 2 and ξ equals 7%. Although they are improvements compared the un-calibrated versions -- Figure 7 shows that both models outperform the models in the Scenario 1 (no calibration [see Figure 6]) -- the models can only predict well early in the simulation when there is little deviation in passenger arrival rate and surrounding traffic speed. There are large observable gaps between the predictions from BusSim-deterministic and BusSim-stochastic as buses reach the end of their routes. The two models were trained with synthetic 'historical' data, but evaluated with 'real-time' data. Recall that there are differences between the 'historical' and 'real-time' data due to the stochastic nature of the system under study (see Figure 3). Therefore a data assimilation procedure is required to prevent the errors gradually increasing throughout the simulation. 11 0100020003000400050006000Time (s)0500010000150002000025000300003500040000Distance (m)BusSim-deterministicBusSim-stochasticReal-timeA PREPRINT - AUGUST 23, 2019 Figure 7: Prediction results from Scenario 2: Parameter calibration 4.5 Scenario 3: Applying a Particle Filter This section applies a PF to the calibrated BusSim-deterministic and BusSim-stochastic, as described in section 3.3. At each time step t, the two models are only provided with the observation vector Ot, and then attempts to use Ot to correct their prediction of future state vectors Xt to XT , where T is the last time step. Figure 8 illustrates the results after the models have been calibrated and have 'real-time' data incorporated (assimilated) into them during runtime. The predicted bus trajectories in Figure 8 fit much closer to the synthetic 'real-time' data than the previous scenarios (Figure 6 and 7). There are still observable gaps between the prediction and the synthetic 'real-time' GPS data, because the underlying models do not know the underlying stochasticiy and dynamicity in the synthetic data, but the improvements (which will be quantified shortly) certainly appear to be substantial. 4.6 Sensitivity analysis In this section, we perform a sensitivity analysis to compare the prediction error in each scenario. The same experiments, as described in Scenario 1 to 3, are repeated at different values of maxDemand and dynamic change rate ξ. To increase the robustness of the comparison, 10 replications have been made for each experiment, and the average Root Mean Squared Error (RMSE) values are reported. RMSE is calculated as the difference in prediction bus location and synthetic 'real-time' bus location: (cid:118)(cid:117)(cid:117)(cid:116) 1 T (cid:16) T(cid:88) k=1 (cid:17)2 yk − yk RM SE = (8) Where yk and yk is the bus location at time k from the model prediction and synthetic 'real-time' data, respectively. Table 3 compares the RMSE from each scenario. It is clear that the Scenario 3 (combination of parameter calibration and data assimilation) outperforms the other two Scenarios. 12 0100020003000400050006000Time (s)0500010000150002000025000300003500040000Distance (m)BusSim-stochasticBusSim-deterministicReal-timeA PREPRINT - AUGUST 23, 2019 Figure 8: Prediction results from Scenario 3: Parameter calibration and Particle Filtering Table 3: Sensitivity analysis of maxDemand and dynamic change rate ξ Values Scenario 1 Scenario 2 Scenario 3 maxDemand Dynamic change rate 0.5 1 1.5 2 2.5 3 3.5 4 4.5 0 2.5 5 7.5 10 12.5 15 17.5 302 313 319 335 340 337 346 338 341 197 203 208 211 218 220 232 235 102 107 112 125 119 127 133 148 145 75 77 82 89 90 93 97 102 24 25 35 49 52 62 66 59 55 41 44 40 39 49 47 45 49 5 Implications This paper presents an integrated framework to reduce uncertainty in ABMs when making predictions in real time, by combining parameter calibration and data assimilation. As discussed in Section 1 and 2, an 'identical twin' approach has been adopted instead of real noisy data to facilitate an effective evaluation of the proposed methods against the synthetic 'ground truth'. The numerical experiment shows that the framework yields more accurate predictions than (i) 13 0100020003000400050006000Time (s)0500010000150002000025000300003500040000Distance (m)BusSim-stochasticBusSim-deterministicReal-timeA PREPRINT - AUGUST 23, 2019 a benchmark scenario (without parameter calibration), and (ii) a scenario with parameter calibration but without data assimilation. In its current form, the framework can provide real time bus locations and arrival times for passenger information systems. The forecasted bus location and arrival information provides key intelligence for waiting passengers [13]. This is beneficial for all public transport passengers, but can be of particular benefit in countries, for example in the Global South [25] where there are frequent delays due to transport systems being complex, heterogeneous or heavily congested. The prediction of bus arrival times is also critical for real-time trip planners. These planning systems propose optimal alternative routes for passengers, or update information on a connecting service that may be unreachable due to delayed buses. Many advanced Intelligent Transport System applications heavily rely on predictions of bus location and arrival times, for example bus control studies such as [11]. A model-based prediction of bus location and arrival time, such as the framework in this paper, would allow bus operators the ability to evaluate and update their transportation infrastructures in real time. 6 Conclusion This paper proposes parameter calibration and data assimilation frameworks to enhance the prediction accuracy in agent-based models (ABM) when the system under study has a stochastic and dynamic nature. This is done in a 'identical twin' approach. We first develop a stochastic and dynamic ABM of bus route, referred to as BusSim-truth. This model is employed to generate synthetic 'historical' and 'real-time' GPS data of bus locations. The 'historical' data is used to train two simpler models of bus route, referred to as BusSim-deterministic and BusSim-stochastic, and evaluate against the 'real-time' data. Similar to the practice, when any simulation model is a simplification of the reality, BusSim-deterministic and BusSim- stochastic are simpler than BusSim-truth, and thus may not be able to produce a prediction similar to the synthetic 'real-time' GPS data under limited data. We propose a solution for this issue by parameter calibration using Cross- Entropy Method (Scenario 2), by a combination of parameter calibration and Particle Filtering (Scenario 3), and show that they outperform the no calibration scenario (Scenario 1) and only Particle Filtering scenario (Scenario 4), at various levels of uncertainty. This paper shows the need for parameter calibration and data assimilation, and particularly the combination of them, to improve the accuracy of model-based prediction using ABMs in real time. Future research direction includes fitting the proposed framework with real data instead of synthetic data. Data Availability This paper does not use any real data. Synthetic data has been generated from one of its models (BusSim-truth model). The source code for all the models, and the used synthetic data are available from https://github.com/ leminhkieu/Bus-Simulation-model. Competing interests We declare we have no competing interests Acknowledgements This project has received funding from the European Research Council (ERC) under the European Union Horizon 2020 research and innovation programme (grant agreement No. 757455), a UK Economic and Social Research Council (ESRC) Future Research Leaders grant (ES/L009900/1) and a ESRC/Alan Turing Joint Fellowship (ES/R007918/1). Appendix A: The BusSim model Figure 2 illustrates the workflow for BusSim-truth. At each current time step t, each Bus agent checks whether the next time step would be larger than the vehicle's scheduled dispatch time δj. If t > δj, we then check whether the bus is on the road (Status equals M OV IN G), or at a stop for passenger dwelling (Status equals DW ELLIN G), or has finished its service (Status equals F IN ISHED), otherwise the bus remains IDLE. 14 A PREPRINT - AUGUST 23, 2019 (9) (11) If the status is M OV IN G, we first check whether the bus is at a bus stop, by comparing the GeoF ence area of each bus stop agent with the bus' location. If the bus is not approaching a bus stop, its current speed vj will be compared with the surrounding traffic speed V . If vj < V , we assume that the bus will speed up with an acceleration rate aj, thus we have: j = vt−dt vt j + aj · dt Therefore for the next time step, the bus will cover a distance of: j = St−dt St j + vt j · dt (10) If the speed already matches the traffic speed V , the bus will maintain the same speed. Or else if the bus is approaching a bus stop, the system will first check if the stop is the last stop. If it is the last stop, then the bus' status will be changed to F IN ISHED and bus speed is changed to zero. If it is not the last stop, the system will change the status of agent Bus j to DW ELLIN G and its speed to zero. The number of boarding and alighting passengers from the bus j, and the time that it will leave the stop are estimated as follows. The number of boarding passenger is proportional to the time gap between the current time (when Bus j approaches the bus stop m) and the last time any bus visits the bus stop m: Bj,m = (cid:98)P o(Arrm · (ta j+1,m − ta j,m)(cid:101) Bj,m ∈ N Equation 11 shows that the number of boarding passengers is estimated using a stochastic Poisson process. A Poisson process is widely adopted in literature to estimate the count of passengers waiting at a public transport stop [7, 42]. Extensions of this stochastic process have been introduced, such as non-homogeneous Poisson process [23], where the arrival rate is time-dependent, but for simplicity we adopt a homogeneous Poisson process for this paper. Equation 11 makes the BusSim-truth model stochastic, because there is randomness in the way the Poisson process generates a number. For more details on the number generation process using stochastic Poisson process (e.g. thinning algorithm), interested readers may refer to [27]. The number of boarding passengers is also limited by the available capacity of the bus: Bj,m = max(cid:0)Bj,m, C − Occm)(cid:1) (12) The number of alighting passengers is proportional to the number of passenger on board (bus occupancy) and the departure rate at the stop m. For simplicity, we assume that Aj,m is the product between the departure rate from bus stop m and the current bus occupancy (the number of passenger on board leaving the last stop): Aj,m = (cid:98)Depm · Occj,m−1(cid:101) Aj,m ∈ N (13) To estimate the amount of time that bus will have to stay at the bus stop m for passenger boarding and alighting, a.k.a. dwell time Dj,m, we adopt the approach in [2] and the Transit Capacity and Quality of Service Manual (TCQSM) [43]: (14) The parameter set [θ1, θ2, θ3] represents the time spent for passenger boarding, alighting, and a fixed value for vehicle stopping and starting, respectively. Equation 14 is the formulation for a single-door bus system, where boarding and alighting occurs sequentially. The departure time of bus j from stop m is calculated from the arrival time ta boarding and alighting, or in other words the dwell time Dm: Dj,m = θ1 + θ2 × Bj,m + θ3 × Aj,m j,m plus the time spent at stops for passenger td j,m = ta j,m + Dj,m (15) j,m, so this is also called the Leave_stop_time, as In BusSim, the bus j is only allowed to leave the bus m at time td can be seen in the Figure 2. If the status of bus j is DW ELLIN G, it is at a stop for passenger boarding and alighting. We then check if the next j,m. If it would, then the bus would start accelerate to leave time step would be larger or equal to the leave stop time td the stop, otherwise it would stay for at least another time interval. Finally, if the status of the bus is F IN ISHED, then we would do nothing. The modelling process then moves to the next Bus agent until the last Bus, then the whole model moves to the next time step until the last time step. BusSim-truth also assumes that parameters dynamically change over time by introducing an additional parameter ξ to represent the change in passenger demand or surrounding traffic speed. For simplicity, we assume that a single, deterministic parameter ξ can model these dynamic changes. In practice, it is possible, and more desirable, to use a time-dependent value of ξ such that dynamic change is better captured, and multiple ξ to model different changes. 15 ξ > 0 represents an increase in passenger demand and traffic speed, and ξ < 0 represents otherwise. In this paper, the change in passenger demand or traffic speed is modelled as: A PREPRINT - AUGUST 23, 2019 V = V ·(cid:0)1 − t T Arrm = Arrm · (1 − t T (cid:1) · 100 ξ · 100 ξ (16) (17) A positive value of ξ in Equation 17 gradually reduces the surrounding traffic speed V and increases the arrival rate Arrm, which would lead to more bus delays and congestion. Appendix B: Cross Entropy Method for Parameter Calibration This Appendix describes the pseudocode for the Cross Entropy Method for Normal distribution [38]. Algorithm 1: Cross-Entropy Method for Normal distribution 1 Set p = (µ1, σ1, µ2, σ2, ..., µK, σK) %Initial distribution parameters 2 Set M %Number of stops 3 Set T % Maximum iteration number 4 Set I % Maximum iteration number 5 Set ρ % Set selection ratio 6 for t from 1 to T do 7 8 9 10 11 12 13 14 Draw y(i) from N (µ, σ) %Draw I samples Compute f i := f (y(i) %Main CEM loop for i from 1 to I do end Sort f i-values %Order by decreasing magnitude γ ← fρ.I %Set threshold Lγ ← {y(i)f (y(i)) ≤ γ %Collect elite samples µ(cid:48) j = 1 Lγ σ(cid:48) j = 1 Lγ µj ← αµ(cid:48) σj ← ασ(cid:48) (cid:80)Lγ (cid:80)Lγ j + (1 − α)µj %Update with step size α j + (1 − α)σj %Update with step size α i=1 µi,j %Update µ i=1 σi,j %Update σ 15 16 17 18 19 end References [1] Michael Balmer, Marcel Rieser, Konrad Meister, David Charypar, Nicolas Lefebvre, and Kai Nagel. Matsim-t: Architecture and simulation times. In Multi-agent systems for traffic and transportation engineering, pages 57 -- 78. IGI Global, 2009. [2] Robert L Bertini and Ahmed M El-Geneidy. Modeling transit trip time using archived bus dispatch system data. Journal of transportation engineering, 130(1):56 -- 67, 2004. [3] Yu Bin, Yang Zhongzhen, and Yao Baozhen. Bus arrival time prediction using support vector machines. Journal of Intelligent Transportation Systems, 10(4):151 -- 158, 2006. [4] E Bonabeau. Agent based modeling: Methods and techniques for simulating human systems. Proceedings of the National Academy of Sciences, 99(90003):7280 -- 7287, 2002. [5] James Carpenter, Peter Clifford, and Paul Fearnhead. Improved particle filter for nonlinear problems. IEE Proceedings-Radar, Sonar and Navigation, 146(1):2 -- 7, 1999. [6] Alberto Carrassi, Marc Bocquet, Laurent Bertino, and Geir Evensen. Data assimilation in the geosciences: An overview of methods, issues, and perspectives. Wiley Interdisciplinary Reviews: Climate Change, 9(5):e535, September 2018. [7] Oded Cats, Wilco Burghout, Tomer Toledo, and Haris Koutsopoulos. Mesoscopic modeling of bus public transportation. Transportation Research Record: Journal of the Transportation Research Board, (2188):9 -- 18, 2010. 16 A PREPRINT - AUGUST 23, 2019 [8] Steven I-Jy Chien, Yuqing Ding, and Chienhung Wei. Dynamic bus arrival time prediction with artificial neural networks. Journal of Transportation Engineering, 128(5):429 -- 438, 2002. [9] Debashish Chowdhury and Rashmi C Desai. Steady-states and kinetics of ordering in bus-route models: connection with the nagel-schreckenberg model. The European Physical Journal B-Condensed Matter and Complex Systems, 15(2):375 -- 384, 2000. [10] Andrew T. Crooks and Sarah Wise. GIS and agent-based models for humanitarian assistance. Computers, Environment and Urban Systems, 41:100 -- 111, 2013. [11] Carlos F Daganzo. A headway-based approach to eliminate bus bunching: Systematic analysis and comparisons. Transportation Research Part B: Methodological, 43(10):913 -- 921, 2009. [12] Annette Eicker, Maike Schumacher, Jürgen Kusche, Petra Döll, and Hannes Müller Schmied. Calibration/data assimilation approach for integrating grace data into the watergap global hydrology model (wghm) using an ensemble kalman filter: First results. Surveys in Geophysics, 35(6):1285 -- 1309, 2014. [13] Yingling Fan, Andrew Guthrie, and David Levinson. Waiting time perceptions at transit stops and stations: Effects of basic amenities, gender, and security. Transportation Research Part A: Policy and Practice, 88:251 -- 264, 2016. [14] Jakob Grazzini, Matteo G. Richiardi, and Mike Tsionas. Bayesian estimation of agent-based models. Journal of Economic Dynamics and Control, 77:26 -- 47, April 2017. [15] Etienne Hans, Nicolas Chiabaut, Ludovic Leclercq, and Robert L Bertini. Real-time bus route state forecasting using particle filter and mesoscopic modeling. Transportation Research Part C: Emerging Technologies, 61:121 -- 140, 2015. [16] Andrew C Harvey. Forecasting, structural time series models and the Kalman filter. Cambridge university press, 1990. [17] A. J. Heppenstall, A.J. Evans, and M H Birkin. Genetic algorithm optimisation of an agent-based model for simulating a retail market. Environment and Planning B: Planning and Design, 34:1051 -- 1070, 2007. Cited by 0018. [18] Scott A Hill. Numerical analysis of a time-headway bus route model. Physica A: Statistical Mechanics and its Applications, 328(1):261 -- 273, 2003. [19] HJC Huijberts. Analysis of a continuous car-following model for a bus route: existence, stability and bifurcations of synchronous motions. Physica A: Statistical Mechanics and its Applications, 308(1):489 -- 517, 2002. [20] Rui Jiang, Mao-Bin Hu, Bin Jia, and Qing-Song Wu. Realistic bus route model considering the capacity of the bus. The European Physical Journal B-Condensed Matter and Complex Systems, 34(3):367 -- 372, 2003. [21] Eugenia Kalnay. Atmospheric Modeling, Data Assimilation and Predictability. Cambridge University Press, 2003. [22] Abbas Khosravi, Ehsan Mazloumi, Saeid Nahavandi, Doug Creighton, and JWC Van Lint. Prediction intervals to account for uncertainties in travel time prediction. IEEE Transactions on Intelligent Transportation Systems, 12(2):537 -- 547, 2011. [23] Le Minh Kieu and Chen Cai. Stochastic collective model of public transport passenger arrival process. IET Intelligent Transport Systems, 12(9), 2018. [24] Augustine Kong, Jun S. Liu, and Wing Hung Wong. Sequential Imputations and Bayesian Missing Data Problems. Journal of the American Statistical Association, 89(425):278 -- 288, March 1994. [25] B Anil Kumar, Lelitha Vanajakshi, and Shankar C Subramanian. Bus travel time prediction using a time-space discretization approach. Transportation Research Part C: Emerging Technologies, 79:308 -- 332, 2017. [26] John M. Lewis, S. Lakshmivarahan, and Sudarshan Dhall. Dynamic Data Assimilation: A Least Squares Approach. Cambridge University Press, Cambridge, 2006. [27] PA W Lewis and Gerald S Shedler. Simulation of nonhomogeneous poisson processes by thinning. Naval research logistics quarterly, 26(3):403 -- 413, 1979. [28] David J. B. Lloyd, Naratip Santitissadeekorn, and Martin B. Short. Exploring data assimilation and forecasting issues for an urban crime model. European Journal of Applied Mathematics, 27(Special Issue 03):451 -- 478, 2016. [29] Yong-Ji Luo, Bin Jia, Xin-Gang Li, Can Wang, and Zi-You Gao. A realistic cellular automata model of bus route system based on open boundary. Transportation Research Part C: Emerging Technologies, 25:202 -- 213, 2012. [30] Nick Malleson, Linda See, Andrew Evans, and Alison Heppenstall. Optimising an Agent-Based Model to Explore the Behaviour of Simulated Burglars. In Vahid Dabbaghian and Vijay Kumar Mago, editors, Theories and Simulations of Complex Social Systems, number 52 in Intelligent Systems Reference Library, pages 179 -- 204. Springer Berlin Heidelberg, 2014. 17 A PREPRINT - AUGUST 23, 2019 [31] Richard J Meinhold and Nozer D Singpurwalla. Understanding the kalman filter. The American Statistician, 37(2):123 -- 127, 1983. [32] Takashi Nagatani. Kinetic clustering and jamming transitions in a car-following model for bus route. Physica A: Statistical Mechanics and its Applications, 287(1):302 -- 312, 2000. [33] Takashi Nagatani. Bunching transition in a time-headway model of a bus route. Physical Review E, 63(3):036115, 2001. [34] D Ngoduy and MJ Maher. Calibration of second order traffic models using continuous cross entropy method. Transportation Research Part C: Emerging Technologies, 24:102 -- 121, 2012. [35] OJ O'loan, MR Evans, and ME Cates. Jamming transition in a homogeneous one-dimensional system: The bus route model. Physical Review E, 58(2):1404, 1998. [36] Juan Jose Pantrigo, Angel Sanchez, Kostas Gianikellis, and Antonio S. Montemayor. Combining Particle Filter and Population-based Metaheuristics for Visual Articulated Motion Tracking. ELCVIA Electronic Letters on Computer Vision and Image Analysis, 5(3):68 -- 83, November 2005. [37] M. Pennisi, R. Catanuto, F. Pappalardo, and S. Motta. Optimal vaccination schedules using simulated annealing. Bioinformatics, 24(15):1740 -- 1742, August 2008. [38] Reuven Rubinstein. The cross-entropy method for combinatorial and continuous optimization. Methodology and computing in applied probability, 1(2):127 -- 190, 1999. [39] Timothy Schoenharl and Greg Madey. Design and Implementation of An Agent-Based Simulation for Emergency Response and Crisis Management. Journal of Algorithms & Computational Technology, 5(4):601 -- 622, 2011. [40] Oliver Talagrand. The Use of Adjoint Equations in Numerical Modelling of the Atmospheric Circulation. In An- dreas Griewank and George F. Corliss, editors, Automatic Differentiation of Algorithms: Theory, Implementation, and Application, pages 169 -- 180. SIAM, Philadelphia, PA, 1991. [41] Tieqiao Tang, Yanfeng Shi, Yunpeng Wang, and Guizhen Yu. A bus-following model with an on-line bus station. Nonlinear Dynamics, 70(1):209 -- 215, 2012. [42] Tomer Toledo, Oded Cats, Wilco Burghout, and Haris N Koutsopoulos. Mesoscopic simulation for transit operations. Transportation Research Part C: Emerging Technologies, 18(6):896 -- 908, 2010. [43] TRB. Transit capacity and quality of service manual. Transit Cooperative Highway Research Program (TCRP) Report 165, 2013. [44] P. Vadakkepat and L. Jing. Improved Particle Filter in Sensor Fusion for Tracking Randomly Moving Object. IEEE Transactions on Instrumentation and Measurement, 55(5):1823 -- 1832, October 2006. [45] Bin Wang, Xiaolei Zou, and Jiang Zhu. Data assimilation and its applications. Proceedings of the National Academy of Sciences, 97(21):11143 -- 11144, 2000. [46] Minghao Wang and Xiaolin Hu. Data assimilation in agent based simulation of smart environments using particle filters. Simulation Modelling Practice and Theory, 56:36 -- 54, 2015. [47] Jonathan A. Ward, Andrew J. Evans, and Nicolas S. Malleson. Dynamic calibration of agent-based models using data assimilation. Royal Society Open Science, 3(4), 2016. 18
1707.08741	1	1707	2017-07-27T07:47:30	Binary Voting with Delegable Proxy: An Analysis of Liquid Democracy	[ "cs.MA", "cs.AI", "cs.CY", "cs.GT" ]	The paper provides an analysis of the voting method known as delegable proxy voting, or liquid democracy. The analysis first positions liquid democracy within the theory of binary aggregation. It then focuses on two issues of the system: the occurrence of delegation cycles; and the effect of delegations on individual rationality when voting on logically interdependent propositions. It finally points to proposals on how the system may be modified in order to address the above issues.	cs.MA	cs	Binary Voting with Delegable Proxy: An Analysis of Liquid Democracy ∗ Zo´e Christoff Davide Grossi Department of Philosophy University of Bayreuth, Germany Department of Computer Science University of Liverpool, UK [email protected] [email protected] The paper provides an analysis of the voting method known as delegable proxy voting, or liquid democracy. The analysis first positions liquid democracy within the theory of binary aggregation. It then focuses on two issues of the system: the occurrence of delegation cycles; and the effect of delegations on individual rationality when voting on logically interdependent propositions. It finally points to proposals on how the system may be modified in order to address the above issues. 1 Introduction Liquid democracy [3] is a form of group decision-making considered to lie between direct and represen- tative democracy. It has been used and popularized by campaigns for democratic reforms (e.g., Make Your Laws1 in the US) and parties (e.g., Demoex2 in Sweden, and Piratenpartei3 in Germany), which used it to coordinate the behavior of party representatives in local as well as national assemblies. At its heart is voting via a delegable proxy, also called transferable or transitive proxy. For each issue submitted to vote, each agent can either cast its own vote, or it can delegate its vote to another agent-a proxy-and that agent can delegate in turn to yet another agent, and so on. This differentiates liquid democracy from standard proxy voting [20, 23], where proxies cannot delegate their vote further. Finally, the agents that decided not to delegate their votes cast their ballots (e.g., under majority rule), but their votes now carry a weight consisting of the number of all agents that, directly or indirectly, entrusted them with their vote. Context Voting by delegable proxy was most probably first outlined in [10]. Analyses of standard (non-delegable) proxy voting from a social choice-theoretic perspective-specifically through the theory of spatial voting-have been put forth in [1] and [14]. To date, little work has focused directly on liquid democracy: [17] provided an empirical study of voting behavior in liquid democracy based on election data from the Liquid Feedback4 platform of the German Piratenpartei; and [22] studied how, in the Liquid Feedback platform, issues to be submitted to vote are selected among user-generated proposals ∗The paper outlines work presented at: the Dynamics in Logic IV workshop, TU Delft, November 2016; the seminars of the Computer Science Departments of the University of Leicester and the University of Oxford, December 2016; the Dutch Social Choice Colloquium, December 2016. The authors wish to thank the participants of the above workshops and seminars for many helpful suggestions. This paper supersedes the earlier technical report [7]. The authors wish also to thank Umberto Grandi for many insightful comments on an earlier version of this paper. Both authors acknowledge support for this research by EPSRC under grant EP/M015815/1. Zo´e Christoff also acknowledges support from the Deutsche Forschungsgemeinschaft (DFG) and Grantov´a agentura Cesk´e republiky (GA CR) joint project RO 4548/6–1. 1www.makeyourlaws.org 2demoex.se/en/ 3www.piratenpartei.de 4www.liquidfeedback.org J. Lang (Ed.): TARK 2017 EPTCS 251, 2017, pp. 134–150, doi:10.4204/EPTCS.251.10 c(cid:13) Z. Christoff & D. Grossi This work is licensed under the Creative Commons Attribution License. Z. Christoff & D. Grossi 135 via proportional rankings.5 However, to the best of our knowledge, no work has so far studied voting by delegable proxy as an aggregation rule in its own sake. We do this in the present paper, studying liquid democracy from the perspective of binary aggregation [11, 13, 15, 12]. Outline The paper starts in Section 2 by introducing some preliminaries on the theory of binary ag- gregation. This preliminary section presents also novel results on binary aggregation with abstentions, which are needed for the analysis developed later in the paper. Section 3 introduces a simple model of liquid democracy based on binary aggregation. Section 4 establishes formal relations between the proposed model of liquid democracy and standard binary aggregation with abstentions. It studies the issue of circular delegations, and the issue of individual (ir)rationality when voting takes place on logi- cally interdependent issues. The section finally moves from the analysis provided to outline two variants of delegable proxy, which: are more resilient against delegation cycles (Section 4.3); better preserve individual rationality when voting on logically interdependent issues (Section 4.4). Section 5 concludes. 2 Binary Aggregation The formalism of choice for the analysis presented in this paper is binary aggregation with abstentions (see, for instance, [11]). This section is devoted to its introduction. 2.1 Opinions and Opinion Profiles A binary aggregation structure (BA structure) is a tuple A = hN, P,γi where: • N = {1, . . . , n} is a non-empty finite set individuals (N = n); • P = {p1, . . . , pm} is a non-empty finite set of issues or propositions (P = m); • γ ∈ L is an (integrity) constraint, where L is the propositional language constructed by closing P under a functionally complete set of Boolean connectives (e.g., {¬, ∧}). A binary opinion is an assignment of acceptance/rejection values (or, truth values) to the set of issues P. Allowing abstention amounts to considering incomplete opinions: an incomplete opinion is a partial function from P to {0, 1}. We will study it as a function O : P → {0, 1, ∗} thereby explicitly denoting the undetermined value "∗" corresponding to abstention. Thus, O(p) = 0 (respectively, O(p) = 1) indicates that opinion O rejects (respectively, accepts) the issue p. Syntactically, the two opinions correspond to the truth of the literals p or ¬p. For p ∈ P we write ±p to denote one element from {p, ¬p}, and ±P to denote Sp∈P {p, ¬p}, which we will refer to as the agenda of A . We say that the incomplete opinion of an agent i is consistent if the set of formulas {p Oi(p) = 1} ∪ {¬p Oi(p) = 0} ∪ {γ} can be extended to a model of γ (in other words, if the set is satisfiable). Intu- itively, the consistency of an incomplete opinion means that the integrity constraint is consistent with i's opinion on the issues she does not abstain about. We also say that an incomplete opinion is closed when- ever the following is the case: if the set of propositional formulas {p Oi(p) = 1} ∪ {¬p Oi(p) = 0} ∪ {γ} logically implies p (respectively, ¬p), then Oi(p) = 1 (respectively, Oi(p) = 0). That is, individual opinions are closed under logical consequence or, in other words, agents cannot abstain on issues whose acceptance or rejection is dictated by their expressed opinions on other issues. The set of incomplete 5Another, somewhat tangential work is [4], which focused on algorithmic aspects of a variant of liquid democracy, called viscous democracy, with applications to recommender systems. 136 Binary Voting with Delegable Proxy: An Analysis of Liquid Democracy opinions is denoted O ∗ and the set of consistent and closed incomplete opinions O ∗ to the latter simply as individual opinions, as they are the ones we focus on. c . We will often refer An opinion profile O = (O1, . . . , On) records the opinion on the elements of P, of every individual in N. Given a profile O the ith projection O is denoted Oi (i.e., the opinion of agent i in profile O). We also denote by O(p) = {i ∈ N Oi(p) = 1} the set of agents accepting issue p in profile O, by O(¬p) = {i ∈ N Oi(p) = 0} the set of agents rejecting p in O, and by O(±p) = O(p) ∪ O(¬p) the set of non- abstaining agents in O. Sometimes we restrict the previous definitions to a coalition C ⊆ N, so that OC(p) (resp., OC(¬p)) denotes the set of agents in C that accept (resp., reject) p. Finally, we write O =−i O′ to denote that the two profiles O and O′ are identical except, possibly, for the opinion of voter i. 2.2 Aggregators An aggregator is a function F : (O ∗ c )N → O ∗, from profiles of closed and consistent incomplete opinions to incomplete opinions. The issue-by-issue strict majority rule (maj) accepts an issue if and only if the majority of the non-abstaining voters accept that issue: 1 maj(O)(p) =  0 ∗ if O(p) > O(¬p) if O(¬p) > O(p) otherwise (1) We will refer to this rule simply as 'majority'. Majority can be thought of as a quota rule. Quota rules in binary aggregation with abstentions are of the following form: accept when the proportion of non-abstaining individuals who accept is above the acceptance-quota; reject when the proportion of non-abstaining individuals who reject is above the rejection-quota; and abstain otherwise:6 Definition 1 (Quota rules). Let A be a BA structure. A quota rule (for A ) is defined as follows, for any issue p ∈ P, and any opinion profile O ∈ (O ∗ c )N : 1 0 ∗ if O(p) ≥ ⌈q1(p) · O(±p)⌉ if O(¬p) ≥ ⌈q0(p) · O(±p)⌉ otherwise (2) F(O)(p) =  where ⌈·⌉ is the cealing function. And, for x ∈ {0, 1}, qx is a function qx : P → (0, 1] ∩ Q assigning a positive rational number smaller or equal to 1 to each issue, and such that, for each p ∈ P: A quota rule is called: uniform if, for all pi, p j ∈ P, qx(pi) = qx(p j); it is called symmetric if, for all p ∈ P, q1(p) = q0(p). qx(p) > 1 − q(1−x)(p). (3) Notice that the definition excludes trivial quota.7 It should also be clear that, by the constraint in c )N → O ∗ as desired.8 Notice finally that if the rule is (3), Definition 1 defines an aggregator of type (O ∗ symmetric, then (3) forces qx(p) > 1 2 , for any given p. 6There are several ways to think of quota rules with abstentions. Instead of a quota being a proportion of non-abstaining agents, one could for instance define rules with absolute quotas instead: accept when at least n agents accept, independently of how many agents do not abstain. In practice, voting rules with abstention are often a combination of those two ideas: accept an issue if a big enough proportion of the population does not abstain, and if a big enough proportion of those accept it. 7Those are quotas with value 0 (always met) or > 1 (never met). Restricting to non-trivial quota is not essential but simplifies our exposition. 8What needs to be avoided here is that both the acceptance and rejection quota are set so low as to make the rule output both the acceptance and the rejection of the issue. Z. Christoff & D. Grossi 137 Example 1. The majority rule (1) is a uniform and symmetric quota rule where q1 and q0 are set to meet the equation ⌈q1(p) · O(±p)⌉ = ⌈q0(p) · O(±p)⌉ =l O(±p)+1 2 m, for any issue p and profile O. This is achieved by setting the quota as 1 one should therefore consider maj as a class of quota rules yielding the same collective opinions. 2N , for each issue p. More precisely 2 < q1(p), q0(p) ≤ 1 2 + 1 N = N+1 Example 2. The uniform and symmetric unanimity rule is defined by setting q1 = q0 = 1. A uniform but asymmetric variant of unanimity can be obtained by setting q1 = 1 and q0 = 1 N . 2.3 Properties of Agendas and Aggregators Definition 2 (simple/evenly negatable agenda). An agenda ±P is said to be simple if there exists no set X ⊆ ±P such that: X ≥ 3, and X is minimally γ-inconsistent, that is: • X is inconsistent with γ • For all Y ⊂ X , Y is consistent with γ (or, γ-consistent). An agenda is said to be evenly negatable if there exists a minimal γ-inconsistent set X ⊆ ±P such that for a set Y ⊆ X of even size, X \Y ∪ {¬p p ∈ Y } is γ-consistent. It is said to be path-connected if there exists p1, . . . , pn ∈ ±P such that p1 =c p2, . . . , pn−1 =x pn where pi =c pi+1 (conditional entailment) denotes that there exists X ⊆ ±P, which is γ-consistent with both pi and ¬pi+1, and such that {p} ∪ X ∪ {γ} logically implies pi+1. We refer the reader to [15, Ch. 2] for a detailed exposition of the above rather technical conditions. We provide just a simple illustrative example here. Example 3. Let P = {p, q, r} and let γ = (p ∧ q) → r. ±P is not simple. The set {p, q, ¬r} ⊆ ±P is inconsistent with γ, but none of its subsets is. Let now P = {p, q, r} and let γ = (r → q) ∧ (q → p). In this case, where issues are ordered by logical entailment, each minimally γ-inconsistent set is of size 2, and the agenda is therefore simple. The trivial example of simple agenda is where γ = ⊤, and the issues are therefore logically independent. We proceed by recalling some well-known properties of aggregators from the judgment and binary aggregation literatures, adapted to the setting of aggregation with abstention:9 Definition 3. Let A be an aggregation structure. An aggregator F : (O ∗ c )N → O ∗ is said to be: unanimous iff for all p ∈ P, for all profiles O and all x ∈ {0, 1, ∗}: if for all i ∈ N, Oi(p) = x, then F(O)(p) = x. I.e., if everybody agrees on a value, that value is the collective value. anonymous iff for any bijection µ : N → N, F(O) = F(Oµ), where Oµ = (cid:10)Oµ(1), . . . , Oµ(n)(cid:11). permuting opinions among individuals does not affect the output of the aggregator. I.e., p-oligarchic iff there exists C ⊆ N (the p-oligarchs) s.t. C 6= /0 and for any profile O, and any value x ∈ {0, 1}, F(O)(p) = x iff Oi(p) = x for all i ∈ C. I.e., there exists a group of agents whose definite opinions always determine the group's definite opinion on p. If F is p-oligarchic, with the same oligarchs on all issues p ∈ P, then it is called oligarchic. monotonic iff, for all p ∈ P and all i ∈ N: for any profiles O, O′, if O =−i O′: (i) if Oi(p) 6= 1 and O′ i(p) ∈ {0, ∗}, then: if F(O)(p) = 0, then F(O′)(p) = 0. I.e., increasing support for a definite collective opinion does not change that collective opinion. i(p) ∈ {1, ∗}, then: if F(O)(p) = 1, then F(O′)(p) = 1; and (ii) if Oi(p) 6= 0 and O′ 9Such adaptation is, in many cases, non-trivial. 138 Binary Voting with Delegable Proxy: An Analysis of Liquid Democracy independent iff, for all p ∈ P, for any profiles O, O′: if for all i ∈ N, Oi(p) = O′ i(p), then F(O)(p) = F(O′)(p). I.e., the collective opinion on each issue is determined only by the individual opinions on that issue. neutral iff, for all p, q ∈ P, for any profile O: if for all i ∈ N, Oi(p) = Oi(q), then F(O)(p) = F(O)(q). I.e., all issues are aggregated in the same manner. responsive iff for all p ∈ P, there exist profiles O, O′ such that F(O)(p) = 1 and F(O′)(p) = 0. I.e., the rule allows for an issue to be accepted for some profile, and rejected for some other. unbiased iff for all p ∈ P, for any profiles O, O′ : if for all i ∈ N, Oi(p) = 1 iff O′ i(p) = 0 (we say that O′ is the "reversed" profile of O), then F(O)(p) = 1 iff F(O′)(p) = 0. I.e., reversing all and only the individual opinions on p (from acceptance to rejection and from rejection to acceptance) results in reversing the collective opinion on p. rational iff for any profile O, F(O) is consistent and closed. I.e., the aggregator preserves the con- straints on individual opinions. Majority is unanimous, anonymous, monotonic, independent, neutral, responsive and unbiased, but it is not rational in general, as witnessed by well-known judgment aggregation paradoxes (cf. [15]). Finally, let us also define the following property. The undecisiveness of an aggregator F on issue p for a given aggregation structure is defined as the number of profiles which result in collective abstention on p, that is: u(F)(p) = {O ∈ O ∗ c F(O)(p) = ∗} . (4) 2.4 Some Results Aggregation by majority is collectively rational under specific assumptions on the aggregation constraint: Proposition 1. Let A be a BA structure with a simple agenda. Then maj is rational. May's theorem [19] famously shows that for preference aggregation, the majority rule is in fact the only aggregator satisfying a specific set of desirable properties. A corresponding characterization of the majority rule is given in standard judgment aggregation (without abstentions): when the agenda is simple, the majority rule is the only aggregator which is rational, anonymous, monotonic and unbiased [15, Th. 3.2]. We give below a novel characterization theorem, which takes into account the possibility of abstentions both at the individual and at the collective level. To the best of our knowledge this is the first result of this kind in the literature on judgment and binary aggregation with abstention. We first prove the following lemma: Lemma 1. Let F be a uniform and symmetric quota rule for a given A . The following holds: 1 q0 ≤ N+1 2N if and only if F = arg minG u(G)(p), for all p ∈ P. 2 < q1 = That is, the quota rule(s) corresponding to the majority rule (Example 1) is precisely the rule that minimizes undecisiveness. We can now state and prove the characterization result: Theorem 1. Let F : (O ∗ c )N → O ∗ be an aggregator for a given A . The following holds: 1. F is a quota rule if and only if it is anonymous, independent, monotonic, and responsive; 2. F is a uniform quota rule if and only if it is a neutral quota rule; Z. Christoff & D. Grossi 139 3. F is a symmetric quota rule if and only if it is an unbiased quota rule; 4. F is the majority rule maj if and only if it is a uniform symmetric quota rule which minimizes undecisiveness. By the above theorem and Proposition 1, it follows that, on simple agendas, majority is the only rational aggregator which is also responsive, anonymous, systematic and monotonic. We conclude by recollecting a well-known impossibility result concerning binary aggregation with abstentions: Theorem 2 ([11, 9]). Let A be a BA structure whose agenda is path connected and evenly negatable. Then if an aggregator F : (O ∗ c )N → O ∗ is independent, unanimous and collectively rational, then it is oligarchic. 3 Binary Liquid Democracy In binary aggregation with delegable proxy, agents either express an acceptance/rejection opinion or delegate the expression of such an opinion to another agent. The section models and studies this type of voting as a form of binary aggregation function. 3.1 Proxy Opinions, Profiles and Delegation Graphs Let a BA structure A be given and assume for now that γ = ⊤, that is, all issues are logically independent. An opinion O : P → {0, 1} ∪ N is an assignment of either a truth value or another agent to each issue in P, such that Oi(p) 6= i (that is, self-delegation is not an expressible opinion). We will later also require proxy opinions to be individually rational, in a precise sense (Section 4.1). For simplicity we are assuming that abstention is not a feasible opinion in proxy voting, but such assumption can be easily lifted in what follows. We call functions of the above kind proxy opinions to distinguish them from standard (binary) opin- ions, and we denote by P the set of all proxy opinions, Pc the set of all individually rational proxy opinions (as defined later in Section 4.1). Finally, P N denotes the set of all profiles of proxy opinions, which we call, proxy profiles. Each proxy profile O induces a delegation graph GO = hN, {Rp}p∈Pi where for i, j ∈ N: iRp j ⇐⇒(cid:26) Oi(p) = j Oi(p) ∈ {0, 1} otherwise if i 6= j ∈ N (5) The expression iRp j stands for "i delegates her vote to j on issue p". Each Rp is a so-called functional relation. It corresponds to the graph of an endomap on N. So we will sometimes refer to the endomap rp : N → N of which Rp is the graph. Relations Rp have a very specific structure and can be thought of as a set of (converging) trees whose roots all belong to cycles (possibly loops). The weight of an agent i w.r.t. p in a delegation graph GO is given by its indegree with respect to R∗ p (i.e., the reflexive and transitive closure of Rp):10 wO C ⊆ N is defined naturally as wO C (p) = ∑i∈C wO i (p) =(cid:12)(cid:12)(cid:8) j ∈ N jR∗ pi(cid:9)(cid:12)(cid:12). The weight of a coalition i (p). This definition of weight makes sure that each 10 We recall that the reflexive transitive closure R∗ of a binary relation R ⊆ N2 is the smallest reflexive and transitive relation that contains R. 140 Binary Voting with Delegable Proxy: An Analysis of Liquid Democracy individual carries the same weight, independently of the structure of the delegation graph. Alternative definitions of weight are of course possible. For all p ∈ P, we consider the function gp : N →℘(N) defined as gp(i) =(cid:8) j ∈ N iR∗ The function associates to each agent i (for a given issue p), the (singleton consisting of the) last agent reachable from i via a path of delegation on issue p, when it exists (and /0 otherwise). Slightly abusing notation we will use gp(i) to denote an agent, that is, the guru of i over p when gp(i) 6= /0. If gp(i) = {i} we call i a guru for p. Notice that gp(i) = {i} iff rp(i) = i, i.e., i is a guru of p iff it is a fixpoint of the endomap rp. p j and jRp j(cid:9) . If the delegation graph GO of a proxy profile O is such that, for some Rp, there exists no i ∈ N such that i is a guru of p, we say that graph GO (and profile O) is void on p. Intuitively, a void profile is a profile where no voter expresses an opinion, because every voter delegates her vote to somebody else. Given a BA structure A , a proxy aggregation rule (or proxy aggregator) for A is a function pv : P N → O ∗ that maps every proxy profile to one collective incomplete opinion. As above, pv(O)(p) denotes the outcome of the aggregation on issue p. 3.2 Proxy Aggregators The most natural form of voting via delegable proxy is a proxy version of the majority rule we discussed in Section 2:11 pvmaj(O)(p) =  1 0 ∗ if ∑i∈O(p) wO if ∑i∈O(¬p) wO otherwise i (p) > ∑i∈O(¬p) wO i (p) > ∑i∈O(p) wO i (p) i (p) (6) Again, the notation O(p) (resp., O(¬p)) denotes the set of voters accepting (resp., rejecting) p in proxy profile O. Intuitively, an issue is accepted by proxy majority in profile O if the sum of the weights of the agents who accept p in O exceeds the majority quota, it is rejected if the sum of the weights of the agents who reject p in O exceeds the majority quota, and it is undecided otherwise. Note that ∑i∈O(p) wO i (p) = {i ∈ NOgi(p) = 1} (and similarly for ¬p), that is, the sum of the weights of the gurus accepting (rejecting) p is precisely the cardinality of the set of agents whose gurus accept (reject) p. It should be clear that for any quota rule F : O ∗ c → O ∗ a proxy variant pvF of F can be defined via an obvious adaptation of (6). 4 Analysis and Extensions In this section we provide an analysis of liquid democracy by highlighting two issues-the failure of rationality in ballots under delegable proxy voting, and the occurrence of delegation cycles-and by em- bedding it in the theory of binary aggregation with abstentions presented in Section 2. We also advance proposals for simple modifications of the delegable proxy voting method in order to address the issues we identify. 11On the importance of majority decisions in the current implementation of liquid democracy by Liquid Feedback cf. [3, p.106]. Z. Christoff & D. Grossi 141 4.1 Individual and Collective Rationality In our discussion so far we have glossed over the issue of logically interdependent issues and collective rationality. The reason is that under the delegative interpretation of liquid democracy developed in the previous sections individual rationality itself appears to be a more debatable requirement than it normally is in classical aggregation. A proxy opinion Oi is individually rational if the set of formulas {γ} ∪np ∈ P Ogp(i)(p) = 1o ∪n¬p ∈ P Ogp(i)(p) = 0o (7) is satisfiable (consistency), and if whenever (7) entails ±p, then ±p belongs to it (closure). That is, the integrity constraint γ is consistent with i's opinion on the issues she does not delegate on, and the opinions of her gurus (if they exist), and those opinions, taken together, are closed under logical consequence. The consistency and closure of (7) capture a highly idealized way of how delegation works: voters are assumed to be able to check or monitor how their gurus are going to vote, and always modify their delegations if an inconsistency arises. So the constraint appears highly unrealistic under a delegative interpretation of liquid democracy. Aggregation via delegable proxy has at least the potential to represent individual opinions as irrational (inconsistent and/or not logically closed). The assumption of individual rationality for proxy opinions, however, is needed in order to establish variants of known binary aggregation results for the case of liquid democracy, to which we turn now. 4.2 Embedding Having defined individual rationality in the previous section, it is possible now to study embeddings from proxy voting to standard aggregation, and vice versa. Aggregation in liquid democracy-as conceived in [3]-should satisfy the principle that the opinion of every voter, whether expressed directly or through proxy, should be given the same weight.12 In other words, this principle suggests that aggregation via delegable proxy should actually be 'blind' for the specific type of delegation graph arising. Making this more formal, we can think of the above principle as suggesting that the only relevant content of a proxy profile is its translation into a standard opinion profile (with abstentions) via a function t : P → O ∗ defined as follows: for any i ∈ N and p ∈ P, t(Oi(p)) = Ogp(i) if gp(i) 6= /0 (i.e., if i has a guru for p), and t(Oi(p)) = ∗ otherwise. Clearly, if we assume proxy profiles to be individually rational, the translation will map proxy opinions into individually rational (consistent and closed) incomplete opinions. By extension, we will denote by t(O) the incomplete opinion profile resulting from translating the individual opinions of a proxy profile O. The above discussion suggests the definition of the following property of proxy aggregators: a proxy aggregator pv has the one man–one vote property (or is a one man–one vote aggregator) if and only if c → O ∗ (assuming the individual rationality of proxy profiles).13 pv = t ◦ F for some aggregator F : O ∗ The class of one man–one vote aggregators can therefore be studied simply as the concatenation t ◦ F where F is an aggregator for binary voting with abstentions, as depicted in Figure 1 (left). 12 "[. . . ] in fact every eligible voter has still exactly one vote [. . . ] unrestricted transitive delegations are an integral part of Liquid Democracy. [. . . ] Unrestricted transitive delegations are treating delegating voters and direct voters equally, which is most democratic and empowers those who could not organize themselves otherwise" [3, p.34-36] 13Not every proxy aggregator satisfies the one man–one vote property. By means of example, consider an aggregator that 1 uses the following notion of weight accrued by gurus in a delegation graph. The weight w(i) of i is ∑ j∈R∗(i) ℓ(i, j) where ℓ(i, j) denotes the length of the delegation path linking j to i. This definition of weight is such that the contribution of voters decreases as their distance from the guru increases. Aggregators of this type are studied in [4]. 142 Binary Voting with Delegable Proxy: An Analysis of Liquid Democracy O t pvF t(O) F O s s(O) Fpv pv F(t(O)) pv(s(O)) Figure 1: Embeddings to and from binary aggregation. Example 4. Proxy majority pvmaj (6) is a one man–one vote rule aggregator. It is easy to check that, for any proxy profile O: pvmaj(O) = maj(t(O)). It follows that for every proxy aggregator pvF = t ◦F the axiomatic machinery developed for standard aggregators can be directly tapped into. Characterization results then extend effortlessly. In particular, Theorem 1 implies the following: Fact 1 (Characterization of proxy majority). A one man–one vote proxy aggregator pv = t ◦ F for a given A is proxy majority pvmaj iff F is anonymous, independent, monotonic, responsive, neutral and minimizes undecisiveness. The fact may well be considered as a theoretical argument in favor of the use of proxy majority in aggregation with delegable proxy as currently done, for instance, in the Liquid Feedback platform. Similarly, we can study an embedding of standard aggregation into voting with delegable proxy. For example, we can define a function s : O ∗ c → Pc from opinion profiles to individually rational proxy profiles as follows. For a given opinion profile O, and issue p consider the set {i ∈ N Oi(p) = ∗} of individuals that abstain in O and take an enumeration σ : {i ∈ N Oi(p) = ∗} → {1, . . . , m} of its elements, with m = {i ∈ N Oi(p) = ∗} . The function is defined as follows: for any i ∈ N and p ∈ P, s(Oi(p)) = Oi(p) if Oi(p) ∈ {0, 1}, s(Oi(p)) = (σ(i) + 1) mod m, otherwise.14 A translation of this type allows to think of standard aggregators F : O ∗ c → O ∗ as the concatenation s ◦ pv, for some proxy aggregator pv, as in Figure 1 (right). The following impossibility result for aggregation with delegable proxy voting can then be obtained as a direct consequence of Theorem 2: Fact 2. Let A be such that its agenda is path connected and evenly negatable. For any proxy aggregator pv, if s ◦ pv is independent, unanimous and collectively rational, then it is oligarchic. 4.2.1 Cycles and Abstentions Proxy aggregators rely on the existence of gurus in the underlying delegation graphs. If the delegation graph Rp on issue p contains no guru, then the aggregator has access to no information in terms of who accepts and who rejects issue p. To avoid bias in favor of acceptance or rejection, such situations should therefore result in an undecided collective opinion. That is for instance the case of pvmaj. However, such situations may well be considered problematic, and the natural question arises therefore of how likely they are, at least in principle. Proposition 2. Let A be a BA structure where γ = ⊤ (i.e., issues are independent) and fix an issue p. If each proxy profile is equally probable (impartial culture assumption), then the probability that a given proxy profile O is such that t(O) is a profile in which every voter abstains tends to 1 e2 as n tends to infinity. 14Notice that since self-delegation (that is, Oi(p) = i) is not feasible in proxy opinions, this definition of s works for profiles where, on each issue, either nobody abstains or at least two individuals abstain. Clearly, a dummy abstaining voter can then be added in profiles where only one individual abstains. Z. Christoff & D. Grossi 143 It follows that for unanimous and one man–one vote proxy aggregators, asymptotically, there is a considerable chance that a profile results in collective abstention. Now contrast this with the probability that all agents abstain on an issue when each voter either expresses a 1 or 0 opinion or abstains (that is, the binary aggregation with abstentions setting studied earlier). In that case the probability that everybody abstains, and therefore the profile is void, clearly tends to 0 as n tends to infinity. Proposition 2 should obviously not be taken as a realistic estimate of the effect of cycles on collective abstention, moreover concrete implementations of delegable proxy voting may be designed to detect and resolve cycles (cf. [24, 17]). Ultimately, theoretical (e.g., game theoretic) models of delegation behavior in voters or, ideally, election data should be used to assess whether delegation cycles ever lead large parts of the electorate to effectively lose representation in the aggregation mechanism. Still, the link we highlight between delegable proxy and collective abstention is, to the best of our knowledge, novel and has escaped so far recognition within the liquid democracy literature.15 4.3 Delegable Proxy with Default Values Motivated by the above analysis, we outline a simple modification of voting via delegable proxy, which requires agents to always submit a substantive opinion on the issues, and at the same time indicate a trustee. In this view, an opinion (called proxy opinion with default) is therefore a function Oi : P → ({0, 1} × N) assigning to every issue an acceptance or rejection value and, at the same time, an individual, which is to be considered the individual the vote is delegated to. Intuitively, each voter expresses an opinion but accepts that opinion to be overruled by the opinion of the individual she entrusts. Note that such individual may well be the voter herself (e.g., Oi(p) = (1, i)). We refer to profiles of such opinions as proxy profiles with default. Let CO(p) = {C ⊆ N C is a Rp-cycle and OC(p) > OC(¬p)} denote the set of cycles of the del- egation graph Rp such that among the agents in the cycle there exists a majority accepting p. The set CO(¬p) is defined in the symmetric way. Now define proxy majority as an aggregator for profiles of proxy opinions with default values: pv′ 1 maj(O)(p) =  if ∑C∈CO(p) wO if ∑C∈CO(¬p) wO otherwise C (p) > ∑C∈CO(¬p) wO C (p) > ∑C∈CO(p) wO C (p) C (p) (8) 0 ∗ where, recall, wO C (p) is the cumulative weight (w.r.t. Rp) of the agents in C. The intuition behind (8) is to use each cycle, and not only loops (i.e., gurus), as sources of information for the proxy aggregator, by attributing to the individuals in a cycle the majority default opinion present in that cycle. 15Delegation cycles are normally criticized for the wrong reason, that is, the fact that hey may be interpreted as to lead to an infinite accrual of voting power: "The by far most discussed issue is the so-called circular delegation problem. What happens if the transitive delegations lead to a cycle, e.g. Alice delegates to Bob, Bob delegates to Chris, and Chris delegates to Alice? Would this lead to an infinite voting weight? Do we need to take special measures to prohibit such a situation? In fact, this is a nonexistent problem: A cycle only exists as long as there is no activity in the cycle in which case the cycle has no effect. As already explained [. . . ], as soon as somebody casts a vote, their (outgoing) delegation will be suspended. Therefore, the cycle naturally disappears before it is used. In our example: If Alice and Chris decide to vote, then Alice will no longer delegate to Bob, and Chris will no longer delegate to Alice [. . . ]. If only Alice decides to vote, then only Alice's delegation to Bob is suspended and Alice would use a voting weight of 3. In either case the cycle is automatically resolved and the total voting weight used is 3." [3, Section 2.4.1] Cf. [2]. We agree that the alleged accrual of infinite voting power is immaterial. However, the fact that the occurrence of a cycle leads to the loss of representation of the voters in the cycle-and of those delegating to them-does not seem to have yet been acknowledged. 144 Binary Voting with Delegable Proxy: An Analysis of Liquid Democracy As one might intuitively expect, this is enough to break the link between delegation cycles and group abstention we identified with Proposition 2. To state the following result we need to adapt the translation function t for proxy profiles, to a translation function t′ translating proxy profiles with default to opinion profiles with abstentions: for any i ∈ N and p ∈ P, t′(Oi(p)) = maj(OC)(p) where C is the cycle reachable from i via Rp. Proposition 3. Let A be a BA structure where γ = ⊤ (i.e., issues are independent) and fix an issue p. If each proxy profile with default is equally probable (impartial culture assumption), then the probability that a given proxy profile with default O is such that t′(O) is a profile in which every voter abstains tends to 0 as n tends to infinity. 4.4 Individually Rational Delegable Proxy Delegable proxy voting can also be studied from a different perspective. Imagine a group where, for each issue p, each agent copies the binary-0 or 1-opinion of a unique trustee.16 Imagine that this group does so repeatedly until all agents (possibly) reach a stable opinion. These new stable opinions can then be aggregated as the 'true' opinions of the individuals in the group, for instance, via majority. The collective opinion of a group of agents, who either express a binary opinion or delegate it to another agent, is (for one man–one vote proxy aggregators) the same as the output obtained from a vote where each individual has to express a binary opinion but gets there by copying the opinion of her trustee (possibly the agent itself). In this perspective, aggregation via delegable proxy can be assimilated to a (stabilizing) process of opinion formation on delegation graphs. The above interpretation of liquid democracy is explicitly put forth in [3].17 Under this 'vote- copying' interpretation, the constraint on individual rationality-consistency and closure of (7)-is, arguably, more easily defendable: each agent will copy opinions coming from her trustees only if con- sistency and closure are preserved. 4.4.1 Boolean DeGroot Processes We briefly develop the above intuition, outlining an opinion diffusion model of delegable proxy which preserves individual rationality in a natural way.18 Definition 4. Fix a BA structure A = hN, P,γi, a profile O ∈(cid:16){0, 1}P(cid:17)N of γ-consistent binary opinions, and a delegation graph G = hN, {Rp}p∈Pi. Consider the stream O0, O1, . . . , On, . . . of opinion profiles recursively defined as follows: • Base: O0 := O 16For simplicity, in this section we assume agents are therefore not allowed to abstain, although this is not a crucial assump- tion for the development of our analysis. 17 "While one way to describe delegations is the transfer of voting weight to another person, you can alternatively think of delegations as automated copying of the ballot of a trustee. While at assemblies with voting by a show of hands it is naturally possible to copy the vote of other people, in Liquid Democracy this becomes an intended principle" [3, p. 22]. 18As we will consider just binary opinions (without abstentions), the concept of individual rationality can be slightly simpli- fied: requiring an opinion to be γ-consistent suffices as in the case of binary opinions without abstentions, consistency implies closedness. Z. Christoff & D. Grossi 145 • Step: for all i ∈ N, p ∈ P, On+1 i (p) :=  where Gp = hN, Rpi. On Rp(i)(p) On i (p) if {γ} ∪np ∈ P ORp(i)(p) = 1o ∪n¬p ∈ P ORp(i)(p) = 0o is consistent otherwise When γ is set to ⊤, the above defines P independent binary processes, one for each issue p. Each of such processes is a Boolean extremal case of a DeGroot stochastic process [8] where opinions are binary, and each agent can trust one and at most one other agent. When the constraint γ is not a tautology, the definition guarantees that at each step individual opinions remain consistent with γ. We call processes defined by the above dynamics individually rational Boolean DeGroot processes (in short, BDPs).19 4.4.2 Stabilization We say that the stream of opinion profiles O0, O1, . . . , On, . . . stabilizes if there exists n ∈ N such that for all m ∈ N, if m ≥ n, then Om = On. We call such profile the limit profile. A BDP that stabilizes can be thought of as an opinion transformation function [18] fG : O → O turning an initial binary profile O into a new binary profile f (O) equal to the limit profile. In this view, individually rational proxy aggregation consists first in an opinion transformation, implemented through a BDP, and then the application of an aggregator (e.g., maj) on the profile of transformed opinions f (O). A BDP that does not converge, can similarly be thought of as mapping the initial profile to a profile involving some level of abstention, where agents connected to some delegation cycle may not end up stabilizing and are therefore considered to abstain. We conclude by establishing conditions for individually rational Boolean DeGroot processes to stabilize. Theorem 3. Fix a BA structure A = hN, P,γi, a profile O of consistent (w.r.t. γ) binary opinions, and a delegation graph G. Then the following holds: if for all p ∈ P, for all C ⊆ N such that C is a cycle in Gp, and all i, j ∈ C: Oi(p) = O j(p), then the individually rational BDP (for O, G and γ) stabilizes in at most k steps, where k = max {diam(Gp)p ∈ P}. When γ = ⊤, the opposite direction also holds, and one can obtain a characterization of the notion of stabilization for BDPs based on properties of the initial opinion profile and of the delegation graph. Theorem 4. Fix a BA structure A = hN, P,γi, a profile O of consistent (w.r.t. γ) binary opinions, and a delegation graph G, and let γ = ⊤. Then the following statements are equivalent: 1. The BDP (for O and G) stabilizes. 2. For all p ∈ P, there is no set of agents S ⊆ N such that: S is a cycle in Gp and there are two agents i, j ∈ S such that Oi(p) 6= O j(p). A special case of Theorem 4 is the case in which Gp contains no cycle of length ≥ 2. In such case, a direct consequence of the theorem is that the process stabilizes from any profile. This is also a corollary of a known stabilization result for DeGroot processes (cf. [16, p.233]). 19Other types of dynamics are of course possible. A recent systematic investigation of opinion diffusion on logically inter- dependent issues is [5]. For a broader study of Boolean DeGroot processes in the context of models of binary opinion diffusion on networks we refer the reader to [6]. 146 Binary Voting with Delegable Proxy: An Analysis of Liquid Democracy 5 Conclusions The paper has shown how delegable proxy voting (liquid democracy) can be understood as an aggregator within the theory of binary aggregation with abstentions, for which we provided a novel characterization theorem of issue-wise majority (Theorem 1). This has allowed us to clarify the impact of cyclical del- egations on individual and collective abstentions (Proposition 2) and to suggest alternative aggregators requiring individuals to reveal a default opinion, which can be shown to better behave in the presence of delegation cycles (Proposition 3). Finally we showed how delegable proxy interferes with individual rationality, a standard tenet of social choice theory. Also in this case we showed how liquid democ- racy could be adjusted-in the form of a stabilizing diffusion process-in order to preserve individual rationality (Theorem 3). Proofs Proof of Proposition 1. If the agenda ±P is simple, then all minimally inconsistent sets have cardinality 2, that is, are of the form {ϕ, ¬ψ} such that ϕ = ¬ψ for ϕ,ψ ∈ P. W.l.o.g. assume ϕ = pi and ψ = p j. Suppose towards a contradiction that there exists a profile O such that maj(O) is inconsistent, that is, maj(O)(pi) = maj(O)(p j) = 1, and ϕ = ¬ψ. By the definition of maj (1) it follows that O(pi) > O(¬pi) and O(p j) > O(¬p j). Since pi = ¬p j by assumption, and since individual opinions are consistent and closed, O(¬p j) ≥ O(pi) and O(¬pi) ≥ O(p j). From the fact that O(pi) > O(¬pi) we can thus conclude that O(¬p j) > O(p j). Contradiction. Proof of Lemma 1. We establish the claim through a series of equivalences. Observe first of all that a uniform and symmetric quota rule F is such that (a) F = arg minG u(G)(p), for all p ∈ P if and only if, (b) for any O ∈ O ∗ c and p ∈ P, u(O)(p) = ∗ if and only if O(p) = O(¬p), that is, an even number of voters vote and the group is split in half. Now, (b) is the case if and only if, (c) the quota of F are set in such a way that ⌈q1(p)O(±p)⌉ = ⌈q0(p)O(±p)⌉ =l O(±p)+1 (c) is the case if and only if, (d) the quota of F are set as 1 defining maj (Example 1). 2 m for any profile O and issue p. In turn 2N , which are the quota 2 < q1(p) = q0(p) ≤ N+1 Proof of Theorem 1. Claim 1 Left-to-right: Easily checked. Right-to-left: Let F be an anonymous, independent, monotonic, and responsive aggregator. By anonymity and independence, for any p ∈ P, and any O ∈ O ∗ c , the only information determining the value of F(O)(p) are the integers O(p) and O(¬p). By responsiveness, there exists a non-empty set of profiles S1 = {O ∈ O ∗F(O)(p) = 1}. Pick O to be any profile in S1 with a minimal value of O(p) O(±p) and call this value q1. Now let O′ be any profile such that O′ =−i O and O′(p) i(p) = 1. By monotonicity, it follows that F(O′)(p) = 1. By iterating this argument a finite number of times we conclude that O(p) O(±p) ≥ q1, we have that F(O)(p) = 1. Given that q1 was defined as a minimal value, we whenever conclude also that if F(O)(p) = 1, then O(p) O′(±p) > q1. This implies that Oi(p) = 0 and O′ O(p±) ≥ q1. The argument for q0 is identical. Claims 2 & 3 follow straightforwardly from the definitions of uniform quota rule (Definition 1) and of neutrality (Definition 3) and, respectively, from the definitions of symmetric quota rules (Definition 1) and of unbiasedness (Definition 3) . Claim 4 Left-to-right. Recall that maj is defined by quota 1 N (Example 1). It is clear that maj is uniform and symmetric. The claim then follows by Lemma 1. Right-to-left. By Lemma 2 < q1 = q0 ≤ 1 2 + 1 Z. Christoff & D. Grossi 147 1 if an aggregator minimizes undecisiveness then its quota are set as 1 define maj (Example 1). 2 < q1 = q0 ≤ 1 2 + 1 N . These quota Proof of Proposition 2. The claim amounts to computing the probability that a random proxy profile O induces a delegation graph Rp that does not contain gurus (or equivalently, whose endomap rp : N → N has no fixpoints) as n tends to infinity. Now, for each agent i, the number of possible opinions on a given issue p (that is, functions O : {p} → {0, 1} ∪ N) is (N\ {i}) ∪ {0, 1} = n + 1 (recall i cannot express "i" as an opinion). The number of opinions in which i is delegating her vote is n − 1. So, the probability that a random opinion of i about p is an opinion delegating i's vote is n−1 n+1 . Hence the probability that a random profile consists only of delegated votes (no gurus), for a fixed issue, is ( n−1 n+1 )n. The claimed value is then established through this series of equations: lim n + 1(cid:19)n n→∞(cid:18) n − 1 = lim = lim n+2 n + 2(cid:19)n n→∞(cid:18) n n !n n→∞ 1 n!n n→∞ 1 n→∞ 1 n )n! 1 + 2 (1 + 2 1 n )n limn→∞(1 + 2 1 e2 = lim = lim = = This completes the proof. Proof of Proposition 3. The claim amounts to computing the probability that a random proxy profile with default opinions O induces a delegation graph Rp (equivalently, an endomap rp : N → N) whose cycles are all hung majorities, that is, whose cycles are all even and exactly half the agents in each cycle accept p. As opinion with defaults consist of both a value x ∈ {0, 1} and a trustee i ∈ N we can treat the probability of each component as independent: the number of all possible proxy profiles with default opinions is, therefore, 2n · nn. First of all, recall that a delegation graph can be represented as a set of trees whose roots are nodes in a cycle, that is, as trees whose roots are elements of a permutation of a subset of N. The number of ways of arranging n elements in trees rooted on m elements (with m > n ≥ 1) is given by the following recursive function (cf. [21]): f (n, m) =(cid:18) n m(cid:19) ∑ 0≤k≤n−m mk f (n − m, k) (9) with f (0, 0) = 1 and f (n, 0) = 0 for any n > 0. So the number nn of all possible delegation graphs equals f (n, k)k! ∑ 1≤k≤n (10) that is, the number of ways of arranging n elements in trees rooted on a permutation of a subset of n (recall that k! is the number of all possible permutations of k elements). Now to obtain the number of 148 Binary Voting with Delegable Proxy: An Analysis of Liquid Democracy ways of arranging n elements in trees rooted on even cycles, each of which is a hung majority we adapt (10) as follows. First we establish the number of delegation graphs (for a given issue) which contain only even cycles, that is: ∑ k≤n and even f (n, k) k! 2k(cid:18)k 2(cid:19) k (11) If each addendum of the above expression is multiplied by 2k, that is the number of possible opinions on p of k agents, one obtains the number of possible proxy profiles with default that determine a delegation graph with only even cycles, with all the possible assignments of opinions x ∈ {0, 1} for the agents in the permutation on which the trees of the graph are rooted: We can then adapt (12) by restricting the subprofiles of opinions of the k agents to hung majorities (i.e., ∑ k≤n and even f (n, k)k!(cid:18)k 2(cid:19) k (12) k (cid:0) k 2(cid:1)). We thus obtain the following value: ∑ k≤n and even f (n, k) k! 2(cid:19)2 2k(cid:18)k k (13) Under the impartial culture assumption, the probability of a proxy profile with default opinions to induce only even cycles with hung majorities is therefore (13) divided by 2n · nn. This quantity approaches 0 as n tends to infinity. Proof of Theorem 3. Assume that for all p ∈ P, for all S ⊆ N such that S is a cycle in Gp, for all i, j ∈ S: Oi(p) = O j(p). Consider an arbitrary i ∈ N. Let ki(p) be the distance from i to the closest agent in a cycle of Gp, and let ki denote max{ki(p)p ∈ P}. We show that for any ki ∈ N, Oki is an opinion which i will not change at any later stage (stable). • If ki = 0: i is its only infuencer, therefore O0 i is stable by assumption. • If ki = n + 1: Assume (IH) that for all agents j such that k j = n, O k j j is stable. This implies that all influencers of i are stable. There are two cases: never be, and therefore On 1. If {γ} ∪np ∈ P ORp(i)(p) = 1o ∪n¬p ∈ P ORp(i)(p) = 0o is not consistent, then it will 2. If {γ} ∪np ∈ P ORp(i)(p) = 1o ∪n¬p ∈ P ORp(i)(p) = 0o is consistent, then for each p, i is already stable. is therefore (by IH) stable. On+1 (p) = O i k j j (p), and On+1 i It follows that after k steps, with k = max {diam(Gp)p ∈ P}, each agent's opinion is stable, and the BDP has therefore stabilized. Proof of Theorem 4. 1) ⇒ 2) We proceed by contraposition. Let p ∈ P, S ⊆ N be a cycle in Gp, i, j ∈ S, and Oi(p) 6= O j(p). Let k be the length of the cycle and d be the distance from i to j. Then Oi(p) will enter a loop of size k: for all x ∈ N, Oxk (p). Therefore, the BDP does not stabilize. 2) ⇒ 1) Assume S ⊆ N be such that S is a cycle in Gp, and for all i, j ∈ S, Oi(p) = O j(p). Then, for all j ∈ S, and all x ∈ N, Ox j(p) = Oi(p) and for all k ∈ N\S with distance d from to i, for all x ∈ N, such that x ≥ d, Ox k(p) = Oi(p). Therefore, the BDP stabilizes. i (p) 6= Oxk+d i Z. Christoff & D. Grossi References 149 [1] Dan Alger (2006): Voting by proxy. Public Choice 126(1-2), pp. 1–26, doi:10.1007/s11127-006-3059-1. [2] Jan Behrens (2015): Cyclical Delegations: Myths or Disaster. Liquid Democracy Journal 3. Available at http://www.liquid-democracy-journal.org/issue/3/The_Liquid_Democracy_Journal-Issue003-02-Circular_Delegations_-_Myth_or_Disaster.html. [3] Jan Behrens, Axel Kistner, Andreas Nitsche & Bjorn Swierczek (2014): Principles of Liquid Feedback. Interaktieve Demokratie. [4] Paolo Boldi, Francesco Bonchi, Carlos Castillo & Sebastiano Vigna (2011): Viscous democracy for social networks. Communicationsofthe ACM 54(6), p. 129, doi:10.1145/1953122.1953154. [5] Sirin Botan (2016): Propositional Opinion Diffusion with Constraints. ILLCMaster of LogicThesis. Avail- able at https://staff.science.uva.nl/s.botan/Botan_MoLThesis.pdf. [6] Zo´e Christoff & Davide Grossi: Stability in Binary Opinion Diffusion. To appear in the Proceedings of LORI'17. [7] Zo´e Christoff & Davide Grossi (2016): Liquid Democracy: An Analysis in Binary Aggregation and Diffusion. Technical Report, University of Liverpool. Available at https://arxiv.org/abs/1612.08048. [8] Morris H. DeGroot (1974): Reaching a Consensus. Journalofthe AmericanStatistical Association 69(345), pp. 118–121, doi:10.1080/01621459.1974.10480137. [9] Franz Dietrich & Christian List (2007): Judgment aggregation without full rationality. Social Choice and Welfare 31(1), pp. 15–39, doi:10.1007/s00355-007-0260-1. [10] Charles L. Dodgson (1884): The Principles of Parliamentary Representation. Harrison and Sons. [11] Elad Dokow & Ron Holzman (2010): Aggregation of binary evaluations with abstentions. Journal of Eco- nomicTheory 145(2), pp. 544–561, doi:10.1016/j.jet.2009.10.015. [12] Ulle Endriss (2016): In F. Brandt, V. Conitzer, U. Endriss, J. Lang & A. D. Procaccia, editors: Handbook of Computational Social Choice, Cambridge University Press, doi:10.1017/CBO9781107446984.018. Judgment Aggregation. [13] Umberto Grandi & Ulle Endriss (2013): Lifting integrity constraints in binary aggregation. Artif.Intell. 199, pp. 45–66, doi:10.1016/j.artint.2013.05.001. [14] James Green-Armytage (2014): Direct voting and proxy voting. Const Polit Econ 26(2), pp. 190–220, doi:10.1007/s10602-014-9176-9. [15] Davide Grossi & Gabriella Pigozzi (2014): sis Lectures on Artificial doi:10.2200/s00559ed1v01y201312aim027. Synthe- Intelligence and Machine Learning, Morgan & Claypool Publishers, Judgment Aggregation: A Primer. [16] Matthew O. Jackson (2008): Social and Economic Networks. Princeton University Press, Princeton, NJ, USA. [17] Christoph Kling, J´erome Kunegis, Heinrich Hartmann, Markus Strohmaier & Steffen Staab (2015): Voting Behaviour and Power in Online Democracy: A Study of LiquidFeedback in Germany's Pirate Party. In: Proceedings of the International Conference on Weblogs and Social Media. Available at https://arxiv.org/abs/1503.07723. [18] Christian List (2010): Group Communication and the Transformation of Judgments: An Impossibility Result. Journalof Political Philosophy 19(1), pp. 1–27, doi:10.1111/j.1467-9760.2010.00369.x. [19] Kenneth O. May (1952): A Set of Independent Necessary and Sufficient Conditions for Simple Majority Decision. Econometrica 20(4), p. 680, doi:10.2307/1907651. [20] James C. Miller (1969): A program for direct and proxy voting in the legislative process. Public Choice 7-7(1), pp. 107–113, doi:10.1007/bf01718736. [21] Paul Purdom & James Williams (1968): Cycle Length in a Random Function. Transactionsof the American MathematicalSociety 133(2), pp. 547–551, doi:10.2307/1994996. 150 Binary Voting with Delegable Proxy: An Analysis of Liquid Democracy [22] Piotr Skowron, Martin Lackner, Markus Brill, Dominik Peters & Edith Elkind (2016): Proportional Rank- ings. Technical Report, University of Oxford. Available at https://arxiv.org/abs/1612.01434. [23] Gordon Tullock (1992): Computerizing politics. MathematicalandComputerModelling 16(8-9), pp. 59–65, doi:10.1016/0895-7177(92)90087-2. [24] Hiroshi Yamakawa, Motohiro Yoshida & Michiko Tsuchiya (2007): Toward Delegated Democracy: International Journal of Human and Social Sciences 1(2), Vote by Yourself, or Trust your Network. doi:10.1.1.193.4169.
1109.1879	1	1109	2011-09-09T02:39:43	A Real-time Localization System Using RFID for Visually Impaired	[ "cs.MA", "cs.NI" ]	Gadgets helping the disabled, especially blind that are in least accessibility of information, use acoustic methods that can cause stress to ear and infringe user's privacy. Even if some project uses embedded Radio Frequency Identification (RFID) into the sidewalk for blind's free walking, the tag memory design is not specified for buildings and road conditions. This paper suggested allocation scheme of RFID tag referring to EPCglobal SGLN, tactile method for conveying information, and use of lithium battery as power source with solar cells as an alternative. Results have shown independent mobility, accidents prevention, stress relief and satisfied factors in terms of cost and human usability.	cs.MA	cs	A Real-time Localization System Using RFID for Visually Impaired Qinghui T.1, Malik M.Y.2, Youngjee H.1, Jinwoo P.1 Dept. of Industrial Engineering and ASRI1, Dept. of EECS2 Seoul National University, Seoul, Korea Abstract Gadgets helping the disabled, especially blind that are in least accessibility of information, use acoustic methods that can cause stress to ear and infringe user ‟s privacy. Even if some project uses embedded Radio Frequency Identification (RFID) into the sidewalk for blind‟s free walking, the tag memory design is not specified for buildings and road conditions. This paper suggested allocation scheme of RFID tag referring to EPCglobal SGLN, tactile method for conveying information, and use of lithium battery as power source with solar cells as an alternative. Results have shown independent mobility, accidents prevention, stress relief and satisfied factors in terms of cost and human usability. Keywords Radio Frequency Identification (RFID), smart white cane, digital Braille, tag design, EPCglobal 1. Introduction This is the era of information technology and everyone has access to information. Lives of human beings are facilitated and they can reaped the benefits of cutting edge (overuse the same words). However, we compare the technical resources available for visually impaired, still „digital divide‟ between their lives and that of the others can be sensed. Specifically, for the blind ‟s walking independently on the road, many researchers have suggested the white canes with audio system and it is most state of the art. But continuously hearing sound output or wearing a headset can be irritating to the system users. Therefore, this resear ch proposed a serialized RFID guide system in real time. We referred to the EPCglobal and suggested information allocation scheme of RFID tag. EPCglobal is a global standards system that combines RFID technology, existing communications network infrastruct ure, and the Electronic Product Code [1]. EPCglobal is to facilitate the exchange of information and object between trading partners in supply chain, so the identification, data capture, and data exchange are the needs. However our suggestion is related to service engineering of social infrastructure, we focus on just identification and data capture with offline concept. Especially information elements of SGLN (Serialized Global Location Number) give a hint in tag scheme. Users could get „where I am‟ from the information of location, roads, and nearby buildings in tactile way, therefore confidently mobility, avoiding ear stress, and guaranteeing privacy are given. 2. Related Works D‟Atri, E. et al. introduced a RFID cane reader and PDA system. PDA communicated with the RFID cane for location information and user has to carry PDA [2]. Ortigosa, N. et al. presented mobility assistance aimed to sonicate to the user the presence of obstacles and free path by using an optical laser distance meter, a stereo-camera and GPS system [3]. Another localization system for blind using GPS network is introduced by S. Chumkamon et al [4]. Main focus of their work is in helping the blind to get know of their location in indoor environments. Yuriko Shiizu et al proposed an indoor guidance arrangement relying on different colors of lines for navigation [5]. This scheme also incorporates a smart cane in which the sensors for color recognition are embedded. Guidance voice allows the user to know about his location in this proposed scheme. For these proposals, user always has to carry some devices along with the white cane and information is transmitted vocally. Additionally, power resources are very limited in these systems. 3. Proposed System Configuration In our paper we considered the use of tactile signals for giving information rather than acoustic signals, and contributed better power source for the operation of the system. Our system is based on RFID passive rags, reader, control unit, Braille display and lithium and solar power source. A reader incorporated in white cane reads the information from the passive tags, then retrieves data from memory and transfers this information to the Braille display so that the visually impaired can read it. Passive RFID tags can be installed on roads, Braille blocks, street lamps, buildings or other signs and location indication boards. Reader can read the information from tag, analyze the data and pass the information to the control unit. Control unit helps in presenting the useful information to the blind people through Braille display. Figure 1: System configuration 4. RFID System EPC Global develops industry-driven standards for the Electronic Product Code™ (EPC) . Most of the tag formations are being designed using EPC Global GEN2 standard. These EPC tag forms are simple and inexpensive. Our proposed system is cost effective and easy to implement because of use of this standard to define our RFID tags. 4.1. RFID Tag Definition This research used EPC Global Serialized Global Location Number-96 (SGLN-96) formation. SGLN-96 contains header, filter value, partition, company prefix, location reference, and extension component fields [6]. Table 1: Company prefix, location reference and extension component allocation 1. Company Prefix (20 bits) 2. Location Reference (21 bits) 3. Extension Component(41bits) Company Direction code 18 bits code 2 bits Main road Sub road Path Building Direction Road Serial number condition number 6 bits 7 bits 8 bits 10 bits 2 bits 5 bits 23 bits As shown in Table 1, we divided company prefix into two parts; company code and direction code. We can define nearly 260 thousand building names using 18 bits of company code. 2 bits of direction code are used to represent relative direction of tag and building. The code values of right and left directions are 00 and 11 respectively, and that of forward and backward direction are 10 and 01. This formation will facilitate figuring of direction in the protocol. Generally, the address definition of building is based on the road layouts . We also used this method to standardize location reference code in SGLN-96. We categorized roads into three types: main road, sub road and path. 6, 7 and 8 bits are respectively allotted to these road categories. Up to 2 million pathways can be defined in the tag code. In extension component part, the first bit‟s value is 1. We allocated next 10 bits to building numbers. Building numbers are allotted from the start of the road, with odd numbers assigned to the left side and even numbers assigned to the right side, as shown in Figure 2. We used 2 bits to indicate the direction of road conditions and the other 5 bits to mark that road condition. Road conditions may refer to several detailed information to help blind pedestrians, like building entrance, stairs, pedestrian crossings, traffic light, turn etc. Remnant 23 bits are used for defining passive tag serial number. These serial numbers increase with respect to the right-hand side of the road. [4]and assume this direction to be positive, 4.2. RFID Tag Layout Design Figure 2 shows a general RFID tag infrastructure layout. The passive RFID tags covered by the protective shield are attached to the sidewalks. When the tag infrastructure is constructed, the tag serial number increases from the start of road to the end which is attached in the road. Although the distance between two tags should be according to the road conditions, we proposed it to be near 8 m. In case of dense road conditions, tags can be positioned closer in order to provide the blind with more information. Figure 2: Real road condition and tag layout 4.3. RFID Tag Analysis Protocol Now we will illustrate our suggested tag protocol. Information of the four tags shown in Figure 2 is shown in Table 2. When a pedestrian walks on Y sub road from left-hand side to right-hand side, the tags are read sequentially and data from them is stored in memory stack in reader. The received data comply with FIFO (first in, first out) rule in the queue. The simplified data analysis algorithm is shown in Figure 3. Table 2: Tag sample Tag code (0 = Null) M shop right 0 Y sub road 0 No 2 forward 0 0030 M shop N company right 0 Y sub road 0 No 4 right entrance 0060 N company N company right 0 Y sub road 0 No 4 forward crosswalk 0070 N company W office right 0 Y sub road 0 No 6 back crosswalk 0071 W office Tag A B C D When the reader came in contact with installed passive opposite values by NOT operation. tag, the tag is verified and its data is stored in reader ‟s memory stack. Tag‟s serial number is compared with the previous entry of stack. As defined earlier, tag serial number is increasing on the right side of the road. If the pedestrian is walking in the positive direction, serial numbers read by the reader increases in value . The reader retrieves information and sends message to the pedestrian. When the pedestrian walks from negative direction, the system can judge the pedestrian‟s direction by decrement in serial numbers. In this case, the system would convert the company prefix and extension component directions in Table 1 to their Figure 3: RFID tag analysis algorithm It solved bidirectional confusion which can be found in previous systems. When tag B is registered by the reader moving from left-hand side to right-hand side, the pedestrian is given information about the entrance of N company and its direction (right-hand side). On the contrary, pedestrian is told that entrance of N company is on left -hand side. 5. Proposed Smart Cane This paper proposed a smart cane, which is responsible for tag reading operation and communicating the data to the user. We have underlined some of the aspects that can enable effective design of the cane. As the smart white cane will perform two significant operations, there is a need of efficient power source for its operation. 5.1 RFID Reader Class 1 (Gen 2) UHF based reader can be deployed in smart cane, because of its large range of reading. They are able to read passive tags within 10 m. Working at much higher frequencies will enable design of small antennas which can easily be embedded in pattern of the PCB. This will help in keeping the reader module size to minimum and can be easily incorporated in the white canes with minor modifications. 5.2 Information display part as Braille part Smart cane is able to communicate the information received by the reader to the user. It uses Braille for transferring the information to the user. User can read the data presented by smart cane conveniently. Braille is the best known communication system for blind people, developed by Louise Braille. In this system characters are exhibited in a pattern of six (or more) dots having two columns of three dots, known as Braille cell. Such cell is shown in Figure 4. A combination of six dots can only have 64 different configurations. For many languages, basic communication can be performed by the use of six dot cells. There are many standards available for Braille cell dimensions, dot size, distance between lines and many other parame ters. Figure 4: Braille cell and representation of English alphabets by Braille cells [8,9] 5.3 Braille Unit For conveying information to the user, Braille display on smart cane uses two lines for displaying the characters. First line will give indication of nearby building, road, street or any other important location around. The second line will provide information about any specific place around that location , like the presence of any pedestrian crossing, community services (elevator, washrooms, etc) and telephone booth. Directions of these places are also shown on the second line. The width of the smart cane at the handling side can be from 110-120 mm. This size is appropriate as it is easy to handle. In common Braille systems dot height is approximately 0.5 mm; the horizontal and vertical spacing between dot centers within a cell is approximately 2.5 mm; the blank space between dots on adjacent cells is 3.75 mm horizontally [10]. If we consider using the common Braille system for displaying data on smart cane nearly 12 characters can be easily shown on a single line, which are enough for conveying basic information. By using two lines we can easily inform the user about the location around him and places of importance and interest surrounding him. 5.4 Control unit between reader and Braille display Our proposed Braille display works on selection of dots and electromechanical switching. Any dot in display system can be chosen to be raised to represent a dot or can be lowered to act as an empty place. In our system reader gets information about location from the passive tags installed and transfers it to a control unit. This control unit is responsible for the mechanism of displaying message on Braille display. The control unit consists of a microcontroller unit (MCU) which can be programmed to facilitate selection of dots using decoders and buffer based on the commands by the MCU. Electrical signals from decoders and buffers can operate electromechanical switches which can raise or lower metallic or plastic pins. By using this simple layout, we can represent data on our Braille display. Figure 5: Outline of the system in smart cane 5.5 RF reader and Braille display power source Choice of the power system for different units in smart cane is very crucial. Our system uses Lithium ion batteries (LIB) as the power source for the RF reader and control unit. They have best energy-to-weight ratios, no memory effect, and a slow loss of charge when not in use. LIB can be recharged easily which will help in minimizing the required maintenance. When the ubiquitous system of the smart cane is not being used, it can be switched off. The ubiquitous activity of smart cane will primarily be used outdoors, which makes it possible to utilize solar energy to charge another cell battery to function like uninterruptible power supply. When the LIB is near depletion, low battery alarm will be given to the user and reader and Braille display power source will be changed from LIB to solar cell battery. Like the cellular phone, LIB can be charged easily and conveniently. 6. Results We evaluate our prototype comparing with previous guide systems based on three factors of context of use including acceptability, usability, and cost benefit. The acceptability tells whether the product will actually be used in real life. The usability represents the user-friendliness of a system. The cost benefit compares unit cost of previous research with ours [12]. For the acceptability factor, our system requires little provision of training to support initial usage and adaptive activities. Setting the RFID tag and power source inside cane, and attaching tactile brail and solar cell outside cane are only different from original white cane. For implementation, installing RFID on the road block, just power on and getting tactile messages are enough. These satisfy the usability factor. Most of the assistance prototypes are equipped with GPS, sensors, or portable nav igation with headset. Those are normally twice expensive than ours just consisting of MCU, lithium battery, and tactile brail. These satisfy the cost benefit factor. 7. Conclusion and future work This research presents tag design and smart white cane for ease in reading information without extra devices. Short evaluation is done according to context of use which considers main user, task and environmental characteristics of the situation in which it will be operated [12]. The prototype is easy to learn, use, and buy or be subsidized from public service, so we concluded the acceptability, usability, and cost benefit are satisfied. Also huge decision opportunities and visibility for the blind in working are given. F uture work will be network environment and compatibility with infrastructure and current standard. Extending network environment to system can enable range to be not only high-value applications of walking decisions, but also tracking risk situations which call for help through the network. Road or building condition changes could be easily updated by the server. Compatibility with existing infrastructure and current standard such as EPCglobal tag scheme are needed for smooth application. References 1. Eeghem, M.V., 2008, “Basics of EPC: Student‟s Handbook v2.0”, 9-12. 2. D‟Atri, E., Medaglia, C.M., Serbanati, A., Ceipidor, U.B., Panizzi, E., D‟Atri A., 2007, “A System to Aid Blind People in the Mobility: A Usability Test and its Results,” 2nd International Conference on Systems. 3. Ortigosa, N., Dunai, L., and et al., 2009, “A Multiple Sensors Prototype for Visually Impaired Subjects Mobility Assistance Using Sound Map,” LivingAll European Conference, January 15 -16, Valencia, Spain. 4. Sakmongkon Chumkamon, Peranitti Tuvaphanthaphiphat, Phongsak Keeratiwintakorn „A Blind Navigation System Using RFID for Indoor Environments‟ Proceedings of ECTI-CON 2008 5. Yuriko Shiizu, Yoshiaki Hirahara, Kenji Yanashima and Kazushige Magatani. „The development of a white cane which navigates the visually impaired‟ Proceedings of the 29th Annual International Conference of the IEEE EMBS. Cité Internationale, Lyon, France August 23-26, 2007. 6. EPC global Inc, 2008, “Tag Data Standard v. 1.4”, 39 -44. 7. Ha, S., 2005, “Traffic signal and Road/road-surface signpost system of base RFID” Proc. of 1th RFID/USN research paper contest, 2005, Seoul, Korea. 8. http://www.indiana.edu/~iuadapts/services/braille.html 9. http://louisbraillebiography.com/ 10. http://www.tiresias.org/research/standards/braille.htm 11. Watanabee, T. & Oouchi, S. (2003) A study of legible braille patterns on capsule paper: diameters of braille dots and their interspaces on original ink-printed paper 12. Maguire, M., 2001, “Context of Us within usability activities”, International Journal Human-Computer Studies, 55, 453-483. APPENDIX A Table A1 shows the bit allocation in EPC SGLN-96 standard by EPCglobal. Table A1: EPC SGLN-96 bit allocation Header Filter Partition Company Location Extension SGLN-96 8 Value 3 Prefix 20-40 3 Reference Component 21-1 41 Table A2 represents some of the Braille standards being used in the world. Table A2: Some Braille standards Horizontal Vertical dot Cell to cell Line to line Dot base Dot height dot to dot to dot (mm) (mm) diameter (mm) (mm) (mm) Electronic Braille 2.4 2.4 French German Small English 2.5 - 2.6 2.5 - 2.6 2.5 2.03 2.5 2.03 6.4 6.0 5.38 >10 10.0 8.46 (mm) 1.2 1.3 - 1.6 1.4 - 1.5 0.8 0.8 - 1.0 ≥0.5 0.33
1803.03491	1	1803	2018-03-09T12:48:03	Valuing knowledge, information and agency in Multi-agent Reinforcement Learning: a case study in smart buildings	[ "cs.MA", "cs.AI", "stat.AP", "stat.ML" ]	Increasing energy efficiency in buildings can reduce costs and emissions substantially. Historically, this has been treated as a local, or single-agent, optimization problem. However, many buildings utilize the same types of thermal equipment e.g. electric heaters and hot water vessels. During operation, occupants in these buildings interact with the equipment differently thereby driving them to diverse regions in the state-space. Reinforcement learning agents can learn from these interactions, recorded as sensor data, to optimize the overall energy efficiency. However, if these agents operate individually at a household level, they can not exploit the replicated structure in the problem. In this paper, we demonstrate that this problem can indeed benefit from multi-agent collaboration by making use of targeted exploration of the state-space allowing for better generalization. We also investigate trade-offs between integrating human knowledge and additional sensors. Results show that savings of over 40% are possible with collaborative multi-agent systems making use of either expert knowledge or additional sensors with no loss of occupant comfort. We find that such multi-agent systems comfortably outperform comparable single agent systems.	cs.MA	cs	Valuing knowledge, information and agency in Multi-agent Reinforcement Learning: a case study in smart buildings Hussain Kazmi KU Leuven and Enervalis Leuven, Belgium Johan Suykens KU Leuven Leuven, Belgium Johan Driesen KU Leuven Leuven, Belgium 8 1 0 2 r a M 9 ] A M . s c [ 1 v 1 9 4 3 0 . 3 0 8 1 : v i X r a ABSTRACT Increasing energy efficiency in buildings can reduce costs and emis- sions substantially. Historically, this has been treated as a local, or single-agent, optimization problem. However, many buildings utilize the same types of thermal equipment e.g. electric heaters and hot water vessels. During operation, occupants in these buildings interact with the equipment differently thereby driving them to diverse regions in the state-space. Reinforcement learning agents can learn from these interactions, recorded as sensor data, to opti- mize the overall energy efficiency. However, if these agents operate individually at a household level, they can not exploit the replicated structure in the problem. In this paper, we demonstrate that this problem can indeed benefit from multi-agent collaboration by mak- ing use of targeted exploration of the state-space allowing for better generalization. We also investigate trade-offs between integrating human knowledge and additional sensors. Results show that sav- ings of over 40% are possible with collaborative multi-agent systems making use of either expert knowledge or additional sensors with no loss of occupant comfort. We find that such multi-agent systems comfortably outperform comparable single agent systems. KEYWORDS Multi-agent reinforcement learning; targeted exploration; energy efficiency, smart buildings; domain knowledge; sensor information 1 INTRODUCTION The theoretical allure of reinforcement learning (RL) as an end to end black box method is obvious. By translating sensory input directly into meaningful control actions, robust optimal systems can be developed in a cost-effective way [3], [4], [6]. In practice however, numerous trade-offs have to be made between quality of control and the cost associated with accomplishing it. These include (1) the extent of information available via sensors, (2) the level and ease of integration of prior human knowledge and (3) the possibility of deploying multiple agents to accelerate learning. These trade-offs are embodied in smart buildings and smart grids, where agents are deployed to perform automated optimal control. The objective for control can vary from case to case but two common ones are to minimize overall energy consumption [2] and peak power consumption [1], while maintaining predefined user comfort bounds. Reducing energy consumption is a local objective where multiple agents act independently, since energy consumed Proc. of the 17th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2018), M. Dastani, G. Sukthankar, E. Andre, S. Koenig (eds.), July 2018, Stockholm, Sweden © 2018 International Foundation for Autonomous Agents and Multiagent Systems (www.ifaamas.org). All rights reserved. https://doi.org/doi in one household does not affect another. Peak shaving, on the other hand, is a multi-agent problem where different agents have to coordinate their energy consumption to reduce simultaneous demand. In this paper we focus on optimizing energy consumption for hot water production, a load that is responsible for well over 10% of the total energy consumed in modern residential buildings [5]. We show that while each agent can act independently to optimize its load as explained in existing literature [2], [8], a coordination mechanism to improve state-space exploration can substantially improve overall efficiency. This collaboration makes use of the insights developed in [7]. Optimizing the hot water system in such settings can be consid- ered an n-player finite, non-zero sum game of hidden information. Here, n-player refers to the fact that individual agents are operating in multiple houses in parallel to optimize their respective rewards. The overall problem is non-zero sum since an agent's strategy does not directly affect other agents or their rewards. Hidden information refers to the fact that in most hot water systems sensing is limited to only a single temperature sensor which is not representative of the system. The framework for optimizing hot water production presented next explores the trade-offs in RL mentioned at the beginning of the paper. Concretely, we investigate and compare the quality of end-to-end control learned using RL for hot water production in smart building communities which employ the same thermal equip- ment (e.g. in large apartment blocks and social houses etc.). To quantify the trade-offs highlighted earlier, we do this for different configurations using additional sensing, human domain knowledge and multiple agents. 2 METHODOLOGY To integrate these three components, we first define a Markov Decision Process (MDP): M = {S, A, T, R}. The structure of the MDP derives from the interactions between a hot water storage vessel, a heating element and the human occupant. The RL agent sends reheat commands to the vessel (via the heating element) following a policy, π, that minimizes energy consumption while maintaining occupant comfort. The control actions, at ϵ A, are thus binary, and the reward stream R(.) that the agent receives is a function of the energy consumed and the impact on occupant comfort. The vessel state, S, is given by a temperature distribution profile and is representative of the energy content in the vessel. The transition function, T(.), defines the next state of the vessel as a function of current state, the agent's action and stochasticity arising from human occupant behavior. The interactions of the agents with the storage vessel are simulated using a model fit to empirical data while occupant behavior is modelled as a stochastic AAMAS'18, July 2018, Stockholm, Sweden Hussain Kazmi, Johan Suykens, and Johan Driesen Figure 1: (a) State-space exploration; (b) MAE for learnt tran- sition model, with different configurations Figure 2: Predicted and observed temperature, snapshot at different time periods, for [top to bottom: Aggregation of RBC agents, SARL(K), MARL(K), MARL(K,I)] time series fit using real world human behavior. The consumption time series are strongly auto- and cross-correlated. To investigate the aforementioned trade-offs, we consider variations involving the following: Information (I): In the default configuration, the storage ves- sel is equipped only with a mid-point temperature sensor. This is not enough to generalize because the temperature distribution is nonlinear and exhibits stratification effects [2]. Additional sensors can facilitate learning of the distribution. Knowledge (K): There are two ways human knowledge can be encoded as prior knowledge for the agents: feature engineering based on available sensor data and constraining the behavior of the vessel model based on thermodynamic laws. This latter can include defining gradients and end point limits on the temperature distribution, its stratification and possible phase inversion. Agency: The structure of the MDP, M, is replicated across all households which share the same thermal equipment. Since dif- ferent agents are driven to different regions of the state-space as a result of different human interactions, learning a shared repre- sentation of the transition function from sensor data can help in generalization by decoupling stochastic human behavior from de- terministic storage vessel behavior. This also makes targeted explo- ration of the state-space a viable alternative to ϵ-greedy strategies. 3 RESULTS AND DISCUSSION Given a reasonably flexible algorithm, the amount of state-space ex- ploration by the agent (if search is suitably diversified) is positively correlated with its generalization. This exploration is illustrated in Fig. 1a with different control strategies which include naïve rule based controllers (RBC) and their aggregation as well as the best single agent reinforcement learner for each household (SARL(K), where K represents domain knowledge), and aggregation of all such agents with targeted exploration (MARL(K)). Finally, we also con- sider both SARL(K) and MARL(K) augmented with extra sensing information (I): SARL(K,I) and MARL(K,I). Figure 3: (a) Energy consumption; (b) Water consumption temperature It is evident from Fig. 1a and 1b that simply increasing the num- ber of agents without adopting a more complex control policy and incorporating domain knowledge does not help the agent explore the state-space (see also Fig. 2). The exploration potential of such strategies also tapers off as the auto-correlated occupant behavior results in the agent visiting the same states repeatedly. This is re- flected in Fig. 1b where the prediction MAE is uniformly high for strategies which explore less. An exception to this is MARL(K,I) which explores the most but has a performance no better than MARL(K); this is because additional sensing provides similar in- formation as human expert knowledge. Fig. 2 reveals that both MARL(K) and MARL(K,I) learn reliable transition functions within two months, a feat that naive aggregation of RBC agents is unable even to after a year. The single agent, SARL(K) configuration suffers from prediction errors mostly at the transition between hot and cold water which has negative implications for the end user. The improvement in learned transition model translates directly into greater rewards for the RL agents over time. This is visualized in Fig. 3a where RL based strategies reduce energy consumption by up to 40%. The savings are highest for the single reinforcement learner with domain knowledge (SARL(K)); however, as mentioned above, these savings come at the cost of reduced occupant comfort (defined as number of hot water draws below 45°C) (Fig. 3b). SARL(k) is the only configuration where this boundary is breached repeatedly as seen in Fig. 3b. This is a direct consequence of learning the incorrect model of the storage vessel and a comparable multi-agent configuration does not suffer from this problem. Valuing knowledge, information and agency in Multi-agent Reinforcement Learning: a case study in smart buildings AAMAS'18, July 2018, Stockholm, Sweden ACKNOWLEDGMENTS This work has been carried out with support from InnoEnergy and VLAIO. REFERENCES [1] Bert Claessens, Stijn Vandael, Frederik Ruelens, Klaas De Craemer, and Bart Beusen. 2013. Peak shaving of a heterogeneous cluster of residential flexibility carriers using reinforcement learning. In Innovative Smart Grid Technologies Europe (ISGT EUROPE), 2013 4th IEEE/PES. IEEE, 1–5. [2] Hussain Kazmi, Simona D'Oca, Chiara Delmastro, Stefan Lodeweyckx, and Ste- fano Paolo Corgnati. 2016. Generalizable occupant-driven optimization model for domestic hot water production in NZEB. Applied Energy 175 (2016), 1–15. [3] Jiwei Li, Will Monroe, Alan Ritter, Michel Galley, Jianfeng Gao, and Dan Jurafsky. 2016. Deep reinforcement learning for dialogue generation. arXiv preprint arXiv:1606.01541 (2016). [4] Volodymyr Mnih, Kavukcuoglu Koray, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller. 2013. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602 (2013). energy consumption information. Energy and buildings 40, 3 (2008), 394–398. [6] Aitor Rovira and Mel Slater. 2017. Reinforcement Learning as a tool to make people move to a specific location in Immersive Virtual Reality. International Journal of Human-Computer Studies 98 (2017), 89. [7] Ming Tan. 1993. Multi-agent reinforcement learning: Independent vs. cooperative agents. In Proceedings of the tenth international conference on machine learning. 330–337. [8] Jiangfeng Zhang and Xiaohua Xia. 2007. Best switching time of hot water cylinder-switched optimal control approach. In Proceedings of AFRICON 2007. 1–7. [5] Luis Pérez-Lombard, José Ortiz, and Christine Pout. 2008. A review on buildings
1001.0115	1	1001	2009-12-31T10:35:55	Developing Artificial Herders Using Jason	[ "cs.MA" ]	This paper gives an overview of a proposed strategy for the "Cows and Herders" scenario given in the Multi-Agent Programming Contest 2009. The strategy is to be implemented using the Jason platform, based on the agent-oriented programming language Agent-Speak. The paper describes the agents, their goals and the strategies they should follow. The basis for the paper and for participating in the contest is a new course given in spring 2009 and our main objective is to show that we are able to implement complex multi-agent systems with the knowledge gained in an introductory course on multi-agent systems.	cs.MA	cs	Developing Artificial Herders Using Jason Niklas Skamriis Boss, Andreas Schmidt Jensen, and Jørgen Villadsen(cid:63) Department of Informatics and Mathematical Modelling Technical University of Denmark Richard Petersens Plads, Building 321, DK-2800 Kongens Lyngby, Denmark Abstract. This paper gives an overview of a proposed strategy for the "Cows and Herders" scenario given in the Multi-Agent Programming Contest 2009. The strategy is to be implemented using the Jason plat- form, based on the agent-oriented programming language AgentSpeak. The paper describes the agents, their goals and the strategies they should follow. The basis for the paper and for participating in the contest is a new course given in spring 2009 and our main objective is to show that we are able to implement complex multi-agent systems with the knowledge gained in an introductory course on multi-agent systems. 1 Introduction This paper describes the work with a multi-agent system consisting of artificial herders attempting to catch cows. The agents will compete in the Multi-Agent Programming Contest 2009 (the scenario "Cows and Herders"). One of our main objectives in the contest has been to gain experience with the development of multi-agent systems using Jason. Our basis for participating in the contest is the course "Artificial Intelligence and Multi-Agent Systems" given in spring 2009 at the Technical University of Denmark. The course provides an introduction to multi-agent systems using Jason as the implementation platform. We hope to show that this introduction is sufficient to be able to implement a more complex multi-agent system, such as the "Cows and Herders" scenario given in the contest. 2 System Analysis and Design Our system consists of three kinds of agents: a herder, a scout and a leader. The leader and the scout are basically herders with extra responsibilities. The scout will initially explore the environment and subsequently act as an ordinary herder. The leader will delegate targets to each of the herders -- including himself. (cid:63) Contact: [email protected] 1 Our system was designed using the Prometheus methodology as a guideline. By this we mean that we have adapted relevant concepts from the methodology, while not following it too strictly (as stated in [3]). It has allowed us to quickly identify the goals and what agents are needed to complete them. Fig. 1. Overview of the system. Figure 1 gives an overview of the system. The diagram distinguishes between the three types of agents, even though the leader and the scout are actually special cases of the herder. This has been done to easily see the different roles each agent plays. All agents know their own position and how many steps of the match that have elapsed. This is used to revise targets, since we do not want them to blindly follow a target. An agent gets a new target by fulfilling the goal get new target. The herders will tell the leader to delegate a target based on the agents current position, while the scout will autonomously decide where to go. We distinguish between the following types of targets. While the agents do not really have an understanding of each concept, it is helpful for us to be able to tell the targets apart. Exploration targets are targets in an area which has yet to be explored. Such target is delegated to the scout, when he has not explored the entire envi- ronment, or a herder, whenever he does not fulfill the criteria for a receiving another type of target. Formation targets are targets behind cows, but within a certain distance from both cows and other herders, so that the group of cows can be controlled and moved (or herded) towards the corral. 2 A switch target is a target next to a switch. The reason for this is that an agent should stand next to a switch in order to trigger it. This target will be delegated whenever an agent is near a closed switch and it is reasonable to open it. This is the case if one or more cows are near the fence or another agent is on one side, while having a target on the other side (thus needing another agent to open the fence, since one agent cannot pass a fence alone). The scenario is quite dynamic since cows are continuously moving and fences can be opened and closed, and all of this must be taken into account. 3 Software Architecture Our strategy and agents are implemented using the Jason platform, which is an implementation of the AgentSpeak language, written in Java. Jason is an effective platform for creating multi-agent systems with a variable number of agents. Combined with internal actions, we have a strong foundation for building a multi-agent system, which not only uses the features of logic programming, but allows us to develop imperative extensions as well. The use of custom architectures in Jason allows us to implement a local simulation, as described in [1]. This eases the testing, as it can be done much faster. As reference implementation we have used an implementation of the 2008 contest made by the authors of Jason. This has helped us getting started, even though the scenario differs in many ways from last year. Our solution to the contest was developed using Eclipse. The implementation will have great focus on the advantages of object oriented programming. This would also ease future expansion of more agents etc. Shared memory could also be modelled by use of references to shared objects used by multiple agents. 4 Agent Team Strategy The agents will be moving around in a partially known environment. At the beginning of a match everything is unknown, except for what lies within the agents' field of view, and as the agents move around they gain knowledge of the environment. The entire map is represented by a graph, where each node in the graph represents a cell in the environment. When objects such as obstacles or cows are discovered etc. the corresponding cell in the graph will be assigned a value of that kind of object. When agents move around they follow paths calculated by our navigation algorithm. We have chosen to represent the environment as a graph, since it makes it is easy to use a graph search algorithms for navigation. The actual paths are calculated using the A* algorithm, which basically is an advanced best-first search as it uses a heuristic to guide the search for optimal paths. A part of our strategy is to try to keep clustered cows together. This means that the agents will have to move around a group of cows to avoid splitting 3 them up. This is ensured be assigning weight to the different cell in the graph. By assigning higher weights to cells occupied by cows and cells adjacent to cows, agents will navigate around a cluster instead of through it. Obstacles are handled slightly different. The algorithm is implemented so that it does not consider cells containing obstacles as valid cells for a path. This ensures that agents do not try to move through obstacles. To optimize the movement of our agents the paths are continuously calcu- lated. This is done since all agents can add new knowledge of obstacles etc. to the graph as they perceive the environment. This ensures that if one agent discovers that a corridor is blocked, then the other agents will try to move around it to get to their target. Experiments have shown that it is more efficient to herd cows in groups. To ensure this the leading agent makes great use of a clustering algorithm. The algorithm works be examining the surroundings of each cow; adjacent cows are grouped together. The strategy for herding the cows will be taken care of by the leading agent. The team leader will coordinate the herding, ensuring that the cows are fleeing the right way and that an agent will open the fence at an appropriate time. Our strategy is mainly towards maximizing our own score. This means that our agents will not try to capture cows already being herded by the opponent deliberately, but it might happen if the leading agent estimates that they are the cows closest to the corral. An agent's beliefs consist of what they perceive and what others tell them to believe. Optimally, we would like that every agent knows the same, i.e. they all have the same beliefs. Unfortunately, since agents can only see a limited area of the environment, this is not directly possible. To ensure that every agent knows the same, any new belief an agent perceives is sent to every other agent. All beliefs are shared immediately, since it does not create much overhead and it is more efficient to share it than consider whether it should be shared. When an agent discovers a static obstacle, every agent should know this, so that their navigation can be adjusted to this new knowledge. If an agent fails to achieve a given goal then we will use the Jason failure handling feature. This is done by implementing a deletion event -!g, which will be executed if a given plan fails [2]. After recovering from a failed plan, we will attempt to reintroduce the goal (+!g) again. 5 Discussion Our strategy is quite dynamic because of our use of path finding and clustering algorithms, which allow the herders to fulfill their goals in any given scenario. However, some of the choices we have made are made on assumptions which may prove to be mistaken when the competition is held. We have decided to have a maximum cluster size (i.e. limit the number of cows in a single cluster), because we believe that the agents may have a hard 4 time herding larger clusters. This may not be true, though, since it could be more efficient to herd as many cows as possible as long as they are clustered. To compute an optimal search it is important to move agents in patterns so that the largest possible area is explored. For example, agents should never move side by side towards the same location, since this would not exploit the full potential of the agents' field of view. Likewise it could prove useful to move agents in patterns that ensures that no cow can remain undetected in the explored area. However, we need to carefully design our algorithms so that they do not take too long time to compute, since the duration of a turn is limited. At the time of writing this article our implementation is complete. However, the contest has been postponed until after the deadline of the article, so we are unable to discuss the results. We have managed to play a single training match against another team, which we won. This match gave us an opportunity to see how our team plays against others. Generally we are quite satisfied with our system, which is able to fulfill the goals of the scenario. Our strategy with a single leader delegating targets lead to a less autonomous approach, but the Jason framework has allowed us to easily implement agents with certain goals and a way to implement plans for handling these goals. 6 Conclusion As discussed our primary strategy will be to maximize our own score rather than prohibiting the opposing team from scoring points. This has been done by optimizing the search for cows and guiding the cows into the corral by using cooperating agents. Likewise all agents will take the positions of the opponents into account when choosing a target. Throughout the project we have considered problems such as navigation, search for objects using multiple start points, clustering, cooperation between agents and multi-agent planning. All planning was implemented using Agent- Speak, while external algorithms such as A*, our clustering algorithm and target delegation were implemented in Java. Despite our limited experience with AgentSpeak and programming intelli- gent multi-agent systems, we have managed to implement a fairly reasonable system, with agents which fulfill the goals of the contest. The ability of Jason to implement custom architectures was a great help during the work. References 1. Rafael H. Bordini, Jomi F. Hubner, and Daniel M. Tralamazza. Developing a Team of Gold Miners Using Jason. Springer-Verlag LNAI 4908, pages 241 -- 245, 2008. 2. Rafael H. Bordini, Michael Wooldridge, and Jomi Fred Hubner. Programming Multi- Agent Systems in AgentSpeak Using Jason. John Wiley & Sons, 2007. 3. Lin Padgham and Michael Winikoff. Developing Intelligent Agent Systems: A Prac- tical Guide. John Wiley & Sons, 2007. 5
1006.5263	1	1006	2010-06-28T04:45:45	Design specifications of the Human Robotic interface for the biomimetic underwater robot "yellow submarine project"	[ "cs.MA", "cs.RO" ]	This paper describes the design of a web based multi agent design for a collision avoidance auto navigation biomimetic submarine for submarine hydroelectricity. The paper describes the nature of the map - topology interface for river bodies and the design of interactive agents for the control of the robotic submarine. The agents are migratory on the web and are designed in XML/html interface with both interactive capabilities and visibility on a map. The paper describes mathematically the user interface and the map definition languages used for the multi agent description	cs.MA	cs	Design specifications of the Human Robotic interface for the biomimetic underwater robot "yellow submarine project" Dr Bheemaiah, Anil K 341/42 17 cross Ideal Homes Township, Rajarajeshwarinagar Bangalore 560098, India [email protected] Abstract: This paper describes the design of a web based multi agent design for a collision avoidance auto navigation biomimetic submarine for submarine hydroelectricity. The paper describes the nature of the map - topology interface for river bodies and the design of interactive agents for the control of the robotic submarine. The agents are migratory on the web and are designed in XML/html interface with both interactive capabilities and visibility on a map. The paper describes mathematically the user interface and the map definition languages used for the multi agent description. Keywords:Human robotic interface, Biomimetic underwater robotics, Distributed Computing. 1 Introduction Multi agents are a distributed approach to the control of robotics in the existence of a front end to the robot for the control and navigation of the robot and for a better exception handling. Agents are used in many domains and fields and is a mature area of software development, predominately web based and web migrating agents, [.Lyle N. Long et al, 3] describes several autonomous vehicle architectures but none with an XML or html web interface, There is much work on independently developed robotic XML languages, like ROBOTML,[ Maxim Makatchev, 1] and [Tingting Fu et al , 2] This paper delineates the specification of the design of the yellow submarine project, without the need for agent autonomous control software. The inspiration drawn from interactive video books leads directly to this application in a single - one to many mapping of agents to robotic navigation problems. 2 Problem Description and Background TheYellowsubmarineproject, [http://www.friendsofwildlife.com/yellowsubmarine/specifications of the yellow submarine design.html] utilizes a biomimetic robot in design that is a thin client that has to be navigated and positioned in the most appropriate location in a river map. The interface for this from the agent communication language layer is a web interface that provides a visual tool displaying the location of the submarines on a map of the river. The map is inherited from Google earth and displays in a much higher resolution the topology of the river area. A drag and click interface is designed to move the robot to the designated coordinates that corresponds to the optimal positioning. Optimal positioning tools are also part of this interface and the subject of another publication.[http://www.friendsofwildlife.com/yellowsubmarine/design specifications of the optimal flow algorithms in the web interface of the biomimetic underwater robotic yellow submarine project.html] This publication delineates the interface to the robotic language layer of the web agent, the design of the web agent and the exception handler. Web interface agent on browser Online navigable map & Exception handler Robotic thin client #1 Autonomous collision avoidance with sensor feedback Robotic Thin Client #2 Autonomous collision avoidance with sensor feedback Robotic Thin Client #3 Autonomous collision avoidance with sensor feedback Fig 1: Schematic of the multi agent design 2.1 Design of the web agent. 2.1.1 The map user interface and click and drag. The map user interface inherits the Google earth interface for the display of the map of the river in rather high resolution with a symbolic representation of the robot defined by a function call in the html code. The coordinates of the submarine are returned and polled every 15 seconds by the GPS_get_coordinates instruction. This is displayed on the map. The robot by another function call can for the event of clicking the robot be dragged to the optimal location from where it is moving and parked by using a menu interface. A map definition language defines the xml markup for the map of the designated robot parking area. The map definition language is a non linear markup indicating in the event:click_on_robot:=get_scale, the scale parameters in the topology that is of significance to the robot parking . Such as MDL: Coordinates_markers MDL:Coordinates_landmarks_passed MDL:Coordinates_lookahead_landmark MDL:Coordinates_flow_obstacles MDL:Flow_obstacles Here the map definition language is a mark up on the map of the flow design algorithms defined in a separate publication for the non linear list of the coordinates of important landmarks, with a look ahead system to aid navigation and location of the robot and the landmarks passed. These land marks are either fed manually on the map or computed from the flow algorithms. The map definition language also stores the vectors as a linear map and marks it in XML on the map of the flows between the landmarks indicated, thus naturally depending on the scale indicated, there would be two or more landmarks or obstacles per flow listed on the MDL. 2.1.2 The interpreter to the robotic language for this interface. The interpreter consists of a menu definition with an event language with event definitions. Event:click_on _robot:=drag robot , place robot. Event : click_on_robot:=compute optimal flow. Fi User- interfa ce Robotic Language Interpreter Thin client hardware and multi sensor fusion, communications and GPS hardware, Auto navigation hardware. g 2 : Design of the three layers of the agent communication model 2.1.3 The exception handler. Exception handling window in the user interface, defined in xml/html in the design of exceptions. 1. Exp_anchor: in case of exp_communication_failure, exp_GPS _failure or exp_sensor/ power failure,exp_ propulsión _failure. 2. Exp: Auto_park: In case of prolonged timeout, auto park in fuel rendezvous terminal 3. Exp_propulsion: in case the robot is unable to anchor. 3 Conclusion and future work This paper is the design of a simple user interface for the positioning and monitoring of submarine robots. It uses the standard web interface and event language for the user interface with and added interpreter for a defined robotic language for the robot that is defined in another publication: Robotic language specifications for the yellow submarine, biomimetic underwater robot alternative hydro electricity project. Future work includes a multi agent interface, improvement of the exception handling and the use of a natural language interface for the robot. References 1. [February 02, 2001] "Human-Robot Interface Using Agents Communicating In An XML-Based Markup Language." By Maxim Makatchev (Department of Manufacturing Engineering and Engineering Management, City University of Hong Kong) and S. K. Tso (Centre for Intelligent Design Automation and Manufacturing, City University of Hong Kong). Pages 270-275 (with 25 references) in Proceedings the Ninth IEEE International Workshop on Robot and Human Interactive Communication [IEEE RO-MAN 2000, Osaka, Japan, September 27-29, 2000]. Piscataway, NJ, USA: IEEE Computer Society, 2. Tingting Fu, Peng Liu, Yigang Wang, Yehua Du, "Integrating Agents and Web Services into Cooperative Design Platform of Vehicle Headlights," snpd, vol. 1, pp.27-32, Eighth ACIS International Conference on Software Engineering, Artificial Intelligence, Networking, and Parallel/Distributed Computing (SNPD 2007), 2007 .Lyle N. Long*, Scott D. Hanford, Oranuj Janrathitikarn, Greg L. Sinsley, and Jodi A. Miller, The Pennsylvania State University, A Review of Intelligent Systems Software for Autonomous Vehicles, Proceedings of the 2007 IEEE Symposium on Computational Intelligence in Security and Defense Applications (CISDA 2007 3.
1807.06103	1	1807	2018-07-16T20:49:12	An agent-based model of an endangered population of the Arctic fox from Mednyi Island	[ "cs.MA", "cs.AI", "q-bio.PE" ]	Artificial Intelligence techniques such as agent-based modeling and probabilistic reasoning have shown promise in modeling complex biological systems and testing ecological hypotheses through simulation. We develop an agent-based model of Arctic foxes from Medniy Island while utilizing Probabilistic Graphical Models to capture the conditional dependencies between the random variables. Such models provide valuable insights in analyzing factors behind catastrophic degradation of this population and in revealing evolutionary mechanisms of its persistence in high-density environment. Using empirical data from studies in Medniy Island, we create a realistic model of Arctic foxes as agents, and study their survival and population dynamics under a variety of conditions.	cs.MA	cs	An agent-based model of an endangered population of the Arctic fox from Mednyi Island Angelina Brilliantova1, Anton Pletenev1, Liliya Doronina2, Hadi Hosseini3 1 Moscow State University, Russia 2 University of Munster, Germany 3 Rochester Institute of Technology, USA [email protected], [email protected], [email protected], [email protected] 8 1 0 2 l u J 6 1 ] A M . s c [ 1 v 3 0 1 6 0 . 7 0 8 1 : v i X r a Abstract Artificial Intelligence techniques such as agent- based modeling and probabilistic reasoning have shown promise in modeling complex biological systems and testing ecological hypotheses through simulation. We develop an agent-based model of Arctic foxes from Medniy Island while utilizing Probabilistic Graphical Models to capture the con- ditional dependencies between the random vari- ables. Such models provide valuable insights in an- alyzing factors behind catastrophic degradation of this population and in revealing evolutionary mech- anisms of its persistence in high-density environ- ment. Using empirical data from studies in Medniy Island, we create a realistic model of Arctic foxes as agents, and study their survival and population dynamics under a variety of conditions. 1 Introduction Studying demographic mechanisms of a population for con- servation risk assessment is often a challenging issue since ecological systems are inherently complex, nonlinear, and include multiple interactions of various components [Wit- mer, 2005]. The complex nature of ecological, biological, and behavioural aspects of species calls for novel interdisci- plinary approaches for understanding the subtle reasons for survival and extinctions of species. In recent years, Artificial Intelligence (AI) techniques such as Agent-Based Modeling (ABM) have shown promise in modeling complex systems under uncertainty, analyzing their characteristics, and allow- ing the testing of more complex biological and ecological hy- potheses through simulation [Grimm and Railsback, 2013; Wellman, 2016; McLane et al., 2011; Yang et al., 2014]. Agent-based modeling have become widely used for ad- dressing eco-evolutionary [Mosser et al., 2015] and epi- demiological [Eisinger and Thulke, 2008; Wang and Spear, 2015] issues, in particular, for studying population dynam- ics and improving conservation management. [Robbins and Robbins, 2004; Stenglein et al., 2015; Carter et al., 2015; Watkins et al., 2015]. The investigation of population dy- namics in endangered Arctic fox subspecies (Vulpes lagopus semenovi, listed in Russian Red Data Book, [Goltsman et al., 1996]) from Mednyi Island (Commander Islands, North Pa- cific) is one of the intriguing examples that requires appli- cation of modern AI approaches. Until the end of the 20th century, the population density remained extremely high as compared to other fox populations: up to a thousand animals on an island of 187 square kilometers. An epizootic of ear mange occurred in 1970–1980 among juveniles that wiped out the majority of cubs and led to a drastic decline in a pop- ulation size. During the next few decades the population had partially recovered, however, it stabilized on a much lower level, hovering around 100 adults (10-15% of the previous size) [Goltsman et al., 2005a; Goltsman et al., 2005b]. To study the effects of several factors on the population dynamics we develop an agent-based model of the popula- tion that incorporates uncertainties through probabilistic vari- ables. In this model, Arctic foxes are represented as agents with particular attributes, interests, and behaviors that can interact with other agents and the environment. This prob- abilistic model sheds light into factors behind catastrophic degradation of the Arctic fox population and reveals evolu- tionary mechanisms allowed for the population persistence in high-density environments. Given the data collected between 1994–2012, we investigate various factors in survival of Arc- tic fox population by empirically studying various scenarios and the interaction between parameters that affect the pop- ulation. In future, this model can provide ecologists with a powerful tool for population level prediction, effective pro- tection, and management of the ecosystems. 2 Background The demography of Mednyi Arctic foxes has been exten- sively investigated throughout recent decades. From 1994 until 2012 approximately 80% of arctic foxes living on the southern part of Mednyi Island were marked with plastic ear-tags and were individually recognized during their life- time. In result, all basic life-history parameters of the is- land population were determined [Goltsman et al., 1996; Goltsman et al., 2005b; Kruchenkova et al., 2009; Doron- ina et al., in prep]. Based on this long-term individual-based field study the attempts to model the population dynamics were made. The results of matrix modeling [Brilliantova, 2017] showed that the population growth rate most notably depends on juvenile survival. Data from tagged animals was used to develop first agent-based model of Mednyi Arctic fox Figure 1: Probabilistic graphical model of the Arctic foxes. Widely supported primary dependencies are depicted in black. (a) Complete probabilistic model of the Arctic foxes (b) Probabilistic model of the Arctic foxes population [Goltsman et al., 2018]. The model presented in this study differs from [Goltsman et al., 2018] in explicitly incorporating probabilistic conditional framework which al- lows for future modification and expansion of the model. 3 The Model This section introduces the formalism required to model the ecosystem. Our agent-based model consists of three primary components: agents, the environment, and processes. In our model, each arctic fox is modeled as an agent i with type θi = (si, ai, ri, rpi, hi) where si ∈ {f, m} indicates the gender, ai denotes the age, ri ∈ {0, 1} is a boolean variable denoting the residency status with 0 representing floaters and 1 resident agents, rpi ∈ {breeder, helper, member of non reproductive group, peripheral animal, single animal} is a categorical variable de- noting the reproductive status of agent i, and hi denotes a home range occupied by the agent i. Agents: The development stages of an agent is modeled by a linear function over time. All animals who survive in each period i+1,∀at grow for 1 year in each time period t, thus, at+1 i. Following the literature on the lifespan of arctic foxes, we set up a limitation for a max(at i) = 12 which exceeds the max- imum observed age of Arctic foxes, 10 years, [Hersteinsson, 1992; Eide et al., 2012]. Therefore, agents with at i > 12 are eliminated at each time t. We further classify the develop- mental stages as age classes Ai: i = at (cid:40)cub, Ai = yearling, adult, if ai = 0 if ai = 1 if ai ∈ [2, 12] (1) Environment: The landmark of the island can be con- sidered as a collection of territories or home ranges. Let H t denote the set of home ranges at time t. Each home range Hj ∈ H t has a set of attributes, that is, Hj = , fj), where xj is the coordinate of home (xj, nt j are inhabitancy parameters indi- range j, and nt cating the total number of agents, the number of agents with j, nm,t j, nm,t , nf,t , nf,t j j j si = m, number of agents with si = f inhabiting Hj at time t respectively; in all cases the number of agents does not ac- j + nf,t . count for cubs. Thus, for all j, t we have nt Each Hj ∈ H t has a corresponding food availability level fj ∈ {poor, medium, rich}, based on proximity and size of bird colonies near location xj. For simplicity, in this pa- per the configuration of home ranges (H t) along the island j ,∀t, t(cid:48). is assumed identical for all periods, that is, H t The total population at time t is denoted by nt, where nt = j = H t(cid:48) j = nm,t j Hj∈H t nt j, thus, n0 is the initial population at t = 0. (cid:80) A Probabilistic Life Cycle The life cycle of animals is heavily influenced by uncertain- ties derived from a variety of environmental factors. For ex- ample, in Mednyi Arctic fox population, survival and disper- sal are strongly affected by the sex of animals, and the sex ratio in litters is influenced by food availability in a home range: in rich habitats females produce more females, while in poor conditions the sex ratio of cubs is biased towards males [Goltsman et al., 2005b; Goltsman et al., 2005a]. To model the inherent conditional dependencies between the random variables, we utilize Probabilistic Graphical Mod- els (PGM) for the Mednyi Arctic fox population (Fig.1a). PGMs are probabilistic models illustrating structured con- ditional dependencies between random variables in complex domains. In Fig.1a several likely dependencies are depicted with red links as the current literature on arctic foxes has yet to explore the causal effects. Therefore, we focus attention on primary dependencies that are widely supported in the lit- erature [Goltsman et al., 2005b; Goltsman et al., 2005a], par- ticularly on the affects of Age (A) and Sex (s) on survival, residency status, and reproduction (Fig.1b). The f and rp parameters are excluded in our current model for simplicity. Survival: Let φi denote the output of survival process for agent i where 1 denotes survival and 0 denotes death of i. Survival process follows a Bernoulli distribution with p(φi = 1Ai, si) as a probability of success. For each agent i, the age- and sex-specific survival is computed using Bayes rule. We assume that Ai and si are independent variables, so we can write: p(φiAi, si) = p(Ai, siφi) × p(φi) = p(Ai, si) p(Aiφi) × p(siφi) × p(φi) p(Ai) × p(si) where marginal probabilities of age class p(Ai) and sex p(si) are estimated using observed data as the ratio of agents in each age class and sex in a population. Conditional probabil- ities of age class p(Aiφi) and sex p(siφi) given survival out- put are estimated as proportions of agents with corresponding age and sex attributes separately in a sample of survivors and dead agents. Although, p(φiAi, si) can be estimated directly from the empirical data, we used Bayes rule (1) to allow for future sensitivity analysis of age- sex-specific survival to the components of the right hand side of the equation, (2) to pro- vide general framework for the future development of a com- plete PGM model (i.e. adding missing factors from Fig.1a for which empirical data is not always available). Residency status and dispersal process: During disper- sal, yearlings and lone adults change their home range trying to find a mate. For each agent the cycle checks its residency Home rangeFood AvailabilityAgeReproductive StatusSexSurvivalResidency statusReproductionAgeReproductionSexSurvivalResidency statusFigure 2: Dispersal algorithm Figure 3: The transitions between different age classes. Diamonds denote conditional and rectangles indicate stochastic processes. Table 1: The parameters and default values of the model parameters default values H 60 pr,a 0.5 pr,y 0.1 b 4 σb 1 psex 0.5 k at t. If At i = adult and status. Let agent i occupies H t nsex(cid:54)=si,t > 0, i.e. there are more than 0 animals of the op- k posite sex in H t k, then agent i is classified as a resident with ri = 1. Otherwise, the agent is classified as a floater with ri = 0 and will be forced to change its home range accord- ing to the following rule: If the set of H t with nsex(cid:54)=si,t > 0 and nsex=si,t = 0 is non-empty, then the agent moves to a randomly selected Hz ∈ H t. Otherwise, the agent randomly j), ∀j. After dispersal, parameters moves to Hz with min(nt z are updated (see Fig.2). k and H t nt, nm,t, nf,t of H t Reproduction: Agents with si = f and Ai = adult or yearling inhabited H t k > 0 reproduce with prob- ability pr,a for Ai = adult and pr,y for Ai = yearling. We assume these probabilities are drawn i.i.d from a fixed dis- tribution. The litter size is drawn from a normal distribution with mean equals b and standard deviation equals σb. A new- born agent i has ai = 0, Ai = cub, hi = Hk, ri = 0, and si is drawn from a Bernoulli distribution with probability psex. k with nm,t 4 Process overview and parameters The model proceeds in annual time steps, starting from win- ter. Within each year or time step, three modules or phases are processed in the following order: winter survival, disper- sal, and reproduction. Within each module, individuals and territories are processed in a random order. The individual life cycle is depicted in (Fig.3). The model parameters were defined according to the field studies conducted on Mednyi Island between 1994–2012 (Table 1). The survival parame- ters are estimated based on the dataset of life stages of an- imals tagged with individual ear tags when they were cubs. The total number of tagged cubs is 517. 4.1 At t = 0 cubs, yearlings, and adults are generated according to the following parameters: n0 = 120 (yearling and adult total number), proportion of adult agents in initial population = 0.24, proportion of yearling agents = 0.15, proportion of Initialization cubs = 0.61. Sex is randomly assigned with 0.5 probabilities; age is assigned uniformly at random from 2 to 8 for adults (maximum observed age of Arctic foxes on Mednyi Island). Each agent is placed randomly on one of the home ranges. 4.2 Empirical Simulations Setup: We implemented our model in Netlogo [Wilensky, 1999] and ran the model with default parameters while vary- ing one of the parameters with others remained at default level (Table 3). The change of parameters in focus was per- formed in regular intervals (∆). We have completed 100 runs for each setting. Each run lasts 50 years or when the popula- tion went extinct. Because the computational costs rise sub- stantially with the population size we limited the total number of agents number of agents excluding cubs (nt) to 500 agents, which corresponds to the assumed ecological capacity of the island. The simulation stops upon reaching the maximum limit. In our analysis, only population-level variables were recorded, i.e. the population size over time nt, the averaged annual population growth (λ), and the proportion of runs with population reaching either extinction or the maximum limit. To allow the population reach its stationary age and sex com- position [Gotelli, 1995] we excluded the first 3 years of each run from λ calculation. We consider the population extinct when nt declines below the 10 agent threshold. Critical mass for survival: We focused on the impact of initial population of adults and yearling n0 on the extinc- tion of agents. In all scenarios, the population level signif- icantly decreases, however the growth rate linearly rises with n0 (Fig.4a). For larger n0 scenarios ( ≥ 220) the median for growth rates were similar (λ = {−0.024;−0.030}) at the initial stages of the simulation (first 10 years) and diverged afterwards. On average, the extinction was equal 100% for scenarios with n0 ∈ {20, 70}, 99% with n0 = 120 (default) and gradually decreases as n0 rises to 2% for n0 = 470. Survival probabilities: For all age class scenarios, λ linearly increases with survival probabilities (e.g. see Fig.4b for cub survival scenarios). The highest impact on population dy- namics is affected first by cubs and then adults survival prob- abilities; a change of +0.1 and more to the default cub or adult survival probabilities leads to a rapid growth of popula- Age class (Ai)adult yearling Presence of opposite sex in HkyesPresence of H with only opposite sexStay in HkMove to one of such HMove to one of H with minimum number of agentsAgent iin home range Hkat tnoyesTable 2: Percentage of extinct and max limit runs. * indicates proportion of runs when the population reaches 500 agents. cubs yearlings adults scenarios % extinct % max limit* % extinct % max limit* % extinct % max limit* +0.05 +0.10 +0.15 +0.20 0% 85% 98% 100% 0% 95% 100% 100% 55% 1% 0% 0% 93% 58% 20% 3% 0% 4% 51% 95% 69% 5% 0% 0% Figure 4: Average growth rate of population in different scenarios (a) Initial agent number scenarios (b) Cub survival scenarios Table 3: Overview of parameters varying in simulations. ∆ indicates the change in parameters. Range Parameter 20-470 n0 ±0.2 of default value Survival of cubs (both sex) Survival of yearlings (both sex) ±0.2 of default value ±0.2 of default value Survival of adults (both sex) ∆ 50 0.05 0.05 0.05 tion (Table 2). For scenarios with +0.05 change to the default cub survival and +0.1 to the default yearling survival, trajec- tories of population dynamics significantly diverge between runs as average growth rate hovers around zero. 5 Discussion We found that cub survival has the most pronounced impact on the population dynamic. It exceeds the effect of adult sur- vival, though cub age class includes only one age cohort (0 year age) as compared to 11 for adults (2-12 year age). These results are consistent with the results of another individual- based model of Mednyi Arctic fox population [Goltsman et al., 2018], with Leslie matrix analysis of this population [Brilliantova, 2017] and with matrix modeling of other pop- ulations of "fast" (i.e, rapidly maturing) mammals [van de Kerk et al., 2013]. The incorporation of AI techniques help reveal the threshold for the population size – around 200 adult and yearling agents - below which the pace of decline accel- erates. We attribute this effect to the growing influence of negative stochastic factors in small populations, i.e. the re- sult of demographic stochasticity [Morris and Doak, 2002]. We intend to investigate this effect in more details as a cru- cial dynamic factor for such small populations. The default parameters – aimed to replicate real characteristics of the pop- ulation - result in a rapid extinction of agents, contradicting empirical evidence of the stable population dynamic for the past 20 years. Such an outcome may be attributed to failure to correctly identify default parameter values and/or simplistic structure of the model which does not incorporate some key factors and links. We note, that increasing the default param- eter of the cub survival probability by 0.05, i.e. the magnitude of error in our estimation, improves the population dynamic with a growth rate increasing to 100. The major shortcomings of our current model deal with the absence of (1) complete environmental factors, e.g. spatial and temporal distribution of food, (2) spatial relationship between home ranges as a factor in dispersal submodel, and (3) the specific social struc- ture of Mednyi foxes – females tend to stay at parental home ranges or move to adjacent ranges, so complex families with one male and several females emerge. We plan to incorpo- rate these features during future development of the model extending further to the proposed final structure (Fig.1a). In the future, we would like to study group formation and terri- tory variations through techniques from cooperative and non- cooperative game theory: the behavior of Arctic foxes, par- ticularly in forming families and packs, does not immediately correlate with food availability and reproductive abilities and seem to follow interesting, but non trivial patterns. Acknowledgments We are immensely grateful to Dr. Mikhail Goltsman, the chief of Medniy Arctic Fox project, who provided insights and expertise to this research. We thank our colleagues from Moscow State Univer- sity, especially Elena Kruchenkova, who carried out field studies on Mednyi Island for over 20 years to collect empirical data used in this paper as well as Vladimir Burkanov and Nature and Biosphere Reserve of Commander Islands for their logistical support. lllllllllllll−0.20−0.15−0.10−0.050.0070120170220270320370420470Initial number of yearling and adult agentsAveraged annual growth ratelllllllllllll−0.20.00.2−0.2−0.15−0.1−0.0500.050.10.150.2Change in survival probabilityAveraged annual growth rateReferences [Brilliantova, 2017] Angelina Brilliantova. A matrix model of an endangered population of the Arctic fox (Vulpes lagopus semenovi) from Mednyi Island (in russian). Mas- ter's thesis, 2017. [Carter et al., 2015] Neil Carter, Simon Levin, Adam Bar- low, and Volker Grimm. Modeling tiger population and territory dynamics using an agent-based approach. Eco- logical Modelling, 312:347–362, 2015. [Doronina et al., in prep] Liliya Doronina, Eliezer Gurarie, and Mikhail Goltsman. Population dynamics and life his- tory parameters in a closed, insular population of arctic fox. (in prep.). [Eide et al., 2012] Nina E. Eide, Audun Stien, Pal Prestrud, Nigel G. Yoccoz, and Eva Fuglei. Reproductive responses to spatial and temporal prey availability in a coastal arctic fox population. Journal of Animal Ecology, 81(3):640– 648, 2012. [Eisinger and Thulke, 2008] Dirk Eisinger Hermann Thulke. eradication of infectious diseases. Ecology, 45(2):415–423, 2008. and Hans- Spatial pattern formation facilitates Journal of Applied [Goltsman et al., 1996] Mikhail Goltsman, Kruchenkova, and David W. Macdonald. nyi Arctic foxes: disease. Oryx, 30(4):251–258, 1996. P. Elena The Med- treating a population imperilled by [Goltsman et al., 2005a] Mikhail Goltsman, P. Kruchenkova, Sergei Sergeev, Paul J. Johnson, and David W. Macdonald. Effects of food availability on dispersal and cub sex ratio in the Mednyi Arctic fox. Behavioral Ecology and Sociobiology, 59(2):198, 2005. Elena [Goltsman et al., 2005b] Mikhail Goltsman, P. Kruchenkova, Sergei Sergeev, Ilja Volodin, and David W. Macdonald. 'island syndrome' in a population of arctic foxes (alopex lagopus) from Mednyi Island. Journal of Zoology, 267(4):405–418, 2005. Elena [Goltsman et al., 2018] Mikhail Goltsman, Elena Sushko, Liliya Doronina, and Elena P. Kruchenkova. An individual-based model of the population dynamics of the arctic fox (vulpes lagopus semenovi) on Mednyi Island (commander islands, north pacific) (in russian). Zool. Z, 97(6):472– 482, 2018. [Gotelli, 1995] Nicholas J. Gotelli. A primer of ecology. Sin- auer Associates Incorporated, 1995. [Grimm and Railsback, 2013] Volker Grimm and Steven F. Individual-based modeling and ecology. Railsback. Princeton university press, 2013. [Hersteinsson, 1992] P Hersteinsson. Demography of the Arctic fox (Alopex lagopus) population in Iceland. In Wildlife 2001: Populations, pages 954–964. Springer, 1992. [Kruchenkova et al., 2009] Elena P. Kruchenkova, Michael Goltsman, Sergei Sergeev, and David W. Macdon- ald. Is alloparenting helpful for Mednyi Island Arctic foxes, Alopex lagopus semenovi? Naturwissenschaften, 96(4):457, 2009. [McLane et al., 2011] Adam J McLane, Christina Semeniuk, Gregory J McDermid, and Danielle J Marceau. The role of agent-based models in wildlife ecology and management. Ecological Modelling, 222(8):1544–1556, 2011. [Morris and Doak, 2002] F. William Morris and F. Daniel Doak. Quantitative conservation biology. 2002. [Mosser et al., 2015] Anna A. Mosser, Margaret Kosmala, and Craig Packer. Landscape heterogeneity and behav- ioral traits drive the evolution of lion group territoriality. Behavioral Ecology, 26(4):1051–1059, 2015. [Robbins and Robbins, 2004] Martha M. Robbins and An- drew M. Robbins. Simulation of the population dynam- ics and social structure of the Virunga mountain gorillas. American Journal of Primatology, 63(4):201–223, 2004. [Stenglein et al., 2015] Jennifer L. Stenglein, Jonathan H. Gilbert, Adrian P. Wydeven, and Timothy R. Van Dee- len. An individual-based model for southern Lake Supe- rior wolves: A tool to explore the effect of human-caused mortality on a landscape of risk. Ecological modelling, 302:13–24, 2015. [van de Kerk et al., 2013] Madelon van de Kerk, Hans de Kroon, Dalia A Conde, and Eelke Jongejans. Carnivora population dynamics are as slow and as fast as those of other mammals: implications for their conservation. PloS one, 8(8):e70354, 2013. [Wang and Spear, 2015] Shuo Wang and Robert C. Spear. Exploring the contribution of host susceptibility to epi- demiological patterns of Schistosoma japonicum infection using an individual-based model. The American journal of tropical medicine and hygiene, 92(6):1245–1252, 2015. [Watkins et al., 2015] A. Watkins, J. Noble, RJ Foster, BJ Harmsen, and CP Doncaster. A spatially explicit agent- based model of the interactions between jaguar popula- tions and their habitats. Ecological modelling, 306:268– 277, 2015. [Wellman, 2016] Michael P Wellman. Putting the agent in agent-based modeling. Autonomous Agents and Multi- Agent Systems, 30(6):1175–1189, 2016. [Wilensky, 1999] Uri Wilensky. Netlogo (and netlogo user manual). Center for connected learning and computer- based modeling, Northwestern University. http://ccl. northwestern. edu/netlogo, 1999. [Witmer, 2005] Gary W. Witmer. Wildlife population mon- itoring: some practical considerations. Wildlife Research, 32(3):259–263, 2005. [Yang et al., 2014] Rong Yang, Benjamin Ford, Milind Tambe, and Andrew Lemieux. Adaptive resource allo- cation for wildlife protection against illegal poachers. In Proceedings of the 2014 international conference on Au- tonomous agents and multi-agent systems, pages 453– 460. International Foundation for Autonomous Agents and Multiagent Systems, 2014.
1706.05254	1	1706	2017-06-13T20:21:00	Collaborative vehicle routing: a survey	[ "cs.MA", "cs.AI", "cs.CY", "math.OC", "physics.soc-ph" ]	In horizontal collaborations, carriers form coalitions in order to perform parts of their logistics operations jointly. By exchanging transportation requests among each other, they can operate more efficiently and in a more sustainable way. Collaborative vehicle routing has been extensively discussed in the literature. We identify three major streams of research: (i) centralized collaborative planning, (ii) decentralized planning without auctions, and (ii) auction-based decentralized planning. For each of them we give a structured overview on the state of knowledge and discuss future research directions.	cs.MA	cs	Collaborative vehicle routing: a survey Margaretha Gansterera,∗, Richard F. Hartla aUniversity of Vienna, Department of Business Administration, Oskar-Morgenstern-Platz 1, 1090 Vienna, Austria Abstract In horizontal collaborations, carriers form coalitions in order to perform parts of their logistics operations jointly. By exchanging transportation requests among each other, they can operate more efficiently and in a more sustain- able way. Collaborative vehicle routing has been extensively discussed in the literature. We identify three major streams of research: (i) centralized collaborative planning, (ii) decentralized planning without auctions, and (ii) auction-based decentralized planning. For each of them we give a structured overview on the state of knowledge and discuss future research directions. Keywords: Logistics, Collaborations, Vehicle routing 1. Introduction The transportation industry is highly competitive and companies need to aim for a maximum level of efficiency in order to stay in business. Fierce competition brings prices down and therefore profit margins have declined to an extremely low level. To increase efficiency, these companies can establish collaborations, where parts of their logistics operations are planned jointly. By collaborative vehicle routing we refer to all kinds of cooperations, which are intended to increase the efficiency of vehicle fleet operations.1 By in- creasing efficiency, collaborations also serve ecological goals. It is well known ∗Corresponding author Email addresses: [email protected] (Margaretha Gansterer), [email protected] (Richard F. Hartl) 1We use the terms collaboration and cooperation interchangeable. In the literature, there is an agreement that collaboration is a strong type of cooperation. However, the boundary between them is vague (Cruijssen et al., 2007c). Preprint submitted to Elsevier June 19, 2017 that transportation is one of the main contributors of CO2 emissions (Ballot and Fontane, 2010). Thus, public authorities are encouraging companies to collaborate. They not only aim at reduced emissions of harmful substances, but also on reduced road congestion, and noise pollution. Moreover, collab- orations in logistics have been shown to increase service levels, gain market shares, enhance capacities, and reduce the negative impacts of the bullwhip effect (Audy et al., 2012). Thus, it is not surprising that collaborative vehicle routing is an active research area of high practical importance. Related reviews by Verdonck et al. (2013) and Cruijssen et al. (2007c) exist. Both are dealing with transportation collaborations. However, Ver- donck et al. (2013) focus on the operational planning of road transportation carriers (i.e. the owners and operators of transportation equipment) only. The perspective of collaborating shippers (i.e. the owners of the shipments) is not taken into account. Furthermore, they do not consider studies on cen- tralized planning (i.e. collaboration in case of full information). We observe that about 45% of the related literature refers to central planning situations. This is an important aspect of collaborative vehicle routing, where a central- ized authority is in charge of allocating requests such that requirements of all collaborators are met. Furthermore, we identify two classes of decentralized settings, which are auction-based and non-auction-based collaborations. Cruijssen et al. (2007c) give an overview on different types of horizontal collaboration, i.e. the levels of integration among collaborators. They do not consider operational planning problems. We find that almost 60% of the related articles were published in the last 3 years. These articles have not been covered in both of the existent reviews. The review by Guajardo and Ronnqvist (2016) deals with cost allocation in collaborative transportation, which is also an important aspect in collab- orative vehicle routing. Because of this very recent survey, we can keep the cost allocation part short. Given the high volume of recent literature on collaborative vehicle routing it is now appropriate to provide a review on the state of knowledge. The contribution of our survey is threefold: 1. We also consider centralized collaborative planning. 2. We survey the literature of the last few years. 3. We give a new and broader classification of articles. The remainder of our survey is organized as follows. The research method- ology used is described in Section 2. Classifications and definitions are pro- 2 vided in Section 3. Centralized collaborative planning is surveyed in Sec- tion 4. Sections 5 and 6 give overviews on decentralized planning with and without auctions, respectively. Each section closes with a discussion on fu- ture research directions. A summarizing conclusion is given in Section 7. 2. Research methodology In our review, we focus on studies where operations research models and solution techniques are applied. Pure empirical studies, not focusing on the operational planning problems, are not considered. However, readers inter- ested in these empirical studies are referred to, e.g., Cruijssen et al. (2007b), Lydeka and Adomavicius (2007), Ballot and Fontane (2010), Schmoltzi and Wallenburg (2011). We also do not consider studies, where the main focus is on general design of coalitions rather, while the transportation planning problems of collaborators are neglected (e.g. Tate, 1996; Verstrepen et al., 2009; Voruganti et al., 2011; Audy et al., 2012; Guajardo and Ronnqvist, 2015). Regarding application fields, we limit our survey to studies on road transportation. Collaborations in rail (e.g. Kuo et al., 2008), naval (e.g. Agarwal and Ergun, 2010), and air (e.g. Ankersmit et al., 2014) transporta- tion lead to interesting planning problems, but these studies do not fit within the scope of this survey. We initially did not omit any study based on its type (primary, secondary; journal articles, proceedings, etc.) or its year of publication. With this scope in mind, we searched library databases, where the major journals in operations research, operations management, management science etc. are covered. Used search terms were basically combinations of • "collaboration", "cooperation", "coalition", "alliance" and • "transportation", "routing", "logistics", "freight", "carrier", "shipper". By this, we identified a first set of relevant studies. At this point we decided to only consider journal publications, in order to survey a reasonable number of articles. For each of the remaining papers we screened the reference list and added articles that had not been find in the first step. Finally, we applied a descendancy approach, which goes from old information to new information, by screening all articles that cite one of the papers that we found relevant. We proceeded until we converged to a final set of publications. 3 3. Classifications and definitions In our review, we distinguish between centralized and decentralized col- laborative planning. In Figures 1- 3, we provide a generalized illustrations of collaborative and non-collaborative settings. In the non-collaborative setting (Figure1), each participant i, i ∈ (A, ..., N ) maximizes his individual profit Pi. This profit depends on his set of requests Ri, the payments pi(Ri) that he gets for his requests Ri, and his costs ci(Ri). The capacity usage Capi of a participant i is limited by his available capacity Li. Figure 1: Generalized illustration of a non-collaborative setting with profit (Pi) maximiz- ing participants i, i ∈ (A, ..., N ), with limited capacity Li. Capi is the capacity usage. In case of centralized planning (Figure 2), the total profit is maximized jointly (which is denoted as ideal model by Schneeweiss, 2003). In decentralized settings (Figure 3), collaborators agree on a mechanism for exchanging subsets of their requests by revealing no or only limited in- formation. ¯R is the set of requests that have been offered for exchange. In our study, the research stream on decentralized planning is further split up in non-auction-based and auction-based studies. All papers are classified based on the model formulation of the routing problem used. Here, the following categories can be identified. 4 ABBBNFigure 2: Generalized illustration of centralized planning with profit (Pi) maximizing participants i, i ∈ (A, ..., N ), with limited capacity Li. Capi is the capacity usage. • Vehicle routing problems (VRP), which give the optimal sets of routes for fleets of vehicles in order to visit a given set of customers (e.g. Gendreau et al., 2008). • Arc routing problems (ARP), which assume that customers are located on arcs that have to be traversed. The capacitated ARP is the arc- based counterpart to the node-based VRP (e.g. Wøhlk, 2008). • Inventory routing problems (IRP), which combine VRP with inventory management (e.g. Bertazzi et al., 2008). • Lane covering problems (LCP), which aim at finding a set of tours covering all lanes with the objective of minimizing the total travel cost (e.g. Ghiani et al., 2008). A lane is the connection between the pickup and the delivery node of a full truckload (FTL) request. • Minimum cost flow problems (MCFP), which aim at sending goods through a network in the cheapest possible way (e.g. Klein, 1967). • Assignment problems (AP), where vehicles are assigned to requests, such that total costs are minimized (e.g. Munkres, 1957). We also indicate whether models assume customers to have delivery time windows (TW) or not. Another frequently appearing extension are pickup 5 ABBBNFigure 3: Generalized illustration of decentralized planning with profit (Pi) maximizing participants i, i ∈ (A, ..., N ), with limited capacity Li. Capi is the capacity usage. ¯R is the subset of requests that have been offered for exchange. 6 ABBBNand delivery (PD) requests. This means that goods have to be picked up at some node and to be delivered to another node, where pickup and delivery locations do not necessarily coincide with a depot. Articles can be further classified based on the investigated types of ship- ment. These can either be FTL or less than truckload (LTL). FTL can of course be seen as a special case of LTL, where the size of customer orders is equal to the vehicle's capacity. Hence, LTL models are applicable for FTL settings as well. However, FTL is often used in the transportation of a sin- gle product, whereas LTL is usually used to transport multiple products in small volumes from depots to customers or from customers to customers. A typical application area for LTL is parcel delivery (Dai and Chen, 2012; Parragh et al., 2008). Figure 4 shows an example of non-collaborative and of collaborative vehicle routes of LTL carriers. An example of collaboration between FTL carriers is displayed in Fig- ure 5. Carrier A drives from customer c to customer a, and from customer a to customer b, while carrier B is serving lanes b-c and b-d. By collaboration the carriers can avoid three empty return trips (Adenso-D´ıaz et al., 2014a). It should be mentioned that b-c can of course also be used by carrier A in a non-collaborative setting. In this case this carrier has only one empty return trip. However, the collaboration with carrier B still improves the situation of carrier A. Players in transportation collaborations might be carriers (also denoted as freight forwarders, logistics service providers, or third party logistics providers) and shippers. Carriers are assumed to be the owners and operators of trans- portation equipment, while shippers own or supply the shipments. Joint routing planning is typically assumed to be done by carriers. When ship- pers consider collaboration, they identify attractive bundles of lanes, helping carriers to reduce empty trips in return of better rates (Ergun et al., 2007b). We observe that the huge majority of papers focuses on carrier-related cooperations. We agree with Cruijssen et al. (2007a), that from the planning perspective it does not matter whether carriers or shippers are in charge of the process. However, in decentralized settings the issue of information asymmetries has to be taken into account. Shippers and carriers typically do not having the same level of information. We therefore find it useful to distinguish whether carriers or shippers are the players in a collaboration. In the papers surveyed, a variety of both, exact and heuristic solution methodologies are represented. An overview on these methodologies is given in Figure 6. 7 Figure 4: Example for non-collaborative and collaborative vehicle routes of three LTL carriers with pickup and delivery requests (Gansterer and Hartl, 2016a). 8 +--+-++-DepotPickupDeliveryABC-+-++-C1C2-C2C1C5+B1B2B3B2B3B1+A1-A2A2A3+-+C3C4C4C5C3-A1A3+-+--+-++-ABC-+-++-C1C2-C2C1C5+B1B2B3B2B3B1+A1-A2A2A3A3+-+A3C3C4C4C5C3-A1A3No collaborationCollaborationFigure 5: Example for non-collaborative and collaborative vehicle routes of two FTL carriers. Dotted arcs are empty return trips (Adenso-D´ıaz et al., 2014a) 9 Carrier AabcdabcdCarrier BNo collaborationCollaboration) 6 1 0 2 ( . l a t e z e d n ´a n r e F , ) 4 1 0 2 ( a t e e P d n a z e d n ´a n r e H , ) 1 1 0 2 ( . l a t e z e d n ´a n r e H , ) 1 1 0 2 ( a t e e P d n a z e d n ´a n r e H , ) 0 1 0 2 ( h t r i w r e i B d n a r e g r e B : t u c - d n a - h c n a r B ) 7 1 0 2 ( u z y u K , ) 2 1 0 2 ( n e h C d n a i a D : e c i r p - d n a - h c n a r B , ) 4 1 0 2 ( . l a t e i a D , ) 2 1 0 2 ( n e h C d n a i a D : n o i t i s o p m o c e d d n a n o i t a x a l e r n a i g n a r g a L ) 7 1 0 2 ( u z y u K : n o i t a r e n e g n m u l o C ) 6 1 0 2 ( n e h C t c a x E ) 4 1 0 2 ( u X d n a g n e W : d e s a b - n o i t i s o p m o c e d s r e d n e B d n a n o i t a x a l e r n a i g n a r g a L s c i t s i r u e h t a M , ) 4 1 0 2 ( r e f p o K d n a g n a W , ) 4 1 0 2 ( . l a t e g n a W : h c r a e s d o o h r o b h g i e n e g r a l e v i t p a d A ) 7 1 0 2 ( r e f p o K d n a a k p o h c S , ) 6 1 0 2 ( . l a t e i L , ) 5 1 0 2 ( r e f p o K d n a g n a W ) a 4 1 0 2 ( . l a t e z a ´ı D - o s n e d A : e r u d e c o r p h c r a e s d e z i m o d n a r y d e e r G ) 6 1 0 2 ( . l a t e o j u a r A - o r e t n i u Q , ) 6 1 0 2 ( . l a t e o v r e u C , ) 1 1 0 2 ( s g i r e D d n a l h a D : n o i t a l u m S i . s e i g o l o d o h t e m n o i t u l o s n o w e i v r e v O : 6 e r u g i F ) 6 1 0 2 ( . l a t e o j u a r A - o r e t n i u Q , ) 5 1 0 2 ( . l a t e u e b a n r e B - z e r ´e P : h c r a e s l a c o l d e t a r e t I ) 3 1 0 2 ( r e d n i b k o o B d n a h a j a r a d a N , ) 1 1 0 2 ( s g i r e D d n a l h a D : d e s a b - h c r a e s l a c o L , ) a 0 1 0 2 ( . l a t e u i L , ) b 7 0 0 2 ( . l a t e n u g r E , ) a 7 0 0 2 ( . l a t e n u g r E : s c i t s i r u e h y d e e r G ) 4 1 0 2 ( h c n o M d n a r e g n e r p S , ) 1 1 0 2 ( . n e h C H d n a i a D : s m e t s y s - t n e g a - i t l u M ) 2 1 0 2 ( h c n o M d n a r e g n e r p S , ) 1 1 0 2 ( . l a t e y e l i a B s r e h t O , ) 4 1 0 2 ( . l a t e o a D , ) b 0 1 0 2 ( . l a t e i u L : s m h t i r o g l a y r a n o i t u l o v E ) 7 1 0 2 ( l t r a H d n a r e r e t s n a G , ) 6 1 0 2 ( . l a t e z e h c n a S ) 2 1 0 2 ( h c n o M d n a r e g n e r p S : n o i t a z i m i t p o y n o l o c t n A ) 6 1 0 2 ( . l a t e u X , ) 1 1 0 2 ( . l a t e y e l i a B : h c r a e s u b a T s c i t s i r u e h a t e M s d o h t e M 10 An important aspect of collaborative operations is how to share the gained benefits. It is shown in Guajardo and Ronnqvist (2016) that most problems in collaborative transportation use sharing methods based on cooperative game theory. The authors identify more than 40 different methods, which they categorize as traditional or ad hoc concepts. However, they show that in the huge majority of studies, one of the following three methods is used: • the well-known Shapley value (Shapley, 1953), which is generally the most applied method (e.g. Kimms and Kozeletskyi, 2016; Vanovermeire and Sorensen, 2014; Engevall et al., 2004). • proportional methods, where each carrier j gets a share αj of the total profit (e.g. Ozener et al., 2013; Berger and Bierwirth, 2010; Frisk et al., 2010 • the nucleolus method initially defined by Schmeidler (1969) (e.g. Agar- wal and Ergun, 2010; Guajardo and Jornsten, 2015; Gothe-Lundgren et al., 1996). In the literature, there are basically two streams of research: (i) articles that focus on the transportation problems, while profit sharing is not taken into account, and (ii) articles dealing with profit sharing, while the trans- portation problem is neglected. There are only a few studies (for instance Krajewska et al., 2008), where both aspects are combined. Thus, our survey aims at building bridges between these two worlds. In our sections on future research, we provide several suggestions how profit sharing aspects should be integrated into the collaborative planning processes. 4. Centralized collaborative planning If collaborative decisions are made by a central authority having full infor- mation, this is referred to as centralized collaborative planning. An example for such a central authority might be an online platform providing services for collaborative decision making (Dai and Chen, 2012). It is obvious that under full information, the decision maker has to tackle a standard optimization problem, since the collaborative aspect is diminished by information disclo- sure. Thus, each transportation planning problem might be interpreted as a collaborative transportation planning problem that a decision maker with full information has to solve. However, to have in this review a reasonable 11 scope, we only survey studies that contribute to the collaborative aspect of transportation planning. An example for such a contribution is, for instance, given in Wang et al. (2014). In this study the central authority does not have full power to find an optimal solution by simply exchanging requests. It rather has to decide on the degree of collaboration taking both outsourcing and request exchange into consideration. It can be observed that that there are two streams of research in this area: • the assessment of the potential benefits of centralized collaborative planning versus non-cooperative settings. In non-cooperative settings players do not show any kind of collaborative efforts. We refer to this research stream as Collaboration Gain Assessment (CGA). • innovative models or innovative solutions approaches for centralized collaborative planning. These studies will be denoted as Methodological Contributions (MC). Table 1 gives an overview on studies contributing to centralized collabo- rative planning. We classify them according to the categories given above, and highlight their main characteristics. 12 . n o i t a m r o f n i l l u f h t i w n o i t a r o b a l l o c r o f s c i t s i r e t c a r a h c c i s a b d n a s e c n e r e f e R : 1 e l b a T x x x x x x x x x x x x x x x x x x x L T F L T L L T L L T L L T F L T L L T L L T L L T F L T L L T F L T L L T L L T L L T L L T L L T L L T L L T L L T L L T L P R V P R V P R V P R V P C L P R A P F C M P R V P C L P R V P R A P R V P R V P R V P R V P R V P R I P R V P R V P R A P A x x x x x x x x x x x x x x x x x x x x x x D P W T t n e m p i h S l e d o M r e i r r a C r e p p h S i s u c o F A G C C M A G C C M C M C M C M A G C C M A G C C M A G C C M A G C A G C A G C A G C A G C C M C M A G C ) 3 1 0 2 ( r e d n i b k o o B d n a h a j a r a d a N ) 6 1 0 2 ( . l a t e s e r r o T - a y o t n o M ) 5 1 0 2 ( . l a t e u e b a n r e B - z e r ´e P ) 6 1 0 2 ( . l a t e o j u a r A - o r e t n i u Q ) 2 1 0 2 ( h c n o M d n a r e g n e r p S ) 6 1 0 2 ( . l a t e z e h c n a S ) 6 1 0 2 ( . l a t e l a s y o S ) 4 1 0 2 ( u X d n a g n e W ) 4 1 0 2 ( . l a t e g n a W ) 2 1 0 2 ( l i r e n a s a v a S d n a z a m l i Y ) a 4 1 0 2 ( . l a t e z a ´ı D - o s n e d A e c n e r e f e R ) a 7 0 0 2 ( . l a t e n e s s j i u r C ) 6 1 0 2 ( . l a t e z e d n ´a n r e F ) 2 1 0 2 ( n e h C d n a i a D ) b 7 0 0 2 ( . l a t e n u g r E ) 6 1 0 2 ( . l a t e s j i u B ) 1 1 0 2 ( a t e e P d n a z e d n ´a n r e H ) 8 0 0 2 ( . l a t e a k s w e j a r K ) a 0 1 0 2 ( . l a t e ) 8 0 0 2 ( i n L i u L ) 7 1 0 2 ( u z y u K 13 4.1. Collaboration gain assessment One of the first studies to systematically assess the potentials of collabo- rative vehicle routing was presented by Cruijssen et al. (2007a). The authors consider a system with multiple companies, each having a separate set of distribution orders. Goods are picked up at a single distribution center and delivered to customer sites. Both, a non-cooperative setting, where each company solves the planning problem independently, and a cooperative set- ting, where routes are planned jointly are investigated. It is shown that joint route planning can achieve synergy values of up to 30%. Many other studies confirm the observation of Cruijssen et al. (2007a), that centralized collaborative planning has the potential to improve total profits by around 20-30% of the non-cooperative solution (e.g. Montoya- Torres et al., 2016; Soysal et al., 2016). A real-world setting, where a local courier service of a multi-national logistics company is investigated by Lin (2008). It is shown that the cooperative strategy, where courier routes are planned jointly, outperforms the non-cooperative setting by up to 20% of travel cost. Joint route planning is generally considered to be done by carriers (e.g. Dai and Chen, 2012; Buijs et al., 2016; Liu et al., 2010a), but also shippers can be involved in joint route planning, as long as they have direct control over the flows of goods (Cruijssen et al., 2007a). However, merging FTL lanes is mostly assumed to be done by shippers (e.g. Adenso-D´ıaz et al., 2014a; Ergun et al., 2007b; Kuyzu, 2017). Here again one might argue, that also carriers can be involved in this type of horizontal collaboration, as it is considered by, e.g., Liu et al. (2010a). Horizontal collaborations not only follow economical but also ecological goals like reduced road congestion, noise pollution, and emissions of harmful substances. Thus, public authorities are encouraging companies to collab- orate. The city of Zurich, for instance, is funding a research project aim- ing at improved cooperation between different transport companies by an IT-based collaboration platform (Schmelzer, 2014). In this spirit, Montoya- Torres et al. (2016) quantify the effect of collaborative routing in the field of city logistics. In order to solve real-world instances from the city of Bogot´a, the centralized problem is decomposed into an assignment and a routing part. By this, the non-cooperative solution can be improved by 25.6% of the travel distance. Many other recent studies account for ecological aspects. P´erez-Bernabeu et al. (2015), for instance, examine different VRP scenarios and show that co- 14 operations can contribute to a noticeable reduction of expected travel costs as well as of greenhouse gas emissions. The VRP with time windows (VRPTW) and with carbon footprint as a constraint is proposed by Sanchez et al. (2016). Using this model, the reduction of carbon emissions in a collaborative set- ting, where different companies pool resources, is investigated. The authors find that the total greenhouse gas emissions can be reduced by 60%, while cost savings were nearly 55%. Soysal et al. (2016) model and analyse the IRP in a collaborative environment, which accounts for perishability, energy use (CO2 emissions), and demand uncertainty. According to their experiments, the cost benefit from cooperation varies in a range of about 4-24%, while the aggregated total emission benefit varies in a range of about 8-33%. Collaboration potentials in stochastic systems has also been assessed by Sprenger and Monch (2012). The authors investigate a real-world scenario found in the German food industry, where products are sent from manu- facturers to customers via intermediate distribution centers. They are the first to show that the cooperative strategy clearly outperforms the non- cooperative algorithms in a dynamic and stochastic logistics system. A large-scale VRPTW is obtained for the delivery of the orders, capacity con- straints, maximum operating times for the vehicles, and outsourcing options. This problem is decomposed into rich VRP sub problems and solved by an algorithm based on ant colony systems. The proposed heuristics is tested in a rolling horizon setting using discrete event simulation. Quintero-Araujo et al. (2016) discuss the potential benefits of collabo- rations in supply chains with stochastic demands. A simheuristic approach is used to compare cooperative and non-cooperative scenarios. The authors find costs reduction around 4% with values rising up to 7.3%. Yilmaz and Savasaneril (2012) study the collaboration of small shippers in the presence of uncertainty. This problem focuses on markets where shippers have random transportation requests for small-volume shipments. The AP, where the coalition decides where to assign an arriving shipper and when to dispatch a vehicle, is proposed and modeled as a Markov decision process. Its performance is compared to a naive and a myopic strategy. The authors find, for instance, that when it is costly to pickup and deliver the shipments to consolidation points, the naive policy is outperformed by the policy of the coalition. While the majority of papers finds that horizontal collaborations can improve the non-cooperative solution by around 20-30%, some authors re- port collaboration profits outside of this range (e.g. Krajewska et al., 2008; 15 Sanchez et al., 2016; Quintero-Araujo et al., 2016). Adenso-D´ıaz et al. (2014a) contribute to this issue by investigating the impact of coalition sizes. Their computational experiments show that benefits are marginally decreas- ing with the size of the partnership. 4.2. Methodological contributions In centralized planning problems, there are several decisions that have to be taken. Typically, not only the routing but also the assignment of cus- tomers to depots has to be considered. In order to approximate optimal solutions even for large real world instances, many authors propose decom- position strategies (e.g. Dai and Chen, 2012; Nadarajah and Bookbinder, 2013; Buijs et al., 2016). While a popular assumption is that in horizontal collaborations the VRP is the underlying planning problem, also collaborative ARP or MCFP have been investigated. Fern´andez et al. (2016) introduce the collaboration un- capacitated ARP. This yields to a profitable ARP, where carriers have cus- tomers that they are not willing or allowed to share, and others that can be exchanged with collaborators. The model is formulated as integer linear program and solved through a branch-and-cut algorithm. The optimal hub routing problem of merged tasks is investigated by Weng and Xu (2014). This problem allows all requests to pass up to two hubs within limited distance. The underlying problem is formulated as multi-depot ARP. Solutions are generated using two heuristics based on Lagrangian relaxation and Benders decomposition. The time-dependent centralized multiple carrier collaboration problem is introduced by Hern´andez and Peeta (2011). The authors assume a setting where carriers either provide or consume collaborative capacity. Capacities are time-dependent but known a priori, and demand is fixed. The problem is modeled as a binary multi-commodity MCFP and solved using a branch-and-cut algorithm. Liu et al. (2010a) define the multi-depot ca- pacitated ARP aiming for a solution with minimized empty movements of truckload carriers. A two-phase greedy algorithm is presented to solve prac- tical large-scale problems. Not only carriers, but also shippers can conduct horizontal collaborations. In this case, they jointly identify sets of lanes that can be submitted to a carrier as attractive bundles. The goal is to offer tours with little or no asset repositioning to carriers. In return, they can get more favorable rates from the carriers. Ergun et al. (2007b) define the shipper collaboration problem, which is formulated as LCP. Solutions are generated by a greedy algorithm. 16 In Kuyzu (2017) the LCP model is extended by a constraint on the number of partners with whom the collaborative tours must be coordinated. Column generation and Branch-and-price approaches are developed for the solution of the resulting LCP variant. In collaborative vehicle routing, typically horizontal cooperations are con- sidered. These refer to collaborative practices among companies acting at the same levels in a market (Cruijssen et al., 2007c). Vertical cooperation on the other hand, indicate hierarchical relationships, meaning that one player is the client of the other. Wang et al. (2014) were the first to present a combination of horizontal and vertical cooperation. They extend the pickup and delivery problem with time windows (PDPTW) to a combination of integrated (verti- cal) and collaborative (horizontal) transportation planning, where both sub- contracting and collaborative request exchange are taken into account. The centralized planning problem is solved using adaptive large neighborhood search (ALNS) and an ALNS-based iterative heuristic. 4.3. Future research directions The area of CGA has been extensively researched. The cost advantages of centralized collaborations have been quantified in several studies, most of them finding potential benefits of 20-30%. Also ecological goals, like re- duction of emissions, have been taken into account. However, most of these studies assume deterministic scenarios. Literature assessing collaboration potentials, when the central authority faces uncertainties, is scarce. Also collaboration gains in more complex, e.g. multi-modal, multi-depot trans- portation systems have yet to be investigated. Centralized authorities typically face huge and highly complex optimiza- tion problems, since they have to plan operations for several interconnected fleets. Thus, sophisticated solution techniques are required. There is a vast field of problems and methods that have not been investigated so far from a collaborative perspective. It could, for instance, be investigated, how a central authority exchanges requests among collaborators, while trying not to redistribute too much. This would lead to a 2-objective problem, which minimizes (i) total cost and (ii) deviation from the decentralized solution. A related question is how the central authority can motivate participants to reveal their data. These incentives might be provided by using smart profit sharing mechanisms or, e.g., side payments. To answer these questions, stud- ies investigating how much information has to be revealed in order to achieve 17 reasonable collaboration profits would be helpful. Finally, since central de- cision makers face huge optimization problems, the application of solution methods for large scale VRP (e.g. Kytojoki et al., 2007) are supposed to fur- ther improve solution quality. For this purpose, advanced processing methods like parallel computing should be taken into account (Ghiani et al., 2003). Also machine learning concepts might be valuable tools to, e.g., tune pa- rameters (Birattari, 2009) or to automatically identify problem structures in collaborative settings. 5. Decentralized planning without auctions If players are not willing to give full information to a central planner, decentralized approaches are needed. In such a decentralized setting col- laborators might cooperate individually or supported by a central authority, which does not have full information. Articles in this area contribute either to the issue of • selecting appropriate collaboration partners, we refer to this as Partner Selection (PS), • requests that should be offered to collaboration partners, which is re- ferred to as Request Selection (RS), • methods for exchanging requests, which is denoted Request Exchange (RE). In Table 2, we categorize all papers dealing with decentralized non-auction- based approaches and give their main characteristics. 18 n o i t a r o b a l l o c d e z i l a r t n e c e d d e s a b - n o i t c u a - n o n r o f s c i t s i r e t c a r a h c c i s a b d n a s e c n e r e f e R : 2 e l b a T D P W T t n e m p h S i l e d o M r e i r r a C r e p p i h S n o i t u b i r t n o c n i a M e c n e r e f e R x x x x x x x x x x x x x x x x x x x L T F L T F L T L L T L L T F L T F L T L L T L L T F L T F L T L L T L L T L L T F P R V P R V P R V P R V P A P C L P F C M P F C M P R A P C L P R V P R V P R V P R V x x x x x x x x x x x x x x S P S R S P E R S P S R S R S R S R E R E R E R E R E R ) b 4 1 0 2 ( . l a t e z a ´ı D - o s n e d A ) 1 1 0 2 ( s g i r e D d n a l h a D ) a 7 0 0 2 ( . l a t e n u g r E ) 4 1 0 2 ( . l a t e o a D ) 1 1 0 2 ( . l a t e y e l i a B ) 6 1 0 2 ( . l a t e o v r e u C ) 4 1 0 2 ( a t e e P d n a z e d n ´a n r e H ) 1 1 0 2 ( . l a t e z e d n ´a n r e H ) 4 1 0 2 ( h c n o M d n a r e g n e r p S ) 4 1 0 2 ( r e f p o K d n a g n a W ) 5 1 0 2 ( r e f p o K d n a g n a W ) 4 1 0 2 ( . l a t e g n a W ) 1 1 0 2 ( . l a t e r e n e z O ) b 0 1 0 2 ( . l a t e i u L 19 5.1. Partner selection The benefit of collaborations of course depends on the partners that form the coalition and the characteristics of their operations. Potential partners might have different requirements, which have to be considered in the joint operational plan (Cuervo et al., 2016). Thus, Adenso-D´ıaz et al. (2014b) propose an a priori index that could be used to roughly predict synergies be- tween potential partners, without the necessity of solving any optimization model. This index is based on the transportation demands of the participat- ing companies. However, average order size and the number of orders seem to be the most influential characteristic on the coalitions profit (Cuervo et al., 2016). A model integrating partner selection and collaborative transportation scheduling is developed by Dao et al. (2014). The mathematical model ba- sically accounts for transportation times, costs, and capabilities of potential partners. 5.2. Request selection Carriers have to decide which of their requests should be offered to collab- oration partners. Typically, carriers do not want to offer all their requests, but to keep some of them to be served with their private fleet. An intuitive solution would be to let the carriers solve a team orienteering problem, and put those requests into the pool, that do not appear in the optimal tour (Archetti et al., 2014). A combination of the request selection and the rout- ing problem of collaborative truckload carriers is introduced by Liu et al. (2010b). It is assumed that carriers receive different kinds of requests, which they have to allocate to either their internal fleet or to an external collabo- rative carrier. The objective is to make a selection of tasks and to route the private vehicles by minimizing a total cost function. The authors develop a memetic algorithm to solve the problem. However, Gansterer and Hartl (2016b) show that in auction-based exchanges, the best request evaluation criteria take geographical aspects into account. It can be assumed, that these criteria are effective in non-auction-based settings as well. Of course, carriers not only decide on requests they want to offer, but on requests they want to acquire. Again, solving a team orienteering problem, which gives the set of valuable requests, would be an intuitive approach. However, this comes with a high computational effort since various different restrictions have to be considered. A request might be valuable for different players in the decentralized system. Thus, a coordinating authority has to 20 find a feasible assignment of requests to carriers. An efficient method to reduce empty backhauls by adding pickup and delivery tasks of partners is proposed by Bailey et al. (2011). Two optimization models are developed, where one is formulated as an integer program, and the other is formulated as a mixed integer program. A greedy heuristic and tabu search, are used to solve these problems. A numerical analysis based on real-world freight data indicates that the percentage of cost savings can be as high as 27%. In Hern´andez and Peeta (2014) a carrier seeks to collaborate with other carriers by acquiring capacity to service excess demand. Carriers first allo- cate requests to their private resources and then find the cost minimizing transport option for excess demand. The problem is addressed from a static perspective, formulated as a MCFP, and solved using a branch-and-cut al- gorithm. In Hern´andez et al. (2011) dynamic capacities are assumed. In vertical collaborations, beneficial requests or tours have to be selected for collaboration partners. Ergun et al. (2007a) investigate vertical collab- oration between shippers and carriers. The authors present optimization technology that can be used by shippers to identify beneficial tours with lit- tle truck repositioning. Timing considerations are a key focus of their study. The effectiveness of their algorithms is shown based on real-world data. 5.3. Request exchange Once collaboration partners and requests have been selected, players have to decide on the exchange mechanism. Due to the inherent complexity, it is not common to exchange sets of unconnected requests, but parts of existing tours. This complexity can be overcome by auction-based systems. We refer to Section 6, where auction-based mechanisms are discussed. However, in non-auctioned-based frameworks, it is reasonable to trade requests being packed in vehicle routes. Wang and Kopfer (2014) propose such an route- based exchange mechanism. Carriers can iteratively generate and submit new routes based on the feedback information from an agent. This information is deduced from the dual values of a linear relaxation of a set partitioning problem. An extension including subcontracting is presented in Wang et al. (2014). The problem becomes even more complex, if dynamics of carrier coali- tions are considered. Two rolling horizon planning approaches, which yield considerably superior results than isolated planning, are proposed by Wang and Kopfer (2015). 21 In the FTL market, typically lanes rather than requests are exchanged. Ozener et al. (2011) consider settings in which several carriers collaborate by means of bilateral lane exchanges with and without side payments. With regard to applicability of request exchange mechanisms for real- world collaborations, some decision support systems have been developed. Dahl and Derigs (2011) present an empirical analysis of the effectiveness of a collaborative decision support system in an express carrier network. They assume carriers to solve a dynamic PDPTW by a local search-based heuristic. They show that with the support of a real-time decision support system based on an adequate compensation scheme, the network is able to perform close to the level obtainable by centralized planning. A decision support system for cooperative transportation planning where several manufacturing companies share their fleets to reduce transportation costs is presented by Sprenger and Monch (2014). 5.4. Future research directions Decentralized non-auction-based systems have advantages and disadvan- tages. They are generally assumed to be less complex than auction-based approaches, since there is, for instance, no need for a bidding procedure. This might be seen as an advantage, but this comes at the price that none of the players has structured information on the collaborators' preferences. This of course leads to relatively low collaboration profits. To overcome this drawback, there are attempts to still get some information by, for instance, doing multiple rounds of exchanges in order to approximate mutual prefer- ences (e.g. Wang and Kopfer, 2014). An interesting research direction might be to compare the performance of such mechanisms with auction-based sys- tems. Also the value of sharing information with collaboration partners has not been investigated so far. 6. Auction-based decentralized planning The decentralized exchange of requests can be organized through auc- tions (e.g. Ledyard et al., 2002), where collaborators submit requests to a common pool. Due to the necessity of a trading mechanism, auctions are generally supposed to be more complex than their conventional (i.e. non- auction-based) counterparts. However, auctions have more potential, since the trading mechanism can be used to indirectly share information of collab- orators' preferences. 22 Table 3: References and basic characteristics for auction-based decentralized planning Model Shipper Carrier Shipment TW PD Reference x VRP Ackermann et al. (2011) VRP x Berger and Bierwirth (2010) x VRP Chen (2016) x VRP Dai and H.Chen (2011) VRP x Dai et al. (2014) x VRP Gansterer and Hartl (2016b) x Gansterer and Hartl (2017) VRP x Krajewska and Kopfer (2006) VRP Li et al. (2016) VRP x VRP Li et al. (2015) x VRP Schopka and Kopfer (2017) Xu et al. (2016) LCP LTL LTL LTL LTL LTL LTL LTL LTL LTL FTL LTL FTL x x x x x x x x x x x x x x x x x x In horizontal collaborations, auctions are used to exchange requests. Thus, collaborators typically have both the roles of buyers and of sellers. A central authority, which is in charge of coordinating the auction process, is called the auctioneer. In combinatorial auctions, requests are not traded individually but are combined to bundles (Pekec and Rothkopf, 2003). This is of particular im- portance in vehicle routing, where a request might not be attractive unless it is combined with other ones. An example for this can be seen in the upper part of Figure 4. Carrier B would probably not be interested in the individual requests C4 or C5, while the bundle (C4, C5) seems to be attractive. A bidding carrier receives the full bundle if the bidding price is accepted. If the bid is rejected, none of the items contained in the package is trans- ferred to this carrier. This eliminates the risk of obtaining only a subset of requests which does not fit into the current request portfolio. Table 3 lists papers on auction-based decentralized planning, and highlights their basic characteristics. An early study on auction-based horizontal transportation collaboration is presented by Krajewska and Kopfer (2006). They are the first to design an auction-based exchange mechanism for collaborating carriers. Ackermann et al. (2011) discuss various goals for a combinatorial request exchange in freight logistics. They give a complete modeling proposal and show the crucial points when designing such a complex system. Berger and Bierwirth 23 (2010) divide the auction process into 5 phases: 1. Carriers decide which requests to put into the auction pool. 2. The auctioneer generates bundles of requests and offers them to the carriers. 3. Carriers place their bids for the offered bundles. 4. Winner Determination Problem: Auctioneer allocates bundles to car- riers based on their bids. 5. Profit sharing: collected profits are distributed among the carriers. In the first phase, participating collaborators can decide either on self- fulfillment, i.e. they plan and execute their transportation requests with their own capacities, or to offer some of them to other carriers. Aiming at network profit maximization, carriers should try to offer requests that are valuable for other network participants. Otherwise, the auction mechanism will not yield improved solutions. However, the identification of requests that are valuable for collaborators is not trivial since the actors do not want to reveal sensitive information. Different selection decisions are illustrated in Figure 7, where carrier A selects traded requests based on his own preferences, while the other carriers offer requests that have a higher probability to be attractive for their collaborators. In such a system, there is of course a high risk of strategic behavior, since carriers might increase individual benefits by offering unprofitable requests, and achieving very attractive ones in return. The first auction phase is investigated by Gansterer and Hartl (2016b). The authors show that the best request evaluation criteria take geographical aspects into account. They clearly dominate pure profit-based strategies. Schopka and Kopfer (2017) investigate pre-selection strategies, which they classify as (i) request potential or (ii) tour potential valuation strategies. Li et al. (2016) assume that carriers in combinatorial auctions have to solve a PDPTW with reserved requests for deciding which requests should be sub- mitted to the auction pool. In the second phase, the requests in the pool are grouped into bundles. These are then offered to participating carriers. If it is assumed that carriers can get a set of requests that exceeds their capacities (outsourcing option), the generation of bundles can be moved to the carriers themselves. The auc- tioneer could then offer the set of requests without grouping them to bundles, while the carriers give their bids on self-created packages of requests. The obvious drawback of this approach is that the auctioneer cannot guarantee 24 Figure 7: Carriers submit request to the pool. Carrier A selects requests based on marginal profits, while carriers B and C take geographical information into account (Gansterer and Hartl, 2016a). to find a feasible assignment of bundles to carriers. This the reason why an outsourcing option has to be included. A simple method to overcome strict capacities (no outsourcing), is to assume that the auctioneer assigns at most one bundle per carrier. Since carriers give their bids on bundles, they can easily communicate whether they are able to handle a specific bundle or not. However, from a practi- cal point of view, offering all possible bundles is not manageable, since the number of bundles grows exponentially with the number of requests that are in the pool. An intuitive approach to reduce the auction's complexity is to limit the number of requests that are traded. Li et al. (2015) do not allow the carriers to submit more than one request to the auction pool. Xu et al. (2016) show effective auction mechanisms for the truckload carrier collabo- ration problem with bilateral lane exchange. Again carriers offer only one lane, which is the one with the highest marginal cost. In the multi-agent framework presented by Dai and H.Chen (2011), there is only one request traded per auction round. Thus, it is a non-combinatorial auction, where carriers act as auctioneers when they want to outsource a request to other 25 Auction pool+--+-++-ABC-+-++-C1C2-C2C1C5+B1B2B3B2B3B1+A1-A2A2A3+-+C3C4C4C5C3-A1A3A1B2B3C4C5carriers, whereas they act as bidders when they want to acquire a request from other carriers. However, limiting the number of offered items, obviously decreases the probability to find good solutions. Gansterer and Hartl (2017) show that, without a loss in solution quality, the set of offered bundles can be efficiently reduced to a relatively small subset of attractive ones. They develop a proxy function for assessing the attractiveness of bundles under in- complete information. This proxy is then used in a genetic algorithms-based framework that aims at producing attractive and feasible bundles. With only a little loss in solution quality, instances can be solved in a fraction of the computational time compared to the situation where all possible bundles are evaluated. Complexity can also be reduced by performing multi-round auctions. These are generally intended to offer subsets of the traded items in mul- tiple rounds. Previously gained information can be used to compose the setting for the next round. By this, the bidders are never faced with the full complexity of the auction pool. A multi-round price-setting based combinatorial auction approach is pro- In each round of the auction, the auctioneer posed by Dai et al. (2014). updates the price for serving each request based on Lagrangian relaxation. Each carrier determines its requests to be outsourced and the requests to be acquired from other carriers by solving a request selection problem based on the prices. Regarding the last auction phase, i.e. profit sharing, we refer to a broad survey presented in Guajardo and Ronnqvist (2016). Chen (2016) propose an alternative to combinatorial auctions for carrier collaboration, which is combinatorial clock-proxy exchange. This exchange has two phases. The clock phase is an iterative exchange based on Lagrangian relaxation. In the proxy phase, the bids that each carrier submits are deter- mined based on the information observed in the clock phase. Future research directions. Auctions can be powerful mechanisms for increas- ing collaboration profits. However, each of the 5 auction phases bears a com- plex and at least partly unsolved decision problem in itself. To make auctions efficiently applicable to real-world settings, many challenging questions still have to be answered. For instance, the strong relationship between the five auction-phases (Berger and Bierwirth, 2010) has not been investigated so far. The majority of studies focuses on one of the five decision phases, while an integrated and practically usable framework is still missing. Also the real- 26 istic aspect that carriers might behave strategically, opens many interesting research questions. In particular, the influence of strategic behavior in the request selection or in the bidding phase are yet to be investigated. At this point, effective profit sharing mechanisms are needed, since these have the potential to impede strategic behavior. So far, only relatively simple auction procedures have been investigated. It might be worth to adopt more complex mechanisms like, e.g., multi-round value-setting auctions (Dai et al., 2014). Also, in the literature there is no structured assessment of the potential of auctions in comparison to optimal solutions. These can of course only be guaranteed, if the auctioneer gets full information. Thus, it is worthwhile to investigate the value of information, i.e. the assessment of types or levels of information that increase solution quality. 7. Conclusion Collaborative vehicle routing is an active research area of high practical importance. In this review paper, we have given a structured overview and classification of the related literature. We identified three major streams of research, which are (i) centralized planning, (ii) non-auction-based decentral- ized planning, and (iii) auction-based decentralized planning. Literature was further classified based on the underlying planning problem and the collab- oration setting. We discussed recent developments and proposed future work directions, which, for instance, are • the application of collaborative frameworks to more complex, e.g. multi- modal, transportation systems, • the investigation of strategic behavior, and effective profit sharing mech- anisms to avoid it, • a comparative study assessing the advantages of auction-based com- pared to non-auction-based systems, • the assessment of the value of information in decentralized exchange mechanisms. In order to produce comparable results, an open access repository of re- lated benchmark instances would be helpful. This would enable a structured 27 investigation of performance gaps between centralized and decentralized ap- proaches. Publicly available data instances are provided by (i) Fern´andez et al. (2016) (http://or-brescia.unibs.it/instances), (ii) Wang et al. (2014), Wang and Kopfer (2015) (http://www.logistik.uni-bremen.de/, and (iii) Gansterer and Hartl (2017) (https://tinyurl.com/y85hcpry). Acknowledgements This work is supported by FWF the Austrian Science Fund (Projectnum- ber P27858-G27) References References Ackermann, H., Ewe, H., Kopfer, H., Kufer, K., 2011. Combinatorial auctions in freight logistics. In: Bose, J., Hu, H., Carlos, C., Shi, X., Stahlbock, R., Voss, S. (Eds.), Computational Logistics. Vol. 6971 of Lecture Notes in Computer Science. Springer Berlin Heidelberg, pp. 1–17. Adenso-D´ıaz, B., Lozano, S., Garcia-Carbajal, S., Smith-Miles, K., 2014a. Assessing partnership savings in horizontal cooperation by planning linked deliveries. Transportation Research Part A: Policy and Practice 66, 268 – 279. Adenso-D´ıaz, B., Lozano, S., Moreno, P., 2014b. Analysis of the synergies of merging multi-company transportation needs. Transportmetrica A: Trans- port Science 10 (6), 533–547. Agarwal, R., Ergun, O., 2010. Network design and allocation mechanisms for carrier alliances in liner shipping. Operations Research 58 (6), 1726–1742. Ankersmit, S., Rezaei, J., Tavasszy, L., 2014. The potential of horizontal collaboration in airport ground freight services. Journal of Air Transport Management 40, 169 – 181. Audy, J.-F., Lehoux, N., D'Amours, S., Ronnqvist, M., 2012. A framework for an efficient implementation of logistics collaborations. International Transactions in Operational Research 19 (5), 633–657. URL http://dx.doi.org/10.1111/j.1475-3995.2010.00799.x 28 Bailey, E., Unnikrishnan, A., Lin, D.-Y., 2011. Models for minimizing back- haul costs through freight collaboration. Transportation Research Records 2224, 51–60. Ballot, E., Fontane, F., 2010. Reducing transportation co2 emissions through pooling of supply networks: perspectives from a case study in french retail chains. Production Planning & Control 21 (6), 640–650. Berger, S., Bierwirth, C., 2010. Solutions to the request reassignment prob- lem in collaborative carrier networks. Transportation Research Part E: Logistics and Transportation Review 46, 627–638. Bertazzi, L., Savelsbergh, M., Speranza, M. G., 2008. Inventory Routing. Springer US, Boston, MA, pp. 49–72. Birattari, M., 2009. Tuning Metaheuristics. Buijs, P., Alvarez, J. A. L., Veenstra, M., Roodbergen, K. J., 2016. Improved collaborative transport planning at dutch logistics service provider fritom. Interfaces 46 (2), 119 – 132. Chen, H., 2016. Combinatorial clock-proxy exchange for carrier collaboration in less than truck load transportation. Transportation Research Part E: Logistics and Transportation Review 91, 152 – 172. Cruijssen, F., Braysy, O., Dullaert, W., Fleuren, H., Salomon, M., 2007a. Joint route planning under varying market conditions. International Jour- nal of Physical Distribution & Logistics Management 37 (4), 287–304. Cruijssen, F., Cools, M., Dullaert, W., 2007b. Horizontal cooperation in logistics: Opportunities and impediments. Transportation Research Part E: Logistics and Transportation Review 43 (2), 129 – 142. Cruijssen, F., Dullaert, W., Fleuren, H., 2007c. Horizontal cooperation in transport and logistics: A literature review. Transportation Journal 46 (3), 22–39. Cuervo, D. P., Vanovermeire, C., Sorensen, K., 2016. Determining collabo- rative profits in coalitions formed by two partners with varying character- istics. Transportation Research Part C: Emerging Technologies 70, 171 – 184. 29 Dahl, S., Derigs, U., 2011. Cooperative planning in express carrier networks an empirical study on the effectiveness of a real-time decision support system. Decision Support Systems 51 (3), 620 – 626. Dai, B., Chen, H., 2012. Mathematical model and solution approach for carriers collaborative transportation planning in less than truckload trans- portation. International Journal of Advanced Operations Management 4, 62–84. Dai, B., Chen, H., Yang, G., 10 2014. Price-setting based combinatorial auc- tion approach for carrier collaboration with pickup and delivery requests. Operational Research 14 (3), 361–386. Dai, B., H.Chen, 2011. A multi-agent and auction-based framework and ap- proach for carrier collaboration. Logistics Research 3 (2-3), 101–120. Dao, S. D., Abhary, K., Marian, R., 2014. Optimisation of partner selection and collaborative transportation scheduling in virtual enterprises using {GA}. Expert Systems with Applications 41 (15), 6701 – 6717. Engevall, S., Gthe-Lundgren, M., Vrbrand, P., 2004. The heterogeneous vehicle-routing game. Transportation Science 38 (1), 71–85. Ergun, O., Kuyzu, G., Savelsbergh, M., 05 2007a. Reducing truckload trans- portation costs through collaboration. Transportation Science 41 (2), 206– 221. Ergun, O., Kuyzu, G., Savelsbergh, M., 2007b. Shipper collaboration. Com- puters & Operations Research 34 (6), 1551 – 1560, part Special Issue: Odysseus 2003 Second International Workshop on Freight Transportation Logistics. Fern´andez, E., Fontana, D., Speranza, M. G., 2016. On the collaboration uncapacitated arc routing problem. Computers & Operations Research 67, 120 – 131. Frisk, M., Goethe-Lundgren, M., Joernsten, K., Ronnqvist, M., 2010. Cost allocation in collaborative forest transportation. European Journal of Op- erational Research 205, 448–4587. 30 Gansterer, M., Hartl, R. F., 2016a. Combinatorial auctions in collaborative vehicle routing. IFORS News (10(4)), 15–16. Gansterer, M., Hartl, R. F., 2016b. Request evaluation strategies for carriers in auction-based collaborations. OR Spectrum 38(1), 3–23. Gansterer, M., Hartl, R. F., 2017. Bundle generation in combinatorial trans- portation auctions. Working paper. URL http://prolog.univie.ac.at/research/PaperMG/Bundles.pdf Gendreau, M., Potvin, J.-Y., Braumlaysy, O., Hasle, G., Løkketangen, A., 2008. Metaheuristics for the Vehicle Routing Problem and Its Extensions: A Categorized Bibliography. Springer US, Boston, MA, pp. 143–169. Ghiani, G., Guerriero, F., Laporte, G., Musmanno, R., 2003. Real-time vehi- cle routing: Solution concepts, algorithms and parallel computing strate- gies. European Journal of Operational Research 151 (1), 1 – 11. Ghiani, G., Manni, E., Triki, C., 2008. The lane covering problem with time windows. Journal of Discrete Mathematical Sciences and Cryptography 11 (1), 67–81. Gothe-Lundgren, M., Jornsten, K., Varbrand, P., 1996. On the nucleolus of the basic vehicle routing game. Mathematical Programming 72 (1), 83–100. Guajardo, M., Jornsten, K., 2015. Common mistakes in computing the nu- cleolus. European Journal of Operational Research 241 (3), 931 – 935. Guajardo, M., Ronnqvist, M., 2015. Operations research models for coali- tion structure in collaborative logistics. European Journal of Operational Research 240 (1), 147 – 159. Guajardo, M., Ronnqvist, M., 2016. A review on cost allocation methods in collaborative transportation. International Transactions in Operational Research 23 (3), 371–392. Hern´andez, S., Peeta, S., 2011. Centralized time-dependent multiple-carrier collaboration problem for less-than-truckload carriers. Transportation Re- search Record: Journal of the Transportation Research Board 2263, 26–34. 31 Hern´andez, S., Peeta, S., 2014. A carrier collaboration problem for less-than- truckload carriers: characteristics and carrier collaboration model. Trans- portmetrica A: Transport Science 10 (4), 327–349. Hern´andez, S., Peeta, S., Kalafatas, G., 2011. A less-than-truckload carrier collaboration planning problem under dynamic capacities. Transportation Research Part E: Logistics and Transportation Review 47 (6), 933 – 946. Kimms, A., Kozeletskyi, I., 2016. Shapley value-based cost allocation in the cooperative traveling salesman problem under rolling horizon planning. EURO Journal on Transportation and Logistics 5 (4), 371–392. Klein, M., 1967. A primal method for minimal cost flows with applications to the assignment and transportation problems. Management Science 14 (3), 205–220. Krajewska, M., Kopfer, H., 2006. Collaborating freight forwarding enter- prises. OR Spectrum 28(3), 301–317. Krajewska, M. A., Kopfer, H., Laporte, G., Ropke, S., Zaccour, G., 11 2008. Horizontal cooperation among freight carriers: request allocation and profit sharing. The Journal of the Operational Research Society 59 (11), 1483–1491. Kuo, A., Miller-Hooks, E., Zhang, K., Mahmassani, H., 2008. Train slot co- operation in multicarrier, international rail-based intermodal freight trans- port. Transportation Research Record: Journal of the Transportation Re- search Board 2043, 31–40. Kuyzu, G., 2017. Lane covering with partner bounds in collaborative truck- load transportation procurement. Computers & Operations Research 77, 32 – 43. Kytojoki, J., Nuortio, T., Braysy, O., Gendreau, M., 2007. An efficient vari- able neighborhood search heuristic for very large scale vehicle routing prob- lems. Computers & Operations Research 34 (9), 2743 – 2757. Ledyard, J., Olson, M., Porter, D., Swanson, J., Torma, D., 2002. The first use of a combined-value auction for transportation services. Interfaces 32 (5), 4–12. 32 Li, J., Rong, G., Feng, Y., 2015. Request selection and exchange approach for carrier collaboration based on auction of a single request. Transportation Research Part E: Logistics and Transportation Review 84, 23 – 39. Li, Y., Chen, H., Prins, C., 2016. Adaptive large neighborhood search for the pickup and delivery problem with time windows, profits, and reserved requests. European Journal of Operational Research 252 (1), 27 – 38. Lin, C., 2008. A cooperative strategy for a vehicle routing problem with pickup and delivery time windows. Computers & Industrial Engineering 55 (4), 766 – 782. Liu, R., Jiang, Z., Fung, R. Y., Chen, F., Liu, X., 2010a. Two-phase heuris- tic algorithms for full truckloads multi-depot capacitated vehicle routing problem in carrier collaboration. Computers & Operations Research 37 (5), 950 – 959, disruption Management. Liu, R., Jiang, Z., Liu, X., Chen, F., 2010b. Task selection and routing prob- lems in collaborative truckload transportation. Transportation Research Part E: Logistics and Transportation Review 46 (6), 1071 – 1085. Lydeka, Z., Adomavicius, B., 2007. Cooperation among the competitors in international cargo transportation sector: Key factors to success. Engi- neering Economics 51 (1), 80–90. Montoya-Torres, J. R., Munoz-Villamizar, A., Vega-Mejia, C. A., 2016. On the impact of collaborative strategies for goods delivery in city logistics. Production Planning & Control 27 (6), 443–455. Munkres, J., 1957. Algorithms for the assignment and transportation prob- lems. Journal of the Society for Industrial and Applied Mathematics 5 (1), 32–38. URL http://www.jstor.org/stable/2098689 Nadarajah, S., Bookbinder, J., 2013. Less-than-truckload carrier collabo- ration problem: Modeling framework and solution approach. Journal of Heuristics 19, 917–942. Ozener, O. O., Ergun, O., Savelsbergh, M., 2011. Lane-exchange mechanisms for truckload carrier collaboration. Transportation Science 45 (1), 1–17. 33 Ozener, O. O., Ergun, O., Savelsbergh, M., 2013. Allocating cost of service to customers in inventory routing. Operations Research 61 (1), 112–125. Parragh, S., Dorner, K., Hartl, R., 2008. A survey on pickup and delivery problems. part ii: Transportation between pickup and delivery locations. Journal fur Betriebswirtschaft 58, 21–51. Pekec, A., Rothkopf, M., 2003. Combinatorial auction design. Management Science 49 (11), 1485–1503. P´erez-Bernabeu, E., Juan, A. A., Faulin, J., Barrios, B. B., 2015. Horizontal cooperation in road transportation: a case illustrating savings in distances and greenhouse gas emissions. International Transactions in Operational Research 22 (3), 585–606. Quintero-Araujo, C. L., Gruler, A., Juan, A. A., 2016. Quantifying Poten- tial Benefits of Horizontal Cooperation in Urban Transportation Under Uncertainty: A Simheuristic Approach. Springer International Publishing, Cham, pp. 280–289. Sanchez, M., Pradenas, L., Deschamps, J.-C., Parada, V., 2016. Reducing the carbon footprint in a vehicle routing problem by pooling resources from different companies. NETNOMICS: Economic Research and Elec- tronic Networking 17 (1), 29–45. Schmeidler, D., 1969. The nucleolus of a characteristic function game. SIAM Journal on Applied Mathematics 17, 1163–1170. Schmelzer, H., 2014. Cooperational platform for urban logistics in Zurich (accessed January 2017). URL https://blog.zhaw.ch/mobine/category/1/city-logistik/ Schmoltzi, C., Wallenburg, C. M., 07 2011. Horizontal cooperations between logistics service providers: Motives, structure, performance. International Journal of Physical Distribution & Logistics Management 41 (6), 552–575. Schneeweiss, C., 2003. Distributed Decision Making. Schopka, K., Kopfer, H., 2017. Pre-selection Strategies for the Collabora- tive Vehicle Routing Problem with Time Windows. Springer International Publishing, Cham, pp. 231–242. 34 Shapley, L., 1953. A value for n-person games. Annals of Mathematical Stud- ies 28, 307–317. Soysal, M., Bloemhof-Ruwaard, J. M., Haijema, R., van der Vorst, J. G., 2016. Modeling a green inventory routing problem for perishable products with horizontal collaboration. Computers & Operations Research, –. Sprenger, R., Monch, L., 2012. A methodology to solve large-scale coopera- tive transportation planning problems. European Journal of Operational Research 223 (3), 626 – 636. Sprenger, R., Monch, L., 2014. A decision support system for cooperative transportation planning: Design, implementation, and performance as- sessment. Expert Systems with Applications 41 (11), 5125 – 5138. Tate, K., 1996. The elements of a successful logistics partnership. Interna- tional Journal of Physical Distribution & Logistics Management 26 (3), 7–13. Vanovermeire, C., Sorensen, K., 2014. Integration of the cost allocation in the optimization of collaborative bundling. Transportation Research Part E: Logistics and Transportation Review 72, 125 – 143. Verdonck, L., Caris, A., Ramaekers, K., Janssens, G. K., 2013. Collaborative logistics from the perspective of road transportation companies. Transport Reviews 33 (6), 700–719. Verstrepen, S., Cools, M., Cruijssen, F., Dullaert, W., 2009. A dynamic framework for managing horizontal cooperation in logistics. International Journal of Logistics Systems and Management 3 4 (5), 228–248. Voruganti, A., Unnikrishnan, A., Waller, S., 2011. Modeling carrier collabo- ration in freight networks. Transportation Letters 3 (1), 51–61. Wang, X., Kopfer, H., 2014. Collaborative transportation planning of less- than-truckload freight. OR Spectrum 36, 357–380. Wang, X., Kopfer, H., 2015. Rolling horizon planning for a dynamic collab- orative routing problem with full-truckload pickup and delivery requests. Flexible Services and Manufacturing Journal 27 (4), 509–533. 35 Wang, X., Kopfer, H., Gendreau, M., 2014. Operational transportation plan- ning of freight forwarding companies in horizontal coalitions. European Journal of Operational Research 237 (3), 1133 – 1141. Weng, K., Xu, Z.-H., 2014. Flow merging and hub route optimization in collaborative transportation. Journal of Applied Mathematics 2014. Wøhlk, S., 2008. A Decade of Capacitated Arc Routing. Springer US, Boston, MA, pp. 29–48. Xu, S. X., Huang, G. Q., Cheng, M., 2016. Truthful, budget-balanced bundle double auctions for carrier collaboration. Transportation Science 0 (0), null. Yilmaz, O., Savasaneril, S., 2012. Collaboration among small shippers in a transportation market. European Journal of Operational Research 218 (2), 408 – 415. 36
1008.2160	1	1008	2010-08-12T16:17:54	An early warning method for crush	[ "cs.MA" ]	Fatal crush conditions occur in crowds with tragic frequency. Event organisers and architects are often criticised for failing to consider the causes and implications of crush, but the reality is that the prediction and mitigation of such conditions offers a significant technical challenge. Full treatment of physical force within crowd simulations is precise but computationally expensive; the more common method of human interpretation of results is computationally "cheap" but subjective and time-consuming. In this paper we propose an alternative method for the analysis of crowd behaviour, which uses information theory to measure crowd disorder. We show how this technique may be easily incorporated into an existing simulation framework, and validate it against an historical event. Our results show that this method offers an effective and efficient route towards automatic detection of crush.	cs.MA	cs	An early warning method for crush Peter J. Harding1, Steve M. V. Gwynne2 and Martyn Amos1,∗ 1 Dept. of Computing and Mathematics, Manchester Metropolitan University, UK. 2 Hughes Associates, Inc., USA. ∗ Corresponding author: [email protected] Abstract Fatal crush conditions occur in crowds with tragic frequency. Event organizers and architects are often criticised for failing to consider the causes and implications of crush, but the reality is that the prediction and mitigation of such conditions offers a significant technical challenge. Full treatment of physical force within crowd simulations is precise but compu- tationally expensive; the more common method of human interpretation of results is computationally "cheap" but subjective and time-consuming. In this paper we propose an alternative method for the analysis of crowd behaviour, which uses information theory to measure crowd disorder. We show how this technique may be easily incorporated into an existing sim- ulation framework, and validate it against an historical event. Our results show that this method offers an effective and efficient route towards au- tomatic detection of crush. 1 Introduction Overloading pedestrian routes can quickly lead to the development of crush con- ditions, as observed in the Hillsborough [20], Station nightclub [4] and Saudi Arabian Hajj [8] incidents, as well as the recent Love Parade tragedy in Ger- many. A more sophisticated understanding of how crush conditions form is therefore critical for the design of tall buildings and other highly-populated, contained regions (such as ships, nightclubs and stadia), as well as for the plan- ning of events and formulation of incident management procedures. A first step towards this deeper understanding is a method for detecting the early-stage formation of crush, which is the problem we address here. The study of crowd evacuation/control scenarios has taken on additional significance in the light of events such as 9/11 . Many tall buildings (such as the World Trade Center towers) were designed alongside the assumption that any necessary evacuation could and would be conducted in a phased manner (e.g. floor-by-floor). One significant factor in building design is the capacity of exit routes (such as corridors and stairwells). Capacities are calculated based 1 on projections of controlled population movement in phased evacuations. If the phased evacuation assumption breaks down (if, for example, occupants of a specific floor refuse to wait their "turn" for fear of catastrophic building failure) then this will have severe implications for overall safety, as exit routes become overloaded. Computer-based simulation studies are often used to analyse the movement of individuals in various scenarios. Such work encompasses the study of histor- ical events [8], the examination of evacuation procedures [6], and the design of aircraft [3]. Existing simulation frameworks include EXODUS [17], PEDFLOW [14] and EVACNET [12] (see [15] for an extensive review), and these offer a range of "real world" features, including exit blockage/obstacles, occupant im- patience and route choice [7]. However, the phenomenon of crush is one that has received relatively little attention so far from the designers of evacuation simulations. Many simulations do not explicitly consider the effects of crush, and those that do factor in crush employ computationally expensive physical force calculations. The two major problems we address are as follows: firstly, the consideration of crush within existing simulation frameworks requires the use of computation- ally intensive Newtonian force calculations. These can drastically slow down simulations, restricting their applicability in the rapid prototyping of building designs and crowd control procedures. The second problem is that the mon- itoring of crush within real crowds is rudimentary, at best, and relies largely on personal observation and interpretation of crowd patterns. This method of crush detection is inherently problematic. We therefore seek a method for the detection of crush conditions that is relatively "cheap" in terms of computational effort, and which may be easily integrated into existing software for crowd monitoring. Such a method will have a significant impact on both simulation-based evacuation studies and real- time analysis of video images (facilitating, for example, the development of automated crush alarms based on CCTV images). In this paper we give a description of our proposed method, which is based on the notion of phase transitions in a system of interacting particles. We show how our method may be easily integrated into an existing simulation framework, and test it using details of an historical event. Our results show that mutual information provides an excellent "early warning" indicator of the emergence of crush conditions. The rest of this paper is organised as follows: in Section 2 we first define the notion of "crush conditions" , and examine how crush has been handled by previous simulation studies. This motivates the search for a new crush detection method, and we show in Section 3 how the concepts of phase transition and mutual information might usefully be applied to the detection (and prediction) of crush. We describe the results of experimental investigations in Section 4, and conclude in Section 5 with a discussion of open questions. 2 2 The problem of crush We first consider the notion of crush conditions. As Fruin observes [2], over- crowding can often lead to injuries and/or fatalities; these may be caused by trampling or falls, but here we are concerned with the particularly common phenomenon of compressive asphyxia (also known as chest compression), which, Fruin argues, is responsible for "virtually all crowd deaths" [2]. This occurs when the torso is compressed by external forces, preventing expansion of the lungs and thus interfering with normal breathing. Difficulty in breathing due to intense pressure levels can often be exacerbated by anxiety and heat, quickly leading to significant physiological problems. Fatal levels of force can emerge within a crowd as a result of pushing, leaning or (less commonly) vertical stacking of bodies. Images of steel barriers bent out of shape (for example, in the aftermath of the Hillsborough disaster [20]) graphically illustrate the extent to which force levels can grow. Fruin reports the results of several studies (either after-the-event forensic tests, or controlled experiments) which suggest that forces exceeding around 1500N could prove fatal [2]. It is therefore an important factor to be considered in simulation studies aimed at improving structural designs or evacuation/control procedures, along with other aspects such as panic or physical obstacles. Crush detection methods used to date in simulation studies may be classified into two generic groups; explicit methods and implicit methods citeharding2008a. The implicit methodology is the traditional approach, and is still highly popular, being the preferred technique in a large number of simulation models (see [16] for an extensive review). It relies on the expert analysis of factors such as population density and environmental considerations, yielding a human interpretation of the output of the simulation to help determine whether or not crush might have occurred. Although subjective, this method is still popular, because it does not require the use of computationally expensive force calculations, relying instead on human expertise and intuition. The explicit modelling of crush conditions incorporates an assessment of crush into the model itself, and therefore requires less human analysis than the implicit approach. Usually based on the calculation of Newtonian force values, and operating in 2-dimensional space, explicit methodologies are used to detect the presence of crush conditions in a much more objective fashion. By simulating the physical force exerted by each individual, they calculate the precise amount of force present within a crowd. Whilst the explicit methodologies offer a measure of the forces acting within a crowd, the calculations needed to assess levels of force require much more computer processing power than an implicit method. Experiments show that the computation time required by a model that explicitly quantifies force can be up to 100 times greater than that required by an implicit model [18]. Given the nature of the current trade-off between precision and computa- tional cost, we therefore seek a relatively "cheap" method (in terms of run time) that will allow us to automatically signal the onset of crush conditions within an evacuation. This will bridge the gap between the two current extremes, allowing 3 architects and policy-makers to quickly and easily incorporate crush into their simulation scenarios. In the next Section, we explain how this may be achieved using Mutual Information. 3 Mutual information for the detection of crush While studying video footage of the 2006 Saudi Arabian Hajj disaster, in which over 340 pilgrims died as the result of a stampede, Helbing et al. noticed distinct transitions in the flow of pedestrians around the time of the significant incident. They observed "a sudden transition from laminar to ... unstable flows" [8]; that is, a sudden "flip" from smooth to irregular flows of human movement. Such transitions are, we believe, key to the early detection of crush, and we now describe our proposed methodology for their detection. Our proposal is that the onset of crush can be detected via the analysis of crowd behaviour. More specifically, by identifying sustained periods of disorder, we may identify the possible onset of crush. By treating analysing this change in observable behaviour using information theory, we qualify the onset of crush conditions without ever explicitly calculating the amount of force present in the simulation. Within a simulation, the two distinct states of a crowd are characterised by the behaviour of individuals. Under "normal" conditions, crowd flow is highly ordered, with the orientation and speed of a specific individual being similar to that of those in their immediate locality. The onset of more turbulent flow sees individuals exhibit a marked change in behaviour, as they change speed and alter course in order to avoid others. We therefore wish to identify these distinct states, and we achieve this by applying statistical analysis techniques to the movement of individuals within crowds. 3.1 Mutual Information Mutual Information (MI) is a probabilistic method for quantifying the inter- dependence of two variables. It has previously been employed as an analytical technique in many areas [1, 11, 21]. More recently, it has been shown that MI may be used to identify a kinetic phase transition in a complex, dynamical system of interacting particles [23]. It is therefore possible to reliably identify the point at which certain particle-based systems move away from a disordered state and begin to exhibit some degree of order [24], and vice versa (this is the phase transition). In the general case, the Mutual Information of two discrete time-series vari- ables, A and B, is defined as: I(A, B) = (cid:88) i,j p(ai, bj) logn p(ai, bj) p(ai)p(bj) (1) where p(ai), p(bj), and p(ai, bj) are the individual probability and joint prob- ability distributions of A and B. In general terms, MI quantifies the interde- 4 pendence of two variables; therefore if A and B are entirely independent, then I(A, B) = 0, but in all all other cases I(A, B) > 0. In the next Section, we show how MI may be integrated with an existing simulation framework, to provide an entirely new metric for the analysis of pedestrian evacuation. 4 Experimental investigations In this Section we describe the results of experiments to investigate the applica- bility of MI as a plausible tool for crush detection. In order to ensure its broad applicability, we first show how MI may be easily integrated into an existing, industry-standard simulation framework. We then validate the technique, by using it to analyse an historical event. By demonstrating that the MI technique correctly detects known incidences of crush within this scenario, we provide support for its adoption as a standard tool. The base simulation environment used is the Fire Dynamics Simulator (FDS) [19], a fluid dynamics-based model of fire and smoke flow. The FDS+Evac module [13] is an evacuation simulation extension for FDS, and is based on the well-known social forces model [9, 10] (SFM) of pedestrian movement. The evacuation module for FDS incorporates the calculation of physical forces, negating the need for additional functionality in this respect. The MI analysis was integrated into the FDS environment as a set of natively coded (FORTRAN 90) libraries. As the technique is entirely passive, i.e. it will not affect the results of the evacuation, there were no concerns regarding the effect this could have on the behaviour of the simulations (although there is clearly a small overhead incurred by the MI calculations). The MI of the system is calculated at every simulation time step, and the results averaged over 100 time steps before being recorded. This equates to one MI reading per second of real-life evacuation time, which gives sufficient granularity. We record the average physical force within a simulation in the same way. 4.1 Experimental validation In order to validate the technique, we choose a well-documented incident that illustrates the significant hazards that an emergency evacuation may present. In 2003, the Station Nightclub (Rhode Island, USA) was the scene of one the worst nightclub fires in recent history, when a pyrotechnic device, used by the rock band Great White, ignited sound insulation foam in the walls and ceiling of the venue. According to the official report into the incident [4], a crush formed at the main escape route within 90 seconds of the start of the fire, trapping patrons inside the club as it filled with smoke. Estimates of the nightclub occupancy vary between 440 and 460; a total of 96 people died during the incident. We select this particular event on the basis of (a) the existence of a sig- nificant amount of professional film footage taken inside the nightclub during 5 Figure 1: (Top) Floorplan of Station nightclub, taken from official report. (Bot- tom) Rendering in FDS+Evac. the incident1, (b) availability of supporting witness evidence and other asso- ciated documentation, and (c) results from substantial simulation tests using FDS as part of the subsequent (extensively documented) formal investigation. We therefore have information on the initial distribution of individuals at the beginning of the incident, visual evidence of crush during the incident, and the final locations of each of the victims, as well as a set of validated simulations with which to compare our own results. We begin by rendering the floor plan of the Station in FDS, using official architectural plans taken from [4] (Figure 1). We use a figure of 450 for the number of agents to be simulated, and their initial distribution is specified 1Ironically, the film crew was present to record a documentary on nightclub safety, after a fatal incident elsewhere four days previously. 6 according to [4] (i.e., with high crowd densities in the Dancefloor and Sunroom areas, and lower densities in other areas). We run two sets of experiments; the first, idealised set is designed to provide baseline evacuation data, and the second set replicates, as closely as possible, the conditions and events in the nightclub during the event. Investigation findings into the spread of the fire suggest that the Stage door became impassable 30 seconds from the start of the incident, so we reflect this fact in our simulation by closing that exit after that period has elapsed. The official investigation was able to identify the exit paths for 248 of the 350 people who escaped from the building. The distribution of evacuees through the three other available exit routes was found to be non-uniform, with estimates of between one-half and two- thirds of patrons attempting to leave via the familiar main exit, rather than the under-utilised (and less familiar) main bar and Kitchen doors. Reports suggest that only 12 people left via the Kitchen door during the evacuation. In order to simulate this distribution of path choices, patrons are assigned a probability of knowledge for each exit route. Exactly 12 evacuees are made aware of the existence of the Kitchen exit, and of the remaining patrons, 100% are given knowledge of the main door, 50% are given knowledge of the main bar door, and 25% are given knowledge of the stage door. On the other hand, the idealised evacuation was structured as follows: there was no blocking of the Stage door, and agents in the simulation had full knowledge of all exit routes. This scenario represents the minimum time it would take to evacuate 450 people from the Station Nightclub, with optimum use made of available exit structures and no hindrance from fire, smoke, or unfavourable environmental conditions. We compare our simulation results with those obtained by the National In- stitute of Standards and Technology (NIST), and detailed in the official investi- gation report [4]. In these experiments, NIST investigators used both Simulex [22] and buildingEXODUS [5] to evaluate both idealised and realistic evacua- tion scenarios. The results obtained were very similar for both packages, so we concentrate on the buildingEXODUS output. Within the "realistic" simulation, occupants were instructed to always select the nearest exit, and the Stage door was also closed after 30 seconds. In the NIST simulation, 91 simulated occupants left via the building front door, which is precisely the number reported in the official investigation. Thirty-five simulated occupants used either the platform door or the kitchen door, which, again, is consistent with the evidence. We therefore conclude that the official NIST simulations provide a sound basis for validating our own simulations. The results of the comparison are depicted in Figure 2. We note only that the results obtained (in terms of leaving profiles over time) are very similar to those reported by NIST, which supports the argument in favour of the soundness of our model. 7 Figure 2: Comparison of leaving profiles between our simulation (FDS) and official NIST results. 4.2 Detection of crush Having established the validity of our simulation in terms of broad outcomes, the next stage is to specifically investigate the emergence of crush, and to see if this is easily detectable using Mutual Information. In order to achieve this, we measure the average force and the level of MI within our simulated population of 450 individuals, for both "real" and "idealised" evacuations. We first consider the results of the force measurements, comparing them with evidence from the investigation. The force measurements for both scenarios are depicted in Figure 3. Across both scenarios the levels of force initially increase as the evacuation commences, but it rapidly decays during the idealised version of events, since evacuees are more uniformly distributed. Force levels drop to zero at around 175s, when everyone has left the building, which is broadly in line with the findings of the NIST idealised situation simulation (195s ± 7s). In the "real" scenario, we observe a sharp initial rise in average force, which initially peaks after around 65 seconds. This is directly in line with the findings of the official investigation, which states that a significant crowd crush occurred by the main entrance (where around a third of the fatalities occurred) at the beginning of the time period 71-102 seconds into the fire. Prior to 1-1/2 minutes into the fire, a crowd-crush occurred in the front vestibule which almost entirely disrupted the flow through the main exit. Many people became stuck in the prone position in the exterior double doors [4, p. xx]. The camera angle shifts away from this door after 0:07:33 (0:01:11 fire time) and does not return to the front door until 0:08:04 (0:01:42 fire time). When the camera returns at 0:08:04 (0:01:42 fire time) a pile-up of occupants is visible. Details regarding how the pile-up 8 0 50 100 150 200 250 300 350 400 450 0 50 100 150 200 250 300 350 400# of Patrons EvacuatedTime (s)FDS IdealNIST IdealFDS ActualNIST ActualFigure 3: Comparison of average force between real and idealised scenarios. occurred are not available from the WPRI-TV video; however, the interruption in flow of evacuating occupants apparent [in Figure 6-3] supports the contention that the disruption may have initiated early during the 31 second period when the camera was pointed elsewhere. [4, p. 182] In Figure 4, we show a screenshot of the simulation after 65 seconds, which graphically illustrates the significant crush around the main entrance and sun- room area (high levels of force are shown in red). The analysis of MI during evacuation is performed using only observable variables, i.e. those with values that could be obtained via direct observation of an evacuation. This is to ensure that our results were not implementation specific, and to maximise the possibility of applying the technique in future to other environments or video-captured data from real-life evacuations. There- fore, the three variables considered for analysis are the 2-dimensional Cartesian coordinates (xi and yi) of each individual, i, together with their heading (Θi). We forego the use of speed within our analysis, as there is often little variation in speed during incidents with high population density. We measure MI using Equation 2, taken from [24]: (cid:88) (cid:88) i,j I(X, Θ) = I(Y, Θ) = I = 2 9 p(xi, θj) log2 p(xi, θj) p(xi)p(θj) p(yi, θj) p(yi)p(θj) p(yi, θj) log2 i,j I(X, Θ) + I(Y, Θ) (2) 0 100 200 300 400 500 600 0 50 100 150 200 250 300 350 400Force (N/m)Time (s)IdealisedActualFigure 4: Screenshot of our fire scenario simulation after 65 elapsed seconds. Our MI measurements are depicted in Figure 5. We expect to see, as the simulations begin, an initial rise in the MI of the system. As evacuees prepare to exit the structure they tend towards alignment, exhibiting similar escape trajec- tories to other evacuees in their locale. In a maximally efficient evacuation this period of high order (and high MI) would be sustained throughout, as evacuees would not alter their course in order to increase their chances of effective egress. However, in an evacuation with a great deal of competition, the order in the system quickly breaks down, as the evacuees reposition themselves in order to increase their probability of escape. MI may therefore may be used as an order parameter, where falling values of MI signify the breakdown of order within a specific evacuation. We observe marked quantitative differences in the MI read- ings between the two simulations. During periods of disorder, MI should tend towards zero, whereas, during ordered segments of the evacuation, MI will rise significantly. 4.3 Idealised scenario In the idealised simulation, we see a sharp initial peak, as individuals all make for the exits at the same time. We then observe a drop, as the evacuees begin to compete for the available exit capacity. An increase in order is seen as one exit route begins to clear, creating the rise in MI at 50 < t < 75, falling back into a state of disorder as the final evacuees clear this (main bar) exit . The MI reading then shows a progressive rise as the final evacuees exit the structure. The sharp drop in MI at the end of the simulation occurs when the number of remaining evacuees falls below some (very low) threshold. 10 Figure 5: Comparison of Mutual Information between idealised and actual sce- narios. 4.4 Realistic scenario The MI readings obtained from the simulation of actual events show a far more disordered evacuation, with an initial rise in MI (signifying order) quickly dis- integrating into disorder. The MI reading at t ≈ 50s approaches zero; this period of highly disordered evacuation remains as the exits to the structure are overwhelmed (see Figure 4). The exit rate of evacuees during this period is extremely low, which is confirmed by the exit profiles (see Figure 2). The MI level slowly rises towards the end of the evacuation, but, notably, the higher levels of order seen in the idealised evacuation are not reached until t ≈ 300s, 5 minutes after the start of the evacuation. 4.5 Correlation analysis We then perform a correlation analysis in order to establish the relationship (if any) between force and Mutual Information. A scatterplot of force versus MI suggests the existence of a statistical association (Figure 6), so we perform a simple linear correlation test. The results of this are as follows: P = 2.2e−16 Rp = −0.571 The P-value obtained is much lower than the standard significance level for a two tailed test (α = 0.01), (P (cid:28) α), which confirms the significance of the result. The correlation coefficient, Rp = −0.571, confirms that there exists a negative correlation between MI and force within an evacuation scenario. 11 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 0 50 100 150 200 250 300 350 400I (bits)Time (s)IdealisedActualFigure 6: Scatterplot of Force versus Mutual Information. 5 Discussion and Conclusions In this paper we have described a novel technique for the analysis of crowd evacuation scenarios. By calculating the Mutual Information of a system of in- teracting individuals, we are able to determine the level of internal force present within a crowd. We have shown that consistently low levels of Mutual Infor- mation are correlated with high levels of force within a crowd. This method removes the need for computationally expensive physical force calculations, and allows planners to quickly and easily incorporate objective measures of crowd disorder and crush into their simulation scenarios. Future work will focus on refinements of the technique, as well as investigation of its "real-world" applica- bility. We are particularly interested in the potential for using our technique to analyse real-time video images, with the eventual aim of developing an on-site automatic early warning system for crush and disorder at large-scale events. References [1] Andrew M. Fraser and Harry L. Swinney. Independent coordinates for strange attractors from mutual information. Physical Review A, 33:1134 -- 1140, 1986. [2] John J. Fruin. The causes and prevention of crowd disasters. In First International Conference on Engineering for Crowd Safety. Elsevier, 1993. [3] E.R. Galea, S.J. Blake, S. Gwynne, and P.J. Lawrence. The use of evac- uation modelling techniques in the design of very large transport aircraft and blended wing body aircraft. The Aeronautical Journal of the Royal Aeronautical Society, pages 207 -- 218, April 2003. 12 0 100 200 300 400 500 0 0.2 0.4 0.6 0.8 1 1.2Force (N/m)MI (bits)Data PointsLine of Best Fit[4] W. Grosshandler, N. Bryner, D. Madrzykowski, and K. Kuntz. Report of the Technical Investigation of the Station Nightclub fire. Technical report, National Institute of Standards and Technology, USA, 2005. Available at http://www.nist.gov/ncst/. [5] S. Gwynne, ER Galea, PJ Lawrence, and L. Filippidis. Modelling occu- pant interaction with fire conditions using the buildingEXODUS evacuation model. Fire Safety Journal, 36(4):327 -- 357, 2001. [6] S. Gwynne, E.R. Galea, C. Lyster, and I. Glen. Analysing the evacuation procedures employed on a thames passenger boat using the maritimeEXO- DUS evacuation model. Fire Technology, 39(3):225 -- 246, 2003. [7] Steve Gwynne, E. R. Galea, M. Owen, P. J. Lawrence, and L. Filippidis. A review of the methodologies used in the computer simulation of evacuation from the built environment. Building and Environment, 34:741 -- 749, 1999. [8] D. Helbing, A. Johansson, and H.Z. Al-Abideen. Dynamics of crowd dis- asters: An empirical study. Physical Review E, 75(4):46109, 2007. [9] D. Helbing and P. Moln´ar. Social force model for pedestrian dynamics. Physical Review E, 51(5):4282 -- 4286, 1995. [10] Dirk Helbing, Illes Farkas, and Tamas Vicsek. Simulating dynamical fea- tures of escape panic. Nature, 407:487 -- 490, 2000. [11] Jaeseung Jeong, John C. Gore, and Bradley S. Peterson. Mutual informa- tion analysis of the EEG in patients with alzheimer's disease. Neurophysi- ology, 112:827 -- 835, 2001. [12] TM Kisko and RL Francis. EVACNET+: A computer program to deter- mine optimal building evacuation plans. Fire Safety Journal, 9(2):211 -- 220, 1985. [13] T. Korhonen, S. Hostikka, S. Heliovaara, H. Ehtamo, and K. Matikainen. FDS+ Evac: Evacuation module for fire dynamics simulator. In Proceedings of the Interflam2007: 11th International Conference on Fire Science and Engineering, pages 1443 -- 1448, 2007. [14] R. Kukla, J. Kerridge, A. Willis, and J. Hine. PEDFLOW: Development of an Autonomous Agent Model of Pedestrian Flow. Transportation Research Record, 1774:11 -- 17, 2001. [15] E.D. Kuligowski. Review of 28 egress models. In R.D. Pea- cock and E.D. Kuligowski, editors, Workshop on Building Occu- pant Movement During Fire Emergencies, pages 68 -- 90. National Institute of Standards and Technology, Gaithersburg, MD, 2005. http://fire.nist.gov/bfrlpubs/fire05/PDF/f05008.pdf. 13 [16] Erica D. Kuligowski and Richard D. Peacock. A review of building evacu- ation models. Technical report, National Institute of Standards and Tech- nology, USA, 2005. [17] M. Owen, E.R. Galea, and P.J. Lawrence. The EXODUS evacuation model applied to building evacuation scenarios. Journal of Fire Protection Engi- neering, 8(2):65, 1996. [18] Michael J. Quinn, Ronald A. Metoyer, and Katharine Hunter-Zaworski. Parallel implementation of the social forces model. In Proceedings of the Second International Conference in Pedestrian and Evacuation Dynamics, pages 63 -- 74, 2003. [19] N.L. Ryder, J.A. Sutula, C.F. Schemel, A.J. Hamer, and V.V. Brunt. Con- sequence modeling using the Fire Dynamics Simulator. Journal of Haz- ardous Materials, 115(1-3):149 -- 154, 2004. [20] L.J. Taylor. The Hillsborough Stadium Disaster. Final Report on Enquiry by Rt. Hon. Justice Taylor, London, UK: Home Office, HMSO, 1989. [21] Philippe Th´evanez. Optimization of mutual information for multiresolution image registration. IEEE Transactions on Image Processing, 9(12):2083 -- 2099, 2000. [22] P.A. Thompson and E.W. Marchant. Testing and application of the com- puter model SIMULEX. Fire Safety Journal, 24(2):149 -- 166, 1995. [23] T. Vicsek, A. Czirok, E. Ben-Jacob, I. Cohen, and O. Schochet. Novel type of phase transition within a system of self-driven particles. Phys. Rev. Lett, 75:1226 -- 1229, 1995. [24] R. T. Wicks, S. C. Chapman, and R. O. Dendy. Mutual information as a tool for identifying phase transitions in dynamical complex systems with limited data. Physical Review E, 75(5), 2007. 14
1808.02720	1	1808	2018-08-08T10:52:47	Memetic Algorithm-Based Path Generation for Multiple Dubins Vehicles Performing Remote Tasks	[ "cs.MA" ]	This paper formalizes path planning problem for a group of heterogeneous Dubins vehicles performing tasks in a remote fashion and develops a memetic algorithm-based method to effectively produce the paths. In the setting, the vehicles are initially located at multiple depots in a two-dimensional space and the objective of planning is to minimize a weighted sum of the total tour cost of the group and the largest individual tour cost amongst the vehicles. While the presented formulation takes the form of a mixed-integer linear program (MILP) for which off-the-shelf solvers are available, the MILP solver easily loses the tractability as the number of tasks and agents grow. Therefore, a memetic algorithm tailored to the presented formulation is proposed. The algorithm features a sophisticated encoding scheme to efficiently. In addition, a path refinement technique that optimizes on the detailed tours with the sequence of visits fixed is proposed to finally obtain further optimized trajectories. Comparative numerical experiments show the validity and efficiency of the proposed methods compared with the previous methods in the literature.	cs.MA	cs	Memetic Algorithm-Based Path Generation for Multiple Dubins Vehicles Performing Remote Tasks Doo-Hyun Cho∗ and Han-Lim Choi† August 9, 2018 Abstract This paper formalizes path planning problem for a group of heteroge- neous Dubins vehicles performing tasks in a remote fashion and develops a memetic algorithm-based method to effectively produce the paths. In the setting, the vehicles are initially located at multiple depots in a two- dimensional space and the objective of planning is to minimize a weighted sum of the total tour cost of the group and the largest individual tour cost amongst the vehicles. While the presented formulation takes the form of a mixed-integer linear program (MILP) for which off-the-shelf solvers are available, the MILP solver easily loses the tractability as the number of tasks and agents grow. Therefore, a memetic algorithm tailored to the presented formulation is proposed. The algorithm features a sophisticated encoding scheme to efficiently. In addition, a path refinement technique that optimizes on the detailed tours with the sequence of visits fixed is proposed to finally obtain further optimized trajectories. Comparative numerical experiments show the validity and efficiency of the proposed methods compared with the previous methods in the literature. 1 Introduction Recent decades have observed significant signs of progress in research on au- tomation/autonomy of unmanned vehicles in many different aspects such as mission planning, resource allocation, motion coordination, path planning, low- level control, sensing, and communication [1, 2]. In particular, multi-agent aspects of a group of unmanned vehicles have been studied to enhance mission performance and resource utilization [3], particularly allowing for heterogene- ity in agent capabilities and characteristics [4, 5, 6]. One crucial decision to fully take advantage of the extended capability of heterogeneous multiple au- tonomous vehicles is to design paths/tours for the agents in such a way that ∗ † Ph.D. Candidate, Department of Aerospace Engineering, KAIST. Associate Professor, Department of Aerospace Engineering, KAIST. 1 optimizes a certain mission performance metric. Limitations in agent motion, payload, and energy [7, 8] makes such decision making necessary in practice at the same time incurs complications in the problem with different kinds of constraints. The problem of this work's interest is to design paths for (unmanned) vehi- cles to complete all the tasks in the mission area. While this type of decision making can naturally be approached in the framework of traveling salesman problem (TSP), the particular problem of this paper features a few more com- plications/sophistications: (a) tasks can be done in a remote manner -- in other words, a task can be treated as done if an agent just passes nearby; (b) agents are subject to non-holonomic motion constraints; (c) agent capabilities are po- tentially heterogeneous, and (d) both the total travel cost and load balancing of workload are considered as the performance metric. The first factor significantly increases the size of decision space -- since an agent can perform a task by passing through many different nearby points and these different options would incur a differing amount of cost, it is not only dif- ficult but also impossible in many cases to define a finite-dimensional decision space. A typical way of handling this indefinite/continuous decision space is to discretize/approximate the problem with the notion of sampling [9, 10]. In other words, instead of considering all the nearby points around the task, generating a finite number of sample nodes and focus on the solutions passing through those sample nodes. This allows for the adoption of richer solution schemes developed in the literature, as then the problem belongs to the category of generalized TSP. While the handling of arbitrary non-holonomic motion constraints requires cal- culation of optimal control solutions to determine agent paths, Dubins vehicle model can be adopted if the agent speed can be regarded as constant and the cost metric is the travel time or the path length. A TSP for a Dubins vehicle can be treated as an asymmetric TSP, which is still much more complicated than the original TSP. Multi-agent aspects of the problem necessitate the extension of the aforementioned framework, particularly giving rise to the discussion on the choice of objective function to be optimized. The heterogeneity of agent ca- pabilities incurs additional complexity -- agent to task compatibility often serves as a constraint in the problem and agent speed and maneuverability often affects the cost calculation. While each of the aforementioned complicating aspects has been addressed in the literature as summarized in Section 2, there has been lit- tle work that combined all of these aspects in a systematic fashion. This paper newly suggests to formalize a variant of the traveling salesman problem, termed generalized, heterogeneous, multi-depot, asymmetric traveling salesman problem (GHMDATSP) to deal with the aforementioned aspects that are meaningful in practice. The key contributions of this work are threefold. First, this work presents a mixed-integer linear program (MILP) formulation that allows for handling real- world instances of GHMDATSP, full version of which has not been presented in the literature. The formulation builds upon sampling-based discretization in GTSP and the Dubins vehicle mode, but particularly takes advantage of the notion of necessarily intersecting neighborhood (NIN), which was first in- 2 troduced in the authors' earlier work [11], to exclude inefficient tours from the solutino space with consideration of non-holonomic motion constraints. Second, a memetic algorithm (MA) tailored to the GHMDATSP formulation that com- putes an optimized tours for a given sent of discretized specification is newly devised; then, a path refinement procedure that further optimizes on the sample nodes is presented to compute a provably better solution than the MA solution. Third, the proposed methods are verified through extensive numerical results in which the performance and computational time are superior to the results from other previously published heuristics. The rest of the paper is organized as follows. Section 2 discusses the relevant literature. Section 3 describes in detail the notation and assumptions used in the problem, and a formulation of the GHMDATSP using a mixed integer linear programming with additional valid constraints in a detail. The novel memetic algorithm to get a near-optimal solution is shown in Section 4. Computational results are provided and discussed in Section 5. Finally, Section 6 gives the conclusion of this study. 2 Literature Review The problem of path planning for multiple unmanned vehicles has been studied in a variety of engineering fields, using numerous kinds of methods. It is beyond the scope of this paper to cover all of the existing studies, so we focus on the traveling salesman problem (TSP) and its variants that are directly related to the problem which is suggested in this work, and the approaches to path planning for nonholonomic vehicles. 2.1 Traveling Salesman Problem and Its Variants There have been tremendous research efforts on the TSP related to the path planning [12, 13]. A classic TSP is a problem of finding a visiting order that minimizes the cost of a closed path (or Hamiltonian path) that visits all the list of tasks exactly once given the locations and the distance between each location [14]. It is a problem that does not consider the motion constraints of the vehicle; in other words, it is a suitable model only for a holonomic vehicle that has no restrictions on its movement. In most cases, however, the motion constraint of the vehicle necessarily exists because of the inherent kinematic characteristic. In particular, a fixed-wing type aerial vehicle only moves forward and can control the change of direction only within a limited range. Therefore, when applying the results of the classical TSP to such a nonholonomic vehicle, a large error can occur if the vehicle cannot follow the given path due to its constraints. The Dubins TSP, a variant of the TSP, assumes that the motion constraints of the vehicle follow the Dubins model [15] when solving the path planning problem [16, 10, 17, 9, 18]. In addition, the classic TSP assumes that the vehicle reaches the specified points exactly, but this is rarely required in real-world applications. It is natural 3 and desirable to perform tasks slightly off the designated location when the tasks are related to data collection such as monitoring (e.g., crop detection and pollution measurement), event detection (e.g., fire and flood detection), and target tracking (e.g., surveillance and reconnaissance) [19, 20, 21, 22, 23]. A model that can be used appropriately in the above situations is a variant of TSP called generalized TSP (GTSP) or TSP with neighborhoods (TSPN) [24]. GTSP is a generalization of TSP where tasks are substituted with areas or sets of points. In the latter case, GTSP can be formulated by creating several sample nodes in the neighborhood area of previously defined tasks. The tasks are assumed to be completed when the solution path visits at least one of the task's nodes. In addition, if the processing time of the task is short, it can be assumed that the vehicle has performed the task by simply passing near the task, without visiting the sample node. It is observed that a TSP with a concept called the intersecting neighborhood (or the necessarily intersecting neighborhood) generated a more effective path than a general TSP when generating a vehicle's path in a situation where remotely executable tasks are densely located [10, 11]. In addition to the GTSP, the TSP has been extended in many ways, one of which is the multiple TSP (MTSP) [25, 26] where multiple vehicles collab- oratively visit the points to fulfill the mission. Similar to the TSP, the MTSP has several variants. Each vehicle can start its tour with a designated starting point, denoted as the depot. If every tour originates from the same point, this problem is called the single depot MTSP. Otherwise, the problem is called the multiple depot MTSP (MDMTSP) [24] if each tour can originate from different points. Another interesting variant of the MTSP is to construct the objective func- tion using min-max [27, 28]. In other words, the objective function is not to minimize the sum of the costs for all the tours, but to minimize the largest of the tour costs allocated to each vehicle. When modeling the objective function in this way, a tour is assigned to each vehicle with almost equal cost. The cost is proportional to time in most cases, so the min-max can be interpreted as minimizing the time required to complete the entire mission. However, when the objective function is set to the min-max, it is often confirmed that the solver is focused on the maximum vehicle cost and generates an unnecessary tour for the rest of the vehicles. Furthermore, the MTSP with vehicles with different characteristics is re- garded as the heterogeneous MTSP [29, 17, 30]. The word heterogeneous can be applied in the sense that the vehicles can differ in the motion constraints from different structures, structural heterogeneity, or different task, functional heterogeneity, due to the sensor characteristics. To our knowledge, mathematical formulations and solution methods for the generalized, heterogeneous, multi-depot, asymmetric traveling salesmen prob- lem (GHMDATSP) have never been studied. The GHMDATSP can be consid- ered as a generalization of the generalized multi-depot traveling salesmen prob- lem (GMDTSP) [31] or heterogeneous multi-depot traveling salesmen problem (HMDTSP) [32, 6], which are known to be NP-Hard. 4 2.2 Approaches to the Problems with Motion Constraints Most of the studies on the aerial vehicle path planning use the Dubins vehicle model for simplicity, and we also approach the problem with the assumption that vehicles obey the above model. The approaches to the DTSP or DTSP with neighborhoods (DTSPN) can be broadly categorized as follows. The first class represents decoupling methods that determine the heading angle of each task after determining the visiting order of the given list of tasks. The second represents transformation methods in which several heading angles for each task are sampled and then the problem is converted to the asymmetric TSP (ATSP). The methods in the third class formulate the problem in the form of mixed inte- ger linear programming, and obtain the optimal solution with exact algorithms such as a branch-and-bound algorithm or a branch-and-cut algorithm. The last class represents the methods that exploit evolutionary techniques, such as the genetic algorithm (GA). The most basic approach to the DTSP is the alternating algorithm (AA) presented in [33]. In the AA, the solution of the classic TSP is fixed as a visiting order. Then, the headings of odd-numbered tasks are set to make the straight line segment with each next even-numbered task, and the remaining parts are set to the optimal Dubins paths. The upper bound of the solution obtained through the AA is known as LTSPκ(cid:100)n/2(cid:101)πρ where LTSP is the cost of the optimal solution of the T SP , κ < 2.658, n is the number of the tasks, and ρ is the minimum turning radius. Similar but slightly more advanced than the AA algorithm, the look-ahead algorithm was presented in [34] to determine the heading of each location with using three successive points. The general steps of the transformation methods are as follows. First, the locations and distances are used to generate a complete graph which represents the original problem. The graph is then converted into the form of the ATSP. After solving the ATSP-formmated graph using a state-of-the-art solver, the output is converted to the original format to obtain the final solution. Much research has been done on the transformation methods, and this is one of the major reasons to take advantage of the existing ATSP solver's superior perfor- mance. In [17], the graph for the heterogenous, multiple depot, and multiple traveling salesman problem (HMDMTSP) was constructed for situations where each of several vehicles has a different turning radius and the heading of each location is given an arbitrary value, and then the graph is converted into the form of the ATSP using the Noon-Bean transformation. In [9], the DTSPN problem was converted into the GTSP with disjoint node sets by generating a number of sample nodes for each task, and finally it was converted into the ATSP. The DTSPN problem was converted to the GTSP in [10] as it was done in [9], and the concept called an intersecting neighborhood was added to handle the densely located tasks efficiently. Similar to [10], the TSPN was handled in [11] by creating a sampling based roadmap, but distances were calculated based on the optimal control approach rather than limiting the dynamics of the vehicle as in the Dubins model. They also borrowed the idea called intersecting neighborhood from [10], modified it to improve the performance and called as 5 the necessarily intersecting neighborhood. The GHMDATSP, which we solved in this paper, was handled in [35] using the transformation method and the necessarily intersecting neighborhood. In each study, the solution of the ATSP was obtained using the LKH algorithm [36]. An exact algorithm has been used to analyze the mathematical character- istics of the problem or to obtain the optimal solution of the instances. The mathematical formulation and its branch-and-cut algorithm for the generalized multiple depot multiple traveling salesmen problem was described in [31]. Sim- ilar to the above, the HMDMTSP with the Dubins vehicle or the Reeds-Shepp vehicle model was analyzed with the suggested branch-and-cut algorithm in [30]. Ideally, the exact algorithm would be most advantageous in terms of optimality, but the size of an instance is quite limited due to its inherent limitations. To overcome the scalability issue, evolutionary computational methods such as the genetic algorithm (GA) and memetic algorithm (MA) have been used in various studies. The possibility of task assignment to multiple unmanned aerial vehicles was confirmed through pure GA in [3]. In [37], the GA was used after the DTSPN with a single vehicle was converted into the GTSP. An MA, which combines the pure GA and the local search strategy to find the local optimal heading in each task region, was suggested in [38] to solve the DTSPN with multiple Dubins vehicles. 3 Statement of the GHMDATSP 3.1 Dubins Vehicle Model xk = vk cos θk, If a fleet of unmanned aerial vehicles is a fixed-wing type, it can be assumed that each vehicle follows a Dubins vehicle dynamics where the vehicles can only move forward. Dubins path refers to the shortest curve in the 2-dimensional Euclidean plane which connects initial and terminal points with given tangents. The mathematical model of this system is as follows: yk = uk. The subscript k ∈ K = {1,··· , m} denotes the index vk sin θk, and θk = γk vk of a vehicle, and m is the number of vehicles in a fleet to be coordinated. xk and yk pair is the position of a vehicle k in a 2-D plane; vk is the speed of a vehicle k, and every vk is a constant during the entire scenario instance; θk, θk are the angular velocity and heading angle of the vehicle k. uk is the control input of the vehicle k to change the heading which varies from -1 to 1. Negative and positive values indicate left and right turns, respectively. If the vehicle takes coordinate turns in a level flight, a normalization constant of the control input max,k − 1 where g is a gravitational acceleration and l2 γk can be assumed as g lmax,k is a maximum load factor of a vehicle k. In this study, the value of uk is set as -1, 0, or 1 to make the length of a curve as short as possible. (cid:113) 6 3.2 Notations for the GHMDATSP The goal of the GHMDATSP is to find tours that minimize some global cost when a list of tasks and a heterogeneous fleet of vehicles are given. Let T = {1,··· , n} be a set of tasks to be visited, and D = {(n + 1)1,··· , (n + 1)m, (n + 2)1,··· , (n + 2)m} be a set of depots and terminals for each vehicle (the initial and final locations of vehicles); we have a heterogeneous fleet of m unmanned vehicles initially located at their own depots. K = {1,··· , m} is used as the notation of a set of vehicles. (n + 1) and (n + 2) in D denote the depot and terminal, respectively. Each cardinality of T , D, and K is denoted as T = n, D = 2m, and K = m. in the instance is denoted by V =(cid:83) In a typical GTSP, a set of sample nodes belonging to the same task is referred to as a cluster. For each task, depot, and terminal, a cluster, which is a set of sample nodes, are independently created and assigned to each vehicle t where k ∈ K and task. A cluster for vehicle k and task t is denoted by V k and t ∈ T ∪ D, V k denotes a union of every task clusters for vehicle k, and Vt denotes a union of every vehicle clusters for task t. A set of every sample node t . Given V , we can define a set of directed edges E between sample nodes belonging to different clusters. Note that there is no edge between sample nodes of different vehicles. For any given nonempty subset S ⊂ V and sample node s, t(S) and t(s) are a set of tasks or a task to which the sample nodes in S or s belong, respectively. We define the ith sample node in the cluster V k t . In the remainder of this paper, we use s ∈ V k t,i for the sake of brevity if possible. t as V k t,i ∈ V k k∈K,t∈T∪D V k instead of V k t The problem can be formulated on a directed graph G = (V, E). Like the general TSP problem, the defined G does not include a self-loop. For each directed edge (s, s(cid:48)) ∈ E when the sample nodes belong to the vehicle k, cs,s(cid:48) is the cost for vehicle k of traversing from s to s(cid:48). The definition of cost depends on the purpose of the mission, but can generally be defined as the time or distance traveled by the vehicle along the path or the amount of fuel used. The cost can be calculated with the model in Section 3.1. 3.3 Necessarily Intersecting Neighborhoods We use the concept called necessarily intersecting neighborhoods (NIN) pro- posed in [11], which is an extension of the Intersecting Regions Algorithm [10]. An instance is illustrated in Figure 1 for the explanation of the NIN. This in- stance has a total of three tasks, and sample nodes s1 and s2 belong to task t1; sample nodes for t2 and t3 are omitted for simplicity. Node s1 and s2 are located at the boundary of the task's neighborhood region Nt1, and the direction of each node is set toward t1. The vehicle has motion constraints based on the Dubins model, assuming that the minimum turning radius is rmin. If the vehicle visits the sample node s1, it necessarily passes through the region of tasks t2 and t3 due to its motion constraints. Similarly, the vehicle necessarily passes t2 if it is set to visit s2. The following is how to verify whether a sample node s passes through an 7 Figure 1: An example of a vehicle passing neighborhoods nec- essarily when entering the sample nodes s1 or s2. In this paper, a set SNIN t arbitrary task t. Draw two circles with a radium rmin that tangent to s, and then check that a region of t intersects both circles simultaneously. If the above is satisfied, the vehicle necessarily passes t when visiting s, or in other words, the neighborhood of t can be expressed as a neighborhood which necessarily are created for each task t intersects s. and sample node s respectively. In the case of SNIN , every sample node except the nodes which belong to the task cluster t is included in SNIN if s necessarily is {s1}. Similarly, in the is {s1, s2}, and SNIN intersects t. In Figure 1, SNIN t3 case of T NIN except the task which already has the sample node s if t is necessarily intersected by s. Therefore, T NIN is {t2, t3}, and T NIN , every task t is included in T NIN is {t2} in Figure 1. and T NIN s s s2 s1 t2 t s t 3.4 Problem Formulation variable x, a sum of edge cost for vehicle k, Costk, is(cid:80) For the formulation, we define the binary decision variables x, y, and yNIN. An element of x, xs,s(cid:48) is defined for each edge (s, s(cid:48)) ∈ E, whose value equals 1 if it is chosen as an element of a tour solution and 0 otherwise. Using the s,s(cid:48)∈V k cs,s(cid:48) · xs,s(cid:48). An element of y, ys is defined whose value is equal to 1 if the node s is visited by a vehicle. Similarly, an element of yNIN, yNIN is defined to be 1 if sample node s necessarily intersects task t. t,s 8 (cid:1872)(cid:2869)(cid:1840)(cid:3047)(cid:3117)(cid:1870)(cid:3040)(cid:3036)(cid:3041)(cid:1872)(cid:2871)(cid:1840)(cid:3047)(cid:3119)(cid:1870)(cid:3040)(cid:3036)(cid:3041): Minimum turning radius(cid:1840): Neighborhood of a task(cid:1872): Task(cid:1871): Sample node(cid:1871)1(cid:1871)2(cid:1872)(cid:2870)(cid:1840)(cid:3047)(cid:3118)The problem is formulated as follows: (cid:0)(cid:80) Minimize: α subject to (cid:1) (cid:33) m (cid:32) (cid:95) s∈NINt ∨ yNIN t,s = ys ys s∈Vt (cid:32)(cid:88) (cid:88) (cid:88) (cid:88) s(cid:48)∈V k s∈V k t s(cid:48)∈V \S ys = 1 xs(cid:48),s + xs,s(cid:48) = 2ys T ∪D\{t(s)} xs(cid:48),s + xs,s(cid:48) ≥ 2ys k∈K Costk + (1 − α) max k∈K Costk (1) (cid:33) yNIN t,s = 1 ∀t ∈ T, s ∈ NINt (2) ∀t ∈ T (3) ∀k ∈ K, t ∈ D (4) ∀k ∈ K, s ∈ V k (5) Vt, s ∈ S (6) ∀S ⊆ (cid:91) t∈T The objective function in Eq. (1) is defined to minimize the linear combi- nation of two terms: a) the mean cost of vehicles and b) the maximum cost of a vehicle from a fleet. The α in Eq.(1) is a coefficient that determines which of the above two terms to focus more on to optimize the problem. Instead of using the total cost sum at the first term, the mean value is used to normalize the size with respect to the second term. Constraints (2) bind all the tasks associated with sample node s. If the sample node s is visited so the value equals 1, then all of the necessarily intersecting tasks related to s are assumed to be visited. Constraints (3) ensure that each target is visited by some vehicle directly vis- iting the generated sample node or indirectly visiting with the concept of NIN. Constraints (4) imply that one of the sample nodes in the depot or the terminal cluster for each vehicle has to be chosen. Constraints (5) is called the degree constraints. If a sample node s of a vehicle k is chosen to be visited and the value of ys equals 1, the in-degree and out-degree of the node is 1 respectively. In other words, one of the edges towards the node s and one of the outward edges from the node s should be selected. Constraints in (6) prevents the generation of subtours of any subset of tasks for each vehicle. 4 Memetic Algorithm based Path Generation The classical genetic algorithm (GA) is a process that repeatedly evolves a population of chromosomes (or solutions) through operators such as selection, crossover, and mutation, and finally obtains a high-quality solution at the end of the iteration (or generation). While borrowing the methodology used in the classical GA, a procedure we propose in this paper consists of modified GA operators, local heuristic methods for chromosomes, and the path refinement 9 Figure 2: A schematic of the memetic algorithm based path generation proce- dure. process. The schematic of the path generation procedure is shown in Figure 2, including related subsection index for each block. The following describes each operator that makes up the procedure. 4.1 Encoding and Decoding Chromosome encoding is the most important part of the memetic algorithm. The genes that make up the chromosome are real numbers made up of the sum of integer and fractional parts. The integer part stores the task cluster index, and the fractional part stores information about the sample node index. We Figure 3: A simple instance of the GHMDATSP. 10 Generate new populationInitial PopulationFitness EvaluationElitisicmImmigrationSelection and CrossoverLocalHeuristicCheck Termination CriteriaNoYesPopulation management4.24.14.5Best solutionPath Refinement4.64.34.44.5Task 1Task 2Task 3Task 4Task 5𝑉(cid:2869),(cid:2869)(cid:2869)𝑉(cid:2870),(cid:2871)(cid:2869)𝑉(cid:2871),(cid:2871)(cid:2869)𝑉(cid:2872),(cid:2870)(cid:2870)𝑉(cid:2873),(cid:2869)(cid:2870)Depot 1𝑉(cid:3005),(cid:2869)(cid:2869)𝑉(cid:3021),(cid:2871)(cid:2869)𝑉(cid:3005),(cid:2870)(cid:2870)𝑉(cid:3021),(cid:2869)(cid:2870)Depot 2(a) An instance with a given (b) Updating with gene 1.1. chromosome. (c) Updating with gene 2.1. (d) Updating with gene 3.1. (e) Updating with gene 4.1. (f) Updating with gene 5.1. (g) Deleting V 1 2,1. (h) Deleting V 2 5,1. (i) Deleting V 1 3,1. Figure 4: An example of decode-NIN operator. 11 Task 1Task 2Task 3Task 4Task 5(cid:1848)(cid:2869),(cid:2869)(cid:2869)(cid:1848)(cid:2870),(cid:2869)(cid:2869)(cid:1848)(cid:2872),(cid:2869)(cid:2870)(cid:1848)(cid:2873),(cid:2869)(cid:2870)Depot 1(cid:1848)(cid:3005),(cid:2869)(cid:2869)(cid:1848)(cid:3021),(cid:2869)(cid:2869)(cid:1848)(cid:3005),(cid:2869)(cid:2870)(cid:1848)(cid:3021),(cid:2869)(cid:2870)Depot 2M4.25.16.211.12.13.1M.114.15.1M.11Chromosome :Task #Value1121314151Basket(cid:1848)(cid:2871),(cid:2869)(cid:2869)Task 1Task 2Task 3Task 4Task 5(cid:1848)(cid:2869),(cid:2869)(cid:2869)Depot 1(cid:1848)(cid:3005),(cid:2869)(cid:2869)(cid:1848)(cid:3021),(cid:2869)(cid:2869)(cid:1848)(cid:3005),(cid:2869)(cid:2870)(cid:1848)(cid:3021),(cid:2869)(cid:2870)Depot 2(cid:1848)(cid:2869),(cid:2869)(cid:2869)Veh. 1 :Task #Value1122324151Basket(cid:1848)(cid:3005),(cid:2869)(cid:2869)Task 1Task 2Task 3(cid:1848)(cid:2869),(cid:2869)(cid:2869)Depot 1(cid:1848)(cid:3005),(cid:2869)(cid:2869)(cid:1848)(cid:3021),(cid:2869)(cid:2869)(cid:1848)(cid:2869),(cid:2869)(cid:2869)Veh. 1 :Task #Value1222324151Basket(cid:1848)(cid:3005),(cid:2869)(cid:2869)(cid:1848)(cid:2870),(cid:2869)(cid:2869)(cid:1848)(cid:2870),(cid:2869)(cid:2869)Task 4Task 5(cid:1848)(cid:3005),(cid:2869)(cid:2870)(cid:1848)(cid:3021),(cid:2869)(cid:2870)Depot 2Task #Value1223324151BasketTask 4Task 5(cid:1848)(cid:3005),(cid:2869)(cid:2870)(cid:1848)(cid:3021),(cid:2869)(cid:2870)Depot 2Task 1Task 2Task 3(cid:1848)(cid:2869),(cid:2869)(cid:2869)(cid:1848)(cid:2870),(cid:2869)(cid:2869)Depot 1(cid:1848)(cid:3005),(cid:2869)(cid:2869)(cid:1848)(cid:3021),(cid:2869)(cid:2869)(cid:1848)(cid:2869),(cid:2869)(cid:2869)(cid:1848)(cid:2870),(cid:2869)(cid:2869)(cid:1848)(cid:2871),(cid:2869)(cid:2869)(cid:1848)(cid:3021),(cid:2869)(cid:2869)Veh. 1 :(cid:1848)(cid:3005),(cid:2869)(cid:2869)(cid:1848)(cid:2871),(cid:2869)(cid:2869)Task #Value1223324152BasketTask 4Task 5(cid:1848)(cid:3005),(cid:2869)(cid:2870)(cid:1848)(cid:3021),(cid:2869)(cid:2870)Depot 2Task 1Task 2Task 3(cid:1848)(cid:2869),(cid:2869)(cid:2869)(cid:1848)(cid:2870),(cid:2869)(cid:2869)Depot 1(cid:1848)(cid:3005),(cid:2869)(cid:2869)(cid:1848)(cid:3021),(cid:2869)(cid:2869)(cid:1848)(cid:2869),(cid:2869)(cid:2869)(cid:1848)(cid:2870),(cid:2869)(cid:2869)(cid:1848)(cid:2871),(cid:2869)(cid:2869)(cid:1848)(cid:3021),(cid:2869)(cid:2869)Veh. 1 :(cid:1848)(cid:3005),(cid:2869)(cid:2869)(cid:1848)(cid:2872),(cid:2869)(cid:2870)4.2(cid:1848)(cid:2872),(cid:2869)(cid:2870)Veh. 2 :4.2(cid:1848)(cid:3005),(cid:2869)(cid:2870)(cid:1848)(cid:2871),(cid:2869)(cid:2869)Task 1Task 2Task 3Task 4Task 5(cid:1848)(cid:2869),(cid:2869)(cid:2869)(cid:1848)(cid:2870),(cid:2869)(cid:2869)(cid:1848)(cid:2872),(cid:2869)(cid:2870)(cid:1848)(cid:2873),(cid:2869)(cid:2870)Depot 1(cid:1848)(cid:3005),(cid:2869)(cid:2869)(cid:1848)(cid:3021),(cid:2869)(cid:2869)(cid:1848)(cid:3005),(cid:2869)(cid:2870)(cid:1848)(cid:3021),(cid:2869)(cid:2870)Depot 24.25.16.21(cid:1848)(cid:2869),(cid:2869)(cid:2869)(cid:1848)(cid:2870),(cid:2869)(cid:2869)(cid:1848)(cid:2871),(cid:2869)(cid:2869)(cid:1848)(cid:3021),(cid:2869)(cid:2869)(cid:1848)(cid:2872),(cid:2869)(cid:2870)(cid:1848)(cid:2873),(cid:2869)(cid:2870)(cid:1848)(cid:3021),(cid:2869)(cid:2870)Veh. 1 :Basket(cid:1848)(cid:3005),(cid:2869)(cid:2869)Veh. 2 :4.2(cid:1848)(cid:3005),(cid:2869)(cid:2870)Task #Value1223324152(cid:1848)(cid:2871),(cid:2869)(cid:2869)Task 1Task 2Task 3Task 4Task 5(cid:1848)(cid:2870),(cid:2869)(cid:2869)(cid:1848)(cid:2871),(cid:2869)(cid:2869)(cid:1848)(cid:2872),(cid:2869)(cid:2870)(cid:1848)(cid:2873),(cid:2869)(cid:2870)Depot 1(cid:1848)(cid:3005),(cid:2869)(cid:2869)(cid:1848)(cid:3021),(cid:2869)(cid:2869)(cid:1848)(cid:3005),(cid:2869)(cid:2870)(cid:1848)(cid:3021),(cid:2869)(cid:2870)Depot 24.25.16.21(cid:1848)(cid:2869),(cid:2869)(cid:2869)(cid:1848)(cid:2870),(cid:2869)(cid:2869)(cid:1848)(cid:2871),(cid:2869)(cid:2869)(cid:1848)(cid:3021),(cid:2869)(cid:2869)(cid:1848)(cid:2872),(cid:2869)(cid:2870)(cid:1848)(cid:2873),(cid:2869)(cid:2870)(cid:1848)(cid:3021),(cid:2869)(cid:2870)Veh. 1 :Basket(cid:1848)(cid:3005),(cid:2869)(cid:2869)Veh. 2 :4.2(cid:1848)(cid:3005),(cid:2869)(cid:2870)Task #Value1122324152X(cid:1848)(cid:2869),(cid:2869)(cid:2869)Task 1Task 2Task 3Task 4Task 5(cid:1848)(cid:2870),(cid:2869)(cid:2869)(cid:1848)(cid:2871),(cid:2869)(cid:2869)(cid:1848)(cid:2872),(cid:2869)(cid:2870)(cid:1848)(cid:2873),(cid:2869)(cid:2870)Depot 1(cid:1848)(cid:3005),(cid:2869)(cid:2869)(cid:1848)(cid:3021),(cid:2869)(cid:2869)(cid:1848)(cid:3005),(cid:2869)(cid:2870)(cid:1848)(cid:3021),(cid:2869)(cid:2870)Depot 24.25.16.21(cid:1848)(cid:2869),(cid:2869)(cid:2869)(cid:1848)(cid:2870),(cid:2869)(cid:2869)(cid:1848)(cid:2871),(cid:2869)(cid:2869)(cid:1848)(cid:3021),(cid:2869)(cid:2869)(cid:1848)(cid:2872),(cid:2869)(cid:2870)(cid:1848)(cid:2873),(cid:2869)(cid:2870)(cid:1848)(cid:3021),(cid:2869)(cid:2870)Veh. 1 :Basket(cid:1848)(cid:3005),(cid:2869)(cid:2869)Veh. 2 :4.2(cid:1848)(cid:3005),(cid:2869)(cid:2870)Task #Value1122324151XX(cid:1848)(cid:2869),(cid:2869)(cid:2869)Task 1Task 2Task 3Task 4Task 5(cid:1848)(cid:2870),(cid:2869)(cid:2869)(cid:1848)(cid:2871),(cid:2869)(cid:2869)(cid:1848)(cid:2872),(cid:2869)(cid:2870)(cid:1848)(cid:2873),(cid:2869)(cid:2870)Depot 1(cid:1848)(cid:3005),(cid:2869)(cid:2869)(cid:1848)(cid:3021),(cid:2869)(cid:2869)(cid:1848)(cid:3005),(cid:2869)(cid:2870)(cid:1848)(cid:3021),(cid:2869)(cid:2870)Depot 24.25.16.21(cid:1848)(cid:2869),(cid:2869)(cid:2869)(cid:1848)(cid:2870),(cid:2869)(cid:2869)(cid:1848)(cid:2871),(cid:2869)(cid:2869)(cid:1848)(cid:3021),(cid:2869)(cid:2869)(cid:1848)(cid:2872),(cid:2869)(cid:2870)(cid:1848)(cid:2873),(cid:2869)(cid:2870)(cid:1848)(cid:3021),(cid:2869)(cid:2870)Veh. 1 :Basket(cid:1848)(cid:3005),(cid:2869)(cid:2869)Veh. 2 :4.2(cid:1848)(cid:3005),(cid:2869)(cid:2870)Task #Value1121314151XXX(cid:1848)(cid:2869),(cid:2869)(cid:2869)first discuss the general chromosome encoding and decoding method and then describe the decoding method which includes the NIN, which is discussed in 3.3. For example, assume that there are two vehicles, five task clusters, and three sample nodes in each task cluster for each vehicle, as shown in Figure 3. The following chromosome M.13 − 1.1 − 2.3 − 3.3 − M − M.21 − 4.2 − 5.1 is decoded for each vehicle as follows: • Vehicle 1: (D,1) - (1,1) - (2,3) - (3,3) - (T,3) • Vehicle 2: (D,2) - (4,2) - (5,1) - (T,1) The first element of each tuple is the task, and the second is the index of the sample node. D is the depot, and T is the terminal cluster. The encoding operator of this study is different from the other studies dealing with GTSP in that it adopts the concept of delimiter to represent multiple vehicles on the chromosome. For a multi-vehicle GTSP instance where the number of vehicles is k, the chromosome has 2k − 1 delimiters, and the gene corresponding to the delimiter is assigned M to the integer part. Here M is a moderately large integer satisfying M > n. Thus, the length of the chromosome of a multi- vehicle GTSP with k vehicles and n tasks is (n + 2k − 1). In the case of the gene corresponding to the delimiter, the roles of the delimiters located at odd- numbered and even-numbered positions are different. In the case of the odd- numbered delimiter, it indicates the depot and terminal cluster of the vehicle. It also contains information about the sample node of the depot and terminal cluster in the fractional part. A delimiter located at an even-numbered position acts as a divider to distinguish the path of the next vehicle from the path of the previous vehicle. In the above chromosome, the first and sixth genes contain the depot sample node and terminal sample node information of vehicles 1 and 2, respectively, and the fifth gene divides the paths of vehicles 1 and 2. When the order of the delimiters in the chromosome is changed through the operators described in 4.4, the sample node visit information may be shifted to the delimiters at even-numbered positions. In this case, the fractional part values of the delimiters that need to be moved are appropriately shifted to another delimiter so that the fractional part exists only in the odd-numbered delimiter. A detailed explanation is given in 4.4.1. The decode operator takes a chromosome as an input and outputs a multi-list variable called tour ={tour 1, ··· , tour m} consisting of m lists containing tour information for each vehicle. In order to apply the NIN discussed in 3.3 to the GA, an additional decoding process is needed in addition to the above, and the decode operator including the additional process is decode-NIN. The basic decoding scheme of decode-NIN is the same as that described above. In addition, the purpose of this decoding operator is to create a reduced tour that removes redundant sample nodes from the tour through the NIN, so that only a minimum sample node can be visited to traverse all task areas. 12 For the convenience of explanation, we explain the decode-NIN using Fig- ure 4. There are 5 tasks and 2 vehicles in a given instance, and there is one sample node for each task cluster. Using the basic decode operator, the given chromosome 1.1− 2.1− 3.1− M.11− M − 4.1− 5.1− M.11 has tour information for vehicle 1 to visit sample nodes in tasks 1, 2, and 3, and for vehicle 2 to visit sample nodes in tasks 4 and 5. Suppose that each task cluster has one basket. Here, basket contains information about how many times the task is visited from all sample nodes of the chromosome. Basically, each task has a de- fault value of 1, because each task is unconditionally visited once by the sample node belonging to its cluster (Figure 4a). After that, we check the additional visit based on the information of the sample node in the chromosome. Since V 1 1,1 goes through tasks 2 and 3 simultaneously, the values of baskets 2 and 3 are increased by one(Figure 4b) and similarly the value of basket 1 is increased because V 1 2,1 passes task 1 at the same time (Figure 4c). A sample node that does not pass other tasks simultaneously, such as V 2 5,1, does not affect the basket value of other tasks. By checking all the sample nodes in the chromosome in this way and adding the values to the basket, the results shown in Figure 4f can be obtained. When the NIN check for all sample nodes in the chromosome is completed, the sample nodes of the tour are deleted as much as possible until all the values of the basket are kept at 1 or more. The sample nodes are deleted in the order of the largest value of the basket. If there are several tasks with the same basket value, a sample node to be erased is determined by the cardinality of T NIN means that the corresponding sample node goes through fewer tasks at the same time and it is better to erase it. In Figure 4f, because the basket value of task 2 is the largest, delete the sample node V 1 2,1 of task 2 from the tour, and decrease the basket value of tasks 1 and 2, which are the tasks passed through V 1 2,1, by one (Figure 4g). Since the task with the largest basket value is 3 and 5, it is needed = ∅, so a sample node to compare the size of each T NIN with a smaller T NIN 5,1 is chosen to be deleted and reducing the basket value by one (Figure 4h). Finally, we delete the sample node V 1 3,1 of task 3 and reduce the basket value of tasks 2 and 3, which are the tasks passed through V 1 3,1, by 1 (Figure 4i). The tour of 4i becomes the final tour of the given chro- mosome because if we reduce the sample node once again, there is a task whose basket value becomes zero. The Algorithm 1 describes a pseudo algorithm of the decode-NIN operator. with the smallest value. This is because a smaller T NIN = 2, and T NIN size V 2 V 2 5,1 s s s . T NIN V 1 3,1 s 4.2 Initialization Two different ways are used to construct the initial population with N chromo- somes. The first is to generate a sequence of which vehicle to visit which task according to the encoding rules in 4.1. The second is to assign the tasks corre- sponding to each Voronoi cell of a vehicle after the region of interest Q ∈ R2 is partitioned into Voronoi cells using the depot locations. The Voronoi diagram V(P, Q) consists of the m disjoint Voronoi cells Vi(P, Q) generated by the set 13 of points P , in other words, V(P, Q) = ∪m i=1Vi(P, Q). An arbitrary task point q ∈ Q belongs to Vi(P, Q) if q − pi ≤ q − pk ∀k ∈ 1,··· , m. After the tasks are assigned to the vehicle, the visiting order is sorted by applying the LKH heuristic [39]. The points in the cell is are given as the input of the ETSP, then its solution is used as the sequence of the initial guess. The sample nodes of each task are chosen randomly. After constructing the initial population, Level-I improvement in 4.4 is applied to complete the initialization. 4.3 Operators for the MA 4.3.1 Reproduction In this study, we apply an elitist strategy to the reproduction operator. This prevents inefficient behavior so the algorithm finds high quality solutions repeat- edly. This is done by moving the fittest chromosome group from the previous generation to the next generation. Through this strategy, the quality of the overall population is promoted as the generation continues, which ensures that a high quality chromosome is selected in the subsequent selection operator. 4.3.2 Selection The selection operator selects two parent chromosomes as the inputs of the crossover operator. Among the various selection operators used in the GA, we used the roulette wheel selection (fitness proportionate selection), which is known as the most representative method. The roulette wheel selection is a method that evaluates the cost of each solution and then selects the chromo- some by adjusting the fitness of the best solution to be κ times the fitness of the worst solution. Here κ is called the selection pressure. The higher the selec- tion pressure, the faster convergence but the higher the likelihood of premature convergence. If the selection pressure is too low, on the other hand, there is a tendency for the average cost of the population not to improve rapidly. κ is an adjustable parameter and is set to 4 in this study. The fitness fi of each chromosome can be obtained using the following equation: fi = cw − ci + cw−cb κ−1 where cw and cb are the worst and best cost in the pool, and ci is the cost of can be expressed as follows: pi = fi(cid:80)N the current chromosome. The probability pi that each chromosome is selected . j=1 fj 4.3.3 Crossover After selecting two different parent chromosomes using the above selection oper- ator, we create a child chromosome through a parameterized uniform crossover. The child chromosome is generated by receiving 60% of genes from parent 1 and 40% of genes from parent 2, but excluding the gene that overlaps with the gene of parent 1 based on task cluster (integer part). If there is a blank space in the child chromosome because of the redundancy, the sequence of task clusters is randomly arranged, and the sample nodes follow the information of parent 1. 14 4.3.4 Immigration After the next generation of populations is populated with reproduction and crossover operators, the remaining number of chromosomes are generated ac- cording to the method described in 4.2. In this study, the role of mutation operator, which is commonly used in the GA, is replaced with an immigration operator to guarantee a diversity of the solution. 4.4 Improvement Heuristics Local improvement heuristics play a major role in enhancing the quality of the solution in the general GA. In the proposed algorithm, the improvement heuris- tic is applied to the initial population or to the chromosome newly generated in the crossover or immigration phase. Improvement heuristics consists of 2-opt and swap methods, each of which is summarized as follows. The 2-opt method consists of a global 2-opt, which applies 2-opt for the entire chromosome, and a local 2-opt applying 2-opt for each tour assigned to the vehicle. The swap con- sists of a task swap that changes the order of two genes across the chromosome and a sample swap that optimizes the sample node information for each gene. The main idea of 2-opt, one of the simplest local search algorithms to solve the TSP problem, is that the tour can be re-arranged by tweaking a part of the tour, thereby improving costs. For the ETSP (the cost from node a to node b is the same as the cost from b to a), the cost of the entire tour can be updated using only the cost variation between new edges caused by exchanging selected edges and destinations of two different edges. However, for the asymmetric TSP, the cost changes according to the direction of edges. Therefore, the cost of all the changed paths must be calculated to update the cost of the entire tour. To cope with this problem, [40] used 3-opt instead of 2-opt to preserve the direction of the path. However, we used 2-opt since the number of task clusters to visit is not very large in the instance handled in this study. In addition, the algorithm is constructed to guarantee the quality of the chromosome through the swap operators. The 2-opt and swap operations are not integrated as used in [41], but are made to work as separate operators. As in [42], level-I improvement is applied when the cost of the chromosome does not belong to the upper rank, and level-II improvement is applied when it belongs to it. We applied the global 2-opt, local 2-opt, and sample swap operators to the chromosome once through Level-I improvement, and we applied the task swap operator 5 times. Level-II improvement is applied to the chromosome until the global 2-opt, local 2-opt, task swap operator fail to improve the cost ten times in succession, and the sam- ple swap operator is applied three times. The algorithm is designed to make the calculation more efficient by applying the improvement operator intensively to the high quality chromosome. 4.4.1 Global 2-opt The global 2-opt operator selects genes a and b at two different positions in the chromosome, then updates the chromosome by reversing the order of a through 15 Figure 5: An example of global 2-opt operator. Figure 6: An example of local 2-opt operator. b. If an odd number of delimiters are included in the selected gene, there is no problem to encode the chromosome after the sequence is inverted. If the number of delimiters is even, however, a chromosome is generated in which the divider is successive or the depot cluster is successive. In this case, the fractional part of the delimiters is rearranged so that there is no problem encoding the chromosome. For example, in Figure 5, the third and eighth positions of a given chromosome are selected (first line) and the chromosome is constructed by rearranging the genes of the corresponding region in the opposite order (second line). At this time, there are two delimiters in the selected interval. By looking at the second and third genes of the chromosome in the second line, the depot clusters are continuous. Therefore, the fractional part existing in the second delimiter of the chromosome is transferred to the third delimiter to construct a chromosome that can be normally encoded. If the cost of the chromosome after the global 2-opt is applied is less than the previous cost, the previous information is replaced. 4.4.2 Local 2-opt Applying 2-opt by selecting any two genes a and b is similar to the global 2- opt operator, but it does not arbitrarily select the entire chromosome. This operator performs 2-opt on the path assigned to each vehicle. Taking Figure 6 as an example, the given chromosome is divided into paths for vehicles 1 and 2 based on the fifth gene. At this point, the upper part of the figure shows a 2-opt example for vehicle 1, and the lower part of the figure shows a 2-opt example for vehicle 2. Similar to the global 2-opt operator, the chromosome is replaced 16 1.5M.146.44.1M3.22.3M.427.15.31.5M.14M.422.33.2M4.16.47.15.31.5M.14M2.33.2M.424.16.47.15.31.5M.146.44.1M3.22.3M.427.15.31.5M.144.16.4M3.22.3M.427.15.31.5M.146.44.1M3.22.3M.427.15.31.5M.146.44.1M7.1M.422.33.25.3when the cost of a chromosome after applying the local 2-opt operator is less than the cost of the previous one. 4.4.3 Task Swap The task swap operator changes position by selecting two different genes. Sim- ilar to the above operators, the costs of chromosomes are compared and then the better one is used to update the population pool. 4.4.4 Sample Node Swap The sample node swap operator updates the chromosome by finding the sample node with the lowest cost for each task cluster. The heuristic is applied sequen- tially from the first gene of a given chromosome. If visiting any sample node s(cid:48) in the same task cluster is less than visiting the existing sample node s, the chromosome gene information is updated. One thing to be kept in mind is that if the previous sample node s has nonempty NIN set T NIN , the tasks in T NIN should be checked to determine whether they are covered by any of the sample nodes in the chromosome if they are not included in T NIN . s s s(cid:48) 4.5 Population Management and Termination Criteria To ensure diversity of the solutions in the GA and not only to enlarge the size of the population, it is necessary to check whether the chromosomes in the pop- ulation actually have different information. If two chromosomes with the same information in the crossover operator are selected as the parent, the information of the child chromosome is almost unchanged from the information of the parent whatever crossover heuristic is used. This means that the convergence rate of the entire GA is slowed down if the chromosome with redundant information in the population is not removed. In order to avoid duplication between chromosomes, it is not enough to check whether the gene values of any two chromosomes are sequentially equal, so whether the information actually contained is the same should be checked. In this study, we examine the duplication of chromosomes simply by their costs. For efficient management, chromosomes are sorted in an ascending order based on the cost for each generation. The algorithm is terminated when the number of generations reaches the given limit or when the cost of the best chromosome in the pool is not improved for finite consecutive times. 4.6 Path Refinement As an additional step in the proposed algorithms, this part suggests a process to refine the paths for each vehicle to improve the quality of the solution. The output of the sampling based methods for the Dubins TSP with neighborhoods is obtained through a limited number of crudely discretized samples in a 3- dimensional space. In other words, even if the optimal solution is obtained 17 for a given roadmap, there is some quality difference from the actual optimum solution with the given conditions. Therefore, to reduce the difference, a local optimization is applied to the outputs from the proposed algorithms. When a local optimization is performed, the parameters are optimized in the continuous state space while it is assumed that the newly refined path follows the sequence of visiting each neighborhood from the given solution. The improvement of the solution quality through this step might be limited since the visiting sequence of the vehicle does not change. However, the global optimal solution of the instance can be obtained if the quality of the given solution from the algorithms above is sufficiently high. In addition, there is an advantage in that the total calculation time can be greatly reduced compared to simply increasing the total number of samples in the instance. Path refinement proceeds as follows. Since the output of the previous step may not have sample nodes for all tasks due to the NIN, the order and state of entry into the neighborhood of each task are sequentially generated based on the path of each vehicle in the given solution. Considering the states of the depots and the terminals, the total number of states to be optimized is n + 2m(cid:48) where n is the number of tasks, and 2m(cid:48) correspond to the number of initial and final states for each vehicle assigned at least a single task. For a local optimization of each task state, the constraints of the state to be optimized and the neighboring states are required. Each is optimized in the direction of decreasing cost where the neighboring states are fixed.1 Similarly, the depot state is optimized with the next state fixed and the terminal state with the previous state fixed. Opti- mization is repeated in an alternating order as follows: odd-numbered states are optimized while others are fixed and then even-numbered states are optimized, and the iteration repeats until the cost of vehicle converges. In the simulation, iteration was performed until the difference between the previous cost and the next cost was less than 0.01%. 5 Numerical Experiments In this section, we discuss the computational results of the formulated problem to address the following questions: 1. How does the performance differ depending on whether NIN and path refinement are applied? 2. How large a problem instance can be handled practically? 3. When each vehicle is heterogeneous, is it possible to generate mission- effective tours considering the characteristics of each vehicle? 4. How do the tours change with the change of the coefficient α in the ob- jective function? 1The 'fmincon' function in MATLAB is used for the local optimization. 18 Table 1: Parameters for simulation. Bold values are for default values. Parameters Number of vehicles Number of sample nodes in each task Task sensing range of vehicle (radius of a task neighborhood for vehicle) (m) Vehicle's velocity and minimum turning radius (m/s, m) Depot location (m, m) Vehicle load factor Metric of cs,s(cid:48) (Cost from s to s(cid:48)) α (Objective function coefficient) Value {1,2,3,4} {1,2,3,4,5,10,20,30,40,50} {100, 150, 200} {(50, 65.9), (60, 94.8), (70, 129.1)} {(110, 230), (1800, 2100), (200, 1500), (1700, 1000)} 4 {length, time} {0, 0.5, 1} There are a number of parameters that can vary the characteristics of prob- lem instance, therefore, first a numerical simulation is performed by changing the number of vehicles and sample nodes in each task cluster while fixing the values of other parameters. After that, the results of the simulation are reported in each section while the parameters that are considered to be important are changed. Detailed parameters are listed in Table 1 with their values. Since one of the main purposes of this paper was to analyze the charac- teristics of the GHMDATSP, we varied the size of the problem by changing the number of vehicles and sample nodes while the number of tasks was fixed for ease of analysis. The instances to be solved are generated from one of the TSPLIB instances called bays29, which contains 29 targets in a 2-dimensional region. Adding m depots and terminals given m vehicles, the total number of clusters is 29+2m for each instance. Sample nodes are randomly generated in the circular neighborhoods of each task with a radius of 150m, and the heading directions in each sample node are also given as random. The sample nodes in the depot and the terminal are located at fixed positions, and also the headings are randomly given. Every vehicle has its unique sample nodes for each cluster, and no nodes are shared between different vehicles. Unless otherwise mentioned, instances are solved by applying the NIN, and the value of the objective function coefficient α is 0.5. We compare the following methods: 1. EA-noNIN, EA-NIN, EA-NIN-PR are the methods based on the MILP implementation. 1.1. EA-noNIN: without considering the NIN. 1.2. EA-NIN: applying NIN. 1.3. EA-NIN-PR: applying path refinement on the solution of EA-NIN. 19 2. MA-noNIN, MA-NIN, MA-NIN-PR are the methods based on the memetic algorithm. 2.1. MA-noNIN: without considering the NIN. 2.2. MA-NIN: applying NIN. 2.3. MA-NIN-PR: applying path refinement on the solution of MA-NIN. 3. OOD is the heuristics by Obermeyer et al. [9]. OOD is a sampling based method, which transforms the problem into the form of the ATSP and solve it using the LKH heuristic. It only can handle a single vehicle prob- lem. 4. ZCXP is the heuristics by Zhang et al. [18]. ZCXP is based on the memetic algorithm and solves the problem in the continuous domain in terms of the location and heading for each task. The capital letters 'EA' for the methods using MILP implementation are borrowed from the word 'exact algorithm', and the detailed description of the corresponding methods is provided in the appendix. All of the MILP based methods were performed on a PC with Intel(R) Xeon(R) CPU E5-2687W v4 @ 3.00GHz and 64.0GB RAM using Gurobi 7.5.1 as a MILP solver. The cal- culation time for each instance is reported in seconds, and the algorithm has an upper limit of 7,200 seconds. The results from the other mehtods were per- formed on a PC with Intel(R) CPU i7-6700K and 16.0GB RAM. The MA based mehtods were implemented in the C#, and the other heuristics (OOD and ZCXP) were implemented in MATLAB environment.2 Before comparing the performance of the above mentioned methods quantitatively, the difference of the results from applying the NIN and the path refinement process is explained qualitatively as follows. 5.1 NIN and Path Refinement The closer the tasks are located, and the larger the turning radius of the vehicle, the more different the characteristics of the solution of the DTSPN from the general ETSP. The dramatic change in the results of each method is shown in this section given the big turning radius (velocity: 70m/s, radius: 129.1m). Figure 7 shows the results for an instance where a single vehicle tours and five sample nodes are created in each cluster. Each result of Figure 7a and Figure 7b is the optimal solution for the given instance through the solver and the results generated from EA-noNIN, EA-NIN, and EA-NIN-PR are shown in order from left to right. Although it is optimal in Figure 7a, to visit one of the nodes of every task exactly, the result made a detour more than necessary since the number samples is small. The result of Figure 7b is quite encouraging in that the detour of the path is significantly reduced via indirect visits compared to the Figure 7a. Each 2The C code of the LKH heuristic was mex-compiled to use in MATLAB environment. 20 (a) Tour result from (b) Tour result from EA-NIN. (c) Tour result from EA-noNIN. EA-NIN-PR. Figure 7: Tours given a single vehicle, varying the method. Velocity: 70m/s. (a) Tour result from (b) Tour result from (c) Tour result from MA-noNIN. MA-NIN. MA-NIN-PR. Figure 8: Tours given four vehicles, varying the method. Velocity: 70m/s. 21 0500100015002000m05001000150020002500m Depot00500100015002000m05001000150020002500m Depot00500100015002000m05001000150020002500m Depot00500100015002000m05001000150020002500m Depot0 Depot1 Depot2 Depot30500100015002000m05001000150020002500m Depot0 Depot1 Depot2 Depot30500100015002000m05001000150020002500m Depot0 Depot1 Depot2 Depot3(a) Comparison of suggested methods. (b) Comparison with other methods. Figure 9: Comparison of objective values for each method. Single vehicle, velocity: 50m/s. Figure 10: Comparison of computational times for each method. indirectly visited task has only a border and is not colored. The result in Figure 7c is obtained by applying the path refinement process to the result of Figure 7b, resulting in a cost reduction of about 15%. It is not the global optimal solution for a given instance, but it can be seen that a satisfactory result can be obtained through an effective heuristic even with a small number of samples. Figure 8 shows the results for an instance where four vehicles are available, and the results from MA-noNIN, MA-NIN, and MA-NIN-PR are shown in order from left to right. Similar to the results in Figure 7, the cost is reduced when NIN and path refinement is applied. 5.2 Comparative experiments Figure 9 shows the comparison of the proposed methods and other methods quantitatively with the mean and the standard deviation of the objective values through the error bar. The instance used in Figure 9 is generated assuming a single vehicle with a speed set to 50m/s. The results of the memetic algorithm 22 123451020304050#sample0.60.70.80.911.11.21.31.41.5Obj104MA-noNINEA-noNINMA-NINEA-NINMA-NIN, PREA-NIN, PR123451020304050#sample0.60.70.80.911.11.21.31.41.5Obj104ZCXPOODMA-NINEA-NINMA-NIN, PREA-NIN, PR123451020304050#sample10-1100101102103104Time (sec)EA-NINZCXPMA-NINOODbased methods (MA-noNIN, MA-NIN, MA-NIN-PR, and ZCXP) are generated ten times for the same instance. The results of the MILP based methods and the OOD are generated once since they generate almost the same outputs for the same instance. In addition, the change of the output according to the number of sample nodes cannot be confirmed from the ZCXP because it is not a sampling based method. In order to ensure convergence of the objective value, the maximum number of generations for the ZCXP is fixed to 10,000 times. Although not shown in the figure, the MA methods proposed in the paper satisfied the terminate condition after about several tens to hundreds of generations, and for the ZCXP it was also confirmed that the objective values almost converged after hundreds of generations. Figure 9a shows the objective values of the proposed methods with the num- ber of sample nodes in the cluster being changed. For the MILP implementation, the maximum number of sample nodes was limited to five due to its complex- ity. The performance of the solution is improved in the order of MA-noNIN, EA-noNIN, MA-NIN, EA-NIN, MA-NIN-PR and EA-NIN-PR. Though it de- pends on the number of sample nodes, it is shown that the objective value can be reduced by up to 50% by the path refinement. Furthermore, it can be seen that each method gradually converges to the near-optimal as the number of sample nodes increases. Figure 9b shows a comparison of the results from the proposed methods and other methods. The performance of the proposed methods is superior to the ZCXP and OOD. The reason why the ZCXP shows poor performance is that the change of visiting sequence depends only on the crossover and mutation operators, which makes it difficult for the chromosome to escape from the local optimals. The OOD shows slightly better performance than MA-noNIN but poorer than the other proposed methods. As in Figure 9a, the result of OOD shows that the objective value gradually decreases as the number of sample nodes increases. Figure 10 shows the comparison of the computational time of each method. Since the difference of each method is large, the time axis is represented by log scale. Because the time complexity of the path refinement is O(n) and the time required for cost convergence was around one to two seconds, the computational time results for the path refinement process were excluded from the comparison. In the case of the MILP implementation, the computational time is the most sensitive to an increase in the number of sample nodes. The computational time of the memetic algorithm does not change much since the terminal condition is constant. Also there is no other factor except the increase in the calculation time of the sample node swap operator (Section 4.4.4) as the number of samples increases. A direct comparison with other methods is difficult since the ZCXP was run in the MATLAB environment, but it differs by about 50 times from the proposed MA based methods which are implemented in C#. The computational time of the OOD increases steeply with the increase of the number of samples since it uses the LKH heuristic whose time complexity is known to be O(n2.2) [39]. Although not shown in the figure, we performed simulations with an increas- ing number of vehicles from one to four, changing the vehicle speed from 50m/s 23 Figure 11: Optimal tours for a heterogeneous fleet. to 60 and 70m/s, and found that the tendency of the objective value and compu- tational time follows the results in Figures 9a and 9b. Table 2 shows the results of the above methods with varying the number of vehicles and sample nodes. The MILP solver could not obtain the optimal solution for most of the cases due to the complexity of the problem, and the suboptimal results are shaded in red in Table 2. As mentioned above, the number of samples is limited up to five for the MILP implementation and the OOD is the heuristic for a single vehicle. The instances not calculated are shaded in gray in Table 2. 5.3 Heterogeneous Motion Constraints To show how the algorithm handles a fleet of structurally heterogeneous vehicles, an instance with different vehicle characteristics are given to be solved. Three vehicles are given as vehicles #0, #1, and #2, with the velocity given as 50m/s, 75m/s, and 100m/s and the sensing range given as 100m, 150m, and 200m respectively, and the cost metric is chosen as 'time'. The dynamics of each vehicle is different and follows the values in Table 1. Vehicle #0 is slow and the remote sensing coverage is small; #2 is fast and the coverage is large; and the characteristics of #1 are about halfway between them. Figure 11 shows the optimal solution of the instance using the EA-NIN method, and the number of nodes per cluster are given as 5. Vehicle #2 handles most of the tasks in the region because of its large coverage and agility, and then #1 handles the remaining tasks. In the solution, no path is assigned to #0 due to its poor performance. 5.4 Various Objective Functions Using the default parameters and (#v, #s) as (4,5), the coefficient value α in eq. (1) is varied to check how the feature of results changes. Figures 12a and 24 500050010001500200025003000m05001000150020002500mDepot #0Depot #1Depot #2Tour - veh. 1Tour - veh. 2sample - veh. 0sample - veh. 1sample - veh. 2Depot - veh. 0Depot - veh. 1Depot - veh. 2(a) Instance - (#v, #s) : (4, (b) Instance - (#v, #s) : (4, (c) Instance - (#v, #s) : (4, 5), α = 1 5), α = 0.5 5), α = 0 Figure 12: Optimal tours with varying α. 12b show the optimal results of given instances when the value of α is 1 and 0, respectively. As mentioned, the formulated problem with α = 1 is a typical min problem, and if α = 0 then the problem becomes a min-max problem. The sum of tour costs in Figure 12a is 7,744, which is smaller than the instance in Figure 12b (9,511), but all of the tasks are assigned to the vehicle #3. The phenomenon in which the task assignment is concentrated only on a few vehicles often occurs in the multiple TSP where the objective function is given to minimize all the mission costs. In case of α = 0, the interesting result is that the objective value (which is the same as the maximum tour cost among the vehicles) is 2,706, which is the same as the instance where α was 0.5. But the total sum of costs given α = 0 is 9,943, which is larger than the case of cost 9,512 when α is 0.5. 6 Conclusion In this paper, the methods and procedure of memetic algorithm based path generation was presented as variants of the TSP, which is called the generalized heterogeneous multiple depot asymmetric traveling salesmen problem (GHM- DATSP), which arises in the context of achieving a mission of remote surveil- lance using a fleet of Dubins vehicles. A mixed-integer linear programming formulation was proposed to solve instances of the GHMDATSP. The suggested procedure incorporates a robust tour improvement heuristics to fit the GHM- DATSP into the classic genetic algorithm. To enhance the performance of the solutions, first we used the concept called the necessarily intersecting neighbor- hood (NIN), which generated effective paths, especially when the tasks were densely located, and second we used the local optimization based path refine- ment process to find the best visiting location and heading of every task region. In order to observe the characteristics of the GHMDATSP with Dubins vehicles and to verify the efficiency of the proposed methods, a wide class of simula- tion was performed on the instances generated from a standard library with diverse variations of the key parameters. The solutions were obtained for up to 25 0500100015002000m05001000150020002500mDepot #0Depot #1Depot #2Depot #30500100015002000m05001000150020002500mDepot #0Depot #1Depot #2Depot #30500100015002000m05001000150020002500mDepot #0Depot 1Depot #2Depot #3Table 2: Computational results, velocity set to 50m/s. Objective cost Time (sec) #v #s EA-NIN, gap(%) MA-NIN EA-NIN, 1 2 3 4 1 2 3 4 5 10 20 30 40 50 1 2 3 4 5 10 20 30 40 50 1 2 3 4 5 10 20 30 40 50 1 2 3 4 5 10 20 30 40 50 12702.4 9942.0 8894.4 8848.2 8458.8 (2.96%) 6255.1 5590.8 5498.6 (9.19%) 4729.5 5447.3 (18.49%) 4420.6 4158.4 (7.14%) 3766.3 (8.40%) 3693.9 (12.26%) 3534.7 (17.27%) 3477.0 3059.1 (5.68%) 2696.8 (8.04%) 2632.2 (8.21%) 2622.7 (15.86%) PR 7358.6 7383.9 7109.5 6788.9 7035.0 4885.9 4379.8 4566.1 4216.2 4682.8 2937.4 3236.2 3271.5 3165.4 3030.0 2549.1 2238.2 2118.6 2172.1 2171.0 12751.4 10118.1 8946.8 8858.6 8616.1 7320.3 7371.9 6967.6 7093.6 6973.9 6383.1 5673.0 5515.0 4812.3 5021.9 4561.3 4384.7 4225.9 4200.5 4196.6 4582.8 4199.0 3718.2 3668.1 3537.5 3311.5 3153.7 3187.4 3104.4 3151.5 3541.2 3076.7 2767.7 2658.4 2535.0 2385.9 2272.8 2201.0 2202.7 2205.8 MA-NIN, PR OOD ZCXP EA MA OOD ZCXP 14037.3 10913.1 9892.7 9800.0 9419.9 8143.0 7846.3 7479.5 7347.4 7182.4 9166.0 5414.8 3873.3 2982.4 0.1 0.1 0.1 0.3 0.2 1.3 6.8 11.3 26.9 40.3 780.8 819.3 809.7 890.4 0.2 3.7 185.9 53.9 7200 41.4 1080.9 7200 2485.0 7200 44.3 7200 7200 7200 7200 1198.9 7200 7200 7200 7200 9.0 12.2 11.1 13.0 14.5 16.6 15.9 17.5 16.6 15.7 12.2 12.8 12.6 14.1 16.4 15.9 18.7 17.5 17.5 20.1 13.0 15.6 15.1 15.5 16.9 17.5 17.5 15.7 21.4 19.9 13.3 17.0 17.6 16.6 18.9 19.6 18.4 20.5 20.9 20.5 7623.7 7351.3 7157.3 6651.1 7085.3 6650.1 6688.9 6671.0 6639.2 6667.8 4652.4 4262.9 4418.8 4254.1 4223.3 4144.6 4095.1 4012.0 4041.2 4069.6 3093.0 3255.7 3054.7 3114.3 3054.3 2982.3 2953.2 3005.4 2956.1 3001.3 2595.4 2237.4 2210.9 2119.0 2092.0 2148.1 2093.5 2042.2 2043.0 2090.3 26 4 vehicles and 29 tasks with different numbers of sample nodes for each cluster. Numerical simulations verified that the proposed methods were superior to the other two state-of-the-art methods in terms of the performance and the compu- tational time. And the comparison with the results from MILP implementation based methods shows the proposed methods can The proposed methods can be further extended to incorporate the limited maximum travel distance for each vehicle. Another possible direction for the research is to reduce the number of tasks by adapting the geometric sensor cover problem. Appendix - MILP Implementation In this section, we describe the major parts of the MILP implementation used to solve the GHMDATSP. The optimal solution can be obtained by providing the formulation developed in Section 3 to an off-the-shelf commercial MILP solver. The problem with the subtour elimination constraints, however, is that the number of these constraints increases exponentially as the number of task increases. It is known that the number of subtour elimination constraints ex- ceeds 1015 for a normal TSP with 50 tasks; therefore generating all subtour elimination constraints is impossible. To handle this issue, the constraints in eq. (6) are generated step by step whenever it is needed, which is called the row generation. Generally, row gen- eration is applied to the constraints that do not occur frequently during the execution of a branch-and-cut algorithm. At the beginning of the solving pro- cess, the constraints are relaxed from the original formulation. Whenever the solver gets a feasible solution of the relaxed problem, the callback procedure is called which checks whether the solution violates any of the constraints in Eq. (6). If the tasks assigned to each vehicle are connected and satisfy the subtour elimination constraints, the solution is regarded as a true solution for the given problem and it is accepted. Otherwise, the solution is abandoned and every subtour which does not include the depot and terminal clusters is put as the ingredient of new constraints. After these constraints are added to the formula- tion, the branch-and-cut algorithm continues to solve the problem. This process is known to solve the traditional traveling salesman problem and its variants in an effective manner. The separation algorithm, which finds every subtour in the feasible solution of the relaxed problem, is provided in Algorithm 2. Whenever the callback is invoked by finding a new candidate incumbent solution that satisfies the constraints of the relaxed problem, the separation al- gorithm takes current decision variables denoted as x∗, y∗, and yNIN. Initially, the list variable unvisited is defined and has an index of all tasks as an element. Then for each vehicle, the sample nodes are checked and are connected through the edges of a closed path passing through the depot and terminal clusters. The tasks of the connected sample nodes are removed from unvisited, as well as the tasks for which their neighborhoods are indirectly visited. If unvisited is an empty set after the above process, an empty set is assigned to the variable S and 27 the procedure is terminated. Otherwise, all the closed paths which are passing through the tasks in unvisited are found and saved in S to generate additional constraints after returning it because in the given solution there are some tasks that a fleet does not visit. References [1] N. Ceccarelli, J. J. Enright, E. Frazzoli, S. J. Rasmussen, and C. J. Schu- macher, "Micro uav path planning for reconnaissance in wind," in American Control Conference, 2007. ACC'07. IEEE, 2007, pp. 5310 -- 5315. [2] S. A. Bortoff, "Path planning for uavs," in American Control Conference, 2000. Proceedings of the 2000, vol. 1, no. 6. IEEE, 2000, pp. 364 -- 368. [3] T. Shima, S. J. Rasmussen, A. G. Sparks, and K. M. Passino, "Multiple task assignments for cooperating uninhabited aerial vehicles using genetic algorithms," Computers & Operations Research, vol. 33, no. 11, pp. 3252 -- 3269, 2006. [4] H.-L. Choi, A. K. Whitten, and J. P. How, "Decentralized task alloca- tion for heterogeneous teams with cooperation constraints," in American Control Conference (ACC), 2010. IEEE, 2010, pp. 3057 -- 3062. [5] S. S. Ponda, L. B. Johnson, A. N. Kopeikin, H.-L. Choi, and J. P. How, "Distributed planning strategies to ensure network connectivity for dy- namic heterogeneous teams," IEEE Journal on Selected Areas in Commu- nications, vol. 30, no. 5, pp. 861 -- 869, 2012. [6] K. Sundar and S. Rathinam, "An exact algorithm for a heterogeneous, mul- tiple depot, multiple traveling salesman problem," in Unmanned Aircraft Systems (ICUAS), 2015 International Conference on. IEEE, 2015, pp. 366 -- 371. [7] G. Cai, J. Dias, and L. Seneviratne, "A survey of small-scale unmanned aerial vehicles: Recent advances and future development trends," Un- manned Systems, vol. 2, no. 02, pp. 175 -- 199, 2014. [8] K. P. Valavanis and G. J. Vachtsevanos, Handbook of unmanned aerial vehicles. Springer Publishing Company, Incorporated, 2014. [9] K. J. Obermeyer, P. Oberlin, and S. Darbha, "Sampling-based path plan- ning for a visual reconnaissance unmanned air vehicle," Journal of Guid- ance, Control, and Dynamics, vol. 35, no. 2, pp. 619 -- 631, 2012. [10] J. T. Isaacs and J. P. Hespanha, "Dubins traveling salesman problem with neighborhoods: a graph-based approach," Algorithms, vol. 6, no. 1, pp. 84 -- 99, 2013. 28 [11] C. H.-J. Jang Dae-Sung and C. Han-Lim, "Optimal control-based uav path planning with dynamically-constrained tsp with neighborhoods," in Inter- national Conference on Control, Automation and Systems, Jeju, Korea, 2017. [12] P. Oberlin, S. Rathinam, and S. Darbha, "Today's traveling salesman prob- lem," IEEE robotics & automation magazine, vol. 17, no. 4, pp. 70 -- 77, 2010. [13] K. Sundar, S. Venkatachalam, and S. G. Manyam, "Path planning for mul- tiple heterogeneous unmanned vehicles with uncertain service times," arXiv preprint arXiv:1702.07647, 2017. [14] G. Dantzig, R. Fulkerson, and S. Johnson, "Solution of a large-scale traveling-salesman problem," Journal of the operations research society of America, vol. 2, no. 4, pp. 393 -- 410, 1954. [15] L. E. Dubins, "On curves of minimal length with a constraint on average curvature, and with prescribed initial and terminal positions and tangents," American Journal of mathematics, vol. 79, no. 3, pp. 497 -- 516, 1957. [16] J. Le Ny, E. Feron, and E. Frazzoli, "On the dubins traveling salesman problem," IEEE Transactions on Automatic Control, vol. 57, no. 1, pp. 265 -- 270, 2012. [17] P. Oberlin, S. Rathinam, and S. Darbha, "A transformation for a heteroge- neous, multiple depot, multiple traveling salesman problem," in American Control Conference, 2009. ACC'09. IEEE, 2009, pp. 1292 -- 1297. [18] X. Zhang, J. Chen, B. Xin, and Z. Peng, "A memetic algorithm for path planning of curvature-constrained uavs performing surveillance of multiple ground targets," Chinese Journal of Aeronautics, vol. 27, no. 3, pp. 622 -- 633, 2014. [19] D.-H. Cho, J.-S. Ha, S. Lee, S. Moon, and H.-L. Choi, "Informative path planning and mapping with multiple uavs in wind fields," arXiv preprint arXiv:1610.01303, 2016. [20] S.-H. Kim, L. Negash, and H.-L. Choi, "Cubature kalman filter based fault detection and isolation for formation control of multi-uavs," in IFAC Sym- posium on Intelligent Autonomous Vehicles, Jun. 2016. [21] H.-L. Choi and J.-S. Ha, "Informative windowed forecasting of continuous-time linear information-based sensor planning," Automatica, vol. 57, pp. 97 -- 104, Jul. 2015. [Online]. Available: http://www.sciencedirect.com/science/article/pii/S0005109815001624 systems for mutual [22] B. Yuan, M. Orlowska, and S. Sadiq, "On the optimal robot routing prob- lem in wireless sensor networks," IEEE Transactions on Knowledge and Data Engineering, vol. 19, no. 9, pp. 1252 -- 1261, 2007. 29 [23] J.-S. Liu, S.-Y. Wu, and K.-M. Chiu, "Path planning of a data mule in wireless sensor network using an improved implementation of clustering- based genetic algorithm," in Computational Intelligence in Control and Automation (CICA), 2013 IEEE Symposium on. IEEE, 2013, pp. 30 -- 37. [24] G. Gutin and A. P. Punnen, The traveling salesman problem and its vari- ations. Springer Science & Business Media, 2006, vol. 12. [25] M. Bellmore and S. Hong, "Transformation of multisalesman problem to the standard traveling salesman problem," Journal of the ACM (JACM), vol. 21, no. 3, pp. 500 -- 504, 1974. [26] G. Laporte and Y. Nobert, "A cutting planes algorithm for the m-salesmen problem," Journal of the Operational Research society, vol. 31, no. 11, pp. 1017 -- 1023, 1980. [27] D. Applegate, W. Cook, S. Dash, and A. Rohe, "Solution of a min-max vehicle routing problem," INFORMS Journal on Computing, vol. 14, no. 2, pp. 132 -- 143, 2002. [28] E. Kivelevitch, K. Cohen, and M. Kumar, "A market-based solution to the multiple traveling salesmen problem," Journal of Intelligent & Robotic Systems, vol. 72, no. 1, p. 21, 2013. [29] R. Baldacci, M. Battarra, and D. Vigo, "Routing a heterogeneous fleet of latest advances and new chal- vehicles," in The vehicle routing problem: lenges. Springer, 2008, pp. 3 -- 27. [30] K. Sundar and S. Rathinam, "Algorithms for heterogeneous, multiple de- pot, multiple unmanned vehicle path planning problems," Journal of In- telligent & Robotic Systems, vol. 88, no. 2-4, pp. 513 -- 526, 2017. [31] -- -- , "Generalized multiple depot traveling salesmen problempolyhedral study and exact algorithm," Computers & Operations Research, vol. 70, pp. 39 -- 55, 2016. [32] S. Salhi, A. Imran, and N. A. Wassan, "The multi-depot vehicle rout- ing problem with heterogeneous vehicle fleet: Formulation and a variable neighborhood search implementation," Computers & Operations Research, vol. 52, pp. 315 -- 325, 2014. [33] K. Savla, E. Frazzoli, and F. Bullo, "On the point-to-point and traveling salesperson problems for dubins' vehicle," in American Control Conference, 2005. Proceedings of the 2005. IEEE, 2005, pp. 786 -- 791. [34] X. Ma and D. A. Castanon, "Receding horizon planning for dubins traveling salesman problems," in Decision and Control, 2006 45th IEEE Conference on. IEEE, 2006, pp. 5453 -- 5458. 30 [35] D.-H. Cho, D.-S. Jang, and H.-L. Choi, "Heterogeneous, multiple depot multi-uav path planning for remote sensing tasks," in AIAA SciTech Fo- rum, 01 2018. [36] K. Helsgaun, "An effective implementation of the lin -- kernighan traveling salesman heuristic," European Journal of Operational Research, vol. 126, no. 1, pp. 106 -- 130, 2000. [37] K. Obermeyer, "Path planning for a uav performing reconnaissance of static ground targets in terrain," in AIAA Guidance, Navigation, and Control Conference, 2009, p. 5888. [38] X. Zhang, J. Chen, B. Xin, and Z. Peng, "A memetic algorithm for path planning of curvature-constrained uavs performing surveillance of multiple ground targets," Chinese Journal of Aeronautics, vol. 27, no. 3, pp. 622 -- 633, 2014. [39] K. Helsgaun, "Solving the equality generalized traveling salesman problem using the lin -- kernighan -- helsgaun algorithm," Mathematical Programming Computation, vol. 7, no. 3, pp. 269 -- 287, 2015. [40] B. Freisleben and P. Merz, "A genetic local search algorithm for solving symmetric and asymmetric traveling salesman problems," in Evolution- ary Computation, 1996., Proceedings of IEEE International Conference on. IEEE, 1996, pp. 616 -- 621. [41] J. Renaud and F. F. Boctor, "An efficient composite heuristic for the sym- metric generalized traveling salesman problem," European Journal of Op- erational Research, vol. 108, no. 3, pp. 571 -- 584, 1998. [42] L. V. Snyder and M. S. Daskin, "A random-key genetic algorithm for the generalized traveling salesman problem," European journal of operational research, vol. 174, no. 1, pp. 38 -- 53, 2006. 31 Algorithm 1 Decode-NIN 1: procedure Decode-NIN (chromosome, NIN) 2: 3: tour ← decode(chromosome) for t := 1 to n do basket[t] ← 1 end for for each sample node s in tour do for each t in T NIN s do basket[t]++ end for end for while every value in basket is bigger than zero do tdel ← argmax(basket) if (cid:12)(cid:12)tdel(cid:12)(cid:12) = 1 then else tdel = tdel tdel ← argmint (cid:12)(cid:12)(cid:12)T NIN V k t,i (cid:12)(cid:12)(cid:12) for t in tdel // k: vehicle visiting t // i: sample node index belonging to t end if s ← sample node belonging to task tdel for each task t in T NIN do s basket[t] - - end for if any element of basket is zero then break while loop end if k ← vehicle visiting s delete s in tour k 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15: 16: 17: 18: 19: 20: 21: 22: 23: 24: 25: 26: 27: 28: end while 29: return tour 30: 31: end procedure 32 Algorithm 2 Separation algorithm - Find subtours 1: procedure FindSubtour (x∗, y∗, yNIN) unvisited = {1,··· , n} for k := 1 to m do 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15: 16: 17: 18: 19: 20: 21: 22: s1 ← Depot sample node of vehicle k (param: y) s2 ← Direct successor of s1 (param: x) while do if s2 belongs to a terminal cluster then break while loop end if t ← task index which contains s2 Remove t in unvisited (cid:54)= ∅ (param: yNIN*) then if T NIN s2 Remove NIN tasks of s2 from unvisited end if s1 ← s2 s2 ← Direct successor of s1 (param: x) end while end for if unvisited is empty then S ← ∅ S ← all closed paths including tasks in unvisited else end if return S 23: 24: 25: end procedure 33
cs/0612014	1	0612	2006-12-04T11:02:28	Going Stupid with EcoLab	[ "cs.MA" ]	In 2005, Railsback et al. proposed a very simple model ({\em Stupid Model}) that could be implemented within a couple of hours, and later extended to demonstrate the use of common ABM platform functionality. They provided implementations of the model in several agent based modelling platforms, and compared the platforms for ease of implementation of this simple model, and performance. In this paper, I implement Railsback et al's Stupid Model in the EcoLab simulation platform, a C++ based modelling platform, demonstrating that it is a feasible platform for these sorts of models, and compare the performance of the implementation with Repast, Mason and Swarm versions.	cs.MA	cs	Going Stupid with EcoLab Russell K. Standish School of Mathematics and Statistics University of New South Wales May 25, 2014 Abstract In 2005, Railsback et al. proposed a very simple model (Stupid Model) that could be implemented within a couple of hours, and later extended to demonstrate the use of common ABM platform functionality. They provided implementations of the model in several agent based modelling platforms, and compared the platforms for ease of implementation of this simple model, and performance. In this paper, I implement Railsback et al’s Stupid Model in the EcoLab simulation platform, a C++ based modelling platform, demonstrating that it is a feasible platform for these sorts of models, and compare the performance of the implementation with Repast, Mason and Swarm ver- sions. 1 Introduction Newcomers to agent based modelling (ABM) will be confused by the variety of different software platforms available to assist in implementing the models. Very few comparative studies between the different platforms have been done, as it is a time consuming task implementing all but the most trivial of models. Furthermore, familiarity with one platform and programming language will lend an automatic advantage in any metrics to that platform over other platforms that the model implementer is less familiar with. In 2005, Railsback et al.[14] proposed a very simple model that could be implemented within a couple of hours, and later extended to demonstrate the use of common ABM platform functionality. They gave it the name “Stupid Model”, partly for fun, but also to reiterate the recommendation of Grimm and Railsback [5] that modelling projects should start with a “ridiculously simplified model”. Railsback et al. implemented their model across a range of ABM platforms: Objective-C and Java Swarm[12], Repast[13] and Mason[9] (both pure Java implementations) and Netlogo. This range of platforms reflects the authors’ collective programming expertise in Objective-C and Java, and with Netlogo having low barrier of entry (Logo was a popular language for teaching school children in the 1980s). 1 EcoLab grew out of a simulation platform supporting a particular class of model, into a general purpose simulation environment using C++[17]. Other C++ agent-based modelling environments exist, eg SymBioSys[11], but none are as general purpose as EcoLab. Other general purpose agent based platforms can be used with C++ models. For instance, with Swarm, C++ code can be linked to Swarm’s objective C library through the shared C language interface, and C++ code can be linked to Repast’s Java library through the Java Native Interface. However, maintaining the interface code quickly becomes prohibitive in the face of evolving models, negating much of the benefits in using a simula- tion platform in the first place. With EcoLab, it is possible to have a similar level of functionality as provided by Swarm or Repast for models implemented in C++, without the interface maintenance overhead. Additionally, EcoLab provides features for distributing the computation over multiple processors in a way that is easier to program than the raw Message Passing Interface (MPI)[16]. With Railsback et al.’s Stupid Model specification, the possibility exists for directly comparing an EcoLab imple- mented agent based model with other platforms for both ease of implementation, and execution performance. Furthermore, the exercise illuminates those parts of EcoLab requiring improvement. 1.1 Why C++ C++[19] is a mature object oriented programming language of more than 20 years standing. It has been widely adopted in industry, consequently open source reference compilers, as well as vendor-tuned optimising compilers exist for most contemporary computer architectures. Because of this popularity, and the availability of compilers, C++ has been extensively deployed for scientific computing since the mid-1990s. In High Performance Computing (HPC), the extreme end of scientific computing, the predominant computing language used for applications is Fortran, with code written in Fortran 77, or increasingly written using the newer Fortran 90 features. However C/C++ applications also make up a substantial fraction of the deployed applications, perhaps as high as 30%, with C++ standing to C in the same relationship as Fortran 90 does to Fortran 77, i.e. typically used as a “better C”1. By contrast, Java[4] has made negligible impact in HPC2. There are several possible reasons for the lack of Java adoption in high performance computing. Firstly, most implementations compile to a virtual machine, and early Java Virtual Machines (JVMs) had performance problems. However, more recent JVMs deploy just in time com- pilation, which closes the performance gap between JVM executed code and natively compiled code. Secondly, certain language features missing in Java (notably operator overloading, and to a lesser extent generic programming) of 1These numbers come from a decade of personal experience at managing the resource allocation process at a High Performance Computing Centre. These general numbers are backed up by anecdotal reports from a number of other people I have corresponded with 2Over the ten years of my personal experience, only one project used Java, out of several hundred that were mostly C/C++ or Fortran. 2 C++ (and Fortran 90 for that matter) assist in writing scientific codes that are closer to the mathematical specification. However, probably the most signif- icant factor is time and innate conservatism of scientific programmers. C++ did not appear significantly in HPC applications until around 15 years after the language was first developed. With only a decade under its belt, Java’s time as an HPC application language might just be beginning[6]. However, for agent based simulation, C++ is not a popular choice, primarily due to its lack of reflection. Reflection is the ability to query an object’s type information at runtime, and in ABM systems like Swarm, reflection is used to implement probes, or the ability to observe all parts of a running simulation from within a graphical user interface[12]. However, with Classdesc, an effec- tive reflection mechanism for C++ is possible[10, 18]. EcoLab uses Classdesc to implement probing, along with automatic checkpointing, the ability to script the model’s initialisation and ongoing computation, and for distributing agents to exploit any parallel computing capability. 2 Method In line with Railsback et al.’s[14] methodology, I implemented Stupid Model using the current EcoLab release, version 4.D21. This is important to give a sense of the maturity of the platform. Otherwise, I might have been tempted to fix up any weaknesses encountered. I followed the the explicit model specification[15] step by step, referring to the Repast Java implementation on the rare occasions the specification was am- biguous. Stupid Model consists of agents called “Stupid Bugs” moving around a Cartesian lattice. No two agents can occupy the same location, so movement involves selecting a cell within a 9 × 9 Moore neighbourhood, testing whether the cell is occupied and moving into the cell if empty. The search procedure is repeated until an empty cell is found. Since different frameworks potentially use different random number algorithms, initialised with a different seed, this introduces indeterminism into model runtimes. In order to reduce the impact of this indeterminism, the density of agents was chosen to be 0.1 (4000 agents in a 200 × 200 world) so that the standard deviation of runtimes was less than 10% of the mean. For measuring application performance, I did both GUI runs, and batch mode runs. In EcoLab, a non-GUI batch run simply involves replacing the “GUI” command from the experiment script, with a call to “simulate”, and comment- ing out any graphical calls (plot, histogram and draw). In Repast, Swarm and Mason, a separate “BatchSwarm” needs to be provided by the programmer, but only the GUI versions of each model were published by Railsback et al. For batch measurements, I commented out the call to addAction that added the display actions. For the Repast implementation, I changed the batch parameter of SimInit::loadModel to true, and timed the run from the command line. With the Mason implementation, I again commented out the display action, and recorded the CPU time so as to discount the delays introduced by hav- 3 ing to click the button. In fact for all platforms, the reported values are the CPU time. For the Objective C Swarm version, I modified the code so that the StupidModelSwarm was directly called from main() rather than indirectly through StupidModelObserverSwarm. I chose to measure the versions 10 and 11 of the Stupid Model. However, the stopping criteria is specified as when the maximum bug size reaches 100. Since bug growth depends on the availability of food, which itself is a function of a random number generator call, and also of the grazing history, this stopping criterion is indeterministic. For the purposes of inter-framework performance comparisons, I changed the stopping criterion to be a fixed number of bug updates (500). In version 10 of Stupid Model, bugs will randomly select a cell within their neighbourhood, and moving to it if the cell is empty, otherwise repeating the selection process. In version 11, all cells in the neighbourhood are iterated over, and the bug moves to the empty cell with the most food. From version 12, bugs can reproduce and die according to random dynamics, so the amount of work per update step will depend on the number of living bugs. Even though these higher version models are more computationally intensive, run times cannot be compared between different platforms due to differences in the order that random numbers are generated. Hence the Stupid 16 mea- surements reported in table 2 should be taken with a certain amount of salt. Nevertheless, I verified that all models executed for 1000 steps, and that the number of Stupid Bugs was roughly the same for each platform (approximately 8-900 after the initial population explosion). Railsback et al. did not do any performance analysis or tuning. For C++ code, performance tuning can deliver big performance improvements. EcoLab can be built with performance counters enabled for the individual TCL commands, and a single run indicated that the initial approach used for evaluating the stopping criterion (evaluating the maximum of the vector of bug sizes in TCL) was very expensive. By implementing a specialised max_bugsize() (all of 4 lines of C++ code) improved performance by about a factor of four. However, for the inter-platform performance comparison, the stopping condition was changed to a fixed number of bug update steps, so this optimisation makes no difference to the performance benchmarks. A more detailed performance profile using the standard GNU/Linux profiling tool gprof, indicated that updating the food availability was a bottleneck, and that cache utilisation could be improved by laying the data contiguously in memory, which is not the case when the data is stored as members of a cell object. This optimisation, which needed some substantial recoding of the model, improved overall performance by a factor of two for model version 16, although it only made about a 10% improvement for version 11. It should be noted that this optimisation technique should also be available for the Java and Objective-C platforms, and presumably may deliver a similar performance boost. All performance benchmarks were run on a 2GHz Intel Pentium M processor with 1GB memory running Slackware Linux 10.0. The Java version used for Repast and Mason was SDK 1.4.2 standard edition. The compiler used for 4 Swarm and EcoLab was GCC 3.4.3. I also did a comparison EcoLab run using the Intel C++ compiler 9.0, but this was more than 50% slower than the GCC compiled code. This somewhat surprising result indicates that icc’s strength lies in vectorising loops that access data contiguously to exploit the inbuilt SSE instructions, but that for more general purpose ABM code, GCC performs better (at least on Linux!). The sourcecode for EcoLab Stupid Model is available from the EcoLab website.[3] 3 Results Similar to all the platforms reviewed by Railsback et al., EcoLab proved capable of implementing all functionality for all versions of Stupid Model. Implementing the first version took longer than any of the remaining versions, as EcoLab does not provide a ready-to-use spatial library. Instead it provides a more general library called Graphcode[18]. Graphcode’s abstraction is a network, or graph of objects, with the links between objects representing data flow. Graphcode can distribute the objects across multiple processors using the Classdesc serialisation library. A cellular space such as found in Swarm or Repast will be a set of objects, each one wired to its neighbours. In such a way, Graphcode can easily represent Cartesian and hexagonal topologies by the way the neighbourhoods are wired. However, the only example using Graphcode provided in the EcoLab was a continuous space example, each cell holding objects located within a certain region of space. Examples of models using different sorts of spatial topologies, as well as a few common cases being supplied as a library would improve the beginner’s experience of EcoLab. In retrospect, it may have been simpler to implement the spatial class on top of a standard vector of cells. This would have gotten the initial model up and running quicker, but limited the model to sequential usage only. By using Graphcode, we enable parallel processing capability. One thing that became clear in this exercise is the need for a smart refer- ence type. Objects like bugs need a reference to the cell in which they inhabit, scheduling lists need references to the bugs that they schedule and so on. Be- cause bugs move from cell to cell, it is better for the cells to have a reference to the bug it contains (if any) rather than for the cell to store the bug itself. In C, the only possibility for references are pointers, which are difficult to serialise properly due to the fact that C makes no guarantees about whether a pointer is valid or not. Substantial care is required to ensure that references remain valid in the event of an object such as a bug being deleted from the system. Classdesc accepts a pragma that asserts that a pointer is either valid or NULL, and whether the pointer chains form cycles or not to allow serialisation, but it’s up to the programmer to ensure software bugs do not invalidate this assertion. C++ also supports static references (eg int&), which are established at the time of the reference’s creation, and then immutable until the reference is destroyed. These references are always valid, however the lack of dynamic control makes them unsuitable for agent based simulations where agents may 5 be dropped or moved, and appropriate references updated. Furthermore static reference cycles cannot be handled with serialisation at all, since the serialisation descriptors cannot distinguish an object from its reference. Whilst it is possible to use EcoLab with a nonserialisable model, one gives up substantial functionality doing so, including the ability to checkpoint/restart the model. What is needed actually is something like Java’s reference type, where ob- jects are created on the heap, and the programmer simply manipulates refer- ences. Once all references to an object have been destroyed, Java’s garbage collector takes care of destroying the object, reclaiming the memory used. It is possible to implement something like this in C++, using operator over- loading to give the resulting type the “look and feel” of a pointer. Such types are usually called smart pointers. The well known Boost library[2] provides a few different versions, some of which are being considered for inclusion in the C++ standard library. EcoLab provides the template ref<T>, which is parameterised by the target type of the reference. Unlike the Boost versions (in which you pass the smart pointer a pointer for it to control), ref has control over the entire lifecycle of the object it points to. The first time a ref object is dereferenced, the target object is created on the heap, and it keeps track of the number of references to the target object, so that once all references to are destroyed, so is the target object. The version of ref supplied in the current EcoLab has a number of deficiencies, however, most notable of which is that it doesn’t provide any way of testing whether the target object exists or not. For the purposes of this exercise, I copied the ref.h header file, and added the necessary functionality. This improved ref.h will be incorporated in future releases of EcoLab. Agents usually need to refer to the environment, or world in which they live. In languages like Java or Objective C, this is simply managed by having the agent store a reference to the world, and/or cell. However, this will set up a reference cycle which will play havoc with model serialisation if the serialisation algorithm doesn’t explicitly account for cycles. EcoLab provides a routine that serialises arbitrary graphs constructed with pointer references. However, it does not currently support the presence of cycles with the ref<> data type. With C++, however, there is a simple workaround. The model is a global variable, and agents can refer to their cell by holding an index into a container of cells stored within this global model. This is the approach I have taken with Stupid Model, and indeed this technique is used in other EcoLab models. However, if the ref<> data type were extended to support serialisation of cyclic graphs, the method deployed in Java and Objective C models can be supported as well. Line counts are often considered a proxy for the amount of effort a program- mer must expend to implement a problem. Table 1 shows the line counts for the 16 different Stupid Model cases for each of the Railsback implementations, as well as the EcoLab implementation. The EcoLab implementation also includes two additional cases, which build upon version 16. The model is parallelised using EcoLab’s MPI-based parallel processing features, and finally, the “field” optimisation whereby the food data is stored in contiguous memory. EcoLab and 6 Version Repast Mason Obj-C Swarm EcoLab 253 259 281 310 322 338 337 320 336 352 358 416 419 432 515 662 753 894 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 parallel field 578 622 865 896 968 1005 1070 1144 1152 1191 1253 1614 1636 1360 1761 2174 158 158 250 256 312 306 359 258 368 381 391 497 484 501 646 753 169 214 263 296 362 316 365 369 383 409 494 670 816 Table 1: Source code line-counts (as reported by the unix command ‘wc’) for the different Stupid Model versions. Makefiles are not included (Swarm & EcoLab), since these are fairly boiler plate code, and fairly negligible. EcoLab counts include the TCL scripts. the two Java platforms seems to need a similar number of lines of code, yet the Swarm implementation needed up to three times the number. Whilst a factor of two or three in source line count is not particularly significant, it does indicate that it takes a bit more effort to implement Swarm models. In table 2, execution times for various stupid model versions is reported. As described in §2, versions 10 & 11 were run in batch mode with as much graphical output turned off as possible. The Java versions performed slightly better for version 10, and the C++ version did better on version 11. However, given the possible range of implementation strategies, one should not read too much into this, except that the myth of Java being slow relative to C++ should be now be firmly laid to rest. The result is broadly in line with other obser- vations that Java implementations tend to be within a factor of 2 of natively compiled applications[1, 7]. The results for Swarm though confirm Railsback et al’s the observation that Objective C performance lags that of the Java (and also now C++) versions. Unfortunately, my knowledge of Objective-C and Swarm internals is not up to the task of explaining this result. In version 16, the full graphical version of the model was run. This included a display of the space, a plot of the number of bugs and a histogram of bug sizes. It should be noted that the Mason implementation lacked the plot and histogram, apparently because this functionality is absent within the Mason 7 Version Repast Mason Obj-C Swarm EcoLab 3.9 14.9 1014 67 10 11 16 field 71 165 402 3.5 32.7 44 3.4 21.3 40.5 Table 2: Execution CPU times (in seconds) for several Stupid Model versions for different platforms. Versions 10 and 11 were performed in batch mode (no graphical output, no GUI control, Mason excepted), version 16 in GUI mode with a plot and histogram. EcoLab’s field version uses raster rather than canvas for display, and omits the expensive histogram widget. All these figures need considerable qualification (see text). toolkit[14] itself, but provided by 3rd party add-ons. One thing that stands out is the slowness of EcoLab. The TCL-based plotting widgets used in EcoLab (also used in Swarm) are slow relative to the equivalent Java offerings. Furthermore, this benchmark displays the space environment using a canvas, which is a high level drawing tool with roughly the same sort of functionality as a standard drawing application (eg. the drawing application in OpenOffice or Xfig). The bugs, predators and empty cells are rendered as coloured squares. The other platforms provide dedicated raster objects for rendering spatial displays. In the “field” version of Stupid Model, instead of representing the model’s objects as squares, a single pixmap object is created on the canvas and manipulated through low level Tk library calls. This amounts to about 40 lines of code, and improves the display performance dramatically. The result listed under the row “field” also omits the expensive histogram functionality (but still displayed the plot of bug numbers). 4 Parallel implementation Having put the extra work into building the space class on top of Graphcode rather than using a simple vector, it raises the question of whether Stupid Model can be effectively parallelised. The first thing that becomes apparent is that Stupid Model as specified is inherently sequential. Two bugs are not allowed to occupy the same spatial location, and movement into a location is performed on a first come first served basis. Since the order in which bugs perform their update move is randomised, the obvious parallel generalisation in a shared memory context is to use locks to prevent two bugs on different processors simultaneously moving to the same lo- cation. However, EcoLab is designed for use with distributed parallel systems, and obtaining the state of a cell located on a remote processor is expensive. In fact, in the MPI transport layer used by EcoLab, such functionality is only supported by “one-sided” communications of MPI 2, a relatively new feature that is not well supported and typically poorly implemented. Instead, the recommended 8 approach in EcoLab is to have separate communication and computation phases, with a snapshot of neighbouring data at the previous timestep supplied to each processor during the communication phase. As Stupid Model is a pedagogical model, there is no one right answer as to respecifying the model for parallelisation. Perhaps the most obvious approach would be to allow multiple bugs to share a single location within the space. This would certainly simplify the code, as additional logic was required to enforce the one-bug-per-location requirement. However, in the spirit of adventure, I propose the following protocol for allowing bugs to migrate from one processor to the next, whilst maintaining the one-bug-per-location property. As in the sequential algorithm, bugs examine their neighbourhood, and choose the cell with the If the destination lies on the current highest food resource as a destination. processor, and the cell is empty, the bug is free to move. If the destination is remote, however, the bug’s desire to move to a remote cell is lodged with an emigration register. Then after all bugs have performed their move, the emigration register is passed to the remote processor, which approves or denies the request depending on whether the destination is already occupied, or an immigration request has already been allowed. The immigration approval list is passed back to the requesting processor, and approved bugs are migrated between processors. The remaining bugs do not move. I coded this solution into the stupid-parallel version, and also the field optimised version stupid-field. None of the other versions are parallel aware code — building them and running them in parallel will only result in the model running on processor 0, with the remaining processors idle. With the stupid-parallel version, it became immediately clear that the Prepare_Neighbours() step dominated the calculation. This highlighted a hitherto unsuspected source of inefficiency in Graphcode’s Prepare_Neighbours() method. To build the list of neighbours to transmit, Graphcode loops over the neighbours of local cells, adding to the list any remote neighbour found. How- ever, this leads to many duplicates, as one cell may be the neighbour of many other cells — for the Stupid Model case, each cell in the transfer list will be duplicated 36 times. In a more common von Neumann neighbourhood of ra- dius 1 there is no duplication, and in the Moore neighbourhood of radius 1 the duplication is only 3 times. In choosing a Moore neighbourhood of radius 4 for their Stupid Model, Railsback et al. unwittingly made this inefficiency blatant. However, even with this inefficiency corrected, Prepare_Neighbours() is still an expensive overhead. The example problem I tested was the same 200 × 200 spatial grid, and so 2 × 200 × 4 × N p cells need to be transferred each time > 1 being the number of processors). This overhead can be amortised step (N by increasing the problem size. p In the stupid-field case, the food_available data is not stored in the cell, but in the additional field data structure, so is not transferred with the cell data during the Prepare_Neighbours() step. In fact, only the food data has any affect on bug movement, so Prepare_Neighbours() is eliminated altogether. In the stupid-field version of the model, we do not transfer the food data, but duplicate the update calculation on the overlap area between two processors. A 9 16 14 12 10 8 6 4 2 0 stupid-parallel stupid-field linear 0 2 4 6 8 10 12 14 16 No. processors Figure 1: Speedup curves for stupid-parallel and stupid-field for a 200 × 200 grid with 4000 stupid bugs moving and growing. Bug reproduc- tion and mortality as well as predation have been turned off. At no stage does stupid-parallel run as fast in parallel as it does sequentially, due to the over- heads of the Prepare Neighbours() step. single Prepare_Neighbours() step is done at the beginning of the model run to ensure access to the food data. Figure 1 shows the speedup curve for both the stupid-parallel and stupid-field model, for the same input script used for the stupid10 and stupid11 bench- marks reported in table 2. The parallel computing experiements were performed on Linux cluster (Be- owulf style) with dual 3GHz Pentium 4 Xeon nodes connected via Gigabit Eth- ernet. Each node has 2GB of memory. 5 Conclusion The aim of this study was to answer the following questions: • is EcoLab suitable for the sorts of agent based models that other more well known platforms are used for • what performance advantages, if any, does the use of C++ provide • what deficiencies are present in EcoLab 10 Stupid Model is a nontrivial, yet fairly simple agent based model that could be implemented without an excessive amount of programming. EcoLab has shown itself to be capable of implementing Stupid Model with about the same sort of effort reported by developers of Repast and Mason versions of the model, and was implemented in around the same number of lines of code. Furthermore, performance was on a par with these Java-based platforms. The main deficiencies encountered were: • A lack of specialised space library, or library of examples in the use of Graphcode for implementing spaces. • A lack of a simple raster object for displaying spaces. The provided canvas functionality is very slow • GUI functionality is slow compared with the Java-based functionality • the smart pointer template ref needs to be improved For addressing the space library issue, I will start with implementing a few well known ABM models to build up a library of practice. Where code appears in common, this can be refactored into a library. To address the GUI performance, a possible future strategy is to develop a Classdesc C++/Java interface to enable C++ coded EcoLab models to run under a Java framework such as Repast. A similar strategy was investigated integrating C++ and Objective C using Classdesc to look at Swarm integration, however it never found practical use and is no longer being maintained[8]. The feasibility of doing this with a Java platform will be the subject of future work. References [1] RF Boisvert, J. Moreira, M. Philippsen, and R. Pozo. Java and numerical computing. Computing in Science & Engineering [see also IEEE Compu- tational Science and Engineering], 3(2):18–24, 2001. [2] Boost C++ Libraries. http://www.boost.org/. [3] EcoLab website. http://ecolab.sourceforge.net. [4] James Gosling, Bill Joy, and Guy L. Steele, Jr. The Java Language Speci- fication. Addison-Wesley, 3rd edition, 2005. [5] V. Grimm and S. F. Railsback. Individual-based Modeling and Ecology. Princeton UP, 2005. [6] Java Grande. http://www.javagrande.org/. [7] J.P.Lewis and Ulrich Neumann. Performance of Java versus C++. http://www.idiom.com/∼zilla/Computer/javaCbenchmark.html, 2003. 11 [8] Richard Leow and Russell K. Standish. Running C++ models un- In Proceedings SwarmFest 2003, 2003. der the Swarm environment. arXiv:cs.MA/0401025. [9] Sean Luke, Claudio Cioffi-Revilla, Liviu Panait, Keith Sullivan, and Gabriel Balan. MASON: A multiagent simulation environment. Simulation, 81:517– 527, 2005. [10] Duraid Madina and Russell K. Standish. A system for reflection in C++. In Proceedings of AUUG2001: Always on and Everywhere. Australian Unix Users Group, 2001. [11] David McFadzean. SimBioSys: A class framework for biological simula- tions. Master’s thesis, Dept. of Computer Science, Calgary, Alberta, 1994. http://www.lucifer.com/david/thesis/. [12] Nelson Minar, Roger Burkhart, Christopher G. Langton, and Manor Aske- nazi. The Swarm simulation system: A toolkit for building multi-agent simulations. Technical Report WP96-06-042, Santa Fe Institute, 1996. http://www.swarm.org. [13] M.J. North, N.T. Collier, and J.R. Vos. Experiences creating three imple- mentations of the Repast agent modeling toolkit. ACM Transactions on Modeling and Computer Simulation, 16:1–25, 2006. [14] S. F. Railsback, S. L. Lytinen, and S. K. Jackson. Agent-based simulation platforms: Review and development recommendations. Simulation, 82:609– 623, 2006. [15] Steve Railsback, Steve Lytinen, and Volker Grimm. StupidModel and ex- tensions: A template and teaching tool for agent-based modeling platforms. http://condor.depaul.edu/slytinen/abm/StupidModel. [16] Marc Snir et al. MPI: the complete reference. MIT Press, Cambridge, MA, 1996. [17] Russell K. Standish and Richard Leow. EcoLab: Agent based mod- In Proceedings SwarmFest 2003, 2003. eling for C++ programmers. arXiv:cs.MA/0401026. [18] Russell K. Standish and Duraid Madina. Classdesc and graphcode: sup- port for scientific programming in C++. International Journal for High Performance Computing and Applications, 2006. submitted. [19] Bjarne Stroustrup. The C++ Programming Language. Addison-Wesley, Reading, Mass., 3rd edition, 1997. 12
cs/0506092	1	0506	2005-06-24T16:46:47	Emergent Statistical Wealth Distributions in Simple Monetary Exchange Models: A Critical Review	[ "cs.MA" ]	This paper reviews recent attempts at modelling inequality of wealth as an emergent phenomenon of interacting-agent processes. We point out that recent models of wealth condensation which draw their inspiration from molecular dynamics have, in fact, reinvented a process introduced quite some time ago by Angle (1986) in the sociological literature. We emphasize some problematic aspects of simple wealth exchange models and contrast them with a monetary model based on economic principles of market mediated exchange. The paper also reports new results on the influence of market power on the wealth distribution in statistical equilibrium. As it turns out, inequality increases but market power alone is not sufficient for changing the exponential tails of simple exchange models into Pareto tails.	cs.MA	cs	Emergent Statistical Wealth Distributions in Simple Monetary Exchange Models: A Critical Review Thomas Lux Department of Economics, University of Kiel, Olshausenstr. 40, 24118 Kiel, Germany Abstract: This paper reviews recent attempts at modelling inequality of wealth as an emergent phenomenon of interacting-agent processes. We point out that recent models of wealth condensation which draw their inspiration from molec- ular dynamics have, in fact, reinvented a process introduced quite some time ago by Angle (1986) in the sociological literature. We emphasize some prob- lematic aspects of simple wealth exchange models and contrast them with a monetary model based on economic principles of market mediated exchange. The paper also reports new results on the influence of market power on the wealth distribution in statistical equilibrium. As it turns out, inequality in- creases but market power alone is not sufficient for changing the exponential tails of simple exchange models into Pareto tails. 1 Introduction Since the days of Vilfredo Pareto, the frequency distribution of wealth among the members of a society has been the subject of intense empirical research. Recent research confirms that power-law behaviour with an exponent between 1 and 2 indeed seems to characterize the right tail of the distribution (Levy and Solomon, 1997; Castaldi and Milakovic, 2005). However, when applied to the entire shape of the empirical distribution, the power law would produce a rather mediocre fit and would be outperformed by other candidate processes like the lognormal or Gamma distributions. As it seems to emerge from the literature, a transition occurs in the data from an exponential shape to power- law behavior somewhere above the 90 percent quantile again. These and other findings should give rise to modelling efforts explaining the remarkably similar wealth distribution of many developed countries. Unfor- tunately, economic theory has been quite silent on this topic for a long time. Until recently, one had to go back the to literature of the fifties and six- 2 Thomas Lux ties (e.g., Champernowne, 1953; Mandelbrot, 1961) to find stochastic models of wealth accumulation in modern societies. Recent advances in computer technology, however, open another avenue for analysis of the emergence of wealth distributions allowing this issue to be studied in a computational agent- based framework. Such a bottom-up approach could, in principle, be helpful in isolating the key mechanisms that apparently lead to a stratification of wealth in advanced economies. As it appears, this path has been pursued re- cently by physicists rather than economists (cf. Bouchaud and M´ezard, 2000; Dragulescu and Yakovenko, 2000; Chakraborty and Chakrabarty, 2000; Sil- ver, Slad and Takamoto, 2002, among others). However, it has been entirely overlooked in the pertinent publications that these models have an important predecessor in the sociological literature. Investigating essentially the same structures already almost twenty years ago, Angle, 1986, might be consid- ered as the first contribution to agent-based analysis of wealth formation. In the following, I will shortly review Angle’s interesting work as the prototypi- cal agent-based model of wealth dynamics, based on particle-like microscopic interactions of agents. I will point out aspects of this class of models (cov- ering most of the econophysics contributions mentioned above) that would be considered to be problematic by economists (section 2). As an alternative framework, I will, then, review the contribution by Silver et. al. (2002) which much better fits into standard economic reasoning, but nevertheless provides a similarly simple formalization of an agent-based exchange model (section 3). Section 4 presents some additional results expanding on the seminal frame- work of Silver et. al. Conclusions are in section 5. 2 Angle’s Surplus Theory of Social Stratification and the Inequality Process In a long chain of papers covering more than 15 years, sociologist John Angle has elaborated on a class of stochastic processes which he first proposed in 1986 as a generating mechanism for the universal emergence of inequality in wealth distributions in human societies. His starting point is evidence he attributes to archeological excavations that inequality among the members of a community is typically first found with the introduction of agriculture and the ensuing prevalence of food abundance: While simpler hunter/gatherer societies appear to be rather egalitarian, production of a “surplus” beyond subsistence level immediately seems to lead to a “ranked society” or some kind of “chiefdom” (Angle, 1986, p. 298). So as soon as there is some excess capacity of food, processes seem to be set into motion from which inequality emerges. Angle, surveying earlier narra- tive work in sociology, sees this as the result of redistribution by which some members of society succeed in grabbing some of the surplus wealth of others. The relevant empirical observations are summarized as follows: Emergent Statistical Wealth Distributions 3 “Proposition 1: Where people are able to produce a surplus, some of the surplus would be fugitive and would leave the possession of the people who produce it. ... Proposition 2: Wealth confers on those who possess it the ability to extract wealth from others. So netting out each person’s ability to do this in a general competition for surplus wealth, the rich tend to take surplus away from the poor.” (Angle, 1986, p. 298). According to Angle, the expropriation of the losers happens via (1) theft, (2) extortion, (3) taxation, (4) exchange coerced by unequal power between the participants, (5) genuinely voluntary exchange, or (6) gift (ibid.). The process he designs as a formalisation of these ideas is a true interacting particle model: in a finite population, agents are randomly matched in pairs and try to catch part of the other’s wealth. A random toss Dt ∈ {0, 1} decides which of both agents is the winner of this conflict. Angle in various papers considers cases with equal winning probabilities 0.5 as well as others with probabilities being biased in favor of either the wealthier or poorer of both individuals. If the winner of this encounter is assumed to take away a fixed proportion of the other’s wealth, ω, the simplest version of the “inequality process” leads to a stochastic evolution of wealth of individuals i and j who had bumped into each other according to: wi,t = wi,t−1 + Dtωwj,t−1 − (1 − Dt)ωwi,t−1, wj,t = wj,t−1 + (1 − Dt)ωwi,t−1 − Dtωwj,t−1. (1) Time t is measured in encounters and one pair of agents from the whole population is chosen for this interaction in each period. Angle (1986) shows via simulations that this dynamics leads to a stationary distribution which can be reasonably well fitted by a Gamma distribution. Angle (1993) provides an argument for why the Gamma distribution approximates the equilibrium distribution of the process for empirically relevant values of its parameters. Later papers provide various extensions of the basic model. While the expo- nential decay of the Gamma distribution might not be in accordance with power law behavior at the high end of the richest individuals, Angle’s model is the first agent-based approach matching several essential features of em- pirical wealth distributions which he carefully lists as desiderada (i.e. stylized facts) for a theory of inequality. Among other properties, he emphasizes the uni-modality with a mode above minimum income which could not be repro- duced by a monotonic distribution function. Angle is also careful to point out that with binned data, realizations of his process would be hard to distinguish from realizations of Pareto random variables which he demonstrates via a few Monte Carlo runs. Unfortunately, Angle’s process might be hard to accept for economists as a theory of the emergence of inequality in market economies. 4 Thomas Lux First, a glance at the list of the six mechanisms for appropriation of another agent’s wealth might raise doubts about their relative importance in modern societies: for most countries of the world, “theft” should perhaps not be the most eminent mechanism for stratification of the wealth distribution. Note also that “genuinely voluntary exchange” is listed only at rank 5 and behind “exchange coerced by unequal power”. However, voluntary exchange is at the heart of economic activity at all levels of development rather than being a minor facet. However, despite being mentioned in the list of mechanisms of redistribu- tion, voluntary exchange is not really considered in Angle’s model in which an agent simply takes away part of the belongings of another. What is more, this kind of encounter would - in its literal sense - hardly be imaginable as both agents would rather prefer not to participate in this game of a burglar economy - at least if they possess a minimum degree of risk aversion. The model, thus, is not in harmony with the principle of voluntary participation of agents in the hypothesized process which economists would consider to be an important requirement for a valid theory of exchange activities. One should also note that another problem is the lack of consideration of the measure- ment of wealth (in terms of monetary units) and the influence of changes of the value of certain components of overall wealth. Despite these problematic features from the viewpoint of economics, An- gle’s model deserves credit as the first contribution in which inequality results as an emergent property of an agent-based approach. A glance at the recent econophysics literature shows that the basic building blocks of practically all relevant contributions share the structure of the inequality process formalized by equation (1). The inequality process is, for example, practically identical to the process proposed by Bouchaud and M´ezard (2000) and isomorphic to almost all other models mentioned above. This recent strand of research on wealth dynamics is, therefore, almost exemplary for the lack of coordination among research pursued on the same topic in different disciplines and for the unfortunate duplication of effort that comes along with it. Interestingly, the above criticism concerning the structure of the exchange process had also been voiced in a review of monetary exchange models devel- oped by physicists by Hayes (2002) who introduced the label of “theft and fraud” economies, but restricted it to variants in which the richer could lose more (in absolute value) than the poor. However, it is not clear why models which introduce a certain asymmetry to avoid this kind of exploitation should not also suffer from the lack of willingness of agents to participate in their ex- change processes. It, therefore, appears that one might wish to reformulate the “burglar economies” in a way that brings elements of voluntary economic ex- change processes into play. While the economics literature has not elaborated on wealth distributions emerging from exchange activities within a group of agents, a huge variety of approaches is available in economics that could be utilized for this purpose. An interesting start has been made in a recent paper by Silver, Slud and Takamoto (2002) which contains a two-good general equi- Emergent Statistical Wealth Distributions 5 librium model of an economy with heterogenous agents. Somewhat ironically, the overall outcome of this model is the same as with the inequality process: the stationary wealth distribution turns out to be a Gamma distribution. 3 An Exchange Economy with Changing Preferences Unlike the framework reviewed in the previous section, the setting of Silver et al. is an extremely familiar one for economists. Their economy consists of two goods, denoted x and y which necessitate the introduction of a relative price p being defined as the current value of a unit of good y in units of good x. Note that with this assumption, considerations of revaluation of wealth components come into play which are altogether neglected in the sociological/physical models. All agents of the economy have their preferences formalized by a so-called Cobb-Douglas utility function: Ui,t = x fi,t i,t · y 1−fi,t i,t . (2) Here, i and t are indices for the individuals and time, respectively. xi,t and yi,t are, therefore, the possessions of good x and y by individual i at time t and fi,t ∈ [0, 1] is a preference parameter which might differ among individuals and, for one and the same individual, might also change over time. Ui,t, then, is utility gained by individual i at time t. Individuals start with a given endowment in t = 0 and try to maximize their utility via transactions in a competitive market where one good is exchanged against the other. Given their possessions of both goods at some time t − 1, it is a simple exercise to compute their demands for goods x and y at time t given the current preference parameter fi,t: xi,t = fi,t(xi,t−1 + ptyi,t−1), yi,t = (1 − fi,t)(cid:18) xi,t−1 pt + yi,t−1(cid:19) . (3) In (3), we have used the standard assumption that agents take the price as given in a competitive market. Note that this market, therefore, dispenses with any assumption of unequal exchange or even exploitation which is so central to the microscopic process of the previous chapter. Summing up demand and supply by all our agents, we can easily calculate the equilibrium price which simultaneously clears both markets: (1 − fi,y) xi,t−1 Pi fi,tyi,t−1 . (4) 6 Thomas Lux pt = Pi After meeting in the market, each agent possesses a different bundle of goods and his wealth can be evaluated as: wi,t = xi,t + ptyi,t. (5) The driving force of the dynamics of the model by Silver et al. is simply the assumption of stochastically changing preferences: all fi,t are drawn anew in each period independently for all individuals. In the baseline scenario, the fi,t are simply drawn from a uniform distribution over [0, 1], but other distri- butions lead to essentially the same results. The dynamics is, thus, generated via the agents’ needs to rebalance their possessions in order to satisfy their new preference ordering. With all agents attempting to change the composi- tion of their “wealth”, price changes are triggered because of fluctuations in the overall demand for x and y. This leads to a revaluation of agents previous possessions, xi,t−1 and yi,t−1, and works like a capital gain or loss. To summarize, we have a model in which all agents are identical except for their random preference shocks and no market or whatsoever power is attributed to anyone. The resulting inequality (illustrated as the benchmark case pm = 0 in Fig. 1) is, therefore, the mere consequence of the eventualities of the history of preference changes and ensuing exchanges of goods. We, therefore, do not have to impose any type of “power” in order to endogeneously generate a stratification of the wealth distribution that - like the model of section 2 - is able to capture all except the very end (the Pareto tail) of the empirical data. 4 Some Extensions of the Monetary Exchange Model The model by Silver et al. demonstrates that stratification of wealth can result from an innocuous exchange dynamics without agents robbing or fleecing each other. It should, therefore, be a promising avenue to supplement the simpler dynamic models in the previous section. In some extensions, we, therefore, tried to explore the sensitivity of this approach to certain changes of its un- derlying assumptions. Among the many sensitivity tests we could imagine, we started with the following variations of the basic framework: • • • replacement of market interaction by pairwise exchange, introduction of agents with higher bargaining power so that the outcome of pairwise matches could differ from a competitive framework, introduction of natural differences among agents of some kind: here we assumed that for part of the population, preference changes are less pro- nounced than for others, Emergent Statistical Wealth Distributions 7 • introduction of savings via a framework which allows for money as an additional component in the utility function. Due to space limitations, we will not provide detailed results on all of these experiments, but will rather confine ourselves to one particularly interesting variant: the introduction of market power. Introducing market power of some sort is certainly interesting in light of the focus of the sociological and physics-inspired literature on issues of power of some individuals over others. Different avenues for implementing market power seem possible. Here, for the sake of a first exploration of this issue, we chose a very simple and extreme one. We assume that part of the population can act as monopolists in pairwise encounters: if they are matched with an agent from the complementary subset of non-monopolists, they can demand the monopoly price. If two non-monopolists are matched, we compute the competitive solution. We do the same when two monopolists meet each other assuming that their potential monopolistic power cancels out. Although this is an almost trivial insight in economics, it should be noted that the monopolist is not entirely free in dictating any price/transaction combination, but has to observe the constraint that the other agent has to voluntarily participate in the transaction. Since the option to not agree on the transaction would leave the monopolist with a zero gain as well, even in this extreme market scenario “exploitation” is much more limited than in a world of “theft and fraud”. Note also that although one could perhaps speak of exploitation (when comparing the monopoly setting with the competitive price), no expropriation is involved whatsoever since even the non-monopolist will still increase his utility by his transaction with the more “powerful” mo- nopolist. As it turns out, allowing for monopoly power indeed changes the resulting wealth distribution. Fig. 1 shows the pdf for (fixed) fractions of monopolists. Varying the proportion of monopolists from 0 (the former competitive sce- nario with pair-wise transactions) to 0.4 we see a slight change in the shape of the distribution. As it happens all distributions still show pronounced expo- nential decline and can be well fitted by Gamma distributions. However, the estimated parameters of the Gamma distribution show a systematic variation. In particular, the slope parameter decreases with the fraction of monopolists, pm. A closer look at the simulation results also shows that the average wealth of monopolists exceeds that of other agents but the difference decreases with increasing pm. Note that the Gini dispersion ratio (G) is a negative function of λ for the Gamma distribution: G = Γ (λ+0.5) π2Γ (λ+1) , so that the increasing in- equality would also be indicated by this popular statistics. 8 Thomas Lux Fig. 1: Kernel estimates of statistical wealth distributions with different frac- tions of monopolistic agents pm. Results are from simulations with 10,000 agents recorded after 5 ∗ 105 trading rounds. The result that monopoly power is not neutral with respect to the distri- bution of wealth is certainly reassuring. However, we may also note that its introduction in the present framework does not lead to a dramatic change of the shape of the distribution. In particular, it does not seem to lead to any- thing like a Pareto tail in place of the exponential tail of the more competitive society. Since we have already chosen the most extreme form of market power in the above setting it seems also unlikely that one could obtain widely dif- ferent results with milder forms of bargaining power. 5 Conclusions and Outlook to Future Research What kind of conclusions can be drawn from this review of different ap- proaches to agent-based models of wealth stratification? First, it is perhaps obvious that this author would like to advocate an approach in line with standard principles of economic modelling. If one is not willing to follow the emphasis of the sociological literature on all types of exertion of power, and Emergent Statistical Wealth Distributions 9 if one tends to the view that wealth is influenced more by legal economic ac- tivity than by illegal theft and fraud, economic exchange should be explicitly incorporated in such models. This would also help to identify more clearly the sources of the changes of wealth. Note that despite the voluntary participation of agents in the exchange economy and the utility-improving nature of each trade, a change in the distribution of wealth comes with it. The difference to earlier models is that the changes in wealth are explained by deeper, under- lying economic forces while they are simply introduced as such in the models reviewed in sec. 2. Market exchange models also allow to consider changes of monetary evaluation of goods and assets as a potentially important source of changes in an individual’s nominal wealth. Unfortunately, monetary exchange so far does not provide an explanation of the power-law characterizing the far end of the distribution. As we have shown above, even an extremely unequal distribution of market power within the population seems not sufficient to replicate this important empirical fea- ture. Following recent proposals in the literature one could try additional positive feedback effects that give agents with an already high level of wealth an additional advantage (West, 2005; Sinha, 2005). In the above model, one could argue that the more wealthy agents would also acquire more bargaining power together with their higher rank in the wealth hierarchy. Whether this would help to explain the outer region, re- mains to be analyzed. However, there are perhaps reasons to doubt that the Pareto feature might be the mere result of clever bargaining. A glance at the Forbes list of richest individuals (analyzed statistically by Levy and Solomon, 1997, and Castaldi and Milakovic, 2005) reveals that the upper end of the distribution is not populated by smart dealers who in a myriad of small deals succeeded to outwit their counterparts. Rather, it is the founders and heirs of industrial dynasties and successful companies operating in new branches of economic activity whom we find there1. The conjecture based on this anec- dotal evidence would be that the upper end of the spectrum has its roots in risky innovative investments. Few of these succeed but the owners behind the succeeding ones receive an overwhelming reward. This would suggest that models without savings and investments should lack a mechanism for a power law tail. One would, therefore, have to go beyond such conservative models and combine their exchange mechanism (which works well for the greater part of the distribution) with an economically plausible process for the emergence of very big fortunes. 1 While the majority of entrants in the Forbes list might fall into that category, a few are, in fact, rather suggestive of “theft and fraud” avenues to big fortunes. 10 Thomas Lux Acknowledgement I would like to thank John Angle, Mishael Milakovic and Sitabhra Sinha for stimulating comments and discussions and Bikas Chakrabarti for raising my interest in the issues explored in this paper. References Angle, J., 1986, The Surplus Theory of Social Stratification and the Size Distribution of Personal Wealth, Social Forces 65, 293-326. Angle, J., 1992, The Inequality Process and the Distribution of Income to Blacks and Whites, Journal of Mathematical Sociology 17, 77-98. Angle, J., 1993, Deriving the Size Distribution of Personal Wealth from “The Rich Get Richer, the Poor Get Poorer”, Journal of Mathematical Sociology 18, 27-46. Angle, J., 1996, How the Gamma Law of Income Distribution Appears In- variant under Aggregation, Journal of Mathematical Sociology 31, 325-358. Bauchaud, J.-P. and M. M´ezard, 2000, Wealth Condensation in a Simple Model of Economy, Physica A 282, 536-545. Castaldi, C. and M. Milakovic, 2005, Turnover Activity in Wealth Port- folios, Working Paper, University of Kiel. Chakraborti, A. and B. Chakrabarti, 2000, Statistical Mechanics of Money: How Saving Propensities Affects its Distribution, European Physical Journal B 17, 167-170. Champernowne, D., 1953, A Model of Income Distribution, Economic Jour- nal 53, 318-351. Dragulescu, A. and V. Yakovenko, 2000, Statistical Mechanics of Money, European Physical Journal B 17, 723-729. Hayes, B., 2002, Follow the Money, American Scientist 90, 2002, 400-405. Levy, M. and S. Solomon, 1997, New Evidence for the Power-Law Distri- bution of Wealth, Physica A 242, 90-94. Mandelbrot, B., 1961, Stable Paretian Random Functions and the Multi- plicative Variation of Income, Econometrica 29, 517-543. Sinha, S., 2005, Pareto-Law Wealth Distribution in an Asset Exchange Econ- omy with Wealth Dependent Asymmetry, Working Paper, Institute of Math- ematical Sciences, Chennai. Silver, J., E. Slud and K. Takamoto, 2002, Statistical Equilibrium Wealth Distributions in an Exchange Economy with Stochastic Preferences, Journal of Economic Theory 106, 417-435. Scafetta, N., B. West and S. Picozzi, 2003, A Trade-Investment Model for Distribution of Wealth, cond-mat 0306579.
1805.05999	2	1805	2018-05-17T09:33:35	Agent Based Rumor Spreading in a scale-free network	[ "cs.MA", "cs.SI" ]	In the last years, the study of rumor spreading on social networks produced a lot of interest among the scientific community, expecially due to the role of social networks in the last political events. The goal of this work is to reproduce real-like diffusions of information and misinformation in a scale-free network using a multi-agent-based model. The data concerning the virtual spreading are easily obtainable, in particular the diffusion of information during the announcement for the discovery of the Higgs Boson on Twitter was recorded and investigated in detail. We made some assumptions on the micro behavior of our agents and registered the effects in a statistical analysis replying the real data diffusion. Then, we studied an hypotetical response to a misinformation diffusion adding debunking agents and trying to model a critic response from the agents using real data from a hoax regarding the Occupy Wall Street movement. After tuning our model to reproduce these results, we measured some network properties and proved the emergence of substantially separated structures like echochambers, independently from the network size scale, i.e. with one hundred, one thousand and ten thousand agents.	cs.MA	cs	Agent Based Rumor Spreading in a scale-free network Mattia Mazzoli1,2, Tullio Re1, Roberto Bertilone1, Marco Maggiora1,3 and Jacopo Pellegrino1,3 1Dipartimento di Fisica, Universitá degli Studi di Torino, 10125 Torino, Italy 2IFISC (CSIC-UIB), Campus UIB, 07122 Palma de Mallorca, Spain 3INFN Sezione di Torino, via P. Giuria 1, 10125 Torino, Italy 8 1 0 2 y a M 7 1 ] A M . s c [ 2 v 9 9 9 5 0 . 5 0 8 1 : v i X r a Abstract In the last years, the study of rumor spreading on social networks produced a lot of interest among the scientific community, expecially due to the role of social networks in the last political events. The goal of this work is to reproduce real-like diffusions of information and misinformation in a scale-free network using a multi-agent-based model. The data concerning the virtual spreading are easily obtainable, in particular the diffusion of information during the announcement for the discovery of the Higgs Boson on TwitterTM was recorded and investigated in detail. We made some assumptions on the micro behavior of our agents and registered the effects in a statistical analysis replying the real data diffusion. Then, we studied an hypotetical response to a misinformation diffusion adding debunking agents and trying to model a critic response from the agents using real data from a hoax regarding the Occupy Wall Street movement. After tuning our model to reproduce these results, we measured some network properties and proved the emergence of substantially separated structures like echochambers, independently from the network size scale, i.e. with one hundred, one thousand and ten thousand agents. Introduction Studying information diffusion attracted the attention of the scientific community in the last decade, thanks to the birth and exponential growth of many social networks like FacebookTM, TwitterTM, InstagramTM, LinkedinTM, etc. (De Domenico et al. 2013; Lotan 2012; Lerman 2016; Gomez-Rodriguez et al. 2014; Zollo et al. 2015; Tambuscio et al. 2016; Serrano et al. 2015; Liu & Chen 2011; Zollo et al. 2017; de C Gatti et al. 2013; Huang et al. 2016; Bessi et al. 2016). The study is interesting and non trivial since it shows a complexity due to a double feedback between topology and users' properties. Indeed it is unknown whether real social networks have been shaped to this structure because of users' interests determining their friendships, or if the users' inter- ests were influenced by their personal network topologic structure, i.e. their friends. To develop our model we chose an innovative approach for the field of information diffusion on networks which is a Multi-Agent based model. Multi-Agent systems are more suitable to investigate this kind of social behavior complexity because, unlike object-oriented systems, agents are capable of performing autonomous actions based on self-interest at run-time. Agents have stronger autonomy and they are social, they can communicate with each other through protocols, be proactive and reactive. Moreover, each of them has its own perception of the environment it lives within. According to their perceptions, agents may decide to autonomously act on the environment, in order to meet their design objectives. For these rationales the agent-based approach seemed to be promising for our pur- pose. (Wooldridge 2009). There are lots of simulation models built to analyze the viral behavior of a diffusion as an emergent property (Serrano et al. 2015; Liu & Chen 2011; de C Gatti et al. 2013). Micro assumptions similar to ours on the agents' behavior like threshold of skepticism, reliability of the news, influence of the neighbors, communication between agents, have been made from other recent studies (Tambuscio et al. 2016; de C Gatti et al. 2013; Huang et al. 2016), but most of them are models which are based on the epidemiologic approach, i.e. SIR models, which have been contrasted in some real data recent analysis (Lerman 2016). One main difference between these models and ours is the existence of emergent debunking behavior agents, which 1 1 THE NETWORK 2 are agents that try to fact-check the information they find. In the SIR models these individuals are usually rep- resented by stiflers, but once they become stiflers they stop interacting with the remaining infected nodes. On the contrary, in our model debunker agents try to make spreaders change their mind on the hoax they spread. As seen in the empirical study reproduced in (Zollo et al. 2017) the behavior of these debunker agents is observed in the polarization of the network in various echo chambers. These echo chambers are resonance bubbles where the information spread from users forms a loop in their friends network. Information in these contexts does not spread uniformely with all the neighbouring nodes of the user but stays trapped in his social circle, made of people who usually share similar contents. These bubbles may be the result of the FacebookTM news feed algorithms which decide which contents have to be shown to the users according to the people with whom they interact more (Facebook 2015) and the phenomenon known as confirmation bias. 1 The network The environment of our agents model is a scale free graph generated with the Barabasi-Albert algorithm. Our network can be formalized as G = (V,E) unweighted and undirected graph, where V is the set of vertices we represent as users, E is the set of the edges, which we represent as the friendship connections on the social network. The number of edges of each node is called his degree k. A scale free graph is a graph constituted of nodes whose distribution of degrees follows a power law function of the form: P(k) ∼= k−γ (1) where P(K) is the probability to find a given degree node in the network, γ is the exponent which stands in the range 2 < γ < 3 (Barabási & Albert 1999). A fundamental aspect of these networks is the presence of many low-degree nodes and specifically few so-called "hubs": nodes with very high degree compared to the size of the net. These properties have been observed in real networks as social networks, the World Wide Web, the network of scientific collaborations, the network of movie actors collaborations (Barabási & Albert 1999) and many more. Social networks as the ones we study, follow the scale free distribution as the study itself mentions (De Domenico et al. 2013). TwitterTM is known to be a directed network due to asimmetric possibility to follow somebody without being followed, but the information diffusion can overpass this limit thanks to platform features like mentions, hashtags and trending topics. The Barabasi-Albert algorithm is the first algorithm that reproduces the structure of scale free networks through two main processes: growth and preferential attachment. The preferential attachment sets a probability for every new node added to the net to set links with the highest degree nodes already present in the graph. (Barabási & Albert 1999). 2 The model To create our model we used the Gama Platform, which is free. First of all we created a scale-free network us- ing the Barabasi-Albert built-in function to reproduce at best the architecture of a real social network. In every simulation we have a brand new network, we do not set a seed to a root for the diffusion of the information, in order to avoid to start everytime with the same fixed network, which could influence the results from the initial conditions. We verified that the built-in function effectively generated a scale-free graph with ten thousand nodes and tested its properties measuring the degree distribution as shown in Fig.1. We fitted the tail of the degree distribution of ten different graphs with N = 104 nodes each, excluding the saturation points for low degrees and we found the mean gamma exponent of our degree distribution. γ = 2.49±0.03sist±0.01stat (2) The sistematic error found is the standard deviation of ten fitted independent gamma values and it results bigger than the statistical error, obtained through the mean of gamma values error in the covariance matrix. We used the standard deviation of the gamma values to know in what range of values our graph generator works and what kind of graphs we can expect. The measure of gamma is then consistent with the expected value of gamma for a scale-free network, being 2 < γ < 3 and the relation P(k) ∼= k−γ, so we are sure to work always 2 THE MODEL 3 Figure 1. Log-log plot of the degree distribution of a single simulation graph with N = 104 nodes, fd is the frequency of the degree classes k. The straight line represents the power law interpolation of the points, which in this case resulted in γ = 2.57±0.02. The statistical error is calculated through the covariance matrix of the fit. with scale free graphs (Barabási & Albert 1999). The γ value measured in (De Domenico et al. 2013) is γ = 2.5, so we work in the same topological conditions. The node species corresponds to our social network users, while the edge species represents the kind of inter- action and relationship between the users. We created two species because we want our model to be flexible for further studies and implementations. Indeed our model could be developed to a dynamic network, which means to kill edge-agents when friendships end and to create new edge-agents when a new friendship arises. We represented a single news as a single in- stance global variable between 0 and 1. The news is accessible with a visualization probability for every agent to simulate the information overflow in the feed. The choice to limit the study to a one-dimensional problem is given by the necessary initial simplification of the model. Further developments will allow for multiple topics inside the news or for multiple news in a single network. Every agent has the possibility to choose whether to spread or not the news, depending on his personal preparation on the topic, which is an individual private threshold randomly assigned at the instantiation of the agents. In our model we represented three different types of diffusion: • spontaneous spreading of the information after direct visualization: happens when news > threshold, which means that the information is reliable enough to the agent; • collective influence: when more than 30% (López-Pintado 2006) of agent's friends are spreaders or a very influent hub agent between the friends of mine shared the information, the agent's threshold decreases, which makes him more gullible. The diffusion induced by friends is an automatic communication we assume to happen between the agents; • communication persuasion: happens when undeployed agents are friends of spreader agents, these send them messages to inform them about the validity of the news. If the interlocutors have a similar prepara- tion on the topic, the undeployed agent's threshold decreases in order to raise his probability of spreading due to augmented faith in his friend; We distinguished our nodes in the network with different colors in order to visualize the evolution of the system during the diffusion. The red ones are the undeployed nodes, i.e. non-spreaders agents, the blue ones are those who spontaneously visualized and believed the information, the green ones are those who have been influenced by their friends, and the yellow ones are those who deployed due to communication. After a fixed time every agent stops spreading and turns off independently from the others, in fact the transmissibility of information diminishes over time as information loses novelty. The probability to retweet information on TwitterTM does not depend on its absolute age, but only the time it first appeared in a user's social feed, as a study demonstrates 3 SPREADING OF A TRUE NEWS 4 (Hodas 2012). Indeed this time of deactivation is taken into account since the news has been spread from the agent and in this model it is constant and fixed for every agent. In this first simulations the agents are not allowed to reactivate. Running the model with different network sizes we found similar times of persistence of the news and observed the same relaxation curves, so we can suppose the model does not depend qualitatively on the scale of the network, but only on the parameters of the simulation. To test the reproducibility of the experiments, we set a seed to the model in order to keep the dynamics similar. We verified the stability of the experiments running the simulations to twenty thousands cycles. We measured the density of activated users, i.e. spreader agents, for every class of degree present in the network, to see how the behaviour of our agents depends on their own connectivity and on the reliability of the news. In Fig.2a we observed that news reliabity (a) (b) Figure 2. Average density of activated agents fa(k) for every class of degree k over twenty iterations in a network with n = 103 nodes. Three different news are spreaded with different resulting dynamics. a) Degree range [1,10], the errorbars represent the standard errors for every class; b) Degree range [10,200]. is more salient for low connected nodes, i.e. agents with few neighbours, whereas high connected nodes in Fig.2b show a common behaviour which seems to be more sensitive to social influence and play a fundamental role in the spreading of the news in the network. High degree classes are taken into account in the final average only if they appear in our iterations. 3 Spreading of a true news We simulated the diffusion of a single news with maximum value of reliability r = 0.99 on a ten thousand nodes network, the probability of visualization of the news is v1 = 10 and every agent has the possibility to spread the news in three different ways of diffusion: spontaneous visualization, collective influence and communication persuasion. We plotted the density of active users versus time and compared the results of the simulation with a SIR rumor spreading model (Zanette 2001) (derived from the epidemiologic SIR model), which is one of the most used models to explain social contagion. Both models are runned over scale free networks generated through the Barabasi-Albert model. The model has been studied analytically through differential equations regulated by two parameters: α to represent the rate of transition to stifler (R), λ to represent the transition rate to spreader (S) (Barrat et al. 2008). At each time step a randomly chosen spreader agent i contacts another element j. I + S → 2S If j is in the ignorant state, it becomes a spreader; S + S → S + R If, on the other hand, j is a spreader or stifler, i becomes a stifler; S + R → 2R (2) (3) (4) 4 SPREADING OF HIGGS'S BOSON DISCOVERY 5 The compartment I represents the ignorant users, S the spreaders, R the stiflers. The results in Fig.3 have been obtained by the average over ten different simulations for both models, both Figure 3. Plot of the MAS (blue) and SIR (red) density of active users da averaged over ten simulations with N = 104 nodes with the relative statistical error bars only. The error bars represent the standard errors of the dynamic at that time. on scale free networks. The error bars represent the standard deviation of the number of activated agents at each time. The SIR model we used follows the equation (2-3-4) described above. The difference between the Multi-Agent based model and the SIR model can be shown in the different diffusion rates in the early times of spread of the news. The news spreading in the SIR models goes strictly viral in the first times, while in the agent based model the diffusion rate is slightly smoother. To fit as best as possible the peak of the spreading in the MAS simulation, we set the SIR model parameters with α = 0.05 and λ = 0.27. Of course this approach is not enough to explain the real dynamics of an information diffusion: as shown in a recent study (Lerman 2016), cognitive limits may explain the difference between information spreading and virus contagion due to friendship paradox (Feld 1991) in social networks and information overflow (Gomez- Rodriguez et al. 2014). 4 Spreading of Higgs's Boson discovery In this phase, we simulated the spread of the announcement of the discovery of Higgs Boson and compared the results with the empirical data measured by (De Domenico et al. 2013) in the paper and with the results obtained by the analysis we made from the free database downloadable from (SnapStanford 2015). Figure 4. Plot of the real density of active users during the discovery announcement of the Higgs Boson on TwitterTM versus time expressed in hours. Data extracted from (De Domenico et al. 2013). 4 SPREADING OF HIGGS'S BOSON DISCOVERY 6 In Fig.4 we can see the number of active users versus time who spreaded the rumor. In our model we used N = 104 agents, as shown in Fig.5, which could communicate and share information between themselves as already said before. The simulation starts with a small reliability information r = 0.45 at time t = 0 and only later on the news is confirmed officially and gains the value of maximum reliability r = 0.99 at time t = 20. (a) (b) (c) Figure 5. Plot of the density of activated agents da versus time averaged over ten simulations obtained with our MAS model with N = 104 nodes for three different values of first visualization: a) v1 = 0.10; b) v1 = 0.05; c) v1 = 0.01. A qualitative representation is given from the number of active users from one simulation in Gama in Fig.6. (a) (b) Figure 6. A qualitative screenshot from the Gama display shows the number of active users Na versus time and the different kinds of spreading versus time we can simulate and observe in our MAS model. a) Diffusion in a network with N = 104 nodes; b) Diffusion in a network with N = 103 nodes. (a) (b) Figure 7. Log-scaled plot (a) Real data density of active users versus time expressed in hours. Picture from the original paper (De Domenico et al. 2013). (b) Fraction of active users in the network na versus time expressed in hours. Screenshot from the Gama display of a single realization in our MAS model with N = 104 nodes. The blue dots are the cumulative spreaders of the news, the red ones are those who spontaneously spreaded the news, the green ones are those who have been influenced from the collectiveness, the yellow ones are those who changed their mind due to communication persuasion. We can observe some similar trends in the curve of the density of active users which represent the different moments of the diffusion in Fig.7. Of course our graph is smaller than the one used in the paper (De Domenico 5 SPREADING OF MISINFORMATION AND CORRECTION 7 et al. 2013), but we can say that the activation dynamics in the graphic are the same, at least from a qualitatively perspective. (a) (b) (c) Figure 8. (a) The graph obtained from the free database of the study, represented in Python with NetworkX describes the interactions during one second in the moment of maximum activity in Twitter. Data extracted from (De Domenico et al. 2013); (b)Screenshot of our network during another moment of maximum activity of our agents. The red dots are those agents who are still undeployed about the news, the green ones are those who have been convinced by their neighbors to share the news, the blue ones are those who spontaneously shared the news, the grey ones are those who stopped sharing and won't reactivate. Network with N = 102 nodes; c) Network with N = 104 nodes. In Fig.8a we show a focus on one hub of the real diffusion network excrated from (De Domenico et al. 2013), while in Fig.8b,c two representations of our model network in two different sizes. 5 Spreading of misinformation and correction After that, we developed a model using the same features but, this time, we introduced a new kind of user who is able to recognize the fakeness of the news and alert his neighbors. The oranges are those who have a threshold-news difference big enough to allow them to contrast the spreading of the misinformation. When an orange agent is aware of the misinformation, he communicates back to those who tried to convince him before. These ones can then be converted to oranges if the communication happens between two users who both have the similar knowledge of the topic. Figure 9. Plot of the real data number of tweets versus time expressed in hours for the NBCTM hoax during an OWS event (Lotan 2012). We simulated with networks of different dimensions, scale-free, as shown in Fig.10 and gave the news with a reliability of r = 0.67, which appeared to be the optimal value to reproduce the observed dynamics in this case. After some time the news happens to be false and his reliability decreases to r = 0.48. The agents recognize the change of the news reliability value and a critical group of users arises. We compared this result 6 ECHO CHAMBERS IN THE NETWORK 8 with the study presented on (Lotan 2012) about the news spreading on Twitter of the protests of Occupy Wall Street in Fig.9. In this case, a false news was spread from the NBC and retweeted from lots of users. Twenty minutes after the correction of the misinformation appeared on Twitter, but the cascade of the correction has been registered to be too weak to contrast the virality of the misinformation. In our model the dynamic is represented by the number of active users versus time. We can notice the similarity in the dynamics of misinformation and correction and see that sadly the correction doesn't take over the misinformation cascade and users keep on sharing a false news. (a) (c) (b) (d) Figure 10. Plot of the number of active users Na versus time in our model. The first news reliability is r = 0.67 but after three cycles turns to be r = 0.48, the first visualization value is v1 = 0.15, the second visualization value is v2 = 0.6. The blues are the users who spread the misinformation, the oranges are those who shared the correction. a) Network with N = 5∗ 103 nodes; b) Network with N = 104 nodes; c) Network with N = 2∗ 104 nodes; d) Network with N = 5∗ 104 nodes. 6 Echo chambers in the network Following the work reported in (Zollo et al. 2017) we studied the emerging properties of the network after a spreading of a news. We analyzed the final threshold distribution of the agents and we made some statistics of it to observe if there was a polarization of the skepticism in the final population for various values of news realiability. We drew nine histograms, one for every different value of news reliability, averaged over ten iterations to visualize the agents skepticism threshold distribution properly in a range from thmin = 0.45 to thmax = 0.9 in Fig.11a. We observe that there is a polarization of the criticism over the news spreaded and this depends on the reliability of the news we gave at the beginning of the simulation. Clearly we see that the critical value of the news reliability stands in a range between r = 0.5 and r = 0.65, so we add in Fig.12 a qualitative representation of a simulation final state of the diffusion graph for a r = 0.50 reliability news. We select the nodes who spreaded or debunked the news only, because we are interested in the properties of a news diffusion graph, which is what we can investigate through the use of online social networks APIs. This representation has been made considering only those links between agents whose thresholds of skepticism did not differ more than ∆th = 0.4, this could be a starting point to consider further links removal and move to dynamic networks. Looking at these picture we can see that a threshold similarity facilitates the communication 6 ECHO CHAMBERS IN THE NETWORK 9 (a) (b) Figure 11. (a) Plot of the final threshold distribution f for nine different values of news reliability averaged over ten simulations and normalized. The histograms show a polarization of the agents skepticism when the news reliability is 0.50 < r < 0.65; (b) Average color assortativity coefficients rc of twenty realizations for different news realiabilities r on two different network sizes. The relative error bars represent the standard errors of the resulting assortativity between agents and the emergence of proto echo-chambers (Bessi et al. 2016). Nonetheless, we have to look at the properties of the network to see if the topology induces any effect on the diffusion process, e.g. the tendency of our agents to be connected to agents with the same features, i.e. assortativity. (a) (b) (c) Figure 12. Qualitative representation of our network final state with r = 0.50. The nodes have been connected to their friends only if their thresholds do not differ more than ∆th = 0.4. The nodes in blue are those who shared the news, the orange ones are the debunkers. Network reproduced with Gephi using the ForceAtlas2 algorithm (Jacomy et al. 2014) to visualize the connections. a) Network with n = 104 nodes; b) Network with n = 102 nodes; c) Network with n = 103 nodes. To understand this we looked at the color of our users, which tells us if they were spreaders or debunkers, we measured the attribute assortative coefficient (Newman 2003) in the final states for various news reliabilities, we averaged the values over ten iterations to have a minimal statistic of it and shown it in Fig.11b. We observe that for contrasted reliability news the averaged assortativity raises while it goes disappearing as the news becomes more reliable. Of course the spreading of a unique news is not enough to reproduce the case studied by Zollo et al. in (Zollo et al. 2017), indeed we cannot say that this proves the existence of the same echo chambers in our model, anyway we know that our agents are strongly influenced by neighbours, so we observe a tendency of our nodes to behave as the nodes with whom they are connected to. REFERENCES Conclusions 10 This work gives a contribution to the field of information diffusion because it offers a new framework of anal- ysis, which takes into account both Multi-Agent based modeling and network science. Our model behaves well with respect to various and different studies on real social networks data, we managed to reproduce some important parameters and features, but first of all we managed to reproduce real phenomena at the macroscopic level. All these phenomena we modeled in every part of this study are the result of an emergent behaviour of our agents, in particular of our hypotheses on the interactions at the microscopic level we made from common experience on the internet or reproduced from other studies on the topic. Concerning the case of the Higgs Boson discovery, we reproduced the diffusion of the news through the mea- sure and the modeling of the density of active users over time, with significant similarity in the quality of the dynamics, considering at the micro level different sets of parameters as the visualization and the news reliabil- ity, which came out to be the most sensitive parameters of our model. We tried to simulate the diffusion of a hoax and its correction in the case of the National Broadcasting Company spreading during the Occupy Wall Street movement in New York. In this case we had the number of tweets per time during a range of time of hours. We reproduced the qualitative dynamics of the event for various network scales simulating the activation of single users with satisfying results for the purpose of this work. We have been surprised to see that other unexpected emergent behaviors arised in our model, as described in empirical studies (Bessi et al. 2016), with- out the need for further assumptions to be made, indeed we observed the emergence of substantially separated structures like echochambers independently of the network size scale. We observed many fluctuations in the results of the simulations due to the architecture of the network, e.g. the presence of hubs, and the stochasticity of the dynamics which contributed to reproduce the complexity of a diffusion dynamic like this. As we have seen agent based models can be very useful instruments to investigate information diffusion on social networks with very few parameters. Despite the complexity of reality, this problem may need further studies to explain how misinformation is often so much more viral than corrections and replicate further phe- nomena based on the same dynamics. For example, further studies may involve raising the heterogeneity of the agents, let them change their links on a temporal network, iterate more news spreadings and check the changes in the graph structure. By means of a machine learning algorithm based on real information diffusion data, it could be possible to tune the parameters on users' real features, simulate and study attacks to few influent spreaders of a network in order to control the outbreak of scientific misinformation. References Barabási, A.-L. & Albert, R. (1999). Emergence of scaling in random networks. science, 286(5439), 509–512 Barrat, A., Barthelemy, M. & Vespignani, A. (2008). Dynamical processes on complex networks. Cambridge university press Bessi, A., Zollo, F., Del Vicario, M., Puliga, M., Scala, A., Caldarelli, G., Uzzi, B. & Quattrociocchi, W. (2016). Users polarization on facebook and youtube. PloS one, 11(8), e0159641 de C Gatti, M. A., Appel, A. P., dos Santos, C. N., Pinhanez, C. S., Cavalin, P. R. & Neto, S. B. (2013). A simulation-based approach to analyze the information diffusion in microblogging online social network. In Simulation Conference (WSC), 2013 Winter, (pp. 1685–1696). IEEE De Domenico, M., Lima, A., Mougel, P. & Musolesi, M. (2013). The anatomy of a scientific rumor. Scientific reports, 3, 2980 Facebook (2015). How news feed works. available at: https://www.facebook.com/help/327131014036297/. Feld, S. L. (1991). Why your friends have more friends than you do. American Journal of Sociology, 96(6), 1464–1477 Gomez-Rodriguez, M., Gummadi, K. P. & Schoelkopf, B. (2014). Quantifying information overload in social media and its impact on social contagions. In ICWSM, (pp. 170–179) REFERENCES 11 Hodas, N. O. (2012). How limited visibility and divided attention constrain social contagion. In In SocialCom. Citeseer Huang, W.-M., Zhang, L.-J., Xu, X.-J. & Fu, X. (2016). Contagion on complex networks with persuasion. Scientific reports, 6, 23766 Jacomy, M., Venturini, T., Heymann, S. & Bastian, M. (2014). Forceatlas2, a continuous graph layout algorithm for handy network visualization designed for the gephi software. PloS one, 9(6), e98679 Lerman, K. (2016). Internet, 8(2), 21 Information is not a virus, and other consequences of human cognitive limits. Future Liu, D. & Chen, X. (2011). Rumor propagation in online social networks like twitter–a simulation study. In Multimedia Information Networking and Security (MINES), 2011 Third International Conference on, (pp. 278–282). IEEE López-Pintado, D. (2006). Contagion and coordination in random networks. International Journal of Game Theory, 34(3), 371–381 Lotan, G. (2012). A tale of three rumors. http://blogs.harvard.edu/truthiness/2012/03/05/541. archived at http://www.webcitation.org/6z0ugjjam. Newman, M. E. (2003). Mixing patterns in networks. Physical Review E, 67(2), 026126 Serrano, E., Iglesias, C. Á. & Garijo, M. (2015). A novel agent-based rumor spreading model in twitter. In Proceedings of the 24th International Conference on World Wide Web, (pp. 811–814). ACM SnapStanford (2015). Higgs twitter dataset https://snap.stanford.edu/data/higgs-twitter.html. archived at http://www.webcitation.org/6z0vwgcwc. Tambuscio, M., Oliveira, D. F., Ciampaglia, G. L. & Ruffo, G. (2016). Network segregation in a model of misinformation and fact checking. arXiv preprint arXiv:1610.04170 Wooldridge, M. (2009). An introduction to multiagent systems. John Wiley & Sons Zanette, D. H. (2001). Critical behavior of propagation on small-world networks. Physical Review E, 64(5), 050901 Zollo, F., Bessi, A., Del Vicario, M., Scala, A., Caldarelli, G., Shekhtman, L., Havlin, S. & Quattrociocchi, W. (2017). Debunking in a world of tribes. PloS one, 12(7), e0181821 Zollo, F., Novak, P. K., Del Vicario, M., Bessi, A., Mozetic, I., Scala, A., Caldarelli, G. & Quattrociocchi, W. (2015). Emotional dynamics in the age of misinformation. PloS one, 10(9), e0138740
1711.10588	1	1711	2017-11-28T22:17:48	Utilitarians Without Utilities: Maximizing Social Welfare for Graph Problems using only Ordinal Preferences - Full Version	[ "cs.MA", "cs.GT" ]	We consider ordinal approximation algorithms for a broad class of utility maximization problems for multi-agent systems. In these problems, agents have utilities for connecting to each other, and the goal is to compute a maximum-utility solution subject to a set of constraints. We represent these as a class of graph optimization problems, including matching, spanning tree problems, TSP, maximum weight planar subgraph, and many others. We study these problems in the ordinal setting: latent numerical utilities exist, but we only have access to ordinal preference information, i.e., every agent specifies an ordering over the other agents by preference. We prove that for the large class of graph problems we identify, ordinal information is enough to compute solutions which are close to optimal, thus demonstrating there is no need to know the underlying numerical utilities. For example, for problems in this class with bounded degree $b$ a simple ordinal greedy algorithm always produces a ($b+1$)-approximation; we also quantify how the quality of ordinal approximation depends on the sparsity of the resulting graphs. In particular, our results imply that ordinal information is enough to obtain a 2-approximation for Maximum Spanning Tree; a 4-approximation for Max Weight Planar Subgraph; a 2-approximation for Max-TSP; and a 2-approximation for various Matching problems.	cs.MA	cs	Utilitarians Without Utilities: Maximizing Social Welfare for Graph Problems using only Ordinal Preferences Full Version Ben Abramowitz Elliot Anshelevich Department of Computer Science Rensselaer Polytechnic Institute Department of Computer Science Rensselaer Polytechnic Institute [email protected] [email protected] September 22, 2018 Abstract We consider ordinal approximation algorithms for a broad class of utility maximization problems for multi-agent systems. In these problems, agents have utilities for connecting to each other, and the goal is to compute a maximum-utility solution subject to a set of constraints. We represent these as a class of graph optimization problems, including matching, spanning tree problems, TSP, maximum weight planar subgraph, and many others. We study these problems in the ordinal setting: latent numerical utilities exist, but we only have access to ordinal preference information, i.e., every agent specifies an ordering over the other agents by preference. We prove that for the large class of graph problems we identify, ordinal information is enough to compute solutions which are close to optimal, thus demonstrating there is no need to know the underlying numerical utilities. For example, for problems in this class with bounded degree b a simple ordinal greedy algorithm always produces a (b + 1)-approximation; we also quantify how the quality of ordinal approximation depends on the sparsity of the resulting graphs. In particular, our results imply that ordinal information is enough to obtain a 2-approximation for Maximum Spanning Tree; a 4-approximation for Max Weight Planar Subgraph; a 2-approximation for Max-TSP; and a 2-approximation for various Matching problems. 7 1 0 2 v o N 8 2 ] A M . s c [ 1 v 8 8 5 0 1 . 1 1 7 1 : v i X r a 1 Introduction Human beings are terrible at expressing their feelings quantitatively. For example, when forming collabo- rations people may be able to order their peers from "best to collaborate with" to worst, but would have a difficult time assigning exact numeric values to the acuteness of these preferences. In other words, even when numerical (possibly latent) utilities exist, in many settings it is much more reasonable to assume that we only know ordinal preferences: every agent specifies the order of their preferences over the alternatives, instead of a numerical value for each alternative. Recently there has been a lot of work using such an implicit utilitarian approach, especially for matching and social choice (see Related Work), in situations where obtaining true numerical utilities may be difficult. Amazingly, as this line of work shows, it is often possible to design algorithms and mechanisms which perform well using only ordinal information. In fact, ordinal algorithms often perform almost as well as omniscient mechanisms which know the true underlying numerical utilities, instead of just the ordinal preferences induced by these utilities. 1 In this work we consider a relatively general network formation setting. All problems considered herein are modeled by an undirected complete graph G = (N ,E), where the (symmetric) weight w(x, y) of each edge (x, y) ∈ E represents the hidden utility of connecting agents x, y ∈ N . The goal is to form a maximum- weight (i.e., maximum utility) graph which obeys some given constraints. For example, constraints may include bounds on the maximum degree, on component size, and many others. This framework includes such problems as matching, group formation, TSP, and many others as special cases; see Section 2 for example constraints and how they lead to different important settings. In many settings we may not have access to the true edge weights w(x, y). Instead, each agent x ∈ N reports a strict preference ordering over the other agents N − {x} with whom it can connect. We assume this ordering to be consistent with the latent weights, so that w(x, y) > w(x, z) implies that x prefers y to z. While it is clearly impossible to form an optimal (i.e., maximum-utility) solution without direct knowledge of the edge weights, we show how to design good approximation algorithms for selecting a maximum weight subgraph of G, subject to a large set of constraints. As usual in this line of work, the measure of performance is simply the sum of the agent utilities. Our paper provides good approximations for a broad class of ordinal analogues to graph optimization problems representing utility maximization for multi-agent systems. Note that unlike all previous work mentioned here, we do not make additional assumptions about the structure of the edge weights: we do not assume either that the agent utilities are normalized (as in [6, 7, 10, 11, 12]), nor that they form a metric space (as in [1, 2, 3, 4, 9, 15, 16, 23]). Thus, our results demonstrate how well one can perform using only ordinal information without additional assumptions. ABC Systems More specifically, we define a class of constraints called ABC Systems which consists of three types of constraints. The first two are the familiar constraints which bound the maximum degree of each node and the maximum component size, or number of nodes in a connected component. The third constraint is a much more general requirement called attachment which only applies to nodes that are already in the same connected component. The maximization problem for an ABC System is to compute a maximum weight subgraph S ⊆ E of an undirected, complete graph G = (N ,E) such that in S every node has degree at most b, every connected component has size (number of nodes) at most c, and S ∈ A for some arbitrary attachment set A. A collection of subgraphs A is an attachment set of G if the following properties hold for all subgraphs F ⊆ E: 1) Heredity: If F ∈ A and F ′ ⊆ F then F ′ ∈ A. 2) Attachment: If F ∈ A and F + e /∈ A for some e = (u, v) /∈ F , then there is a (u, v)-path in F . 1 Note that all three of our constraints possess the heredity property which enables greedy heuristics, like the Ordinal Greedy algorithm we introduce in Section 3.1, to construct valid solutions. The intuition behind the attachment property is that if F ∈ A but F + e /∈ A, then e must have both endpoints in the same component. Therefore, the number of such edges whose addition would violate A within any component of size x is bounded by x·min (b,x−1) The utility maximization objective for ABC Systems encompasses a wide variety of well-known problems central to algorithm design. The examples we address in this paper include Max Weight b-Matching, Maximum Weight Spanning Tree, Maximum Traveling Salesperson, and Max Weight Planar Subgraph. Our results also encapsulate many other interesting optimization problems for ABC Systems which we will not discuss directly, like finding the maximum weight subgraph with minimum girth k, maximum cycle length l, or which excludes a variety of graph minors (including all 2-edge-connected minors). As we show, all such problems are amenable to knowing only secondhand ordinal information, instead of the true numerical utilities. − (x − 1), where x is bounded by c. 2 2 Maximization Problem Ordinal Greedy Omniscient Greedy Best Known ABC System AB System Spanning Tree Planar Subgraph Traveling Salesperson b-Matching b + 1 b + 1 max{2, d + 1} 2 4 2 2 max{2, d + 1} 1 3 2 2 - - 1 72/25 [8] 9/7 [19] 1 Table 1: Here we compare our results for Ordinal Greedy, known results for Omniscient Greedy, and the best known polynomial-time algorithm with full-information. All of our bounds are tight except for the one on Planar Subgraph. 1.1 Our Contributions Most algorithmic techniques for maximizing utility for the full-information setting do not translate to the ordinal information setting. These typically rely on non-local information, like comparisons between weights of non-adjacent edges, or comparing the total weights of sets of edges. This is not possible using only ordinal information. Even the fundamental, and well-studied [24, 14, 22], Omniscient Greedy algorithm, which adds edges in strictly non-increasing order of their weight, cannot be executed using only ordinal information. Instead, we focus on the natural Ordinal Greedy algorithm (defined in Section 3.1), which adds edges iteratively as long as the edge (x, y) being added is the most preferred edge for both x and y out of all the possible edges which could be added at that time. Ordinal Greedy has some very nice properties: in addition to being natural and providing high-utility solutions (as we prove in this paper), it also always creates pairwise stable solutions: no pair of agents would have incentive to destroy some of their links and form a new link connecting them. Note that the performance of Ordinal Greedy can be very different from Omniscient Greedy: see Example 1 in Section 3.1 for intuition of why this must be. In this paper, we analyze the performance of Ordinal Greedy for many ABC Systems (see Table 1). We first prove that for general ABC Systems, Ordinal Greedy always produces a solution with weight at most factor b + 1 away from optimum, and that this factor is tight. In other words, for general problems including all those in Table 1, as long as the number of connections for each node must be bounded by some small b, then using only ordinal information it is possible to compete with the best possible solution, and thus with any algorithm which knows the true numerical utilities. Such results tell us that when b is small, there is no need to find out the hidden edge weights/utilities; knowing the ordinal preferences is good enough. Second, we show that by relaxing the component size constraint (i.e., setting c to be unbounded) we can achieve significant improvements. For convenience, we call ABC Systems with the component size constraint relaxed AB Systems. We prove that as long as any solution formed by such an AB System is guaranteed to be at most d-sparse, then Ordinal Greedy forms a solution within a factor of d + 1 from optimum. Since the sparsity of a graph is at most half the average degree in any subgraph, we know that d ≤ b 2 + 1)-approximable, giving us a factor of 2 improvement over general ABC Systems. 2 . Therefore an AB System is at worst ( b This result is more powerful than it may first appear, as many important constraints yield sparse solutions. For example, since all tours and trees are 1-sparse, Ordinal Greedy provides a 2-approximation for both Maximum Traveling Salesperson and Maximum Weight Spanning Tree. And since all planar graphs are 3-sparse, we obtain a 4-approximation for Max Weight Planar Subgraph which uses only ordinal information. Lastly, we consider Max Weight b-Matching, in which the only constraint is that each agent can be matched with at most b others. This is simply an ABC System with unbounded c and A being all possible 1We use the "+" and "-" notation when adding or removing a single edge to or from a set. 3 sets. We prove that our approximation factor drops to a constant 2, regardless of the value of b. To prove the ordinal approximations above, we first demonstrate that for any ABC System (and any graphic system with the heredity property defined earlier) Ordinal Greedy achieves its worst approximation on an instance with weight function w : E → {0, 1}. We use this fact heavily to establish our approximation bounds, and believe it to be of independent interest. Note that similar results for Omniscient Greedy have relied critically on the fact that it selects edges in strictly non-increasing order by weight. Clearly this does not and cannot hold for the ordinal setting, as it is even possible for the minimum weight edge of the graph to be selected before the maximum weight edge. Because of this, our proofs require completely new approaches and techniques. 1.2 Related Work Historically, it has been common to approach problems in the ordinal setting with a normative view by designing mechanisms which satisfy axiomatic properties, like stability or truthfulness. These axiomatic properties are useful in many applications, but do not provide a quantitative measure of the quality of a solution. The notion of distortion and the implicit utilitarian framework were first introduced by [21] in the context of voting to provide such a measure. Since then the distortion, or approximation factor, of various ordinal utility maximization mechanisms has been studied, particularly for matchings [3, 4, 5, 12, 9] and social choice [1, 2, 6, 7, 10, 11, 15, 16, 23]. Our work is unlike that in social choice, since we consider network formation problems where agent preferences are expressed over one another. In the context of matchings, [3, 4] develop various matching algorithms as a black-box to provide approximations for a variety of matching and clustering problems under the implicit utilitarian view. Additionally, [12] and [9] provide results for one-sided matchings and [5] consider bipartite matchings. However, all previous work on approximation for utility maximization mentioned above either assumes the underlying weights form a metric space [1, 2, 3, 4, 5, 15, 16, 23, 9] or are normalized [6, 7, 10, 11, 12], with only two exceptions. The first exception is that Maximum Traveling Salesperson yields a 2-approximation without the metric assumption [4]. We prove this result as part of a much more general theorem using much more general techniques. The second is the result developed in the full-information setting by [20] and affirmed in the ordinal setting by [3], that Max Weight Matching yields a 2-approximation without any assumptions on the weights. We generalize this result to all b-Matchings instead of only b = 1. The work most similar to ours is [3] which bounds the distortion of ordinal mechanisms for several problems, including Maximum Traveling Salesperson, but relies heavily on the assumption that the weights obey the triangle inequality. For the Ordinal Greedy algorithm, [3] show the metric assumption implies that any two edges which are both most preferred by their respective endpoints at some iteration must be within a factor of 2 of one another, even if they are not adjacent. By contrast, this non-local information is unavailable to us in our model. Our paper is unique in that we identify a large class of problems for which assumptions on the weights are unnecessary to achieve good approximations to optimum with only ordinal information. The Omniscient Greedy algorithm has been studied extensively. In fact, it is known to be optimal on exactly the set of independence systems (any system with the heredity property) which are matroids [14, 22], which includes Maximum Spanning Trees. [17] showed that Omniscient Greedy provides good approximations for many independence systems, including matching and symmetric TSP. [13] further demonstrated that Omniscient Greedy provides a tight 3-approximation for Max Weight Planar Subgraph. These results were later reformulated as k-extendibility by [18], who applies this idea to a diverse set of problems, including b-Matching. Unfortunately, the proofs for all results just mentioned rely critically on Omniscient Greedy selecting edges in strictly non-increasing order, making them untenable in the ordinal setting. No ordinal algorithm can yield optimum solutions, even for matroids. However, as we show in Table 1, our results compete well with the best known polynomial-time algorithms for ABC Systems. 4 2 Model and Problem Statements The input for all problems in this paper is a set N of agents (nodes) of size n, and a strict preference ordering for each x ∈ N over the edges adjacent to x. The preference orderings reported by each agent are induced by a set of hidden symmetric weights w(x, y) = w(y, x) for all x, y ∈ N . The set of hidden weights corresponds to an undirected, complete graph G = (N ,E) with non-negative weight function w : E → R+. The transitive relation of the individual preference orderings for all agents determines a partial ordering σ over all edges; note that some pairs of edges may end up being incomparable in σ (see Example 1). The preference ordering σ is said to be consistent with the hidden weights if ∀x, y, z ∈ N , if x prefers y to z then it must be that w(x, y) ≥ w(x, z). If an edge e1 is known to be at least as large as edge e2 according to this partial ordering σ, then we will say that e1 dominates e2 in σ. The problems we consider are optimization problems where for an instance (w, σ) the objective is to compute the subgraph of G with maximum total edge weight, subject to a set of constraints, knowing only σ. For weight function w, we let OP Tw be the optimal solution for the weights prescribed by w, and we let w(OP Tw) be the total weight of the optimal solution evaluated by w. Likewise, we use S to denote our constructed solution and w(S) its weight. Our approximation factor for a problem is therefore α = max (w,σ) , where S is any solution returned by our algorithm for (w, σ). w(OP Tw) w(S) Recall the definition of ABC Systems. Given constraints A, b, c we can say without loss of generality that b ≤ c − 1, because if any node has c or more neighbors in a component, this component would have to be of size greater than c. When b = c − 1 this effectively removes the node degree constraint. Similarly, when c = n, the component size is effectively unbounded. Therefore, 1 ≤ b < c ≤ n. Likewise, when A = all subgraphs of G, this effectively annuls the attachment set constraint. Some specific problems which we consider in this paper are as follows. Max ABC: c ≤ n, b < c, A = any attachment set of G Max AB: c = n, b < c, A = any attachment set of G Maximum Spanning Tree: c = n, b = c − 1, A = all acyclic subgraphs of G Maximum Traveling Salesperson: c = n, b = 2, A = all subgraphs of G without non-Hamiltonian cycles Max Weight Planar Subgraph: c = n, b = c − 1, A = all planar subgraphs of G Max Weight b-Matching: c = n, b < c, A = all subgraphs of G Max Weight Matching: c = n, b = 1, A = all subgraphs of G 3 Algorithmic Framework In this section we define the Ordinal Greedy algorithm and reveal some of its salient properties. Rather than limit ourselves only to ABC Systems, in this section we consider general graphic independence systems. An independence system for our setting is a pair (E,L) where E corresponds to the set of edges in some graph and L is a collection of subsets of E such that if F ∈ L and F ′ ⊆ F then F ′ ∈ L. The sets in L are called independent. It is easy to see that all ABC Systems are independence systems because their three constraints possess this heredity property. Let B denote the set of all subgraphs in which all nodes have degree at most b and let C denote the set of all subgraphs in which all connected components have size at most c. Our Max ABC problem can be restated as: Given a graph G = (N ,E), attachment set A, degree limit b and component size limit c, compute the maximum weight subgraph in L = A ∩ B ∩ C. 5 3.1 The Ordinal Greedy Algorithm In an ordinal setting, algorithms only have access to a partial ordering, or set of preference orderings, which provide strictly local information about the preferences of each agent. This precludes the use of algorithms which require comparisons between the weights of non-adjacent edges. In fact, it is not difficult to see that no ordinal algorithm can be guaranteed to compute the optimal solution for even simple settings, e.g., forming a matching [3]. However, the Ordinal Greedy algorithm defined below performs well in this setting because it relies on strictly local information. Ordinal Greedy starts from the empty set and builds up a sequence of intermediate solutions by adding locally optimal edges at each iteration which do not violate a set of constraints, i.e., preserve independence. To understand how this heuristic is applied to the ordinal setting, we must formalize what it means for an edge to be locally optimal. Definition 1. Undominated Edge Given a set E of edges, (u, v) ∈ E is undominated if for all (u, x) and (v, y) in E, w(u, v) ≥ w(u, x) and w(u, v) ≥ w(v, y). At this point it is important to make several observations. First, every edge set E has at least one undominated edge, because its maximum weight edge must be undominated. However, there may be undominated edges which are not globally maximum. Second, for any edge set E it is straightforward to find at least one undominated edge using only the partial ordering σ (see [3] for details). Undominated edges are either of the form (u, v) where u and v are each other's most preferred neighbor, or form cycles in which each subsequent node is the first choice of the previous one, and thus all edges in the cycle have the same weight. What follows is a general purpose Ordinal Greedy algorithm, which starts from the empty set and iteratively selects undominated edges from the set of remaining edges which do not violate the constraints in question. The algorithm uses the partial ordering σ to determine which edges are undominated at each iteration. The algorithm concludes when there are no edges left which can be added to the subgraph without violating the constraints, so the final solution S is maximal in this sense. Algorithm 1: Ordinal Greedy Input: Edge set E, partial ordering σ, collection of valid subgraphs L Initialize S = ∅, E = E ; while E 6= ∅ do Pick an undominated edge e = (u, v) ∈ E and add it to the intermediate solution: S ← S + e ; Remove e from E ; Remove all edges f from E such that S + f /∈ L ; end Output: Return S We refer to the iteration at which an edge e = (u, v) is removed from E as the critical iteration of e. When the inputs to our algorithm (E,L) characterize an ABC System, there are exactly four cases which may occur at the critical iteration of edge e: 1) e is added to the ordinal greedy solution S 2) e is removed from E because S + e /∈ A 3) e is removed from E because S + e /∈ B (where B = sets of edges with any degree ≤ b) 4) e is removed from E because S + e /∈ C (where C = sets of edges with any component size ≤ c) For cases 2-4 we say e was eliminated due to A,B, or C. If an edge e = (u, v) was eliminated due to A, the attachment property implies there must be a (u, v)-path in the intermediate solution at its critical iteration. In other words, u and v are already in the same connected component in S at this iteration. If e = (u, v) was eliminated due to B, either u or v must already have degree exactly b at this iteration. 6 If e = (u, v) was eliminated due to C, u and v must already be in disjoint connected components whose cumulative size is greater than c. Note that in these three cases, an edge can only be eliminated if at least one adjacent edge of equal or greater weight has already been added to the intermediate solution, and all adjacent edges already added to the intermediate solution must be of equal or greater weight (since only undominated edges are added to our solution). There are limiting values of A, b, and c for which elimination due to these constraints cannot occur. When c = n, no edge can be eliminated due to C because all nodes can be in the same connected component. Additionally, when A is the set of all subgraphs of G, no edge can be eliminated due to A. If adding an edge would violate more than one constraint, we say that it was eliminated in order of priority C → B → A. For example, when b = c− 1, no edge can be eliminated due to B, because for a node to reach degree b the size of its component must be exactly c and we say that any incident edge would be eliminated due to C. As the following sections show, the approximation factor of Ordinal Greedy for an ABC System depends on which cases of elimination can occur. Notice that the performance of the Ordinal Greedy algorithm can deviate significantly from the Om- niscient Greedy algorithm in the full-information setting (which we call "Omniscient Greedy" because it knows the underlying edge weights and can choose the edge with maximum weight at each iteration). Consider the following example. Example 1. Suppose the graph G = (N ,E) is constructed as follows. Let N = {u1, ...uk, v1, ...vk}. Let w(ui, vi) = 1 + ǫ for i ≤ k for some infinitesimal ǫ. Let w(ui, uj) = 1 for all i 6= j. Let w(vi, vj) = ǫ for all i 6= j. Let all other edges have weight 0. Consider the ABC System corresponding to finding a Maximum-Weight Spanning Tree. It is clear that Omniscient Greedy will find the optimum solution with weight w(OP Tw) = k(1 + ǫ) + (k − 1). Now consider Ordinal Greedy. Suppose Ordinal Greedy begins by selecting (ui, vi) for i ≤ k, which are all undominated at the beginning of the algorithm. Once these edges have been selected, edges of the form (ui, uj) and (vi, vj) become undominated for i 6= j. Now, if an edge (ui, uj) or (vi, vj) is selected, the other must be eliminated at that iteration, since taking it would form a cycle. Notice that since we only have access to ordinal information, there is no possible way for Ordinal Greedy to tell which of these edges is better: they are both edges which are most preferred by their endpoints, even though one secretly has weight 1 and the other only ǫ. In other words, these edges are incomparable in the partial preference order σ. Suppose Ordinal Greedy proceeds by selecting (vi, vi+1) for i < k. Then the Ordinal Greedy solution formed has weight w(S) = k(1 + ǫ) + (k − 1)ǫ. This example shows that (in the limit) it is not possible for Ordinal Greedy to always result in solutions better than a factor of 2 away from optimum, even though Omniscient Greedy can easily compute the true optimum solution. As we show in this paper, however, despite its knowledge handicap, Ordinal Greedy can often produce surprisingly good results. 3.2 Properties of Ordinal Greedy For any independence system in the full-information setting, the Omniscient Greedy algorithm has been shown to achieve its worst approximation on an instance with a binary weight function ¯w : E → {0, 1} [17]. However, previous proofs have relied crucially on the fact that Omniscient Greedy selects edges in strictly non-increasing order by weight, which is not possible with only ordinal information. We offer a new proof to show that even in the ordinal setting, Ordinal Greedy always achieves its worst approximation factor on an instance with a binary weight function ¯w : E → {0, 1} for any graphic independence system. This theorem will allow us to prove approximation bounds for ABC and AB Systems later in this paper. Theorem 1. For any graphic independence system (E,L), for any instance (w, σ) with weight function w : E → R+ and partial ordering σ consistent with w, there exists an instance ( ¯w, σ) with weight function ¯w : E → {0, 1} such that σ is consistent with ¯w and the worst-case ratio of the optimal solution to an Ordinal Greedy solution is at least as large as for (w, σ). 7 Proof Sketch. Before we begin the proof, we provide a short proof sketch. Suppose on instance (w, σ) the ratio between the optimal solution OP Tw and solution S constructed by Ordinal Greedy is w(OP Tw) w(S) = δ. Our goal is to construct a binary weight function ¯w such that ¯w(OP T ¯w) ¯w(S) ≥ δ. When δ is infinite, constructing ¯w is straightforward, so we only consider finite values of δ. First we create a weight function w by raising the weights of all edges not in S as much as possible without altering the weights of the edges of S, such that σ remains consistent with w. Since Ordinal Greedy selects S and none of its edge weights have changed, and the edge weights of OP Tw cannot have decreased, then w(OP Tw) w(S) ≥ δ. From w we carefully create ¯w by proving that there must exist a subset of edges to which we can assign weight 1 and let all other edges have weight 0, such that σ is consistent with ¯w and ¯w(OP T ¯w) (cid:3) ¯w(S) ≥ ¯w(OP Tw) ¯w(S) ≥ δ. Proof. Recall that a partial ordering σ is consistent with weight function w if for all x, y, z ∈ N , if x prefers y to z in σ then w(x, y) ≥ w(x, z). We will now show that for any instance (w, σ) where w : E → R+ for which Ordinal Greedy provides a δ-approximation for δ > 0 in the worst case, there exists an instance ( ¯w, σ) where ¯w : E → {0, 1} for which Ordinal Greedy provides no better than a δ-approximation in the worst case. Given any weight function w we now construct a binary weight function ¯w such that the approximation factor is at least as large and σ is still consistent with ¯w. Observation 1. For an independence system (E,L) the solution S computed by Ordinal Greedy depends only on σ, not the edge weights. Therefore, if σ is consistent with w and w, the possible solutions S are the same for instances (w, σ) and ( w, σ). However, w(S) and w(S) may differ. . ¯w(S) ¯w(S) w(S) ≤ ¯w(OP Tw) w(S) ≤ ¯w(OP T ¯w) Therefore, given a worst possible solution S constructed by Ordinal Greedy for (w, σ), our goal is to take the weight function w and construct a binary weight function ¯w such that w(OP Tw) , and σ is still consistent with ¯w (and thus S can still be produced by Ordinal Greedy for the instance with weights ¯w). Recall that OP Tw is the optimum (maximum-weight) solution for weights w. Since by definition ¯w(OP T ¯w) ≥ ¯w(OP Tw), it is enough to show that δ = w(OP Tw) Suppose there is some edge e /∈ S, such that no edge in S is known to be greater than or equal to it in the partial ordering σ. If such an edge exists, we can let ¯w be the weight function such that e and all edges known to be greater than or equal to e in σ have weight 1, and all other edges have weight 0. Clearly, σ remains consistent with ¯w. To see this, consider any x, y, z ∈ N such that x prefers y to z. Then either both w(x, y) and w(x, z) dominate e in σ, so ¯w(x, y) = ¯w(x, z) = 1, or neither do, in which case ¯w(x, y) = ¯w(x, z) = 0. The only other case is that w(x, y) dominates e and w(x, z) does not (since we know that w(x, y) dominates everything that w(x, z) does), and then 1 = ¯w(x, y) > ¯w(x, z) = 0. In all cases, ¯w(x, y) ≥ ¯w(x, z), so σ is consistent with ¯w. This means S remains the same set of edges, and since ¯w(S) = 0, the approximation factor becomes unbounded. This means for any instance where there is some edge e /∈ S for which no edge in S dominates it in σ, we can always create a weight function ¯w with an approximation factor at least as large. Therefore, for the rest of the proof we assume that for every edge e /∈ S there exists some edge in S known to be at least as great by the partial ordering σ. For our greedy solution S, fix an ordering {s1, s2, ..., sm} over the edges of S in non-increasing order by weight so that w(s1) ≥ w(s2) ≥ ... ≥ w(sm). Construct weight function w by increasing the weight of each edge not in S to be equal to the weight of the smallest-weight edge si ∈ S known to be greater than or equal to it in the partial ordering σ. Note that by our assumption above, such an edge si always exists. Claim 1. σ is consistent with w. Proof. Consider any two adjacent edges (x, y) and (x, z) where x prefers y to z. Let (u, v) ∈ S be a smallest edge known by σ to have weight at least w(x, y). Then (u, v) is also known to have weight at least w(x, z) since x prefers y to z. Therefore, the smallest edge of S known to have weight at least w(x, z) is either (u, v) or has weight smaller than w(u, v). After increasing the weights, w(x, y) = w(u, v) ≥ w(x, z). Therefore if x prefers y to z in σ, then w(x, y) ≥ w(x, z). By definition, σ remains consistent with w. 8 The above process of forming w forms an assignment of edges: consider every edge e to be assigned to the smallest edge in si ∈ S known to be larger than or equal to it in the partial ordering, where w(e) = w(si). Now alter these assignments so that if w(si) = w(si+1) then all edges with this weight, including si, are assigned to si+1. Let ri be the number of edges of OP Tw assigned to si. Note that if w(si) = w(si+1) then ri = 0. Lemma 1. m P i=1 ri · w(si) ≥ δ m P i=1 w(si) Proof. On the left side of the inequality, the product ri · w(si) denotes the total weight of the edges of OP Tw assigned to si after having their weight increased. This sum over i ≤ m computes w(OP Tw), the total weight of the optimal solution over w, evaluated by w. On the right hand side, the summation yields the total weight of the edges in the greedy solution, w(S) multiplied by δ. Since σ is consistent with w by Claim 1 the greedy solution S remains the same, and since none of the weights of edges of S were altered w(S) = w(S). By construction, w(OP Tw) ≥ w(OP Tw) because the weights of edges of OP Tw could only have been increased. And so w(OP Tw) ≥ w(OP Tw) w(S) = δ · w(S). w(S) We now demonstrate that we can alter the weights of w to create a binary weight function ¯w : E → {0, 1} such that ¯w(OP Tw) ≥ δ · ¯w(S). All changes to the weights keep σ consistent with ¯w, so that S remains a solution of Ordinal Greedy. Lemma 2. There exists some k ≤ m such that P i=1 ri ≥ δk. k k Proof. Suppose to the contrary that ri < δk for all k ≤ m. We show by induction that this implies w(si), which yields a contradiction to Lemma 1. Specifically, we will show that for every P i=1 m ri w(si) < δ i=1 P j ≤ m, m P i=1 j X i=1 ri w(si) ≤ δ j−1 X i=1 j w(si) + [ X i=1 ri − δ(j − 1)] w(sj). (1) m m i=1 m m i=1 P ri < δm, then When applied to j = m, this gives us the result that ri−δ(m−1)] w(sm). ri− δ(m− 1) < δ, and thus the right hand side of the above inequality is strictly Since less than δ Pm i=1 w(si), which gives us a contradiction with Lemma 1, as desired. Note that here we use the fact that w(sm) > 0 without loss of generality; if this were not the case then we can make the same argument for j being the largest integer such that w(sj) > 0. Thus all that is left is to prove Inequality (1). i=1 w(si)+[ ri w(si) ≤ δ Pm−1 P P i=1 P i=1 We proceed by induction. The base case for j = 1 is trivially true. Now assume that Inequality (1) holds for j, and we will prove it for j + 1. Then, j+1 X i=1 ri w(si) ≤ δ j−1 X i=1 j w(si) + [ X i=1 ri − δ(j − 1)] w(sj) + rj+1 w(sj+1). j Let ξ = δj − P i=1 ri. Since by our assumption j P i=1 ri < δj, we know that ξ > 0. Suppose in the right-hand side of the above inequality, we increase the coefficient of w(sj) by ξ, and decrease the coefficient of w(sj+1) by ξ. Since w(sj) ≥ w(sj+1), this only makes the quantity larger. Thus, we obtain that: 9 j+1 X i=1 ri w(si) ≤ δ j X i=1 j+1 w(si) + [ X i=1 ri − δj] w(sj+1), as desired. This proves Inequality (1) for every j ≤ m, and thus completes the proof of this Lemma. k k−1 P i=1 k P i=1 P i=1 ri ≥ δk, then and all other edges have weight 0. Note that rk > 0 since if rk = 0 and We now use this value of k to construct the binary weight function ¯w. Take the smallest k such that ri ≥ δk and let ¯w be the weight function where for all i ≤ k, ¯w(si) = 1, ¯w(e) = 1 for all e assigned to si, ri ≥ δ(k − 1). We now argue that σ is still consistent with ¯w. Consider any adjacent edges (x, y), (x, z) such that x prefers y to z, and suppose that (x, y) is assigned to edge si, while (x, z) is assigned to edge sj. First consider the case when w(si) 6= w(sj). It must be that i < j, since the set of edges dominating (x, y) is a subset of edges dominating (x, z), and edges are assigned to a smallest edge of S dominating them. Then ¯w(x, y) ≥ ¯w(x, z), as desired. If instead w(si) = w(sj), then by construction of our assignment we have that i = j, so ¯w(x, y) = ¯w(x, z). Therefore σ is consistent with ¯w because for any adjacent edges (x, y), (x, z) such that x prefers y to z it must be that ¯w(x, y) ≥ ¯w(x, z). since σ is consistent with ¯w. By definition of ¯w, we know that ¯w(OP Tw) = Pk to our choice of k, we thus have that ¯w(OP Tw) ≥ δ · ¯w(S). This concludes our proof that for any instance (w, σ) where w : E → R+ for which Ordinal Greedy provides a δ-approximation for δ > 0 in the worst case, there exists an instance ( ¯w, σ) where ¯w : E → {0, 1} for which Ordinal Greedy provides no better than a δ-approximation in the worst case. Thus we now have a binary weight function ¯w such that S is a possible solution of Ordinal Greedy, i=1 ri, and ¯w(S) = k. Due Another nice property is that every solution constructed by Ordinal Greedy is pairwise stable. Pairwise stability means that no pair of agents x, y has incentive to collude to add edge (x, y) by each giving up some of their edges in the Ordinal Greedy solution S. Either this exchange would decrease the total utility of one of the agents, or adding (x, y) is infeasible even after sacrificing the other edges. Here we assume that the utility of a node x in solution S is simply the total weight of edges in S incident on x. Theorem 2. Any solution S constructed by Ordinal Greedy on an independence system is pairwise stable. Proof. Let (x, y) /∈ S and let Fx and Fy be any set of edges in S adjacent to x and y excluding (x, y). If x and y can improve their individual utilities by adding (x, y) at the expense of removing all of the edges in Fx ∪ Fy, this means w(x, y) > w(Fx) and w(x, y) > w(Fy). However, this clearly implies w(x, y) is larger than the weight of each individual edge in Fx and Fy. If (x, y) /∈ S, then its critical iteration must have occurred before any of the adjacent edges in Fx and Fy were added to the Ordinal Greedy solution. Therefore S − Fx − Fy + (x, y) cannot be a feasible solution. 4 Ordinal Approximation for ABC Systems In this section we bound the worst-case performance of Ordinal Greedy compared to the optimal solution for any ABC System. We use αABC to denote the approximation factor, or the ratio of the optimal solution to the worst possible Ordinal Greedy solution for any ABC System. Unlike Example 1 in Section 3.1 for the maximum spanning tree problem, Ordinal Greedy does not provide a constant approximation factor for all ABC Systems. However, it does always provide a finite approximation which depends on the degree limit b. To simplify notation, since the optimal solution here is only evaluated using the same weight function used to generate it, we refer to the total weight of the optimal solution w(OP Tw) as w(OP T ). Here we show that αABC ≤ b + 1 for any ABC System and provide a family of examples where w(OP T ) w(S) = b + 1 to show that b + 1 is a tight bound on the approximation factor. 10 In later sections, we explore classes of ABC Systems in which Ordinal Greedy achieves a better worst-case approximation. Note that this result is quite general. As we discussed, ABC Systems include many varied constraints, some quite difficult to approximate. Our result in this section states that, even for extremely complex A and constraints on component size c, as long as the maximum allowed degree of any node is small, then it is possible to form a good approximation to the true optimum solution while only knowing ordinal information instead of the true edge weights. Theorem 3. For any ABC System, the Ordinal Greedy algorithm always produces a solution within a factor of (b+1) of the optimal solution, and this bound is tight. Proof Sketch. Before we begin the detailed proof, we provide a short proof sketch. We proceed via a charging argument. We wish to charge the weight of the edges of OP T to the edges of S such that all edges of OP T are fully charged somewhere, and no edge of S receives a charge greater than b + 1 times its weight. However, unlike Omniscient Greedy in the full-information setting, we cannot assume that any eliminated edge of OP T has weight smaller than all edges of the Ordinal Greedy solution S which were added before its critical iteration. This prohibits us from using the methods in previous work. Thankfully, due to Theorem 1 we know that if Ordinal Greedy produces a solution within a factor of (b + 1) of optimal for all instances with binary weight functions, then this holds for all instances. We therefore assume that all weights are {0, 1}, and can now charge any edge of OP T to any weight 1 edge of S, but must ensure that no weight is charged to any edges of S with weight 0. To ensure no edge of S is charged more than b+1 times its weight, we look at the connected components of S with only weight 1 edges, and charge all edges of OP T to these components. Specifically, we design the following charging scheme. Let (u, v) be an edge of OP T where w(u, v) = 1. Let Pu and Pv be the connected components containing u and v in the subgraph of S containing only weight 1 edges. We charge the weight of w(u, v) between Pu and Pv based on what occurs at the critical iteration of (u, v). If (u, v) ∈ S then Pu = Pv, so charge its full weight to this component. If (u, v) was eliminated due to A, charge its full weight to either Pu or Pv arbitrarily. Note that while the attachment property of A ensures that u and v are in the same connected component in S at this iteration, this does not imply that Pu = Pv because all (u, v)-paths in S may contain a weight 0 edge. If (u, v) was eliminated due to B, one of its endpoints must have a degree of exactly b in S at this iteration, so charge its full weight to the component containing this endpoint. If (u, v) was eliminated due to C, we split the charge between Pu and Pv based on the size of the connected components in S containing u and v at this iteration. Let qu and qv be the sizes of the qu−1 connected components in S containing u and v at the critical iteration of (u, v). Charge qu+qv−2 to Pu and qv−1 qu+qv−2 to Pv. In all four cases we have ensured the full weight of the edge of OP T has been charged between the components containing its endpoints. The rest of the proof involves arguing that each such component Pu of size p is charged a total of at most (b + 1)(p − 1) using the above charging scheme. Since such a component must contain at least p − 1 edges with weight 1, this completes the proof of the upper bound because it shows that the total number of edges in OP T with weight 1 is at most b + 1 times the number of such edges in S. We then provide a family of examples to show this bound is tight. Omniscient Greedy has the same worst-case solution as Ordinal Greedy on this family of examples, so our ordinal approximation competes well despite its knowledge handicap. Note that in the example yielding the lower bound of αABC ≥ b+1, all edges are eliminated due to C. The following section demonstrates that when the component size constraint is relaxed by allowing c = n, the approximation factor improves significantly. (cid:3) Proof. First, we prove the upper bound αABC ≤ b + 1. Since all ABC Systems are independence systems, by Theorem 1 it is sufficient to show that α ≤ b + 1 for all instances with integral weight functions w : E → {0, 1} to provide an upper bound on α for all instances. We proceed via a charging argument. Let w be a binary weight function w : E → {0, 1}, so all edges in our graph are either weight 1 or weight 0. Given the Ordinal Greedy solution S constructed for some 11 instance (w, σ) we consider the subgraph ¯S ⊆ S with only weight 1 edges and partition ¯S into connected components. We then construct a charging scheme which charges the full weight of all the edges in OP T to these components. The total charge over all these components represents the total weight of OP T , while the total weight of these components is equal to the weight of the greedy solution. If our charging scheme guarantees that the full weight of all edges of OP T has been charged to the components, and no component of ¯w is charged more than (b + 1) times the total weight of its edges, then we can sum over these components to show that w(OP T ) ≤ (b + 1)w( ¯S) = (b + 1)w(S). Let P be the set of connected components in ¯S. Let P ∈ P be any connected component in this subgraph of size p ≥ 1 and total weight w(P ). We make two obvious, but critical observations about each component P . Observation 2. For any component P , its total weight w(P ) is at least p− 1, the weight of its maximum spanning tree. Observation 3. There is a path between any two nodes in S with only weight 1 edges if and only if the nodes are in the same component P . Claim 2. There exists a charging scheme which charges the full weight of all edges of OP T to the components of P such that the total charge to any component P of size p is at most (b + 1)(p − 1). Proof. We need only charge the edges of OP T such that w(u, v) = 1 because the weight 0 edges do not contribute to the value of w(OP T ). Charging Scheme: Let (u, v) be an edge of OP T where w(u, v) = 1. Let Pu and Pv be the connected components containing u and v in the subgraph ¯S ⊆ S containing only weight 1 edges. We charge the weight of w(u, v) between Pu and Pv based on what occurs at the critical iteration of (u, v). If (u, v) ∈ S then Pu = Pv, so charge its full weight to this component. If (u, v) was eliminated due to A, charge its full weight to either Pu or Pv arbitrarily. Note that while the attachment property of A ensures that u and v are in the same connected component in S at this iteration, this does not imply that they are in the same connected component in P because all (u, v)-paths in S may contain a weight 0 edge. If (u, v) was eliminated due to B, one of its endpoints must have a degree of exactly b in S at this iteration, so charge its full weight to the component containing this endpoint. If (u, v) was eliminated due to C, we split the charge between Pu and Pv based on the size of the connected components in S containing u and v at this iteration. Let qu and qv be the sizes of the connected components in S containing u and v at the critical iteration of (u, v). Charge qu+qv−2 to Pv. In all four cases we have ensured the full weight of the edge of OP T has been charged between the components containing its endpoints. qu−1 qu+qv−2 to Pu and qv−1 Case 1) p > b Case 2) p ≤ b The weight of each edge (u, v) ∈ OP T has been charged exclusively to the components containing its endpoints, Pu and Pv. Therefore, to determine the maximum possible charge to any component P , we bound the charge from edges of OP T with one or both endpoints in P and show this is at most (b+1)(p−1). All nodes have at most b adjacent edges in OP T , so the maximum charge to any component P with p nodes is b · p. If p > b then the total charge on P is at most b · p ≤ b · p + (p − (b + 1)) = (b + 1)(p − 1). Let u0 be the first node in P to have an edge of weight 0 added adjacent to it in S at some iteration of Ordinal Greedy. By definition, at any time before the critical iteration of this weight 0 edge u0 cannot be in the same component in S as any edge of weight 0. If there is no node in P with an adjacent weight 0 edge in S, let u0 be any arbitrary node in P . Let a Type 2 edge be an edge of OP T with weight 1 which is incident to u0, but not to any other node in P . Let a Type 1 edge be all edges of OP T of weight 1 which are not Type 1, including all edges with both endpoints in P and those with a single endpoint in P which is not u0. Since all nodes have degree at most b in OP T and there are (p − 1) nodes other than u0, it is clear that Type 1 edges cumulatively contribute a charge of at most b · (p − 1) to P . Here we show that Type 2 12 edges contribute a total charge of at most p − 1, limiting the total charge to any component P to at most b · (p − 1) + (p − 1) = (b + 1)(p − 1). Let (u0, v) be an edge of OP T where w(u0, v) = 1. The critical iteration of (u0, v) must be before the critical iteration of the first weight 0 edge incident to u0 in S. This is because at the iteration the weight 0 edge was added to S it had to be undominated, so (u0, v) could not still have been adjacent to it in the set of available edges E. Let Pv denote the component containing v. We look at the four cases of charging based on the critical iteration of (u0, v) to show that Type 2 edges contribute a charge of at most (p − 1) to P . If (u0, v) ∈ S, then clearly this is Type 1 because P = Pv and so it has already been charged to P . Likewise, if (u0, v) was eliminated due to A then its endpoints must be in the same component P = Pv. This is because the attachment property ensures that u and v are in the same connected component in S at this iteration, and by the definition of u0 it cannot yet have a path to any edge of weight 0. Therefore any path from u0 to v in S must contain only weight 1 edges, meaning P = Pv. If (u0, v) was eliminated due to B then its full weight is either charged to P or to Pv. For its weight to be charged to P , u0 must have degree b at this iteration. For (u0, v) to have weight 1 this means all b of the edges incident to u0 in S must have weight 1 because otherwise they could not have been undominated before the critical iteration of (u0, v). However, this would mean that all neighbors of u0 are in P , so p > b. Therefore, if p ≤ b there can be no Type 2 edges charged to P which were eliminated due to B. We can now see that the only Type 2 edges charged to P are those which are eliminated due to C. For each of these edges P is charged qu0 −1 qu0 +qv−2. For (u0, v) to be eliminated due to C this means the combined sizes of the disjoint components in S containing u0 and v must be at least c, otherwise the edge (u0, v) In other words, qu0 + qv > c so qu0 + qv ≥ c − 1. And since no would still be a valid edge to add. weight 0 edge may yet be adjacent to the component in S containing u0 at this iteration, we know that c−1 , so the charge from each Type 2 edge eliminated due to C is at most qu0 ≤ p. Therefore, p−1 c−1 . And since there can be at most b Type 2 edges, the total charge they contribute to P is at most b p−1 c−1 = (p− 1) b c−1 ≤ p− 1 because b ≤ c− 1 (the maximum degree can never be more than the component size). This leaves us with a total charge to P of at most b(p − 1) + (p − 1) = (b + 1)(p − 1) when p ≤ b. Together with Case 1, we have shown that the total charge to each component P ∈ P of size p at most (b + 1)(p − 1). This concludes the proof of Claim 2. qu0 +qv−2 ≤ p−1 qu0 −1 By summing the charge over all components P ∈ P we get w(OP T ) ≤ (b + 1) P P ∈P (p − 1) ≤ (b + 1) P w(P ) = (b + 1)w(S) from the above claim. In other words, αABC = max w,σ P ∈P To show that the above result is tight, consider the following ABC System. This system represents the problem of hedonic coalition formation with additive separable symmetric preferences (b = c − 1 and A = all subgraphs of G), where agents are partitioned into coalitions (cliques) and each agent's total utility is the sum of its utility for being matched with all other agents in its coalition. w(OP T ) w(S) ≤ b + 1. Example 2 Suppose that n = c2, b = c − 1, and A = all subgraphs of G. In other words, the only constraint is that all components must be of size at most c = √n. Label the nodes uij for i ∈ [1, c] and j ∈ [1, c]. Let w(ui1, uij) = 1 for all j > 1. Let w(ui1, uk1) = 1 + ǫ for all k = i + 1 for some infinitesimal ǫ. Let all other edges have weight 0. For appropriate choices of preferences σ, Ordinal Greedy may select each of the edges (ui1, uk1) for all k = i + 1 before selecting any others, creating a path of length c − 1. Thus, all of the weight 1 edges are eliminated due to C. The optimal solution is to select each of the edges (ui1, uij) = 1 for all j > 1. This yields w(S) = (c − 1)(1 + ǫ) while w(OP T ) = b · c because OP T consists of c stars with b edges each. Therefore as ǫ → 0, w(OP T ) c−1 = b + 1 because b = c − 1. As α is an upper bound on the ratio between the optimal solution and the greedy solution for any instance, we have α ≥ b + 1. w(S) = b · c 13 This concludes our proof of Theorem 3. 5 AB Systems and Important Special Cases In this section, we bound the performance of Ordinal Greedy on ABC Systems where c = n, effectively removing the component size constraint. We then discuss some common examples of maximization prob- lems on AB Systems, including Max Spanning Tree, Max TSP, and Max Planar Subgraph. To improve our bound from b + 1 we invoke the notion of sparsity. Definition 2. Sparsity A graph S is d-sparse if for all subgraphs F ⊆ S containing V (F ) nodes and E(F ) edges E(F ) any d < d there exists a subgraph F ⊆ S such that E( F ) > d. V ( F ) V (F ) ≤ d and for Suppose our attachment set A and degree limit b imply that any feasible solution must be d-sparse. Note that this sparsity is implied by our constraints, and is not a separate constraint. Our main result in this section is that, for any graph collection which is guaranteed to be d-sparse, ordinal information is enough to produce good approximations. Specifically, we prove a bound of d + 1 for such settings. Since the sparsity corresponds to an upper bound on average degree of the nodes, it is always true that d ≤ b 2, and so when c = n, this immediately reduces the approximation factor from b + 1 to b 2 + 1. Even for large b, however, there are many natural classes of graphs that are always sparse, including planar graphs, scale-free graphs, graphs of small arboricity or treewidth, and many others. As we discuss in the next section, this result allows us to provide extremely strong guarantees for many important problems. Theorem 4. For any ABC System where the components can be of any size and the constraints imply that any feasible solution must be d-sparse, the Ordinal Greedy algorithm always produces a solution within a factor of max{2, (d + 1)} of the optimal solution, and this bound is tight. Proof Sketch. As with our proof of Theorem 3 for general ABC Systems, we only need to consider instances with weights {0, 1} due to Theorem 1. However, the charging schemes and proofs for ABC Systems and AB Systems differ significantly. To lower the approximation factor from b+1 to max{2, d+1}, we have to be more selective about where we charge the edges of OP T . For simplicity, we first assign the edges of OP T to their endpoints, before considering the total charge to all the nodes in any component. Since OP T is d-sparse, the edges of OP T can be assigned to their endpoints such that each node is assigned at most d edges. We then take such an assignment and for each edge of OP T eliminated due to B we change its assignment, if necessary, to the node which caused its elimination. Let P be a component of the subgraph of S containing only weight 1 edges, and suppose p = P. Then, similarly to the proof of Theorem 3, we must show that this component will be charged at most max{2, d + 1}w(P ), but unlike before, components may be charged more than (d + 1)(p − 1) if w(P ) > p − 1. Now we consider two cases based on whether any node in a component was charged an edge of OP T eliminated due to B. If there such a node in a component, then it must be possible to distribute the charge on the nodes over the edges of the component directly so that no edge is charged more than max{2, d + 1} times its weight. If there is no such node, then we show that at least one node in the component must be charged at most p − 1 and the rest are charged at most d(p − 1), cumulatively providing a charge at most (d + 1)(p − 1) which can then be distributed over the edges in the component. Once again, we provide a family of examples to show that this bound is tight. (cid:3) Proof. In other words, we will show that αAB = max{2, d + 1}. For AB Systems, the fact that c = n means no edge is eliminated due to C. This leaves us with only 3 cases which can occur at the critical iteration of an edge. We begin by creating an assignment of edges of OP T to the nodes to construct our charging scheme. We use the same notation as in the proof of Theorem 3. 14 As before, we know that the weight of a component P is at least that of its maximum spanning tree, w(P ) ≥ p− 1. However, in certain cases, our strategy is different than the ABC Systems proof. We ensure that each component P is charged at most max{2, d + 1}w(P ), but a component may be charged more than max{2, d + 1}(p − 1). As in the proof of Theorem 3, for each edge (u, v) of OP T , we charge its weight between the components containing its endpoints. We need only to charge the edges of OP T such that w(u, v) = 1, because the weight 0 edges do not contribute to the value of w(OP T ). Claim 3. There exists an assignment of edges of OP T to their endpoints such that each node is assigned at most d edges. Proof. Let Nd be a set of dn nodes containing d duplicates of each node in N . Let O be a set of nodes where each node corresponds to an edge of OP T . Construct a bipartite graph by building an edge from each node in O to each of the 2d nodes in Nd corresponding to the endpoints of the edge of OP T it represents. Consider any subset of nodes R ⊆ O corresponding to a subset of edges in OP T . Let V (R) be the set of endpoints in N of all edges of OP T represented in R. Let R′ be the set of all edges of OP T with both endpoints in V (R), so that R ⊆ R′. Since R′ can have sparsity at most d, it follows that R′ ≤ d · V (R). Therefore, if we consider the d · V (R) nodes in Nd corresponding to the duplicates of V (R), which all have at least one edge to a node in R, we have that R ≤ R′ ≤ d · V (R). By Hall's Condition, we can create a perfect matching between O and Nd. Take this perfect matching as described above, and assign each edge of OP T to the endpoint in N corresponding to the node in Nd to which they are matched. Since there d duplicates of each node in Nd, each node in N can be assigned at most d edges. d is fractional the same result holds using a similar argument by duplicating the nodes in O. Note that while the proof above is written assuming that d is an integer to duplicate the nodes, when However, this assignment is not sufficient for our charging scheme. There may be too much assigned to small components P , and even to nodes which have no adjacent weight 1 edges in S at all would be charged d, when they should not be charged at all. Therefore, we take this assignment and alter it to create our charging scheme so that every component P is charged at most (d + 1)w(P ). Take the (possibly fractional) assignment of edges of OP T to their endpoints from Claim 3, and for every edge of OP T which was eliminated due to B change its assignment, if necessary, to be entirely to the endpoint which had degree b at its critical iteration. Given this new assignment, our charging scheme is now simple. For each node in P charge the weight of the edges of OP T assigned to it to P . We now show that each component P has been charged at most max{2, d + 1}w(P ). Let a b-node be defined as a node with exactly b adjacent edges of weight 1 in S, and therefore b adjacent edges in P . By construction, only b-nodes can receive a charge greater than d and at most b, while all other nodes are charged at most d. If a node has degree b in S but is not a b-node because one or more of its adjacent edges in S has weight 0, then any edges of OP T assigned to it which were eliminated due to B must have weight 0. The is because if any of the eliminated edges had weight 1, an edge of weight 0 could not be undominated at an iteration when the weight 1 edge is still adjacent to it in E. Therefore, a node which is not a b-node can have at most d edges of weight 1 assigned to it. Case 1) P contains at least one b-node Instead of showing that the total charge on P is bounded by max{2, d + 1}w(P ), in this case it is simpler to think of the charge from the edges of OP T as assigned to each particular node. The charge on P is the cumulative charge on all of the nodes in P , which we distribute over the edges of P . For any component P , if the total charge to P can be distributed over its edges so that each edge is charged at most max{2, d + 1}, then the total charge to P is at most max{2, d + 1}w(P ). By summing over all P ∈ P we get OP T (w) ≤ max{2, d + 1}w(S). 15 All b-nodes have charge at most b, which we can distribute so that each of its b adjacent edges is charged at most 1. Now select one b-node and consider the maximum spanning tree of P . For all nodes which are not b-nodes distribute all d of their charge to their adjacent edge on the path to the selected b-node in this maximum spanning tree. If any edge is between two b-nodes, it is charged at most 2. If any edge has exactly one b-node as an endpoint, it may be charged at most 1 from this endpoint and d from the other for a total of d + 1. For any edge between two nodes which are not b-nodes, it is only charged from one of its endpoints, which has charge at most d. Since every edge of P has at most max{2, d + 1} charge, the total charge over P is at most max{2, d + 1}w(P ). Let u0 be the first node in P to have an edge of weight 0 added adjacent to it in S at some iteration of the Ordinal Greedy algorithm. If there is no node in P with an adjacent weight 0 edge in S, let u0 be any node in P . Case 2) P does not contain any b-nodes We show that the total charge to any component P is at most max{2, d + 1}(p − 1) by showing that there are at most d(p − 1) edges of OP T adjacent to the nodes of P , excluding those that have u0 as their only endpoint in P . And there are at most p− 1 edges of OP T which have u0 as their only endpoint in P . Since there are no b-nodes, all nodes in P must be charged at most d. Clearly, u0 could only be charged by edges which were eliminated due to A or are included in P . This means that for all edges of OP T of weight 1 charged to u0, their other endpoint must be in P . Therefore the total charge on u0 is at most min{d, p − 1}. Note that we cannot assume these edges have already been charged to P . over at least (p− 1) edges of P . We have that d(p−1)+min{d,p−1} each component P is charged at most (d + 1)(p − 1) ≤ (d + 1)w(P ) ≤ max{2, d + 1}w(P ). The sum of the total charge to the nodes of P is at most d(p−1)+min{d, p−1} which we can distribute ≤ d + 1. Our charging scheme ensures that Together with Case 1, we sum over P ∈ P and get w(OP T ) ≤ max{2, d + 1}w(S), as desired. We now show that the above bound is tight. Let N = {u1, ..., uk, v1, ..., vk}. Let w(ui, uj) = 1 for all i 6= j, w(ui, vi) = 1 + ǫ for all i ≤ k for some infinitesimal ǫ, and let all other edges have weight 0. Let c = n, b = c − 1 and A = all planar subgraphs such that no cycle may contain an edge (ui, vi) for any i ≤ k. This implies d ≤ 6k−6 For appropriate preference orderings σ, Ordinal Greedy selects the edges (ui, vi) for all i ≤ k first. Now edges (vi, vj) with weight 0 may be undominated. If any edge (vi, vj) or (ui, uj) is selected or eliminated the other must be eliminated at that iteration, but no ordinal algorithm can decide optimally between these edges because they are not adjacent. k because any planar graph on n nodes has at most 3n − 6 edges. 2k = 3k−3 p−1 Suppose Ordinal Greedy selects all edges (vi, vj) where j = i+ 1, causing all edges (ui, uj) with weight 1 to be eliminated due to A, and proceeds by selecting additional (vi, vj) edges until the solution is maximally planar. This yield w(S) = k(1 + ǫ), whereas the optimal solution selects a maximal planar subgraph of edges (ui, uj). The edges (ui, uj) have a combined weight of 3k − 6 because any clique on k nodes has a maximally planar triangulation of size 3k − 6. The edges (ui, vi) have combined weight k(1 + ǫ), for a total of w(OP T ) = 4k − 6 + ǫk. Therefore as ǫ → 0, αAB ≥ w(OP T ) k + 1, which asymptotically approaches 3k−3 w(S) = 4k−6 k = 3k−6 The example for Max Weight Matching in Section 6 provides a lower bound example where d = 1 2, so k + 1 = d + 1 as k → ∞. d < 1 and α ≥ 2. 5.1 Important Cases of AB Systems Theorem 4 establishes that for AB Systems in which solutions are always sparse, ordinal algorithms don't perform much worse than ones which know the true underlying edge weights. While our result in the previous section is quite general, it is worth noting how it applies to many important problems which happen to be special cases of AB Systems. Since all tours and cycles are 1-sparse, and all planar graphs are at most 3-sparse, we immediately arrive at the following corollaries: 16 Corollary 1. Ordinal Greedy always computes a 2-approximation for Maximum Weight Spanning Tree, and this bound is tight. See Example 1 for example demonstrating the lower bound for Maximum Spanning Tree is tight. Corollary 2. Ordinal Greedy always computes a 2-approximation for Maximum Traveling Salesperson, and this bound is tight. Proof. Let the set of agents be N = {u1, ..., uk, v1, ..., vk−3}. Let w(ui, ui+1) = 1 + ǫ for all i ≤ k − 1 for some infinitesimal ǫ. Let w(vi, ui+1) = 1 and w(vi, ui+2) = 1 for all i ≤ k − 3. Let w(u1, uk−1) = 1 and w(u2, uk) = 1 and w(u1, uk) = 1. Let all other edges have weight 0. Suppose the Ordinal Greedy algorithm begins by selecting (ui, ui+1) = 1 + ǫ for all i < k. This creates a path of length k− 1 which causes all weight 1 edges in the graph to be eliminated due to A or B. Ordinal Greedy may then proceed by selecting some set of weight 0 edges, such as (u1, v1), (uk, vk−3), and (vi, vi+1) for i < k−3. Therefore, w(S) = (k−1)(1+ǫ). Meanwhile, the optimal solution is to select all of the weight 1 edges of the graph: (vi, ui+1) and (vi, ui+2) for all i ≤ k − 3 as well as (u1, uk−1), (u2, uk), and (u1, uk). This yields w(OP T ) = 2k − 3. Therefore, αT SP ≥ w(OP T ) (k−1)(1+ǫ) which asymptotically approaches 2 as k → ∞ and ǫ → 0. Corollary 3. Ordinal Greedy always computes a 4-approximation for Max Weight Planar Subgraph. w(S) = 2k−3 More generally, the same arguments can be applied to any problem where the goal is to find maximum- weight subgraphs with some excluded minor, finding maximum-weight graphs with small treewidth or arboricity, as well as a variety of other graph problems. 6 b-Matching For any ABC System where c = n and A = all subgraphs of G, the only constraint is that each node must have degree at most b. This is equivalent to the well-known problem of Max Weight b-Matching. In this case, the approximation provided by Ordinal Greedy improves greatly over general AB Systems. In fact, it provides a strict 2-approximation regardless of the value of b. Theorem 5. For any ABC System on graph G, where c = n and A = all subgraphs of G, Ordinal Greedy always constructs a solution within a factor of 2 of the optimal solution. This bound is tight. Proof. Assume ¯w : E → {0, 1}. We proceed in the same way as the proof of Theorem 4. For each weight 1 edge (u, v) ∈ OP T eliminated due to B, charge its full weight to its endpoint which has degree b. All b adjacent edges to this node in the greedy solution must have weight 1 to have been undominated before the critical iteration of (u, v), so it is a b-node. Since no node can have more than b adjacent edges in the optimal solution, this means that the total charge on any node is bounded by the number of edges of weight 1 adjacent to this node in the greedy solution. Therefore, all nodes can distribute their charge over their adjacent edges in S such that no edge in the greedy solution is charged more than 2. Consider the following minimal example to show that the factor of 2 is tight. Suppose b = 1. Let N = {u1, u2, v1, v2}. Let w(u1, u2) = w(v1, v2) = w(u1, v1) = 1. Let w(u2, v2) = 0. For suitable σ, Ordinal Greedy selects (u1, v1) first because it is preferred by both u1 and v1, and the only other edge it can subsequently add is (u2, v2). This yields w(S) = 1 and w(OP T ) = 2. Note that b = 1 is the problem Max Weight Matching, and our result generalizes the results from [3] and [20]. 17 7 Conclusion and Further Directions In this paper we identify a large class of problems we call ABC Systems for which ordinal preference information is sufficient for algorithms to provide good approximations to optimal, even without access to cardinal utilities. Previous work has shown that if agent preferences form a metric space, approximations for TSP and matching can improve in expectation [3]. It remains to be seen how Ordinal Greedy performs on ABC Systems in expectation and how much the approximation factors for general ABC or AB Systems improve when this metric assumption holds. Also along the lines of previous work, it would be interesting to investigate whether truthful ordinal algorithms for ABC and AB Systems can compete with our non- truthful algorithm, much as [4] did for the problems first approached in [3]. Lastly, we have seen that all solutions produced by Ordinal Greedy are pairwise stable, but it is unknown for our problems whether all pairwise stable solutions produce a good approximation to optimum (although it is easily seen to be true for MST and TSP). Acknolwedgments This work was partially supported by NSF award CCF-1527497. References [1] Elliot Anshelevich, Onkar Bhardwaj, and John Postl. Approximating optimal social choice under metric preferences. In Proceedings of the Twenty-Ninth AAAI Conference on Artificial Intelligence, pages 777 -- 783, 2015. [2] Elliot Anshelevich and John Postl. Randomized social choice functions under metric preferences. Journal of Artificial Intelligence Research, 58:797 -- 827, 2017. [3] Elliot Anshelevich and Shreyas Sekar. Blind, greedy, and random: algorithms for matching and clus- tering using only ordinal information. In Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence, pages 383 -- 389, 2016. [4] Elliot Anshelevich and Shreyas Sekar. Truthful mechanisms for matching and clustering in an ordinal world. In International Conference on Web and Internet Economics (WINE), 2016. [5] Elliot Anshelevich and Wennan Zhu. Tradeoffs between information and ordinal approximation for bipartite matching. In International Symposium on Algorithmic Game Theory (SAGT), 2017. [6] Craig Boutilier, Ioannis Caragiannis, Simi Haber, Tyler Lu, Ariel D Procaccia, and Or Sheffet. Optimal social choice functions: A utilitarian view. Artificial Intelligence, 227:190 -- 213, 2015. [7] Felix Brandt, Vincent Conitzer, Ulle Endriss, Ariel D Procaccia, and J´erome Lang. Handbook of computational social choice. Cambridge University Press, 2016. [8] Gruia Calinescu, Cristina G Fernandes, Howard Karloff, and Alexander Zelikovsky. A new approxi- mation algorithm for finding heavy planar subgraphs. Algorithmica, 36(2):179 -- 205, 2003. [9] Ioannis Caragiannis, Aris Filos-Ratsikas, Søren Kristoffer Stiil Frederiksen, Kristoffer Arnsfelt Hansen, and Zihan Tan. Truthful facility assignment with resource augmentation: An exact analysis of serial dictatorship. In International Conference on Web and Internet Economics (WINE), 2016. [10] Ioannis Caragiannis, Swaprava Nath, Ariel D Procaccia, and Nisarg Shah. Subset selection via implicit utilitarian voting. Journal of Artificial Intelligence Research, 58:123 -- 152, 2017. 18 [11] Ioannis Caragiannis and Ariel D Procaccia. Voting almost maximizes social welfare despite limited communication. Artificial Intelligence, 175(9-10):1655 -- 1671, 2011. [12] George Christodoulou, Aris Filos-Ratsikas, Søren Kristoffer Stiil Frederiksen, Paul W Goldberg, Jie Zhang, and Jinshan Zhang. Social welfare in one-sided matching mechanisms. In International Con- ference on Autonomous Agents and Multiagent Systems, pages 30 -- 50. Springer, 2016. [13] ME Dyer, LR Foulds, and AM Frieze. Analysis of heuristics for finding a maximum weight planar subgraph. European Journal of Operational Research, 20(1):102 -- 114, 1985. [14] Jack Edmonds. Matroids and the greedy algorithm. Mathematical programming, 1(1):127 -- 136, 1971. [15] Ashish Goel, Anilesh K Krishnaswamy, and Kamesh Munagala. Metric distortion of social choice rules: Lower bounds and fairness properties. In Proceedings of the 2017 ACM Conference on Economics and Computation, pages 287 -- 304. ACM, 2017. [16] Stephen Gross, Elliot Anshelevich, and Lirong Xia. Vote until two of you agree: Mechanisms with In Proceedings of the Thirty-First AAAI Conference on small distortion and sample complexity. Artificial Intelligence, pages 544 -- 550, 2017. [17] Bernhard Korte and Dirk Hausmann. An analysis of the greedy heuristic for independence systems. Annals of Discrete Mathematics, 2:65 -- 74, 1978. [18] Juli´an Mestre. Greedy in approximation algorithms. In European Symposium on Algorithms, volume 4168, pages 528 -- 539. Springer, 2006. [19] Katarzyna E Paluch, Marcin Mucha, and Aleksander Madry. A 7/9-approximation algorithm for the maximum traveling salesman problem. Approximation, Randomization, and Combinatorial Optimiza- tion. Algorithms and Techniques, pages 298 -- 311, 2009. [20] Robert Preis. Linear time 1/2-approximation algorithm for maximum weighted matching in general graphs. In STACS, volume 99, pages 259 -- 269. Springer, 1999. [21] Ariel D Procaccia and Jeffrey S Rosenschein. The distortion of cardinal preferences in voting. In In The Tenth International Workshop on Cooperative Information Agents (CIA): Lecture Notes in Computer Science (LNCS), volume 4149, pages 317 -- 331. Springer, 2006. [22] Richard Rado. A theorem on independence relations. The Quarterly Journal of Mathematics, (1):83 -- 89, 1942. [23] Piotr Krzysztof Skowron and Edith Elkind. Social choice under metric preferences: Scoring rules and stv. In Thirty-First AAAI Conference on Artificial Intelligence, pages 706 -- 712, 2017. [24] Andrew Vince. A framework for the greedy algorithm. Discrete Applied Mathematics, 121(1):247 -- 260, 2002. 19
1110.2765	1	1110	2011-10-12T19:22:19	Multi-Issue Negotiation with Deadlines	[ "cs.MA", "cs.AI" ]	This paper studies bilateral multi-issue negotiation between self-interested autonomous agents. Now, there are a number of different procedures that can be used for this process; the three main ones being the package deal procedure in which all the issues are bundled and discussed together, the simultaneous procedure in which the issues are discussed simultaneously but independently of each other, and the sequential procedure in which the issues are discussed one after another. Since each of them yields a different outcome, a key problem is to decide which one to use in which circumstances. Specifically, we consider this question for a model in which the agents have time constraints (in the form of both deadlines and discount factors) and information uncertainty (in that the agents do not know the opponents utility function). For this model, we consider issues that are both independent and those that are interdependent and determine equilibria for each case for each procedure. In so doing, we show that the package deal is in fact the optimal procedure for each party. We then go on to show that, although the package deal may be computationally more complex than the other two procedures, it generates Pareto optimal outcomes (unlike the other two), it has similar earliest and latest possible times of agreement to the simultaneous procedure (which is better than the sequential procedure), and that it (like the other two procedures) generates a unique outcome only under certain conditions (which we define).	cs.MA	cs	Journal of Artificial Intelligence Research 27 (2006) 381-4 17 Submitted 3/06; published 11/06 Multi-Issue Negotiation with Deadlines Shaheen S. Fatima Michael Wooldridge Department of Computer Science, University of Liverpool, Liverpool L69 3BX, U.K. Nicholas R. Jennings School of Electronics and Computer Science, University of Southampton, Southampton SO17 1BJ, U.K. S .S .FAT IMA@C SC . L IV.AC .UK M . J .WOOLDR IDGE@C SC . L IV.AC .UK NR J@EC S . SOTON .AC .UK Abstract This paper studies bilateral multi-issue negotiation between self-interested autonomous agents. Now, there are a number of different procedures that can be used for this process; the three main ones being the package deal procedure in which all the issues are bundled and discussed together, the simultaneous procedure in which the issues are discussed simultaneously but independently of each other, and the sequential procedure in which the issues are discussed one after another. Since each of them yields a different outcome, a key problem is to decide which one to use in which circumstances. Specifically, we consider this question for a model in which the agents have time constraints (in the form of both deadlines and discount factors) and information uncertainty (in that the agents do not know the opponent’s utility function). For this model, we consider issues that are both independent and those that are interdependent and determine equilibria for each case for each procedure. In so doing, we show that the package deal is in fact the optimal procedure for each party. We then go on to show that, although the package deal may be computationally more com- plex than the other two procedures, it generates Pareto optimal outcomes (unlike the other two), it has similar earliest and latest possible times of agreement to the simultaneous procedure (which is better than the sequential procedure), and that it (like the other two procedures) generates a unique outcome only under certain conditions (which we de fine). 1. Introduction Negotiation is a key form of interaction in multiagent systems (Maes, Guttman, & Moukas, 1999; Sandholm, 2000). It is a process in which disputing agents decide how to divide the gains from cooperation. Since this decision is made jointly by the agents themselves (Rosenschein & Zlotkin, 1994; Raiffa, 1982; Pruitt, 1981; Fisher & Ury, 1981; Young, 1975; Kraus, 2001), each agent can only obtain what the other is prepared to allow them. Now, the simplest form of negotiation involves two agents and a single-issue. For example, consider a scenario in which a buyer and a seller negotiate on the price of a good. To begin, the two agents are likely to differ on the price at which they believe the trade should take place, but through a process of joint decision-making they either arrive at a price that is mutually acceptable or they fail to reach an agreement. Since agents are likely to begin with different prices, one or both of them must move toward the other, through a series of offers and counter offers, in order to obtain a mutually acceptable outcome. However, before the agents can actually perform such negotiations, they must decide the rules for making offers and counter offers. That is, they must set the negotiation protocol (Lax & Sebenius, 1986; Osborne & Rubinstein, 1990; Rosenschein & Zlotkin, 1994; Kraus, Wilkenfeld, & Zlotkin, 1995; Lomuscio, Wooldridge, & Jennings, 2003). c(cid:13)2006 AI Access Foundation. All rights reserved. FAT IMA , WOOLDR IDGE , & JENN ING S On the basis of this protocol, each agent chooses its strategy (i.e., what offers it should make during the course of negotiation). For competitive scenarios with self-interested agents, each partic- ipant defines its strategy so as to maximise its individual ut ility. Furthermore, for such scenarios, an agent’s optimal strategy depends very strongly on the information it has about its opponent (Fatima, Wooldridge, & Jennings, 2002, 2004). For example, the strategy that a buyer would use if it knew the seller’s reserve price differs from the one it would use if it did not. From all of this, it can be seen that the outcome of single-issue negotiation depends on four key factors (Harsanyi, 1977): the negotiation protocol, the players’ strategies, the players’ preferences over the possible outcomes, and the information that the players have about each other. However, in most bilateral negotiations, the parties involved need to settle more than one issue. For example, agents may need to come to agreements about objects/services that are characterised by attributes such as price, delivery time, quality, reliability, and so on. For such multi-issue negotiations, the outcome also depends on one additional factor: the negotiation procedure (Schelling, 1956, 1960; Fershtman, 1990), which spec- ifies how the issues will be settled. Broadly speaking, there are three ways of negotiating multiple issues (Keeney & Raiffa, 1976; Raiffa, 1982): • Package deal: This approach links all the issues and discusses them together as bundle. • Simultaneous negotiation: This involves settling the issues simultaneously, but independently, of each other. • Sequential negotiation: This involves negotiating the issues sequentially, one after another. Now, these three different procedures have different properties and yield different outcomes to the negotiators (Fershtman, 2000). So the key question to answer is: which of them is best? Here, since we are concerned with self-interested agents, our notion of the optimal procedure is the one that maximises an agent’s individual return. However, such optimality is only part of the story; given our motivations we are also concerned with the Pareto optimality of the solutions for these procedures (because Pareto optimality ensures that utility does not go wasted), the computational complexity of the procedures (because for scenarios with information uncertainty, the agents need to compute their equilibrium offers during the process of negotiation, as opposed to the complete in- formation scenario where the strategies can be precompiled), the actual time of agreement (because for scenarios with information uncertainty, this time depends on an agent’s beliefs about its oppo- nent and an agreement may not occur in the first time period), a nd the uniqueness of the solutions they generate (because this allows the agents to know their actual shares). One immediate observation in this vein is that the package deal gives rise to the possibility of making tradeoffs across issues. Such tradeoffs are possible when different agents value different issues differently. For example, if there are two issues and one agent values the first more than the second, while the other agent values the second issue more than the first, then it is possible to make tradeoffs and thereby improve the utility of both agents relative to the situation without tradeoffs. In contrast, for the simultaneous and sequential approaches, the issues are settled independently and so there is no scope for such tradeoffs between them. Moreover, we seek to answer the above question about optimality for the types of situation that are commonly faced by agents in real-world contexts. Thus, we consider negotiations in which there are: 1. Time constraints. Agents have time constraints in the form of both deadlines and discount factors. Here we view deadlines as an essential element since negotiation cannot go on in- (Livne, 1979). Likewise, discount definitely, rather it must end within a reasonable time limit 382 MULT I - I S SUE N EGOT IAT ION W ITH D EADL INE S factors are essential since the desirability of the good being traded often declines with time. This happens either because the good is perishable or due to inflation. Moreover, the strategic behaviour of agents with deadlines and discount factors differs from those without (see Ru- binstein, 1982, for single issue bargaining without deadlines and Sandholm & Vulkan, 1999; Ma & Manove, 1993; Fershtman & Seidmann, 1993; Kraus, 2001, for bargaining with dead- lines and discount factors). For instance, the presence of a deadline induces each negotiator to play a strategy that ensures the best possible agreement before the deadline is reached. Likewise, the presence of a discount factor means that reaching an agreement today is not the same as reaching it tomorrow. Hence, the agents try to reach an agreement sooner rather than later. 2. Uncertainty about the opponent’s negotiation parameters. The information that agents have about their negotiation opponent is likely to be uncertain (see Fudenberg & Tirole, 1983; Fudenberg, Levine, & Tirole, 1985; Rubinstein, 1985, for single issue bargaining with uncer- tainty). Moreover, in some bargaining situations, one of the players may know something of relevance that the other does not. For example, when bargaining over the price of a second hand car, the seller knows its quality, but the buyer does not. Such situations are said to have asymmetry in information between the players (Muthoo, 1999). On the other hand, in sym- metric information situations both players have the same information. Again, agents have to operate in both situations and so we analyse both cases. 3. Interdependence between issues. The issues under negotiation may be independent or inter- dependent. In the former case, an agent’s utility from an issue depends only on the agreement that is reached on it, not on how the other issues are settled. In the latter case, an agent’s utility from an issue depends not only on the agreement that is reached on it but also on how the other issues are settled (Bar-Yam, 1997; Klein, Faratin, Sayama, & Bar-Yam, 2003). Both situations are common in multiagent systems and so again we analyse both cases. Thus we study five different settings: i) complete informati on setting (C I ), ii) a setting with independent issues and symmetric uncertainty about the agents’ utilities (SUI ), iii) a setting with independent issues and asymmetric uncertainty about the agents’ utilities (AUI ), iv) a setting with interdependent issues and symmetric uncertainty about the agents’ utilities (SUD ), and v) a setting with interdependent issues and asymmetric uncertainty about the agents’ utilities (AUD ). Our methodology is to first derive equilibria for each of the p rocedures in each of the above settings, From this, we can determine which of them is optimal. As we will see, this analysis shows that, for all the settings, the package deal is the best. We then go on to analyse the procedures in terms of other performance metrics. Specifically, we show th at, in all the settings, only the package deal generates a Pareto optimal outcome. We also show that although the package deal may be com- putationally more complex than the other two procedures, it has similar earliest and latest possible times of agreement to the simultaneous procedure (which is better than the sequential procedure), and it (like the other two procedures) generates a unique outcome only in certain situations (which we define). The key results of our study are summarised in Tabl e 1. There has previously been some formal comparison of different procedures to find the optimal one (see Section 7 for details). However, all this work has at least one of the following major limitations. First, it has focused on comparing procedures for negotiation without deadlines. Note that existing work has obtained equilibrium for negotiation with deadlines, but only for the single 383 FAT IMA , WOOLDR IDGE , & JENN ING S Information setting Package deal Simultaneous Sequential C I SUI , SUD AUI , and AUD Time of agreement tc For the cth issue For the cth issue For the cth partition tc = 1 tc = 1 tc = c for 1 ≤ c ≤ m for 1 ≤ c ≤ m for 1 ≤ c ≤ µ For the cth issue For the cth issue For the cth partition te te te c = ts c = 1 c = 1 c tl tl c = ts tl c = min(2r − 1, n) c = min(2r − 1, n) c + min(2r − 1, n) for 1 ≤ c ≤ m for 1 ≤ c ≤ m for 1 ≤ c ≤ µ Time to O(mn) O(M n) O(M n) C I 2 )) O(Sz πz r3T (n − T 2 )) O(Sz πz r3T (n − T SUI and SUD O(m πr3T (n − T compute 2 )) equilibrium AUI and AUD 2 ) T 2 ) T 2 ) T O(mπr3 (n − T O(Sz πz r3 (n − T O(Sz πz r3 (n − T 2 ) 2 ) 2 ) Pareto C I , SUI ,SUD , optimal? AUI , and AUD C I SUI ,SUD , AUI , and AUD If ¬C1 If ¬C3 ∨ C4 Unique equilibrium? Yes No If C2 If C5 No If C2 If C5 c the earliest c denotes the start time for the cth partition, te Table 1: A summary of key results. ts c the latest possible time of agreement). possible time of agreement, and tl issue case (Sandholm & Vulkan, 1999; Stahl, 1972), and a special type of the sequential procedure for multiple issues (Fatima et al., 2004). See Section 7 for details. Second, it has focussed only on independent issues and asymmetric information settings. Third, it has only focused on finding the optimal procedure, but has not considered the additional solution properties of different procedures. Given this, our paper makes a threefold contribution. First, we obtain the equilibrium for each procedure when there are deadlines. Second, we analyse multiple issues that are both independent and interdependent. Moreover, we analyse both symmetric and asymmetric information settings. Finally, on the basis of the equilibrium for different procedures, we provide the first comprehensive comparison of their solution properties (viz. time complexity, Pareto optimality, uniqueness, and time of agreement). When taken together, the results clearly indicate the choices and tradeoffs involved in choosing a negotiation procedure in a wide range of circumstances. This knowledge can be used by a system designer who is responsible for designing the mechanism that should be used to moderate the negotiation encounters and by the agents themselves if they can choose how to arrange their interactions. Furthermore, this knowledge also tells the agents what their equilibrium offers are during negotiation. The remainder of the paper is organised as follows. We begin by giving a brief overview of single-issue negotiation in Section 2. In Section 3, we study the three multi-issue procedures for the setting with complete information and where the issues are independent. This study is undertaken to provide a foundation for Sections 4, 5, and 6, which treat the information about the agents’ utilities as uncertain. More specifically, in Section 4, we a nalyse a scenario with symmetric uncer- 384 MULT I - I S SUE N EGOT IAT ION W ITH D EADL INE S tainty about the opponent’s utility. In Section 5, we analyse a scenario with asymmetric uncertainty about the opponent’s utility. Sections 4 and 5 both deal with independent issues. In Section 6, we extend the analysis to interdependent issues. Section 7 discusses the related literature and Section 8 concludes. Appendix A provides a summary of notation employed throughout the paper. 2. Single-Issue Negotiation Assume there are two agents: a and b. Each agent has time constraints in the form of deadlines and discount factors. Since we focus on competitive scenarios with self-interested agents, we model negotiation using the ‘split the pie game’ analysed by Osborne and Rubinstein (1994), Binmore, Osborne, and Rubinstein (1992). We begin by introducing this complete information game. Let the two agents be negotiating over a single issue (i). This issue is a ‘pie’ of size 1 and the agents want to determine how to divide it between themselves. There is a deadline (i.e., a number of rounds by which negotiation must end). Let n ∈ N+ denote this deadline. The agents use Rubinstein’s alternating offers protocol (Osborne & Rubinstein, 1994), which proceeds through a series of time periods. One of the agents, say a, starts negotiation in the first time period (i.e., t = 1) by making an offer (xi ), that lies in the interval [0, 1], to b. Agent b can either accept or reject the offer. If it accepts, negotiation ends in an agreement with a getting a share of xi and b getting yi = 1 − xi . Otherwise, negotiation proceeds to the next time period, in which agent b makes a counter-offer. This process of making offers continues until one of the agents either accepts an offer or quits negotiation (resulting in a conflict). Thus, there a re three possible actions an agent can take during any time period: accept the last offer, make a new counter-offer, or quit the negotiation. An essential feature of negotiations involving alternating offers is that the pie is assumed to shrink with time (Rubinstein, 1982). Specifically, it shrin ks at each step of offer and counteroffer. This shrinkage models a decrease in the value of the pie (representing the fact that the pie perishes with time or there is inflation). This shrinkage is represent ed with a discount factor denoted 0 < δi ≤ 1 for both1 agents. At t = 1, the size of the pie is 1, but in all subsequent time periods t > 1, the pie shrinks to δ t−1 . i We denote the set of real numbers by R and the set of real numbers in the interval [0, 1] by R1 . Then let [xt i ] denote the offer made at time period t where xt i and y t i denote the share for agent a i , y t and b respectively. Then, for a given pie, the set of possible offers is: i = δ t−1 i + y t i ≥ 0, and xt i ≥ 0, y t i ] : xt i , y t {[xt i } i ∈ R1 and y t where xt i ∈ R1 . Each player’s utility function is defined over the set R. Let ua : i R1 × N+ → R and ub i : R1 × N+ → R denote the utility functions of the two agents. At time t, if i = δ t−1 i respectively (where xt a and b receive a share of xt i and y t ), then their utilities are: i + y t i i , t) = (cid:26) xt i (xt ua i 0 i , t) = (cid:26) y t i (y t ub i 0 1. Having a different discount factor for different agents only makes the presentation more involved without leading to any changes in the analysis of the strategic behaviour of the agents or the time complexity of finding the equilibrium offers. Hence we have a single discount factor for both agents. if t ≤ n otherwise if t ≤ n otherwise 385 FAT IMA , WOOLDR IDGE , & JENN ING S The conflict utility (i.e., the utility received in the event that no deal is struck) is zero for both agents. Note that δ is not shown explicitly in an agent’s utility function but is implicit. This is because, during any time period t, xt i and y t i denote a’s and b’s actual shares respectively (not the i = δ t−1 . In other words δ is included in an agent’s share. This ratios of their shares) where xt i + y t i will become clearer when we show the agents’ shares in Expression 1. For the above setting, the agents reason as follows in order to determine what to offer. Let agent a denote the first mover (i.e., at t = 1, a proposes to b how to split the pie). To begin, consider the case where the deadline for both agents is n = 1. If b accepts, the division occurs as agreed; if not, neither agent gets anything (since n = 1 is the deadline). Here, a is in a powerful position and is able to propose to keep 100 percent of the pie and give nothing to b 2 . Since the deadline is n = 1, b accepts this offer and agreement takes place in the first time period. Now, consider the case where the deadline is n = 2. In the first round, the size of the pie is 1 but it shrinks to δi in the second round. In order to decide what to offer in the firs t round, a looks ahead to t = 2 and reasons backwards3 . Agent a reasons that if negotiation proceeds to the second round, b will take 100 percent of the shrunken pie by offering [0, δi ] and leave nothing for a. Thus, in the first time period, if a offers b anything less than δi , b will reject the offer. Hence, during the first time period, agent a offers [1 − δi , δi ]. Agent b accepts this and an agreement occurs in the first time period. In general, if the deadline is n, negotiation proceeds as follows. As before, agent a decides what to offer in the first round by looking ahead as far as t = n and then reasoning backwards. This decision making leads a to make the following offer in the first time period: j=0 [(−1)j δj j=0 [(−1)j δj i ], 1 − Σn−1 [Σn−1 i ]] (1) Agent b accepts this offer and negotiation ends in the first time peri od. Note that the equilibrium outcome depends on who makes the first move. Since we have two a gents and either of them could move first, we get two possible equilibrium outcomes. On the basis of the above equilibrium for single-issue negotiation with complete information, we first obtain the equilibrium for multiple issues and then det ermine the optimal negotiation procedure for the various settings that we have previously described. 3. Multi-Issue Negotiation with Complete Information As mentioned in Section 1, the existing literature does not provide an analysis of all the multi-issue procedures for negotiation with deadlines. Hence, we begin by analysing the complete information setting. From this base, we can then extend to the case where there is information uncertainty. Here a and b negotiate over m > 1 independent issues (Section 6 deals with interdependent issues). These issues are m distinct pies and the agents want to determine how to split each of them. Let S = {1, 2, . . . , m} denote the set of m pies. As before, each pie is of size 1. Let the discount factor for issue c, where 1 ≤ c ≤ m, be 0 < δc ≤ 1. For each issue, let n denote each agent’s 2. It is possible that b may reject such a proposal. In practice, a will have to propose an offer that is just enough to induce b to accept. However, to keep the exposition simple, we assume that a can get the whole pie by making the 100 percent proposal. 3. This backward reasoning method is adopted from (Stahl, 1972). Our model is a generalisation of (Stahl, 1972); during time period t, an agent in our model can propose any offer between zero and δ t−1 (because the size of the pie is δ t−1 ), but a player in (Stahl, 1972) is given a fixed number of alter natives to choose from. 386 MULT I - I S SUE N EGOT IAT ION W ITH D EADL INE S deadline. In the offer for time period t (where 1 ≤ t ≤ n), agent a’s (b’s) share for each of the m 1 (y t ∈ Rm issues is now represented as an m element vector xt ∈ Rm 1 ). Thus, if agent a’s share for c). The shares for a and b are together c , then agent b’s share is y t issue c at time t is xt c = (δ t−1 c − xt represented as the package [xt , y t ]. We define an agent’s cumulative utility using the additive fo rm. There are two reasons for this. First, it is the most common form for cumulative utilities in traditional multi-issue utility theory (Keeney & Raiffa, 1976). Second, additive cumulative utilities are linear and so the problem of making tradeoffs becomes computationally tractable4 . The functions U a : Rm 1 × Rm 1 × N+ → R and U b : Rm 1 × N+ → R give the cumulative utilities for a and b respectively at time t. These 1 × Rm are defined as follows: if t ≤ n otherwise U a ([xt , y t ], t) = (Σm c (xt c ua c=1ka c , t) 0 U b ([xt , y t ], t) = (Σm c=1kb c ub c(y t c , t) 0 where ka ∈ Rm + denotes an m element vector of constants for agent a and kb ∈ Rm + that for b. Here R+ denotes the set of positive real numbers. These vectors indicate how the agents value different c+1 , then agent a values issue c more than issue c + 1. Likewise issues. For example, if ka c > ka for agent b. In other words, the m issues are perfect substitutes (i.e., all that matters to an agent is its total utility for all the m issues and not that for any subset of those Varian, 2003; Mas-Colell, Whinston, & Green, 1995). In all the settings we study, the issues will be perfect substitutes. c , and δc c , kb Each agent has complete information about all negotiation parameters (i.e., n, m, ka for 1 ≤ c ≤ m). For this complete information setting, we now determine the equilibrium for the package deal, the simultaneous procedure, and the sequential procedure. if t ≤ n otherwise (2) (3) 3.1 The Package Deal Procedure In this procedure, the agents use the same protocol as for single-issue negotiation (described in Sec- tion 2). However, an offer for the package deal includes a proposal for each issue under negotiation. Thus, for m issues, an offer includes m divisions, one for each issue. Agents are allowed to either accept a complete offer (i.e., all m issues) or reject a complete offer. An agreement can therefore take place either on all m issues or on none of them. As per single-issue negotiation, an agent decides what to offer by looking ahead and reasoning backwards. However, since an offer for the package deal includes a share for all the m issues, agents can now make tradeoffs across the issues in order to maximise their cumulative utilities. For c denote c and bt c ] where at 1 ≤ c ≤ m, the equilibrium offer for issue c at time t is denoted as [at c , bt the shares for agent a and b respectively. We denote the equilibrium package at time t as [at , bt ] 4. Using a form other than the additive one will make the function nonlinear. Consequently an agent’s tradeoff problem becomes a global optimization problem with a nonlinear objective function. Due to their computational complexity, such nonlinear optimization problems can only be solved using approximation methods (Horst & Tuy, 1996; Bar- Yam, 1997; Klein et al., 2003). Moreover, these methods are not general in that they depend on how the cumulative utilities are actually defined. In order to overcome this dif ficulty, we used the additive form for defining cumulative utilities. Consequently, our tradeoff problem is a linear optimization problem, the exact solution to which can be found in polynomial time (as shown in Theorems 1 and 2). Although our results apply to the above defined additive cumulative utilities, in Section 6.4 we discuss how they would hold for nonlinear utilities. 387 FAT IMA , WOOLDR IDGE , & JENN ING S 1 ) is an m element vector that denotes a’s (b’s) share for each of the m 1 (bt ∈ Rm where at ∈ Rm issues. Also, for 1 ≤ t ≤ n, δ t−1 ∈ Rm 1 is an m element vector that represents the sizes of the m pies at time t. The symbol 0 denotes an m element vector of zeroes. Note that for 1 ≤ t ≤ n, at + bt = δ t−1 (i.e., the sum of the agents’ shares (at time t) for each pie is equal to the size of the pie at t). Finally, for time period t (for 1 ≤ t ≤ n) we let a(t) (respectively b(t)) denote the equilibrium strategy for agent a (respectively b). As mentioned in Section 1, the package deal allows agents to make tradeoffs. We let TRAD EO FFA (TRAD EO FFB ) denote agent a’s (b’s) function for making tradeoffs. Given this, the following theo- rem characterises the equilibrium for the package deal procedure. IF a’s TURN IF b’s TURN IF b’s TURN IF a’s TURN Theorem 1 For the package deal procedure, the following strategies form a Nash equilibrium. The equilibrium strategy for t = n is: a(n) = (cid:26) OFFER [δn−1 , 0] ACCEPT b(n) = (cid:26) OFFER [0, δn−1 ] ACCEPT For all preceding time periods t < n, if [xt , y t ] denotes the offer made at time t, then the equilibrium strategies are defined as follows: a(t) = (cid:26) OFFER tradeoffa(ka , kb , δ, ub(t), m, t) IF a’s TURN If (U a ([xt , y t ], t) ≥ ua(t)) ACCEPT else REJECT IF b’s TURN b(t) = (cid:26) OFFER tradeoffb(ka , kb , δ, ua(t), m, t) IF b’s TURN If (U b ([xt , y t ], t) ≥ ub(t)) ACCEPT else REJECT IF a’s TURN where ua(t) = U a ([at+1 , bt+1 ], t + 1) and ub(t) = U b ([at+1 , bt+1 ], t + 1). An agreement takes place at t = 1. Proof: We look ahead to the last time period (i.e., t = n) and then reason backwards. To begin, if negotiation reaches the deadline (n), then the agent whose turn it is takes everything and leaves nothing for its opponent. Hence, we get the strategies a(n) and b(n) as given in the statement of the theorem. In all the preceding time periods (t < n), the offering agent proposes a package that gives its opponent a cumulative utility equal to what the opponent would get from its own equilibrium offer for the next time period. During time period t, either a or b could be the offering agent. Consider the case where a makes an offer at t. The package that a offers at t gives b a cumulative utility of U b ([at+1 , bt+1 ], t + 1). However, since there is more than one issue, there is more than one package that gives b a cumulative utility of U b ([at+1 , bt+1 ], t + 1). From among these packages, a offers the one that maximises its own cumulative utility (because it is a utility maximiser). Thus, the problem for a is to find the package [at , bt ] so as to: maximise such that Σm c at c=1ka c c=1(δ t−1 c − at Σm c)kb c = ub(t) 0 ≤ at for 1 ≤ c ≤ m c ≤ 1 388 (4) MULT I - I S SUE N EGOT IAT ION W ITH D EADL INE S This tradeoff problem is similar to the fractional knapsack problem (Martello & Toth, 1990; Cor- men, Leiserson, Rivest, & Stein, 2003), the optimal solution for which can be generated using a greedy approach5 (i.e., by filling the knapsack with items in the decreasing or der of value per unit weight). The items in the knapsack problem are analogous to the issues in our case. The only differ- ence is that the fractional knapsack problem starts with an empty knapsack and aims to fill it with items so as to maximise the cumulative value, while an agent’s tradeoff problem can be viewed as starting with the agent having 100 per cent of all the issues and then aiming to give away portions of issues to its opponent so that the latter gets a given cumulative utility, while the resulting loss in its own utility is minimised. Thus, in order to find how to sp lit the m issues, agent a considers c for 1 ≤ c ≤ m because ka c is the utility that a needs to give up in order increase b’s utility ka c /kb c /kb by one. Since a wants to maximise its own utility and give b a utility of U b ([at+1 , bt+1 ], t + 1), it divides the m pies such that it gets the maximum possible share for those issues for which ka c is c /kb c is low. Thus, high and gives to agent b the maximum possible share for those issues for which ka c /kb c . It then does a begins by giving b the maximum possible share for the issue with the lowest ka c /kb c and repeats this process until b gets a cumulative the same for the issue with the next lowest ka c /kb utility of U b ([at+1 , bt+1 ], t + 1). In order to facilitate this process of making tradeoffs, the individ- c+1 . The function tradeoffa takes six ual elements of kb are arranged such that ka c+1/kb c > ka c /kb parameters: ka , kb , δ , ub(t), m, and t and uses the above described greedy method to solve the maximisation problem given in Equation 4 and return the corresponding package. If there is more than one package that solves Equation 4, then tradeoffa returns any one of them (because agent a gets equal utility from all such packages and so does agent b). The function tradeoffb for agent b is analogous to that for a. On the other hand, the equilibrium strategy for the agent that receives an offer is as follows. For time period t, let b denote the receiving agent. Then, b accepts [xt , y t ] if ub(t) ≤ U b ([xt , y t ], t), oth- erwise it rejects the offer because it can get a higher utility in the next time period. The equilibrium strategy for a as receiving agent is defined analogously. Hence we get the eq uilibrium strategies (a(t) and b(t)) given in the statement of the theorem. In this way, we reason backwards and obtain the offers for t = 1. The first mover makes this offer and the other agent accepts it. An agreement therefore occurs in the first time period. (cid:3) Theorem 2 For the package deal procedure, the time taken to determine an equilibrium offer for t = 1 is O(mn) where m is the number of issues and n is the deadline. Proof: We know from Theorem 1 that the time to compute the equilibrium offer for t = n is linear in the number of issues (see strategies a(n) and b(n)). Consider a time period t < n. During this time period, the function tradeoffa is used to make tradeoffs. The time complexity of tradeoffa (which uses the greedy approach described in the proof of Theorem 1) is O(m) (Martello & Toth, 1990; Cormen et al., 2003). This function needs to be repeated for every time period from the (n − 1)th to the first. Hence the time complexity of finding an offer fo r the first time period is O(mn). (cid:3) 5. The time complexity of this approach is O(m) (Martello & Toth, 1990), where m denotes the number of items. Note that the greedy method for the fractional knapsack problem takes O(m) time regardless of whether the coefficients c and kb c (for 1 ≤ c ≤ m) in Equation 4 are positive or negative (Martello & Toth, 1990). In the present setting ka (as we mentioned at the beginning of Section 3) these coeffici ents are all positive. However, we will come across negative coefficients when we deal with interdependent issu es in Section 6. 389 FAT IMA , WOOLDR IDGE , & JENN ING S Theorem 3 The package deal procedure generates a Pareto optimal outcome. Proof: Recall that we consider competitive negotiations. Hence, for an individual issue c (where 1 ≤ c ≤ m), an increase in one agent’s utility results in a decrease in that of the other. However, for the package deal procedure, an agent considers its cumulative utility from all m issues. Con- sequently, during the process of backward reasoning, at time t < n, the agent that makes tradeoffs maximises its own cumulative utility without lowering that of its opponent (with respect to what the opponent would offer in the next time period). Hence the equilibrium outcome for the package deal is Pareto optimal. (cid:3) Theorem 4 For a given first mover, the package deal procedure has a uniqu e equilibrium outcome if the following condition is false: C1 . There exists an i and a j (where 1 ≤ i ≤ m and 1 ≤ j ≤ m) such that (i 6= j ) and (ka j ). i /kb i = ka j /kb Proof: Consider a time period t < n and let a denote the offering agent. Recall from Theorem 1 that a splits the m issues in the increasing order of ka i . Thus, for a given i and j , if ka j , i /kb i /kb i = ka j /kb then agent a is indifferent between which of the two issues (i or j ) it splits up first. For example, if 2 = 0.5. If a 2 = 4, then ka 1 = 2, and kb 2 = 2, kb 1 = 1, ka m = 2, n = 2, δ = 0.5, ka 2 /kb 1 = ka 1 /kb is the offering agent at t = 1, it can offer (1, 0) for issue 1 and (1/4, 3/4) for issue 2. This gives a cumulative utility of 1.5 to a and 3 to b. Alternatively a can offer (0, 1) for issue 1 and (3/4, 1/4) for issue 2 since this also results in the same cumulative utilities to a and b. On the other hand, if ka j , then a splits issue i first if ka j and issue j first i /kb i 6= ka j /kb i /kb i < ka j /kb if ka j . In other words, there is only one possible equilibrium offer that a can make at j /kb i > ka i /kb any time t < n. Likewise there is one possible equilibrium offer that b can make at any time t < n. Since there is a unique offer for each time period, the equilibrium outcome is unique. (cid:3) Note that the uniqueness we refer to in Theorem 4 is with respect to a given first mover. If the first mover changes, then the equilibrium outcome may change , as the following example illustrates. 2 = 1. If a is the offering agent at 1 = 2, and kb 2 = 2, kb 1 = 1, ka Let m = 2, n = 2, δ = 0.5, ka t = 1, its equilibrium offer is (1/4, 3/4) for the first issue and (1, 0) for the second. This results in a cumulative of 2.25 to a and 1.5 to b. In contrast, if b is the offering agent at t = 1, its equilibrium offer is (0, 1) for the first issue and (3/4, 1/4) for the second. This results in a cumulative utility of 1.5 to a and 2.25 to b. In the following discussion, we use the term unique to mean unique with respect to a given first mover. 3.2 The Simultaneous Procedure For this procedure, the m issues are partitioned into µ > 1 disjoint subsets. For 1 ≤ c ≤ µ, let Sc denote the cth partition where ∪µ c=1Sc = {1, . . . , m}. The issues within each subset are settled using the package deal. Negotiation for each of the µ partitions starts at t = 1. Thus, for µ = m, all m issues are settled simultaneously and independently of each other. At the other extreme, we have only one partition (i.e., µ = 1) which is the package deal procedure described in Section 3.1. Since the issues in each subset (i.e., each Sc ) are settled using the package deal, the equilibrium for each of these µ partitions is obtained from Theorem 1. Consequently, we get the following results. First, an agreement for each issue occurs in the first round. T his is because negotiation for each partition starts at t = 1. Also, from Theorem 1, we know that an agreement for the package deal 390 MULT I - I S SUE N EGOT IAT ION W ITH D EADL INE S occurs at t = 1. Hence, for the simultaneous procedure, an agreement for each partition (and hence each issue) occurs in the first time period. Second, for the simultaneous procedure, the time taken to determine an equilibrium offer for t = 1 is Σµ c=1O(Sc n) where Sc is the number of issues in the cth partition and n is the deadline. This is explained as follows. Since the time taken to find the e quilibrium offer for t = 1 for the package deal (i.e., for µ = 1) is O(mn) (see Theorem 2), the time taken to compute the equilibrium offer for t = 1 for the cth partition is O(Sc n). Hence, for all µ partitions, the time complexity is Σµ c=1O(Sc n) which is equal to O(M n), where M denotes the number of issues in the largest partition. Third, it follows from Theorem 4 that the simultaneous procedure has a unique equilibrium outcome if the following condition C2 is true: C2 . There is no partition c (where 1 ≤ c ≤ µ) for which the condition C1 is true. Finally, as Theorem 5 shows, the simultaneous procedure may not generate a Pareto optimal outcome. Theorem 5 The simultaneous procedure may not generate a Pareto optimal outcome. Proof: The package deal allows tradeoffs to be made across all the m issues, while the simultaneous procedure allows tradeoffs to be made across issues within each partition but not across partitions. Hence the simultaneous procedure may not generate a Pareto optimal outcome. We show this with a counter example. Consider the case where n = 2, δ = 0.5, m = 3, µ = 2, S1 = {1, 2}, S2 = {3}, 3 = 0.25. Let a denote the first mover. From 2 = 0.5, and kb 1 = 1, kb 3 = 3, kb 2 = 2, ka 1 = 1, ka ka Theorem 1, we know that in the equilibrium for partition S1 , agent a gets a share of 0.25 for issue 1 and 1 for issue 2, and b gets a share of 0.75 for issue 1 and nothing for issue 2. For partition S2 , each agent gets a share of 1/2. Thus, a’s cumulative utility from all three issues is 3.75 and that of b is 0.875. Now consider the case where all three issues are discussed using the package deal. Here, µ = 1 and all other parameters remain the same. In the equilibrium outcome for this procedure, a gets a cumulative utility of 5.125 and b gets 0.875. This means that the procedure with µ = 2 does not generate a Pareto optimal outcome. (cid:3) 3.3 The Sequential Procedure For this procedure, the m issues are partitioned into µ > 1 disjoint subsets. For 1 ≤ c ≤ µ, let Sc denote the cth partition where ∪µ c=1Sc = {1, . . . , m}. The µ partitions are negotiated sequentially, one after another. The issues within a subset are settled using the package deal. Negotiation for the first partition starts at time t = 1. If negotiation for the cth (for 1 ≤ c ≤ µ) partition ends at tc , then negotiation for the (c + 1)th partition starts at time tc + 1. Each player gets its share for all the issues in a partition as soon as the partition is settled. Thus, for µ = m, all m issues are settled in sequence. At the other extreme, we have only one partition (i.e., µ = 1) which is the package deal procedure described in Section 3.1. Since the issues in each subset (i.e., each Sc ) are settled using the package deal, the equilibrium for each of these µ subsets is obtained from Theorem 1 by substituting the appropriate negotiation start times for each partition. 391 FAT IMA , WOOLDR IDGE , & JENN ING S Theorem 6 For the sequential procedure, the equilibrium time of agreement for the cth partition (for 1 ≤ c ≤ µ) is Tc = c. Proof: From Theorem 1, we know that an agreement for the package deal occurs in the first time period. Hence, negotiation for each partition ends in the same time period in which it starts (i.e., negotiation for the cth partition starts at t = c and results in an agreement in the same time period). The time taken to settle all the m issues is therefore µ. (cid:3) Note that the time complexity of the sequential procedure (i.e., the time to compute equilibrium offers) is the same as that for the simultaneous procedure. Also, like the simultaneous procedure, the equilibrium outcome for the sequential procedure may not be Pareto optimal. Finally, the condition for the equilibrium outcome for the sequential procedure to be unique is the same as that for the simultaneous procedure. 3.4 The Optimal Procedure Having obtained the equilibrium outcomes for the three multi-issue procedures, we now compare them in terms of the utilities they generate for each player. Then the procedure that gives a player the maximum utility is its optimal one. Note that, for the sequential procedure, the equilibrium outcome strongly depends on the order in which the partitions are settled. This ordering is called the negotiation agenda. There are two ways of defining the agenda (Fershtman, 1990): exogenously or endogenously. If the agenda is determined before the actual negotiation over the issues begins, then it is said to be exogenous. On the other hand, for the endogenous agenda, the agents decide what issue they will settle next during the process of negotiation. The agenda that gives an agent the maximum utility between all possible agendas is its optimal one (Fatima et al., 2004). Our objective here is not to determine the optimal agenda, but to consider a given agenda and compare the equilibrium outcome for the sequential procedure for that agenda with the outcomes for the simultaneous and the package deal procedures, in order to find the optimal procedure. The following theorem characterises this procedure. Theorem 7 Irrespective of how the m issues are split into µ > 1 partitions, the package deal is optimal for both parties. Proof: In order to compare an agent’s utility from different procedures, it is important to take into account who initiates negotiation. For the package deal, the first mover makes an offer on all the issues. Hence we compare an agent’s utilities for the three procedures, given the agent that will be the first mover for all the three procedures for all the issues . We first show that the outcome for the package deal is no worse t han that for the simultaneous procedure. Consider the simultaneous procedure for any µ > 1. For this procedure, for t ≤ n, the offering agent makes tradeoffs across the issues in each partition independently of the other parti- tions. Now consider the package deal procedure (i.e., with µ = 1 partitions). For this procedure, the offering agent makes tradeoffs across all m issues. Since the difference between the procedure with µ = 1 and the one with µ > 1 is that the former makes tradeoffs across all m issues while the latter does not, each agent’s utility from the former procedure is no worse than its utility from the latter. We now show that for a given µ (where µ > 1), for each agent, the outcome for the simultaneous procedure is better than that for the sequential one (irrespective of the agenda for the sequential procedure). We do this by considering each of the µ partitions. 392 MULT I - I S SUE N EGOT IAT ION W ITH D EADL INE S Time of agreement (tc ) Time to compute equilibrium Pareto optimal? Unique equilibrium? Package deal Simultaneous For the cth issue For the cth issue tc = 1 tc = 1 for 1 ≤ c ≤ m for 1 ≤ c ≤ m O(mn) O(M n) Sequential For the cth partition tc = c for 1 ≤ c ≤ µ O(M n) Yes If ¬C1 No If C2 No If C2 Table 2: A comparison of the outcomes for the three multi-issue procedures for the complete infor- mation setting (C I ). • Partition c = 1. Since negotiation for the first partition starts at t = 1 for both the simulta- neous and the sequential procedures, the outcome for this partition is the same for µ = 1 and µ > 1. Hence, for the first partition, an agent gets equal utility f rom the two procedures. • Partition c > 1. Let agent a denote the first mover for partition c (for 2 ≤ c ≤ µ) for both simultaneous and sequential procedures. Also, let U a sim and U a seq denote a’s cumulative utility for this partition from the equilibrium outcome for the simultaneous and the sequential procedures respectively. Likewise, let U b sim and U b seq denote b’s cumulative utility for this partition from the equilibrium outcome for the simultaneous and the sequential procedures respectively. Now for the simultaneous procedure, negotiation for each partition starts in the first time period. An agreement for each partition also occurs in the fir st time period. On the other hand, for the sequential procedure, negotiation for the cth partition starts in the cth time period and results in an agreement in the same time period (see Theorem 6). Since each pie shrinks with time, agent a’s cumulative utility U a sim is greater than U a seq , and agent b’s cumulative utility sim is greater than U b seq . U b Thus, the simultaneous procedure is better than the sequential one for both agents. Furthermore (as shown above), the outcome for the package deal is no worse than that for the simultaneous procedure for both agents. Therefore, for each agent, the package deal is the optimal procedure. (cid:3) These results are summarised in Table 2. For the above analysis, the negotiation parameters n, δc , c , and kb c (for 1 ≤ c ≤ m) were common knowledge to the agents. However, this is unlikely to ka be the case for most encounters. Therefore we now extend this analysis to incomplete information scenarios with uncertainty about utility functions6 . In Section 4, we focus on the symmetric infor- mation setting where each agent is uncertain about the other’s utility function. Then, in Section 5, we examine the asymmetric information setting where one of the two agents is uncertain about the other’s utility function, but the other agent knows the utility function of both agents. 6. There are two other sources of uncertainty: uncertainty about the negotiation deadline and uncertainty about discount factors. Future work will deal with uncertainty about discount factors. However, for independent issues, we analysed the case with symmetric uncertainty about deadlines in (Fatima, Wooldridge, & Jennings, 2006). The extension of this work to the case of interdependent issues is another direction for future work. 393 FAT IMA , WOOLDR IDGE , & JENN ING S 4. Multi-Issue Negotiation with Symmetric Uncertainty about the Opponent’s Utility In this symmetric information setting, each agent is uncertain about its opponent’s utility function: for 1 ≤ c ≤ m, agent a (b) is uncertain about kb c (ka c ). Specifically, let K denote a vector of r vectors + (for 1 ≤ i ≤ r) consists of m constant positive real numbers. These where each vector Ki ∈ Rm r vectors are the possible values for ka ∈ Rm + and kb ∈ Rm + . In other words, there are r types7 for agent a and r types for agent b. Let P a : N+ → R1 denote the discrete probability distribution function for ka and P b : N+ → R1 that for kb . The domain for these two functions is [1..r ]. In other words, for 1 ≤ i ≤ r , P a (i) (P b (i)) is the probability that agent a (b) is of type i. For 1 ≤ c ≤ m, let Kic denote the cth element of vector Ki . In this setting, the vector K and the functions P a and P b are common knowledge to the nego- tiators. Also, each agent knows its own type, but not that of its opponent. In addition, each agent knows r , δ , n, and m. r different cumulative Since there are r types for agent a and r types for agent b, we define utility functions for each of the two agents. If agent a (b) is of type i (for 1 ≤ i ≤ r) then its utility 1 × N+ → R) from the division specified by the package 1 × N+ → R (U b 1 × Rm i : Rm 1 × Rm U a : Rm i [xt , y t ] at time t is: i ([xt , y t ], t) = (Σm c=1Kicua c (xt c , t) U a 0 i ([xt , y t ], t) = (Σm c=1Kicub c (y t c , t) U b 0 Note that, as before, the issues are perfect substitutes. For this setting, we determine the equi- librium outcomes for each of the three multi-issue procedures and then compare them. if t ≤ n otherwise if t ≤ n otherwise (5) (6) 4.1 The Package Deal Procedure We know from Theorem 1 that the equilibrium outcome for the complete information setting de- pends on ka c and kb c (for 1 ≤ c ≤ m). However, in this setting, there is uncertainty about ka c and c . Hence we use the standard expected utility theory (Neumann & Morgenstern, 1947; Fishburn, kb 1988; Harsanyi & Selten, 1972) to find an agent’s optimal stra tegy. Before doing so, however, we first introduce some notation. For 1 ≤ i ≤ r , we let a(i, t) denote the equilibrium strategy for an agent a of type i for the time period t. Analogously, b(i, t) denotes the equilibrium strategy for an agent b of type i for the time period t. Note that for 1 ≤ i ≤ r , if [at , bt ] is the package offered at time t in equilibrium, then at + bt = δ t−1 (i.e., for each pie, the sum of the shares of the two agents is equal to the size of the pie at time t). Also, for 1 ≤ i ≤ r , we let a(i, j, t) denote the equilibrium strategy for an agent a of type i for the time period t, assuming that b is of type j . Analogously, b(i, j, t) denotes the equilibrium strategy for an agent b of type i for the time period t, assuming that a is of type j . Also, let eua(i, t) denote the cumulative utility that an agent a of type i expects to get from b’s equilibrium offer at time t (i.e., a is the receiving agent and b the offering agent at t). Likewise, eub(i, t) denotes the cumulative utility that an agent b of type i expects to get from a’s equilibrium offer at time t (i.e., b is the receiving agent and a the offering agent at t). We let eua(i, j, t) denote agent a’s expected cumulative utility from its own equilibrium offer at time t if a is of type i, 7. An agent’s type indicates which of the r vectors it corresponds to. 394 MULT I - I S SUE N EGOT IAT ION W ITH D EADL INE S assuming that b is of type j . Note that this is a’s utility when it is the offering agent at t. And let eub(i, j, t) denote agent b’s expected cumulative utility from its own equilibrium offer at time t if b is of type i and assuming that a is of type j . Note that this is b’s utility when it is the offering agent at t. Recall that in this setting, each agent only knows its own type, but not that of its opponent. Since there are r possible types, there are r possible offers an agent can make at any time period (one offer corresponding to each of the opponent’s types). Among these r offers, the one that gives an agent the maximum expected cumulative utility is its optimal offer. If the cth offer (1 ≤ c ≤ r) gives an agent the maximum expected cumulative utility, then we say that the optimal choice for the agent is c. For time period t, we let opta(i, t) (optb(i, t)) denote the optimal choice for agent a (b) of type i. At t = n, the offering agent gets everything and the opponent gets zero utility. Thus, for t = n, we have the following: eua(i, n) = 0 for 1 ≤ i ≤ r for 1 ≤ i ≤ r Kic δ t−1 c eua(i, j, n) = eub(i, n) = 0 m Xc=1 m Xc=1 Note that for t = n, eua(i, j, n) and eub(i, j, n) do not depend on j because in the last time period, the offering agent gets 100 percent of all the m pies. For all preceding time periods t < n, we have the following: for 1 ≤ i ≤ r and 1 ≤ j ≤ r for 1 ≤ i ≤ r and 1 ≤ j ≤ r eub(i, j, n) = Kic δ t−1 c (10) (9) (7) (8) eua(i, t) = eua(i, θ , t + 1) for 1 ≤ i ≤ r where θ = opta(i, t + 1) for 1 ≤ i ≤ r where λ = optb(i, t + 1) (11) (12) (14) (13) eub(i, j, t) = eua(i, j, t) = F b (i, j, e, t) × P a (e) F a (i, j, e, t) × P b (e) for 1 ≤ i ≤ r and 1 ≤ j ≤ r for 1 ≤ i ≤ r and 1 ≤ j ≤ r eub(i, t) = eub(i, λ, t + 1) r Xe=1 r Xe=1 The function F a takes four parameters: i, j , e, and t, and returns the utility that an agent a of type i gets from offering the equilibrium package for time t, assuming that agent b is of type j where in fact it is of type e. Obviously, agent b accepts a’s offer at t if U b e (a(i, j, t), t) ≥ eub(e, γ , t + 1) where γ = optb(e, t + 1). Otherwise, agent b rejects a’s offer and negotiation proceeds to the next round in which case a’s expected utility is EUA(i, t + 1). Hence, F a is defined as follows: F a (i, j, e, t) = (cid:26) U a i (a(i, j, t), t) eua(i, t + 1) where the strategy a(i, j, t) for t = n is defined as follows: A(i, j, n) = (cid:26) OFFER [δn−1 , 0] ACCEPT 395 if U b e (a(i, j, t), t) ≥ eub(e, γ , t + 1) where γ = optb(e, t + 1) otherwise if a’s turn otherwise FAT IMA , WOOLDR IDGE , & JENN ING S and for all preceding time periods t < n it is defined as: A(i, j, t) = (cid:26) OFFER tradeoffa1(K, δ, eub(j, t), i, j, m, t, P a , P b ) if U a i ([xt , y t ], t) ≥ EUA(i, t) ACCEPT else REJECT where [xt , y t ] denotes the offer made at t and the function8 TRAD EO FFA1 is defined as follows. Like TRAD EO FFA , the function TRAD EO FFA1 solves the following maximisation problem: if a’s turn otherwise maximise such that c=1Kicat Σm c c=1 (δ t−1 Σm c − at c )Kj c = eub(j, t) 0 ≤ at for 1 ≤ c ≤ m c ≤ 1 (15) where i denotes a’s type and j that of b. However, the difference between TRAD EO FFA1 and TRAD EO FFA arises when there is more than one package that maximises a’s cumulative utility (i.e., c ) while giving b a cumulative utility of eub(j, t). If there is more than one such package, c=1Kicat Σm then in Theorem 1, it does not matter which of these packages a offers to b (because both agents have complete information). Hence, TRAD EO FFA can return any one such package. However, in the present setting, there is uncertainty. Therefore, if there is more than one package that maximises a’s cumulative utility while giving b a cumulative utility of eub(j, t), then TRAD EO FFA1 returns the package that maximises a’s expected cumulative utility. For instance, let [at , bt ] be one such package that maximises a’s cumulative utility. Then a’s expected cumulative utility from [at , bt ] (i.e., eua(i, j, t)) is as given in Equation 13 where: if U b e ([at , bt ], t) ≥ eub(e, γ , t + 1) where γ = optb(e, t + 1) F a (i, j, e, t) = (cid:26) U a i ([at , bt ], t) eua(i, t + 1) otherwise Obviously, if there is more than one package that maximises a’s expected cumulative utility and gives b a utility of eub(j, t) then TRAD EO FFA1 returns any one such package. We now turn to agent b. For this agent, F b , B(i, j, t), and tradeoffb1 are defined analogously as follows: F b (i, j, e, t) = (cid:26) U b i (b(i, j, t), t) eub(i, t + 1) where the strategy b(i, j, t) for t = n is defined as follows: B(i, j, n) = (cid:26) OFFER [0, δn−1 ] ACCEPT and for all preceding time periods t < n it is defined as: B(i, j, t) = (cid:26) OFFER tradeoffb1(K, δ, eua(j, t), i, j, m, t, P a , P b ) if U b i ([xt , y t ], t) ≥ EUB(i, t) ACCEPT else REJECT 8. A method for making tradeoffs has been proposed by Faratin, Sierra, and Jennings (2002) for an incomplete infor- mation setting, but this method differs from ours. Also, Faratin et al. only present a method for making tradeoffs, but they do not show that the resulting offer is in equilibrium. In contrast, our method shows that the resulting offer is in equilibrium. e (b(i, j, t), t) ≥ eua(e, α, t + 1) where α = opta(e, t + 1) if U a otherwise if b’s turn otherwise if b’s turn otherwise 396 MULT I - I S SUE N EGOT IAT ION W ITH D EADL INE S Thus, the optimal choice for agent a (i.e., opta(i, t)) and that for agent b (i.e., optb(i, t)) are defined as follows: eua(i, j, t) eub(i, j, t) (16) (17) for 1 ≤ i ≤ r for 1 ≤ i ≤ r opta(i, t) = arg maxr j=1 optb(i, t) = arg maxr j=1 Note that the offering agent’s optimal choice for t = n does not depend on its opponent’s type since the offering agent gets all the pies. We compute the optimal choice for the first time period by reas oning backwards from t = n. At t = 1, if an agent a of type i is the offering agent, then it offers the package that corresponds to agent b being of type opta(i, 1). Likewise, if an agent b of type i is the offering agent, then it offers the package that corresponds to agent a being of type optb(i, 1). However, since opta(i, 1) and optb(i, 1) are obtained in the absence of complete information, If an agreement does not occur an agreement may or may not take place in the first time period. at t = 1, then the agents need to update their beliefs as follows. Let T a t ⊆ {1, 2, . . . , r} denote 1 = {1, 2, . . . , r} and T b the set of possible types for agent a at time t. For t = 1, we have T a 1 = {1, 2, . . . , r}. Assume that an agent a of type i makes an offer at t = 1. If the offer that a makes gets rejected, then it means that b is not of type opta(i, 1) and so a updates its beliefs about b using Bayes’ rule. Now, on the basis of a’s offer at t = 1 (say [x1 , y1 ]), agent b can infer the possible types for agent a. Thus, agent b too updates its beliefs using Bayes’ rule. The belief update rules for time t are as defined below. UPDATE BELIEFS: Agent a puts all the weight of the posterior distribution of b’s type over T b t − {optb(i, t)} using Bayes’ rule. Agent b puts all the weight of the posterior distribution of a’s type over K using Bayes’ rule where K ⊆ {1, 2, . . . , r} is the set of possible types for a that can offer [xt , y t ] in equilibrium. The belief update rule for the case where b offers at t = 1 is analogous to the above case where a offers at t = 1. Thus if the offer at t = 1 gets rejected, then negotiation goes to the next round. At t = 2, the offering agent (say an agent a of type i) finds opta(i, 2) with its updated beliefs. This process of updating beliefs and making offers continues until an agreement is reached. In Section 3, we used the concept of Nash equilibrium because the agents had complete infor- mation. However, in the current setting, each agent is uncertain about its opponent’s type and so an agent’s optimal strategy depends on its beliefs about its opponent. Hence we use the concept of sequential equilibrium (Kreps & Wilson, 1982; van Damme, 1983) for this setting. Sequential equilibrium is defined in terms of two elements: a strategy profile and a system of beliefs. The strategy profile comprises of a pair of strategies, one for ea ch agent. The belief system has the fol- lowing properties. Each agent has a belief about its opponent’s type. In each time period, an agent’s strategy is optimal given its current beliefs (during the time period) and the opponent’s possible strategies. For each time period, each agent’s beliefs (about its opponent) are consistent with the offers it received. Using this concept of sequential equilibrium, the following theorem characterises the equilibrium for the package deal procedure. Theorem 8 For the package deal procedure, the following strategies form a sequential equilibrium. The equilibrium strategies for t = n are: a(i, n) = (cid:26) OFFER [δn−1 , 0] ACCEPT 397 IF a’s TURN IF b’s TURN FAT IMA , WOOLDR IDGE , & JENN ING S IF b’s TURN IF a’s TURN IF a’s TURN IF b’s TURN IF b’s TURN IF a’s TURN OFFER tradeoffa1(K, δ, eub(ψ , t), i, ψ , m, t, P a , P b ) If offer gets rejected UPDATE BELIEFS RECEIVE OFFER and UPDATE BELIEFS If (U a i ([xt , y t ], t) ≥ eua(i, t)) ACCEPT else REJECT b(i, n) = (cid:26) OFFER [0, δn−1 ] ACCEPT for 1 ≤ i ≤ r . For all preceding time periods t < n, if [xt , y t ] denotes the offer made at time t, then the equilibrium strategies are defined as follows: a(i, t) =   OFFER tradeoffb1(K, δ, eua(φ, t), i, φ, m, t, P a , P b ) b(i, t) =  If offer gets rejected UPDATE BELIEFS RECEIVE OFFER and UPDATE BELIEFS  If (U b i (xt , y t ], t) ≥ eub(i, t)) ACCEPT else REJECT for 1 ≤ i ≤ r . Here, ψ = opta(i, t) and φ = optb(i, t). The earliest possible time of agreement is t = 1 and the latest possible time of agreement is t = min(2r − 1, n). Proof: At time t = n, the offering agent takes all the pies and leaves nothing for its opponent. The opponent accepts this and we get a(i, n) and b(i, n). Now consider a time period t < n. Recall that during negotiation for the complete information setting (see Section 3.1), at time t < n, the offering agent proposes a package that gives its opponent a cumulative utility equal to what the opponent would get from its own equilibrium offer for the next time period. However, for the current incomplete information setting, an agent knows its own type but not that of its opponent. Hence, for this scenario, at time t < n, the offering agent (say a) proposes a package that gives b an expected cumulative utility equal to what b would get from its own equilibrium offer for the next time period (i.e., eub(ψ , t)). This package is determined by the tradeoffa1 function. Likewise, if b is the offering agent at time t, then it makes tradeoffs using tradeoffb1 and offers a an expected cumulative utility eua(φ, t). We obtain the equilibrium offer for t = n − 1 and then reason backwards until we obtain the equilibrium offer for t = 1. However, since these offers are computed in the absence of complete information (i.e., on the basis of expected utilities), an agreement may or may not take place at t = 1. If an agreement does not take place at t = 1, then negotiation proceeds as follows. Consider a time period t such that 1 ≤ t < n. Let [xt , y t ] denote the offer made at time t. The agent that receives the offer (say agent a) updates its beliefs using Bayes’ rule: put all the weight of the posterior distribution of b’s type over K where K ⊆ {1, 2, . . . , r} is the set of possible types for b that can offer [xt , y t ] in equilibrium. If the proposed offer ([xt , y t ]) gets rejected, then the offering agent (say agent b of type i) updates its beliefs using Bayes’ rule: put all the weight of the posterior distribution of a’s type over T a t − {optb(i, t)}. The belief update rule for the case where agent a offers at time t are analogous to the above rule. These belief update rules when incorporated in the agents’ strategies give a(i, t) and b(i, t) as shown in the statement of the theorem. We now show that the beliefs specified above are consistent. D uring any time period t < n, let the strategy profile ( a(i, t), b(i, t)) assign probability 1 − ǫ to the above specified posterior beliefs and probability ǫ to the rest of the support for the opponent’s type. As ǫ → 0, the above strategy pair converges to (a, b). Also, the beliefs generated by the strategy pair converge to the beliefs described above. Given these beliefs, the strategies a and b are sequentially rational. 398 MULT I - I S SUE N EGOT IAT ION W ITH D EADL INE S The earliest possible time of agreement is t = 1. We show this with the following example. Let n = 2, m = 2, r = 2, δ = 1/2, and K = [1, 2; 5, 1]. Let agent a be the offering agent at time t = 1. Assume that a is of type 1 (i.e., ka = [1, 2]). Let P b (1) = 0.1 and P b (2) = 0.9. Since r = 2, agent a can play two possible strategies at time t = 1: one that corresponds to the case where b is of type 1 and the other that corresponds to the case where b is of type 2. For the former case, a’s equilibrium 4 ] for the second one. Hence eua(1, 1, 1) = 1.5. offer at t = 1 is [0, 1] for the first issue and [ 3 4 , 1 For the latter case, a’s equilibrium offer at t = 1 is [ 2 5 ] for the first issue and [1, 0] for the second 5 , 3 issue. Hence eua(1, 2, 1) = 2.16. Since eua(1, 2, 1) > eua(1, 1, 1), opta(1, 1) = 2 and a plays the latter strategy. Now if b is in fact of type 2, then it accepts a’s offer at t = 1. But if b is in fact of type 1, it rejects a’s offer at t = 1 since it can get a higher utility at t = 2. An agreement therefore occurs at t = 2. Thus, the earliest possible time of agreement is t = 1. Now consider the case where an a of type i offers at t = 1 but an agreement does not occur at this time. When a’s offer gets rejected, it knows that b is not of type opta(i, 1). Thus the number of possible types for b is now reduced to r − 1. This happens every time a makes an offer (i.e., every alternate time period) but it gets rejected. When negotiation reaches time period t = 2r − 1, there is only one possible type for b. Likewise, there is only one possible type for agent a. An agreement therefore takes place at t = 2r − 1. However, if n < 2r − 1 then an agreement occurs at t = n (see a(i, n) and b(i, n)). In other words, if an agreement does not occur at t = 1, then it occurs at the latest by t = min(2r − 1, n). (cid:3) As we mentioned earlier, if there is more than one package that solves Equation 15, then tradeoffa1 returns the one that maximises a’s expected cumulative utility. Let paij t (where i denotes a’s type and j that of b) denote the set of all possible packages that tradeoffa1 can return at time t. The set pbij t for agent b is defined analogously. Theorem 9 For a given first mover, the package deal procedure has a uniqu e equilibrium outcome if the condition C3 is false or C4 is true. C3 . There exists an i, j , c, and d, such that (c 6= d) and (i 6= j ) and (Kic /Kj c = Kid/Kj d ) where 1 ≤ i ≤ r , 1 ≤ j ≤ r , 1 ≤ c ≤ m, and 1 ≤ d ≤ m. C4 . pa ij t = 1 and pb ij t = 1 where 1 ≤ i ≤ r , 1 ≤ j ≤ r , i 6= j , and 1 ≤ t ≤ n. Proof: Let i denote agent a’s type and j denote b’s type where i 6= j , 1 ≤ i ≤ r , and 1 ≤ k ≤ r . Note that if a and b are of the same type, they have similar preferences for different issues. So i 6= j because the agents gain from making tradeoffs when they are of different types. The rest of the proof for the condition C3 follows from Theorem 4. Consider C4 . If C3 is true, then we know that, at time t, tradeoffa1 returns that package that solves Equation 15 and maximises a’s expected cumulative utility. Hence if paij t contains a single element, then there is only one possible package that tradeoffa1 can return. Likewise, if pbij t contains a single element, then there is only one possible package that tradeoffb1 can return. If there is only one possible offer for each time period 1 ≤ t ≤ n, then the equilibrium outcome is unique. (cid:3) In order to determine the time complexity of the package deal, we first find the complexity of the tradeoffa1 function. As we mentioned before, tradeoffa1 differs from tradeoffa when there is more than one package that solves the maximisation problem of Equation 15. We know from Theorem 9 that there is more than one such package if the condition C3 is true. We also 399 FAT IMA , WOOLDR IDGE , & JENN ING S know from Theorem 1 that using the greedy approach, tradeoffa considers the m issues in the increasing order of Kic/Kj c where i denotes a’s type and j denotes b’s type. Let S ij p ⊆ S denote a set of issues (where 0 ≤ D ij < m, 1 ≤ p ≤ D ij , i denotes a’s type, and j denotes b’s type) such that: and: S ij p > 1 for 1 ≤ p ≤ D ij ∀c,d∈S ij p Kic Kj c = Kid Kj d In other words, S ij p is a set of issues such that if c and d belong to S ij p then Kic/Kj c = Kid/Kj d , and D ij is the number of sets that satisfy this condition. So if D ij = 0 then it means that there is only one package that solves Equation 15. But if D ij > 0 then there is more than one package that solves Equation 15 and from among these tradeoffa1 must find the one that maximises a’s expected cumulative utility. For example if the set of issues is S = {1, 2, 3, 4}, r = 2, K1 = {5, 6, 7, 8}, 1 = 3. So while making tradeoffs, 1 = {2, 3, 4}, and S 12 and K2 = {9, 6, 7, 8}, then D12 = 1, S 12 in any order because for all the three issues it needs to give up the a can consider the issues in S 12 1 1 can be ordered same amount of utility in order to increase b’s utility by 1. The three issues in S 12 in 3! different ways resulting in 3! different packages. From among these 3! different packages, tradeoffa1 must find the one that maximises a’s expected cumulative utility. In general, for D ij > 1, let π ij denote the number9 of possible packages tradeoffa1 needs to consider where π ij is: Dij Yp=1 In other words, if a’s type is i and b’s type is j , then there are π ij packages that solve Equation 15 and from among these tradeoffa1 must find the one that maximises a’s expected cumulative utility. So if D ij = 0, then π ij = 1. Let π be defined as: S ij p ! π ij = π = max 1≤i≤r,1≤j≤r,i6=j π ij (18) In other words, π is the maximum number of packages that tradeoffa1 will have to search to find the one that maximises a’s expected cumulative utility (considering all possible types of a and all possible types of b). Note that, as before, a and b are of different types (i.e., i 6= j in Equation 18) because the agents gain from making tradeoffs when they are of different types. The time complexity of tradeoffa1 depends on π . Theorem 10 The time complexity of tradeoffa1 is O(m π). Proof: We know from Theorem 2 that the time complexity of finding any o ne package that solves Equation 15 is O(m). However, if there is more than one package that solves Equation 15 then tradeoffa1 returns the one that maximises a’s expected cumulative utility. The time to compute a’s expected cumulative utility from any one such package is O(m). The maximum number of such packages for which a needs to find its expected cumulative utility is π . Thus the time complexity of tradeoffa1 is O(m π). (cid:3) 9. Note that π ij is defined in terms of the factorial of p ≪ m. S ij p , but S ij p is independent of m and it is assumed that S ij 400 MULT I - I S SUE N EGOT IAT ION W ITH D EADL INE S Corollary 1 If D ij = 0 for 1 ≤ i ≤ r , 1 ≤ j ≤ r , and i 6= j , then the time complexity of tradeoffa1 is the same as the complexity of tradeoffa. Proof: If D ij = 0 for 1 ≤ i ≤ r , 1 ≤ j ≤ r , and i 6= j , then π ij = 1 and so π = 1. So the time complexity of tradeoffa1 is O(m). (cid:3) Theorem 11 The time complexity of computing the equilibrium offers for the package deal proce- dure is O(m π r3T (n − T 2 )) where T = min(2r − 1, n). Proof: Let a denote the agent that offers at t = 1 and assume that n is even (the proof for odd n is analogous). We begin with the last time period and then reason backwards. Since n is even and a starts at t = 1, it is b’s turn to offer in the last time period. For t = n, the time taken to find eub(i, j, t) (for a given i and j ) is O(m) (see Equation 10). Hence, the time taken to find eub(i, j, t) for all possible types of b (i.e., 1 ≤ j ≤ r) is O(mr). Note that, at this stage, eub(i, t − 1) is known for 1 ≤ i ≤ r (see Equation 12). Now consider the time period t = n − 1. Since n is even, it is a’s turn to offer at t = n − 1. In order to find a(i, t), we first need to find ψ where ψ = opta(i, t). From Equation 16 we know that, for a given i, the time to find opta(i, t) depends on the time taken to find eua(i, j, t) which in turn depends on the time to find fa (i, j, e, t) (see Equation 13). The time taken for fa (i, j, e, t) depends on the time taken for a(i, j, t). For a given i and a given j , the time taken to find a(i, j, t) is the time taken by the function tradeoffa. Since eub(j, t) is already known at time t, the time taken by tradeoffa1 is O(m π) (see Theorem 10). The time taken to find fa (i, j, e, t) is therefore O(m π). Given this, the time to find eua(i, j, t) (for a given i, j , and t) is O(m πr). Hence, for a given i, the time to find ψ = opta(i, t) is O(m πr2 ). At this stage, EUB(ψ , t) is known (see the last sentence in the first paragraph of this proof). Consequently , for a given i, the time to find a(i, t) is O(m πr2 ). Recall that each agent knows only its own type and not that of its opponent. Hence we need to determine a(i, t) for all possible types of a (i.e., for 1 ≤ i ≤ r). This takes O(m πr3 ) time. Note that at this stage eua(i, j, t) is known for all possible values of i and all possible values of j (where 1 ≤ i ≤ r and 1 ≤ j ≤ r). Now consider the time period t = n − 2 when it is b’s turn to offer. For t = n − 2 and a given i, the time to find optb(i, t) is O(m πr2 ) and so the time to find optb(i, t) for all possible types of b (i.e., for 1 ≤ i ≤ r) is O(m π r3 ). In the same way, the time required to do all the necessary computation for each time period t < n is O(m πr3 ). Hence, the total time to find the equilibrium offer for the fir st time period is O((n − 1)m π r3). However, as noted previously, an agreement may or may not occur in the first time period. If an agreement does not take place at t = 1, then the agents update their beliefs and compute the equilibrium offer for t = 2 with the updated beliefs. The time to compute the equilibrium offer for t = 2 is O((n − 2)m π r3 ). This process of updating beliefs and finding the equilibrium offer is repeated at most T = min(2r − 1, n) times. Hence the time complexity of the 2 )) (see Cormen et al., 2003, – page 47 package deal is ΣT i=1O((n − i)m π r3) = O(m πr3T (n − T – for details on how to simplify an expression of the form ΣT i=1O((n − i)m πr3 )). (cid:3) Theorem 12 The package deal procedure generates a Pareto optimal outcome. Proof: This follows from Theorem 3. The difference between the complete information setting of Theorem 3 and the current incomplete information setting is that for the former setting the agents maximise their cumulative utilities, whereas in the current setting they maximise their expected the offering agent maximises its expected cumulative utilities. Specifically, for every time period, 401 FAT IMA , WOOLDR IDGE , & JENN ING S cumulative utility from all the m issues such that its opponent’s expected cumulative utility is equal to what the opponent would get from its own equilibrium offer for the next time period. Hence, for the current setting, the equilibrium offer for every time period is Pareto optimal. (cid:3) 4.2 The Simultaneous Procedure Recall that for this procedure, the µ > 1 partitions are discussed in parallel but independently of each other. The offers made during the negotiation for any one partition do not affect the offers for the others. Specifically, negotiation for each partitio n starts at t = 1 and each partition is settled using the package deal procedure. Since each partition is dealt with separately, the results of Theorem 8 apply directly to each of the µ partitions. Let πc denote π for the cth partition. Then, from Theorem 11, we know that the time taken for the cth (for 1 ≤ c ≤ µ) partition is O(Sc πc r3T (n − T 2 )). Let the partition for which Sc πc is highest be denoted Sz . Then the time complexity of the simultaneous procedure is O(Sz πz r3T (n − T 2 )). Also, from Theorem 5, it follows that the simultaneous procedure may not generate a Pareto optimal outcome. Finally, from Theorem 9 we know that the simultaneous procedure has a unique equilibrium outcome if the following condition is satisfied : C5 . If there is no partition c (where 1 ≤ c ≤ µ) for which the condition (¬C3 ∨ C4 ) is false. 4.3 The Sequential Procedure For this procedure, the µ > 1 partitions are discussed independently and one after another. Also, for 1 ≤ c ≤ µ, negotiation on the cth partition starts in the time period that follows an agreement on the (c − 1)th partition. Since the package deal is used for each partition, the following results are obtained on the basis of Theorem 8. First, Theorem 8 applies to each of the µ > 1 partitions. Thus, for the sequential procedure, if negotiation for the cth (for 1 ≤ c ≤ µ) partition starts at time tc , then it ends at the earliest at time tc and at the latest by tc + min(2r − 1, n). Second, it follows from Theorem 11 that the time taken for the sequential procedure is O(Sz πz r3T (n − T 2 )). Third, the sequential procedure may not generate a Pareto optimal outcome (see Theorem 5). Finally, the conditions for uniqueness are the same as those for the simultaneous procedure. 4.4 The Optimal Procedure Having obtained the equilibrium outcomes for the three procedures for the above defined incomplete information scenario, we now compare them in terms of the expected utilities they generate to each player. Again, the procedure that gives a player the maximum expected utility is the optimal one. Theorem 13 The package deal is optimal for each agent. Proof: The proof for this is the same as Theorem 7. The only difference between the complete information setting of Theorem 7 and the current incomplete information setting is that for the package deal procedure for the former setting (during time period t < n), the offering agent pro- poses a package that maximises its own cumulative utility, while giving its opponent a cumulative utility equal to what the opponent would get from its own equilibrium offer in the next time period. On the other hand, for the current incomplete information setting, the offering agent proposes a package that maximises its own expected cumulative utility while giving its opponent an expected 402 MULT I - I S SUE N EGOT IAT ION W ITH D EADL INE S Time of agreement Simultaneous Package deal Earliest: 1 Earliest: 1 Latest: min(2r − 1, n) Latest: min(2r − 1, n) for all m issues for all m issues Sequential For the cth partition te c = ts c c = ts tl c + min(2r − 1, n) for 1 ≤ c ≤ µ O(Sz πz r3T (n − T 2 )) Time to compute equilibrium Pareto optimal? Unique equilibrium? O(m πr3T (n − T 2 )) O(Sz πz r3T (n − T 2 )) Yes If ¬C3 ∨ C4 No If C5 No If C5 Table 3: A comparison of the expected outcomes for the three multi-issue procedures for the sym- c denotes the start time for the metric information setting (for the sequential procedure, ts cth partition, te c the earliest possible time of agreement, and tl c the latest possible time of agreement). cumulative utility equal to what the opponent would get from its own equilibrium offer in the next time period. Also, for each agent, the package deal maximises the expected cumulative utility from all the m issues (since tradeoffs are made across all the m issues). But the simultaneous procedure maximises each agent’s expected cumulative utility for each partition (i.e., the simultaneous proce- dure does not make tradeoffs across partitions). Hence each agent’s expected cumulative utility for all the m issues is higher for the package deal relative to the simultaneous procedure. Furthermore, irrespective of how the m issues are partitioned into µ partitions, we know that the simultaneous procedure is better than the sequential one for each agent (see Theorem 7). Hence, the package deal is optimal for each agent. (cid:3) These results are summarised in Table 3. 5. Multi-Issue Negotiation with Asymmetric Uncertainty about the Opponent’s Utility In some bargaining situations, one of the players may know something of relevance that the other may not know. For example, when bargaining over the price of a second hand car, the seller knows its quality but the buyer does not. Such situations are said to have asymmetry in information between the players (Muthoo, 1999). Our asymmetric information setting differs from the symmetric one explored in the previous section in that one of the two agents (say a) has complete information, but the other (say b) is uncertain about a’s utility function: for 1 ≤ c ≤ m, agent b is uncertain about c . Here, K , P a , P b , n, r , and m are as defined in Section 4. The negotiation parameters K , P a , ka P b , r , δ , n, and m are common knowledge to the negotiators. Furthermore, a knows its own type and that of b, while b knows its own type but not that of a. Finally, the definitions for the cumulative utility functions remain the same as in Section 4. For this setting, we now determine the equilibrium for each of the three multi-issue procedures. 403 FAT IMA , WOOLDR IDGE , & JENN ING S 5.1 The Package Deal Procedure We extend the analysis of Section 4 to the current setting as follows. It is clear that for the last time period (t = n), the utilities eua(i, t) and eub(i, t) are as per Section 4. Let ¯j denote b’s actual type. Recall that agent a now knows ¯j . Hence on the basis of Equation 13 for the SUI setting, we get eua(i, j, t) for the current asymmetric information setting as follows: eua(i, j, t) = F a (i, j, ¯j , t) for 1 ≤ i ≤ r and 1 ≤ j ≤ r (19) On the other hand, since agent b is uncertain about a’s type, the definitions for eub(i, t) and eub(i, j, t) are as given in Section 4. Also, the definitions for F a , F b , a(i, j, t), b(i, j, t), opta(i, t), and optb(i, t) for all time periods remain the same as in Section 4. Finally, in this setting, belief updating does not apply to agent a because it has complete in- formation. Only agent b updates its beliefs about a. This is done in the same way described in Section 4. Because of b’s uncertainty, we use the concept of sequential equilibrium in this setting as well. The following theorem characterises the equilibrium for the package deal procedure. IF a’s TURN IF b’s TURN IF b’s TURN IF a’s TURN OFFER tradeoffa1(K, δ, eub(¯j , t), i, ¯j , m, t, P a , P b ) RECEIVE OFFER If (U a i ([xt , y t ], t) ≥ eua(i, t)) ACCEPT else REJECT Theorem 14 For the package deal procedure the following strategies form a sequential equilib- rium. The equilibrium strategies for t = n are: a(i, n) = (cid:26) OFFER [δn−1 , 0] ACCEPT b(i, n) = (cid:26) OFFER [0, δn−1 ] ACCEPT for 1 ≤ i ≤ r . For all preceding time periods t < n, if [xt , y t ] denotes the offer made at time t, then the equilibrium strategies are defined as follows: a(i, t) =   OFFER tradeoffb1(K, δ, eua(φ, t), i, φ, m, t, P a , P b ) b(i, t) =  If offer gets rejected UPDATE BELIEFS RECEIVE OFFER and UPDATE BELIEFS  If (U b i (xt , y t ], t) ≥ eub(i, t)) ACCEPT else REJECT for 1 ≤ i ≤ r . Here, ¯j denotes agent b’s type and φ = optb(i, t). The earliest possible time of agreement is t = 1 and the latest possible time is t = min(2r − 1, n). Proof: As Theorem 8. The only difference is that a now knows b’s type (¯j ). Hence this information is used as a parameter for tradeoffa1. The earliest possible time of agreement is t = 1. We show this with the following example. Let n = 2, m = 2, r = 2, δ = 1/2, and K = [1, 2; 5, 1]. Let b (i.e., the agent with uncertain information) be the offering agent at time t = 1. Assume that b is of type 2 (i.e., kb = [5, 1]). Let P a (1) = 0.9 and P a (2) = 0.1. Since r = 2, b can play two possible strategies at time t = 1: one that corresponds to the case where a is of type 1 and the other that corresponds to the case where a is of type 2. For the former case, b’s equilibrium offer at t = 1 is [0, 1] for the first issue IF a’s TURN IF b’s TURN IF b’s TURN IF a’s TURN 404 MULT I - I S SUE N EGOT IAT ION W ITH D EADL INE S 4 ] for the second. Hence eub(1, 1, 1) = 4.725. For the latter case, b’s equilibrium offer and [ 3 4 , 1 5 ] for the first issue and [1, 0] for the second one. Hence eub(1, 2, 1) = 3. Since at t = 1 is [ 2 5 , 3 eub(1, 1, 1) > eub(1, 2, 1), optb(1, 1) = 1 and b plays the former strategy. Now if a is in fact of type 1, then it accepts b’s offer at t = 1. But if a is in fact of type 2, it rejects b’s offer at t = 1 since it can get a higher utility at t = 2. An agreement therefore occurs at t = 2. Thus, the earliest possible time of agreement is t = 1. Now consider the case where an agent b of type i offers at t = 1 but an agreement does not occur at this time. When b’s offer gets rejected, it knows that a is not of type optb(i, 1). Thus the number of possible types for a is now reduced to r − 1. This happens every time b makes an offer (i.e., every alternate time period) but it gets rejected. When negotiation reaches time period t = 2r − 1, there is only one possible type for a. Since a knows b’s type, an agreement therefore takes place at t = 2r − 1. However, if n < 2r − 1 then an agreement occurs at t = n (see a(i, n) and b(i, n)). In other words, if an agreement does not occur at t = 1, then it occurs at the latest by t = min(2r − 1, n). (cid:3) Note that the latest possible time of agreement for the asymmetric information setting is the same as that for the symmetric information setting of Theorem 8. This is because, in the asymmetric setting, although a knows b’s type, b is uncertain about a’s type. Also, it takes 2r − 1 time periods for b to come to know a’s actual type. Hence, the earliest and latest time of agreement is the same for both settings. Theorem 15 The time complexity of computing the equilibrium offers for the package deal proce- 2 (n − T dure is O(m π r3 T 2 )) where T = min(2r − 1, n). Proof: Let a denote the agent that offers at t = 1 and assume that n is even (the proof for odd n is analogous). We begin with the last time period and then reason backwards. Since n is even and agent a starts at t = 1, it is b’s turn to offer in the last time period. For t = n, the time taken to find eub(i, j, t) (for a given i and j ) is O(m) (see Equation 10). Hence, the time taken to find eub(i, j, t) for all possible types of b (i.e., 1 ≤ j ≤ r) is O(mr). Note that, at this stage, eub(i, t − 1) is known for 1 ≤ i ≤ r (see Equation 12). Now consider the time period t = n − 1. Since n is even, it is a’s turn to offer at t = n − 1. In order to find a(i, t), we first need to find ψ where ψ = opta(i, t). From Equation 16 we know that, for a given i, the time to find opta(i, t) depends on the time taken to find eua(i, j, t) which, in turn, depends on the time to find fa (i, j, e, t) (see Equation 19). The time taken for fa (i, j, e, t) depends on the time taken for a(i, j, t). For a given i and a given j , the time taken to find a(i, j, t) is the time taken by tradeoffa1. Since eub(j, t) is already known at time t, the time taken by the function tradeoffa1 is O(m π) (as Theorem 2). The time taken to find fa (i, j, e, t) is therefore O(m π). Given this, the time to find eua(i, j, t) (for a given i, j , and t) is O(m π) since b’s type is known to both agents – see Equation 19. Hence, for a given i, the time to find ψ = opta(i, t) is O(m πr). At this stage, EUB(ψ , t) is known (see the last sentence in the first paragraph of this proof). Consequently, for a given i, the time to find a(i, t) is O(m πr). Recall that b does not know a’s type. Hence we need to determine a(i, t) for all possible types of a (i.e., for 1 ≤ i ≤ r). This takes O(m π r2 ) time. Note that at this stage eua(i, j, t) is known for all possible values of i and all possible values of j (where 1 ≤ i ≤ r and 1 ≤ j ≤ r). Now consider the time period t = n − 2 when it is b’s turn to offer. The only difference between the computation for t = n − 1 and t = n − 2 is that for the former case, the time to find eua(i, j, t) (for a given i, j , and t) is O(m π) since b’s type is known to both agents. However for the latter 405 FAT IMA , WOOLDR IDGE , & JENN ING S case, the time to find eub(i, j, t) (for a given i, j , and t) is O(m πr) since a’s type is not known to b (see Equation 14). Consequently, for a given i, the time to find b(i, t) is O(m πr2 ). So the time to determine b(i, t) for all possible types of b (i.e., for 1 ≤ i ≤ r) is O(m πr3 ) time. Note that at this stage eub(i, j, t) is known for all possible values of i and all possible values of j (where 1 ≤ i ≤ r and 1 ≤ j ≤ r). In the same way, the time required to do all the necessary computation for each odd time period t < n is O(m π r2 ), while that for each even time period is O(m πr3 ). Hence, the total time to find the equilibrium offer for the first time period is O(m πr3 ( n−1 2 )). However, as noted previously, an agreement may or may not occur in the first time period. If an ag reement does not take place at t = 1, then the agents update their beliefs and compute the equilibrium offer for t = 2 with the updated beliefs. The time to compute the equilibrium offer for t = 2 is O(m πr3 ( n−2 2 )). This process of updating beliefs and finding the equilibrium offer is repeat ed at most T = min(2r − 1, n) times. i=1O(m πr3 ( n−i Hence the time complexity of the package deal is ΣT 2 ). (cid:3) 2 )) = O(m πr3 (n − T 2 ) T Theorem 16 The package deal procedure generates a Pareto optimal outcome. Proof: As per Theorem 12. (cid:3) Theorem 17 For a given first mover, the package deal procedure has a uniqu e equilibrium outcome if ¬C3 ∨ C4 is true. Proof: As per Theorem 9. (cid:3) 5.2 The Simultaneous Procedure Theorem 14 applies to each of the µ > 1 partitions. Hence, from Theorem 15, we know that the time taken for the cth (for 1 ≤ c ≤ µ) partition is O(Sc πc r3 ( n−T 2 ). Hence, the time complexity 2 ) T of the simultaneous procedure is O(Sz πz r3 (n − T 2 ) T 2 ). Also, from Theorem 5, it follows that the simultaneous procedure may not generate a Pareto optimal outcome. Finally, from Theorem 17 we know that the simultaneous procedure has a unique equilibrium outcome if the condition C5 is true. 5.3 The Sequential Procedure First, Theorem 14 applies to each of the µ > 1 partitions. Thus, for the sequential procedure, if negotiation for the cth (for 1 ≤ c ≤ µ) partition starts at time tc , then it ends at the earliest at time tc and at the latest by tc + min(2r − 1, n). Second, it follows from Theorem 15 that the time taken 2 ) T for the sequential procedure is O(Sz πz r3 (n − T 2 ). Third, the sequential procedure may not generate a Pareto optimal outcome (see Theorem 5). Finally, the conditions for uniqueness are the same as those for the simultaneous procedure. 5.4 The Optimal Procedure It follows from Theorem 13 that, for each agent, the optimal procedure is the package deal. These results are summarised in Table 4. 6. Multi-Issue Negotiation for Interdependent Issues For the independent issues case of Section 4, an agent’s utility for issue c (for 1 ≤ c ≤ m) depends only on its share for that issue and is independent of other issues. However, in many cases, an 406 MULT I - I S SUE N EGOT IAT ION W ITH D EADL INE S Time of agreement Simultaneous Package deal Earliest: 1 Earliest: 1 Latest: min(2r − 1, n) Latest: min(2r − 1, n) for all m issues for all m issues Sequential For the cth partition te c = ts c c = ts tl c + min(2r − 1, n) for 1 ≤ c ≤ µ O(Sz πz r3 (n − T 2 ) T 2 ) Time to compute equilibrium Pareto optimal? Unique equilibrium? O(m πr3 T 2 (n − T 2 )) O(Sz πz r3 (n − T 2 ) T 2 ) Yes If ¬C3 ∨ C4 No If C5 No If C5 Table 4: A comparison of the expected outcomes for the three multi-issue procedures for the asym- c denotes the start time for the metric information setting (for the sequential procedure, ts cth partition, te c the earliest possible time of agreement, and tl c the latest possible time of agreement). if t ≤ n otherwise agent’s utility from an issue depends not only on its share for the issue, but also on its share for others (Klein et al., 2003). Given this, in this section we focus on such interdependent issues. Specifically, we model interdependence between the issues a s follows. Consider a package [xt , y t ]. For this package, for an agent a of type i, the utility from issue c at time t is now of the form: ic ([xt , y t ], t) = (Kicxc + Σm j=1χij (xc − xj ) ua 0 and that for an agent b of type i, it is: ic ([xt , y t ], t) = (Kicyc + Σm j=1χij (yc − yj ) ub 0 where Kic denotes a constant positive real number and χij a constant real number that may be either positive or negative. As before, an agent’s cumulative utility is the sum of its utilities from the individual issues: i ([xt , y t ], t) = (Σm c=1 U a 0 i ([xt , y t ], t) = (Σm c=1 U b 0 Here ¯K denotes a vector analogous to the vector K except that the individual elements of the latter are all constant positive real numbers, while those of the former may be positive or negative. Note that in Equations 5 and 6, all the coefficients are positi ve (i.e., Kic > 0 for 1 ≤ i ≤ r and 1 ≤ c ≤ m). But in Equations 22 and 23, the coefficient ( ¯Kic ) may be a positive or a negative real number. if t ≤ n otherwise if t ≤ n otherwise ¯Kicxt c ¯Kicy t c if t ≤ n otherwise (20) (21) (22) (23) 407 FAT IMA , WOOLDR IDGE , & JENN ING S The above cumulative utility functions are linear (see Pollak, 1976; Charness & Rabin, 2002; Sobel, 2005, for other forms of utility functions for interdependent preferences10 ). As mentioned before, we chose the linear form for reasons of computational tractability. In this setting the vector ¯K and the functions P a and P b are common knowledge to the nego- tiators. Also, each agent knows its own type, but not that of its opponent. In addition, each agent knows r , δ , n, and m. In other words, there is symmetric uncertainty about the opponent’s utility (as we will see in Section 6.4, the results for the asymmetric case can easily be obtained from the following analysis for the symmetric case). 6.1 The Package Deal Procedure For the cumulative utilities defined in Equations 22 and 23, T heorem 18 characterises the equilib- rium for the package deal. IF a’s TURN IF b’s TURN IF b’s TURN IF a’s TURN OFFER tradeoffa1( ¯K , δ, eub(ψ , t), i, ψ , m, t, P a , P b ) If offer gets rejected UPDATE BELIEFS RECEIVE OFFER and UPDATE BELIEFS If (U a i ([xt , y t ], t) ≥ eua(i, t)) ACCEPT else REJECT Theorem 18 For the package deal procedure, the following strategies form a sequential equilib- rium. The equilibrium strategies for t = n are: a(i, n) = (cid:26) OFFER [δn−1 , 0] ACCEPT b(i, n) = (cid:26) OFFER [0, δn−1 ] ACCEPT for 1 ≤ i ≤ r . For all preceding time periods t < n, if [xt , y t ] denotes the offer made at time t, then the equilibrium strategies are defined as follows: a(i, t) =   OFFER tradeoffb1( ¯K , δ, eua(φ, t), i, φ, m, t, P a , P b ) b(i, t) =  If offer gets rejected UPDATE BELIEFS RECEIVE OFFER and UPDATE BELIEFS  If (U b i (xt , y t ], t) ≥ eub(i, t)) ACCEPT else REJECT for 1 ≤ i ≤ r . Here, ψ = opta(i, t) and φ = optb(i, t). The earliest possible time of agreement is t = 1 and the latest possible time is t = min(2r − 1, n). Proof: As Theorem 8. The only difference between the independent issues setting of Theorem 8 and the present interdependent issues one is in terms of the definition for cumulative utilities: in Equations 5 and 6, all the coefficients are positive (i.e., Kic > 0 for 1 ≤ i ≤ r and 1 ≤ c ≤ m). But in Equations 22 and 23, the coefficient ( ¯Kic ) may be a positive or a negative real number. However, the greedy method (given in Theorem 1) for solving the fractional knapsack problem of Equation 15 works for both positive and negative coefficient s (Martello & Toth, 1990; Cormen et al., 2003). Hence, the proof of Theorem 8 applies to this setting as well. (cid:3) IF a’s TURN IF b’s TURN IF b’s TURN IF a’s TURN 10. Although in (Pollak, 1976; Charness & Rabin, 2002; Sobel, 2005) these forms are discussed in the context of how an agent’s utility depends on the utility of other agent’s, they may equally well be interpreted for the case where an agent’s utility for an issue depends on its share for other issues. 408 MULT I - I S SUE N EGOT IAT ION W ITH D EADL INE S Theorem 19 The time complexity of computing the equilibrium offers for the package deal proce- dure is O(m π r3T (n − T 2 )) where T = min(2r − 1, n). Proof: As Theorem 11. Since the method for making tradeoffs is the same as that for the setting with symmetric uncertainty and independent issues (i.e., SUI ), the time complexity is the same as in Theorem 11. (cid:3) It is obvious that Theorems 9 and 12 extend to this setting as well. 6.2 The Simultaneous Procedure It follows from above that all the results of Section 4.2 apply to this setting as well. 6.3 The Sequential Procedure It also follows from above that the results of Section 4.3 apply to this setting as well. 6.4 The Optimal Procedure It follows from Theorem 13 that the package deal remains the optimal procedure even if the issues are interdependent. The results for this setting are the same as those in Section 4 and are summarised in Table 3. Finally, consider the asymmetric information setting of Section 5 but in the current context of interdependent issues. From the above analysis for symmetric uncertainty with interdependent issues, it is clear that the method for making tradeoffs remains the same irrespective of whether the information is symmetric or asymmetric. Consequently, for the case of asymmetric information with interdependent issues, we get the same results as those in Section 5. Recall that this analysis was done for linear cumulative utilities. We now discuss how our results would hold for more complex utility functions that are non-linear11 . For cumulative utilities that are nonlinear, the tradeoff problem becomes a global optimization problem with a nonlinear objective function. Due to their computational complexity, such nonlinear optimization problems can only be solved using approximation methods (Horst & Tuy, 1996; Bar-Yam, 1997; Klein et al., 2003). In contrast, our tradeoff problem is a linear optimization problem, the exact solution to which can be found in polynomial time (as shown in Theorems 1 and 2). Although our results apply to linear cumulative utilities, it is not difficult to s ee how they would hold for the nonlinear case. First, the time of agreement for our case would hold for other (nonlinear) functions. This is because this time depends not on the actual definition of th e agents’ cumulative utilities but on the information setting (i.e., whether or not the information is complete). Second, let O(ω) denote the time complexity of TRAD EO FFA1 for nonlinear utilities for the package deal with µ = 1, and O(ωc) that for the cth partition. Also, let Sz denote the partition for which O(ωz ) is the highest between all partitions. Then, we know from Theorem 11 that the time complexity of the package deal for the setting with symmetric uncertainty is O(ωr3T (n− T 2 )). Consequently, the time complexity of both the simultaneous and the sequential procedures is O(ωz r3T (n − T 2 )). Third, while the package deal outcome for our additive cumulative utilities is Pareto optimal, the package deal outcome for nonlinear utilities may not be Pareto optimal. This is because (as stated above) 11. Note that bilateral bargaining for which the players’ utility functions are nonlinear has been studied by Hoel (1986) in the context of a single issue as opposed to the multi-issue case which is the focus of our study. 409 FAT IMA , WOOLDR IDGE , & JENN ING S nonlinear optimization problems can only be solved using approximation methods while the linear optimization problem can be solved using an exact method (as in proof of Theorem 1). Finally, since the conditions for a unique solution depend on the actual definition of cumulative utilities, the conditions given in Tables 1 2, 3, and 4 may not hold for other forms of utility functions. 7. Related Work Since Schelling (1956) first noted the fact that the outcome o f negotiation depends on the choice of negotiation procedure, much research effort has been devoted to the study of different procedures for negotiating multiple issues. For instance, Fershtman (1990) extended the model developed by Rubinstein (1982), for splitting a single pie, to sequential negotiation for two pies. However, this model assumes complete information, imposes an agenda exogenously, and then studies the relation between the agenda and the outcome of the sequential bargaining game. In more detail, for two pies of different sizes, he analyses the effect of going first on th e large and the small pie. A number of researchers have also studied negotiations with an endogenous agenda (Inderst, 2000; In & Serrano, 2003; Bac & Raff, 1996). In Inderst (2000) players have discount factors, but no deadlines. For independent issues, this work assumes complete information and studies three different negotiation procedures: package deal, simultaneous, and sequential negotiation with endogenous agenda. Their main result is that the package deal is the optimal procedure and that for each procedure there exist multiple equilibria. In and Serrano (2003) extend this work by finding conditions under which the equilibrium becomes unique. Note that our work differs from both of these in that we analyse negotiations with both discount factors and deadlines, which we consider to be much more common with automated negotiations. Moreover, we do this for both independent and interdependent issues without making the complete information assumption. Bac and Raff (1996) also developed a model that has an endogenous agenda. They extended the model developed by Rubinstein (1985) for single pie bargaining with incomplete information by adding a second pie. In this model, the players have discount factors, but no deadlines. The size of the pie is known to both agents and the discounting factor is assumed to be equal for all the issues for both agents. Also, there is asymmetric information: one of the players knows its own discounting factor and that of its opponent, while the other player knows its own discounting factor, but is uncertain of its opponent’s. In more detail, this factor can take one of two values, δH with probability x, and δL with probability 1 − x. These probabilities are common knowledge. For this model, the authors determine the equilibrium for the package deal and the sequential procedure. They show that, under certain conditions, the sequential procedure can be the optimal one. However, there are three key differences between this model and ours. First, we analyse both symmetric and asymmetric information settings, while Bac and Raff analyse only the latter. Second, the negotiators in our model have a deadline, while in Bac and Raff they do not. Again, we believe our analysis covers situations that often occur in automated negotiation settings. Finally, Bac and Raff focus on independent issues, but we analyse both independent and interdependent issues. A slightly different approach (from the above ones) was taken by Busch and Horstmann (1997). Again, they extended the model developed by Rubinstein (1985), but by adding a preliminary period in which the agents bargain over the agenda. The outcome of this stage is then used as the agenda for negotiating over the issues. In this complete information model, there are two pies for bargaining. Furthermore, these two issues become available for negotiation at different time points. The players have discount factors or fixed time costs, but no deadlines. S ince there are two issues, there are two 410 MULT I - I S SUE N EGOT IAT ION W ITH D EADL INE S possible agendas. The outcome for these two agendas is compared with that for the package deal. Their main result is that the players may have conflicting pre ferences over the optimal agenda. Note that a key difference between this model and ours is that all the issues in our model are available from the beginning, while in their model the two issues become available at different time points. Furthermore, Busch and Horstman assume complete information, while we do not. From all the models mentioned above, perhaps the one that is closest to ours is the one developed by Inderst (2000). Unlike our work, Inderst assumes complete information and independent issues. Also, it does not model player deadlines, while we do. However, Inderst does model players’ time preferences as discount factors. Also, just like our model, all the issues for negotiation are available at the beginning of negotiation. In terms of results, Inderst shows that the package deal is the optimal procedure. Our study also shows that the package deal is the optimal procedure for both agents. Finally, our work provides a detailed analysis of the attributes of the different procedures (such as the time of agreement, the time complexity, the Pareto optimality, and the conditions for uniqueness), while Inderst does not. In summary, all the aforementioned models for multi-issue negotiation differ from ours in at least one of three major ways. The players in our model have both discount factors and deadlines, but a general characteristic of the above models is that the players only have discount factors but no deadlines12 . Negotiation with deadlines has been studied by Sandholm and Vulkan (1999) (in the context of a single issue) and by Fatima et al. (2004) for the sequential procedure with µ = m. Given this, our contribution lies firstly in finding the equilibriu m for all the three procedures. Second, we analyse both asymmetric and symmetric information settings, while previous work analyses only the former. Third, we analyse both independent and interdependent issues while previous work focuses primarily on independent issues. Furthermore, the existing literature does not compare the different multi-issue procedures in terms of their attributes (viz. time complexity, Pareto optimality, uniqueness, and time of agreement). By considering these, our study allows a more informed choice to be made about a wider range of tradeoffs that are involved in determining which is the most appropriate procedure. Finally, we would like to point that in Fatima et al. (2006), we considered independent issues and carried out the same study as we do in this work, but in a symmetric information setting with uncertainty about the negotiation deadline (as opposed to uncertainty over the agents’ utility func- tions that is the focus of this work). The key result of (Fatima et al., 2006) is similar to the result of our current work, namely that the optimal procedure in (Fatima et al., 2006) is the package deal. 8. Conclusions and Future Work This paper studied bilateral multi-issue negotiation between self-interested agents in a wide range of settings. Each player has time constraints in the form of deadlines and discount factors. Specifically, we considered both independent and interdependent issues and studied the three main multi-issue procedures for conducting such negotiations: the package deal, the simultaneous procedure, and the sequential procedure. We determined equilibria for each procedure for two different information settings. In the first, there is symmetric uncertainty about the opponent’s utility. In the second, there is asymmetric uncertainty about the opponent’s utility. We analysed both settings for the case of independent and interdependent issues. For each setting, we compared the outcomes of the 12. (Fatima et al., 2004) studies a multi-issue model with deadlines, but it focuses on determining the equilibrium for one speci fic sequential procedure: the one in which each partiti on has a single issue. 411 FAT IMA , WOOLDR IDGE , & JENN ING S different procedures and showed that the package deal is optimal for each agent. We then compared the three procedures in terms of four attributes: the time complexity of the procedure, the Pareto optimality of the equilibrium solution, the uniqueness of the equilibrium solution, and the time of agreement (see Table 1). In more detail, our study shows that the package deal is in fact the optimal procedure for each party. We also showed that although the package deal may be computationally more complex than the other two procedures, it generates Pareto optimal outcomes (unlike the other two procedures), it has similar earliest and latest possible times of agreement as the simultaneous procedure (which is better than the sequential procedure), and that it (like the other two procedures) generates a unique outcome only under certain conditions (which we defined). There are several interesting directions for extending the current analysis. First, in this work, we modelled the players’ time preferences in the form of discount factors which is the most common basis for such analysis. However, existing literature (Busch & Horstman, 1997) shows that the outcome for negotiation with discount factors can differ from the outcome for negotiation with fixed time costs. It will, therefore, be interesting to exten d our results to negotiations with fixed time costs. Second, our present work analysed the setting with uncertainty about utility functions. Generalisation of our results to scenarios with other sources of uncertainties such as the agents’ discount factors is another direction for future work. Acknowledgements We are grateful to Sarit Kraus for her detailed comments on earlier versions of this paper. We also thank the anonymous referees; their comments helped us to substantially improve the readability and accuracy of the paper. 412 MULT I - I S SUE N EGOT IAT ION W ITH D EADL INE S Appendix A. Summary of Notation a, b The two negotiating agents. n Negotiation deadline for both agents. m Total number of issues. S The set of m issues. Sc A subset of S (Sc ⊆ S ). M Number of issues in the largest partition. µ Number of partitions for the simultaneous and sequential procedures. δc Discount factor for issue c (for 1 ≤ c ≤ m). δ An m element vector that represents the discount factor for the m issues. xt An m element vector that denotes a’s share for each of the m issues at time t. y t An m element vector that denotes b’s share for each of the m issues at time t. [xt , y t ] The package offered at time t. c Agent a’s share for issue c in the equilibrium offer for time period t. at c Agent b’s share for issue c in the equilibrium offer for time period t. bt at An m element vector that denotes a’s share for each of the m issues in equilibrium at time t. bt An m element vector that denotes b’s share for each of the m issues in equilibrium at time t. [at , bt ] The equilibrium package offered at time t. i Cumulative utility function for agent a of type i. U a i Cumulative utility function for agent b of type i. U b ua(t) Agent a’s cumulative utility from the equilibrium offer for time t. ub(t) Agent b’s cumulative utility from the equilibrium offer for time t. a(i, j, t) Agent a’s equilibrium offer for time t if a is of type i assuming b is type j . b(i, j, t) Agent b’s equilibrium offer for time t if b is of type i assuming a is type j . a(i, t) Equilibrium strategy for an agent a of type i at time t. b(i, t) Equilibrium strategy for an agent b of type i at time t. eua(i, t) Cumulative utility that an agent a of type i expects to get from b’s equilibrium offer at time t (i.e., a is the receiving agent and b the offering agent at t). 413 FAT IMA , WOOLDR IDGE , & JENN ING S eub(i, t) Cumulative utility that an agent b of type i expects to get from a’s equilibrium offer at time t (i.e., b is the receiving agent and a the offering agent at t). eua(i, j, t) Agent a’s expected cumulative utility from its equilibrium offer for time t if a is type i and assuming that b is type j . eub(i, j, t) Agent b’s expected cumulative utility from its equilibrium offer for time t if b is type i and assuming a is type j . r Number of types for agent a (and also the number of types for agent b). t Set of possible types for agent a at time t. T a t Set of possible types for agent b at time t. T b P a The probability distribution function for ka . P b The probability distribution function for kb . K A vector of r vectors each element of which is in turn a vector of m positive reals. p ⊆ S where i denotes a’s type and j that of b) such that S ij p A subset of S (S ij S ij p > 1 and = Kid Kic . ∀c,d∈S ij Kjc Kjd p tradeoffa Agent a’s function for making tradeoffs in the complete information setting. tradeoffb Agent b’s function for making tradeoffs in the complete information setting. tradeoffa1 Agent a’s function for making tradeoffs in the four incomplete information settings: SUI , SUD , AUI , AUD . tradeoffb1 Agent b’s function for making tradeoffs in the four incomplete information settings: SUI , SUD , AUI , AUD . π Maximum number of packages that tradeoffa1 (or tradeoffb1) will have to search to find the one that maximises a’s (or b’s) expected cumulative utility (considering all possible types of a and b). paij t The set of all possible packages that tradeoffa1 can return at time t (i denotes a’s type and j that of b). pbij t The set of all possible packages that tradeoffb1 can return at time t (i denotes a’s type and j that of b). References Bac, M., & Raff, H. (1996). Issue-by-issue negotiations: the role of information and time preference. Games and Economic Behavior, 13, 125–134. Bar-Yam, Y. (1997). Dynamics of Complex Systems. Addison Wesley. 414 MULT I - I S SUE N EGOT IAT ION W ITH D EADL INE S Binmore, K., Osborne, M. J., & Rubinstein, A. (1992). Noncooperative models of bargaining. In Aumann, R. J., & Hart, S. (Eds.), Handbook of Game theory with Economic Applications, Vol. 1, pp. 179–225. North-Holland. Busch, L. A., & Horstman, I. J. (1997). Bargaining frictions, bargaining procedures and implied costs in multiple-issue bargaining. Economica, 64, 669–680. Charness, G., & Rabin, M. (2002). Understanding social preferences with simple tests. The Quar- terly Journal of Economics, 117(3), 817–869. Cormen, T. H., Leiserson, C. E., Rivest, R. L., & Stein, C. (2003). An introduction to algorithms. The MIT Press, Cambridge, Massachusetts. Faratin, P., Sierra, C., & Jennings, N. R. (2002). Using similarity criteria to make trade-offs in automated negotiations. Artificial Intelligence Journal , 142(2), 205–237. Fatima, S. S., Wooldridge, M., & Jennings, N. R. (2002). The influence of information on negoti- ation equilibrium. In Agent Mediated Electronic Commerce IV, Designing Mechanisms and Systems, No. 2531 in LNCS, pp. 180 – 193. Springer Verlag. Fatima, S. S., Wooldridge, M., & Jennings, N. R. (2004). An agenda based framework for multi- issue negotiation. Artificial Intelligence Journal , 152(1), 1–45. Fatima, S. S., Wooldridge, M., & Jennings, N. R. (2006). On efficient procedures for multi-issue ne- gotiation. In Proceedings of the Eighth International Workshop on Agent Mediated Electronic Commerce (AMEC), pp. 71–85, Hakodate, Japan. Fershtman, C. (1990). The importance of the agenda in bargaining. Games and Economic Behavior, 2, 224–238. Fershtman, C. (2000). A note on multi-issue two-sided bargaining: bilateral procedures. Games and Economic Behavior, 30, 216–227. Fershtman, C., & Seidmann, D. J. (1993). Deadline effects and inefficient delay in bargaining with endogenous commitment. Journal of Economic Theory, 60(2), 306–321. Fishburn, P. C. (1988). Normative thoeries of decision making under risk and uncertainty. In Bell, D. E., Raiffa, H., & Tversky, A. (Eds.), Decision making: Descriptive, normative, and prescriptive interactions. Cambridge University Press. Fisher, R., & Ury, W. (1981). Getting to yes: Negotiating agreement without giving in. Houghton Mifflin, Boston. Fudenberg, D., Levine, D., & Tirole, J. (1985). Infinite hori zon models of bargaining with one sided incomplete information. In Roth, A. (Ed.), Game Theoretic Models of Bargaining. University of Cambridge Press, Cambridge. Fudenberg, D., & Tirole, J. (1983). Sequential bargaining with incomplete information. Review of Economic Studies, 50, 221–247. Harsanyi, J. C. (1977). Rational behavior and bargaining equilibrium in games and social situa- tions. Cambridge University Press. Harsanyi, J. C., & Selten, R. (1972). A generalized Nash solution for two-person bargaining games with incomplete information. Management Science, 18(5), 80–106. 415 FAT IMA , WOOLDR IDGE , & JENN ING S Hoel, M. (1986). Perfect equilibria in sequential bargaining games with nonlinear utility functions. Scandinavian Journal of Economics, 88(2), 383–400. Horst, R., & Tuy, H. (1996). Global optimazation: Deterministic approaches. Springer. In, Y., & Serrano, R. (2003). Agenda restrictions in multi-issue bargaining (ii): unrestricted agendas. Economics Letters, 79, 325–331. Inderst, R. (2000). Multi-issue bargaining with endogenous agenda. Games and Economic Behav- ior, 30, 64–82. Keeney, R., & Raiffa, H. (1976). Decisions with Multiple Objectives: Preferences and Value Trade- offs. New York: John Wiley. Klein, M., Faratin, P., Sayama, H., & Bar-Yam, Y. (2003). Negotiating complex contracts. IEEE Intelligent Systems, 8(6), 32–38. Kraus, S. (2001). Strategic negotiation in multi-agent environments. The MIT Press, Cambridge, Massachusetts. Kraus, S., Wilkenfeld, J., & Zlotkin, G. (1995). Negotiation under time constraints. Artificial Intelligence Journal, 75(2), 297–345. Kreps, D. M., & Wilson, R. (1982). Sequential equilibrium. Econometrica, 50, 863–894. Lax, D. A., & Sebenius, J. K. (1986). The manager as negotiator: Bargaining for cooperation and competitive gain. The Free Press, New York. Livne, Z. A. (1979). The role of time in negotiation. Ph.D. thesis, Massachusetts Institute of Technology. Lomuscio, A., Wooldridge, M., & Jennings, N. R. (2003). A classification scheme for negotiation in electronic commerce. International Journal of Group Decision and Negotiation, 12(1), 31–56. Ma, C. A., & Manove, M. (1993). Bargaining with deadlines and imperfect player control. Econo- metrica, 61, 1313–1339. Maes, P., Guttman, R., & Moukas, A. (1999). Agents that buy and sell. Communications of the ACM, 42(3), 81–91. Martello, S., & Toth, P. (1990). Knapsack problems: Algorithms and computer implementations. John Wiley and Sons. Chapter 2. Mas-Colell, A., Whinston, M. D., & Green, J. R. (1995). Microeconomic Theory. Oxford University Press. Muthoo, A. (1999). Bargaining Theory with Applications. Cambridge University Press. Neumann, J. V., & Morgenstern, O. (1947). Theory of Games and Economic Behavior. Princeton: Princeton University Press. Osborne, M. J., & Rubinstein, A. (1990). Bargaining and Markets. Academic Press, San Diego, California. Osborne, M. J., & Rubinstein, A. (1994). A Course in Game Theory. The MIT Press. Pollak, R. A. (1976). Interdependent preferences. American Economic Review, 66(3), 309–320. 416 MULT I - I S SUE N EGOT IAT ION W ITH D EADL INE S Pruitt, D. G. (1981). Negotiation Behavior. Academic Press. Raiffa, H. (1982). The Art and Science of Negotiation. Harvard University Press, Cambridge, USA. Rosenschein, J. S., & Zlotkin, G. (1994). Rules of Encounter. MIT Press. Rubinstein, A. (1982). Perfect equilibrium in a bargaining model. Econometrica, 50(1), 97–109. Rubinstein, A. (1985). A bargaining model with incomplete information about time preferences. Econometrica, 53, 1151–1172. Sandholm, T. (2000). Agents in electronic commerce: component technologies for automated nego- tiation and coalition formation.. Autonomous Agents and Multi-Agent Systems, 3(1), 73–96. Sandholm, T., & Vulkan, N. (1999). Bargaining with deadlines. In AAAI-99, pp. 44–51, Orlando, FL. Schelling, T. C. (1956). An essay on bargaining. American Economic Review, 46, 281–306. Schelling, T. C. (1960). The strategy of conflict . Oxford University Press. Sobel, J. (2005). Interdependent preferences and reciprocity. Journal of Economic Literature, XLIII, 392–436. Stahl, I. (1972). Bargaining Theory. Economics Research Institute, Stockholm School of Eco- nomics, Stockholm. van Damme, E. (1983). Refinements of the Nash equilibrium concept . Berlin:Springer-Verlag. Varian, H. R. (2003). Intermediate Microeconomics. W. W. Norton and Company. Young, O. R. (1975). Bargaining: Formal theories of negotiation. Urbana: University of Illinois Press. 417
1710.00709	1	1710	2017-10-02T15:03:00	Synthesising Evolutionarily Stable Normative Systems	[ "cs.MA", "cs.GT" ]	Within the area of multi-agent systems, normative systems are a widely used framework for the coordination of interdependent activities. A crucial problem associated with normative systems is that of synthesising norms that effectively accomplish a coordination task and whose compliance forms a rational choice for the agents within the system. In this work, we introduce a framework for the synthesis of normative systems that effectively coordinate a multi-agent system and whose norms are likely to be adopted by rational agents. Our approach roots in evolutionary game theory. Our framework considers multi-agent systems in which evolutionary forces lead successful norms to prosper and spread within the agent population, while unsuccessful norms are discarded. The outputs of this evolutionary norm synthesis process are normative systems whose compliance forms a rational choice for the agents. We empirically show the effectiveness of our approach through empirical evaluation in a simulated traffic domain.	cs.MA	cs	Auton Agent Multi-Agent Syst manuscript No. (will be inserted by the editor) Synthesising Evolutionarily Stable Normative Systems Javier Morales · Michael Wooldridge · Juan A. Rodr´ıguez-Aguilar · Maite L´opez-S´anchez 7 1 0 2 t c O 2 ] A M . s c [ 1 v 9 0 7 0 0 . 0 1 7 1 : v i X r a Received: date / Accepted: date Abstract Within the area of multi-agent systems, normative systems are a widely used framework for the coordination of interdependent activities. A crucial problem associated with normative systems is that of synthesising norms that effectively accomplish a coordi- nation task and whose compliance forms a rational choice for the agents within the system. In this work, we introduce a framework for the synthesis of normative systems that effec- tively coordinate a multi-agent system and whose norms are likely to be adopted by rational agents. Our approach roots in evolutionary game theory. Our framework considers multi- agent systems in which evolutionary forces lead successful norms to prosper and spread within the agent population, while unsuccessful norms are discarded. The outputs of this evolutionary norm synthesis process are normative systems whose compliance forms a ra- tional choice for the agents. We empirically show the effectiveness of our approach through empirical evaluation in a simulated traffic domain. Keywords Norms · Normative systems · Norm synthesis · Evolutionary algorithm 1 Introduction Within both human and agent societies, normative systems (norms) have been widely studied as mechanisms for coordinating the interplay between autonomous entities [8,25]. Given a society, norms can resolve coordination problems by guiding the decision-making of its individuals, restricting their behaviour once some preconditions are fulfilled. Javier Morales and Michael Wooldridge Department of Computer Science, University of Oxford. Oxford, United Kingdom E-mail: [email protected] E-mail: [email protected] Juan A. Rodr´ıguez-Aguilar Artificial Intelligence Research Institute (IIIA-CSIC). Campus de la UAB, Bellaterra, Spain E-mail: [email protected] Maite L´opez-S´anchez Department of Applied Mathematics and Computer Science, Universitat de Barcelona. Barcelona, Spain E-mail: [email protected] 2 Javier Morales et al. In the literature on norm research, normative systems are typically represented as sets of soft constraints on the behaviour of agents, who can autonomously decide whether or not to comply with them. Often, agents face a choice between norm compliance, which allows them to achieve the social welfare at an individual cost, and infringement, which enables them to achieve better individual results at the cost of jeopardising social welfare. Thus, norm compliance is a concern in the area of normative systems, which could be summarised by means of the following questions: 1. Will the individuals of a society comply with the norms of a given normative system? 2. If not, what type of normative systems will they comply with? How to synthesise them? Accordingly, much work in the literature in norm research has focused on the problem of norm compliance [24,13,6], and particularly on how to synthesise norms that discourage non-compliant behaviour [6,5]. Along these lines, some works like [2,23,28,4,30] have taken inspiration on the framework of evolutionary game theory (EGT) [29] to understand the process whereby societies come to adopt norms. They consider a setting in which agents repeatedly play a game (e.g., the Prisoner's Dilemma [20]) by using different strategies. Strategies that are seen to be successful prosper and spread within the agent society through an evolutionary game theoretic process whereby agents tend to adopt successful strategies with higher probabilities than unsuccessful ones. A norm is regarded as a behavioural regu- larity that emerges within the society: a norm is said to have been established once a major- ity of agents adopt the same strategy, which everyone prefers to conform, on the assumption that everyone else does. Such a strategy has the property that complying with it is a rational choice for the agents, and hence it is said to be evolutionarily stable (ESS): once the agents adopt it, no agent can benefit from deviating. Although EGT has been proven to be useful to predict which norms can be evolutionarily stable, most of the works on EGT and norms make strong assumptions that are inconvenient when synthesising norms for multi-agent systems (MAS). First, they consider that agents play a single game – and hence, a single norm can be synthesised. In fact, agents in a MAS typically engage in a wide variety of interaction situations (games) that may require different norms. As an example, humans have designed different traffic rules to coordinate drivers in a variety of situations (e.g., when entering a junction, or when overtaking a vehicle). Hence, MAS coordination may require employing sets of norms instead of a single one. Second, these works assume a deterministic setting in which the game that the agents can play along with its payoffs are known beforehand. However, some systems may have a certain degree of non-determinism that makes it impossible to assume the outcomes of agents' interactions. For instance, one may not be able to ensure that a car will not have an accident once it stops at a red light – the brakes may fail, or another car may hit it from behind. Against this background, this paper contributes to the state of the art by introducing a framework for the synthesis of evolutionarily stable normative systems (ESNS) for non- deterministic settings. Our framework incorporates ideas from EGT. It carries out an evo- lutionary process whereby the agents of a MAS can ultimately adopt sets of coordination norms that are evolutionarily stable (and hence, whose compliance is rational for them). Our framework assumes that the potential coordination situations and their outcomes are unknown beforehand. Agents are permitted to interact, and our framework discovers these situations at runtime, modelling them as games, and empirically computing their payoffs. Norms that successfully coordinate the agents prosper and spread, and agents ultimately adopt evolutionarily stable norms. We provide empirical evaluation of our framework in a simulated traffic domain, and we show that it can synthesise ESNSs that avoid car collisions. Synthesising Evolutionarily Stable Normative Systems 3 Our framework provides a valuable tool for decision support for policy makers. Given a society, it opens the possibility of: 1. Synthesising an ESNS that successfully accomplishes a coordination task. 2. Predicting whether the agents will comply with a given normative system or not, and if they do not, to anticipate the type of normative systems that they will comply with. The remainder of the paper is organised as follows. Section 2 provides the necessary background to understand EGT. Section 3 describes our framework, whereas Section 4 il- lustrates its empirical evaluation. Section 5 reviews the state of the art in norm synthesis, and Section 6 provides some concluding remarks and outlines possible future research. 2 Background: evolutionary game theory We start by describing the framework of evolutionary game theory (EGT) [29] and the key concept of evolutionarily stable strategy (ESS). EGT combines population ecology with classical game theory. It considers a population of agents that repeatedly engage in strategic pairwise interactions by adopting different (pure) strategies. An ESS is a strategy that, if adopted by a majority of agents, no agent could benefit from using any alternative strategy – namely, the fitness (i.e., the average payoff) of an agent using that strategy is higher than the fitness of any agent using alternative strategies. EGT provides a model of the underlying process whereby strategies change in a popula- tion. It assumes that successful strategies "reproduce", growing in frequency with higher probabilities than less successful strategies. Evolutionarily stable strategies are attractor points of such a natural selection process whereby agents can converge to adopting an ESS. Figure 1 graphically illustrates the EGT model. It considers an initial population of agents Pt that adopt different strategies to play a game (Figure 1.1). First, agents are paired off randomly and play the game (Figure 1.2). Each strategy has a certain fitness that quanti- fies the average payoff to an agent that adopts the strategy to play against other strategists. Strategies are then replicated (Figure 1.3), growing in frequency proportionally to their rel- ative fitness with respect to the average fitness of the population. Then, a new population Pt+1 is generated that reflects the changes in strategy frequencies (Figure 1.4). Such popu- lation is then employed to repeat the process, which ends once the population remains stable between generations (that is, the frequencies of each strategy remain unchanged). Next, we detail the equations employed by EGT to perform strategy replication, also known as the replicator dynamics. 2.1 Replicator dynamics We illustrate the replicator dynamics by making use of the first game analysed by May- nard Smith in [29]: the Hawk-Dove game. In this game, two individuals compete for some resource of value V . The players can adopt one of two different strategies: – Hawk (aggressive behaviour): Fight to get the resource until either getting injured or the opponent backs down. – Dove (collaborative behaviour): Back down if the opponent shows aggressive be- haviour. Share the resource if the opponent shows collaborative behaviour. 4 Javier Morales et al. H H V /2 − C/2 D 0 D V V /2 Fig. 1: EGT model, composed of four phases: (1) a population Pt is generated in which each mass of strategists has a certain size; (2) agents are randomly paired and play the game; (3) each mass of strategists grows in terms of their fitness when playing against different strate- gists; (4) a new population Pt+1 is generated in which each mass of strategists has grown in numbers proportional to its fitness. Table 1: Payoff matrix of the Hawk-Dove game. When a Hawk (H) meets a Hawk, both get the resource half of the times with an injury cost. When a Hawk meets a Dove (D), the Hawk takes the resource and the Dove takes noth- ing. When a Dove meets a Dove, they share the resource. The payoff matrix of this game is depicted in Table 1. When a Hawk (H) meets a Hawk, both engage in conflict and have 50% chance to get the resource and 50% chance to get injured. Hence, both get payoff V /2 less an injury cost C/2. When a Hawk meets a Dove (D), the Hawk takes the whole resource and the Dove takes nothing (payoffs V and 0, respectively). When two Doves encounter, they equally split the resource (payoff V /2). Consider a population of agents that adopt either strategy H or strategy D. Let F (H ) ∈ [0, 1] be the frequency of Hawk strategists in the population, and F (D) ∈ [0, 1] the fre- quency of Dove strategists. Note that F (H ) + F (D) = 1. Let us denote as ρ(H , D) the payoff to a Hawk when playing against a Dove, and analogously for other strategy pairs. We assume that each strategist has an initial fitness f0. The fitness of each strategy will depend on: (1) the payoff to an agent when encountering either a Hawk or a Dove, and (2) the prob- ability to encounter each one of these, which actually is a representation of the frequency of strategists of each type. Then, the fitness f of each strategy can be computed as: f (H ) = f0(H ) + F (H ) · ρ(H , H ) + F (D) · ρ(H , D) f (D) = f0(D) + F (H ) · ρ(D, H ) + F (D) · ρ(D, D) (1) (2) f0(H ) and f0(D) being the initial fitness of Hawks and Doves, respectively. In this manner, the fitness of a Hawk is computed as the summation of its initial fitness, the probability of encountering a Hawk times the payoff to the Hawk when that happens, and the probability of encountering a Dove times the payoff to the Hawk when that happens. The fitness of a Dove is computed analogously. Agents reproduce in numbers proportional to their fitnesses. In the next generation, the frequency of Hawks and Doves is updated in terms of their relative fitnesses with respect to the average fitness of the Hawk-Dove population. Then, if Hawks perform better than average they will grow in frequency, and if Doves perform worse than average they will decrease in frequency. Formally, the frequencies of Hawks and Doves are updated as: = F (H ) + F (H ) ·(cid:2)f (H ) − θ(cid:3) = F (D) + F (D) ·(cid:2)f (D) − θ(cid:3) (cid:48) (cid:48) F (H ) F (D) where θ is the weighted average fitness of the Hawk-Dove population, computed as: θ = F (H ) · f (H ) + F (D) · f (D) In biology, replication models the natural process whereby the fittest individuals are more likely to survive and to reproduce than less fit ones. In economic settings (such as (3) (4) (5) Population PtPopulation Pt+1Interaction(game playing)Replication1234Synthesising Evolutionarily Stable Normative Systems 5 multi-agent systems), replication provides a model of imitation [14,7] whereby the agents tend to imitate strategists that appear to perform well, i.e., have a higher fitness, thereby adopting their strategies over time. Then, if a strategy is fitter than the average, agents will be more likely to adopt it than to adopt a less fit one. As previously mentioned, the replication process can eventually lead the population to a point of equilibrium in which the frequencies of each strategy do not change over time because their fitnesses are equal. When this happens, the population can be either monomor- phic (a majority of agents adopt the same strategy) or polymorphic (the agents adopt a variety of strategies). If the population composition can be restored after a disturbance1, then it is said that the population is in an evolutionarily stable state. If such population is monomor- phic, then it is said that the strategy adopted by its agents is an ESS. Next, we detail the conditions for a strategy to be evolutionarily stable. 2.2 Evolutionarily stable strategies A strategy is evolutionarily stable if, once a majority of agents adopt it, their fitness is higher than that of any possible mutant strategist. Otherwise, agents may be tempted to switch to alternative strategies, and the strategy would be unstable. As an example, consider a popula- tion composed mainly of Hawks with a small proportion of Doves. That is, F (H ) (cid:39) 1 and F (D) (cid:28) 1. If Hawk is an evolutionarily stable strategy, then it must satisfy that either: 1. Hawk is a best response to itself. That is, Hawks must perform better than Doves when playing against Hawks. or 2. Hawk is a best response to Dove. In other words, Hawk is not necessarily a best re- sponse to itself, but Hawks must perform better against Doves than they perform against themselves. Formally, this amounts to satisfying either condition 6 or condition 7 below. ρ(H , H ) > ρ(D, H ) ρ(H , H ) = ρ(D, H ) and ρ(H , D) > ρ(D, D) (6) (7) It is obvious that Dove is not an ESS, since a population of Doves can be invaded by a Hawk mutant. The only evolutionarily strategy is Hawk, as long as the injury cost C is lower than the value of the resource V (so that it is worth getting injured in order to obtain the resource). If the injury cost is greater than the value of the resource, then there is no ESS. 3 Evolutionary norm synthesis In this section we introduce our framework for the synthesis of evolutionarily stable nor- mative systems for non-deterministic settings – hereafter, referred to as our "evolutionary norm synthesis system", or ENSS for shorter. In Section 3.1, we start by providing a gen- eral overview of the ENSS operation. Then, we provide some basic definitions and formally define our problem in Section 3.2. Finally, we describe in detail how the ENSS performs evolutionary norm synthesis in Section 3.3. 1 Provided that the disturbance is not too large. For example, a small number of mutant strategists joins the scenario, and after some time they are "eliminated" by dominant strategists. 6 Javier Morales et al. 3.1 Evolutionary norm synthesis system (ENSS) Our framework is intended to synthesise norms that achieve coordination in a MAS and are evolutionarily stable. With this aim, we will assume that our system considers no previous knowledge about the potential interactions of the agents, neither about their outcomes. In- stead, it will learn these situations from the observation of the agents' activities at runtime, and will synthesise norms to coordinate them. Likewise EGT (Section 2), our system will endow each norm with: (i) a fitness value that quantifies its utility to coordinate the agents in a coordination setting; and (ii) a frequency value that stands for the proportion of agents that have adopted the norm. The ENSS enacts an evolutionary process whereby the fittest norms prosper and increase in frequency, and the less fit ones are ultimately eliminated. Eventually, the agents converge to adopting a set of norms that are evolutionarily stable. Figure 2 illustrates the evolutionary process implemented by our ENSS. It starts with an initial agent population Pt whose agents employ no norms to coordinate. Then, it repeatedly performs the following tasks: – Game recognition. The ENSS identifies new situations that require coordination by ob- serving agents' interactions, and keeps track of them as games in a Games Base (GB). For each recognised game, the ENSS creates norms prescribing different coordination strategies, and sends each norm to different agents. The result will be a heterogeneous population whose agents use different strategies to coordinate. – Payoff learning. Our ENSS continuously monitors the agents' game play in order to cumulate evidence about the performance (the payoffs) of each norm in terms of the frequency with which it successfully coordinates the agents. – Norm replication. The ENSS computes norms' fitnesses based on their average payoffs up to a given time, and replicates norms in numbers proportional to their fitness: the frequency of those norms fitter than average will increase, while that of those norms less fit than average will decrease. The output will be a new population Pt+1 in which the size of the set of agents adopting each norm is proportional to its frequency. The ENSS will repeatedly perform these tasks until either it converges, or the MAS stops running. We say that the ENSS has converged to a stable population once the frequency of each norm remains unchanged from population Pt to population Pt+1 for a given num- ber of iterations I. Upon convergence, if a large majority of agents have adopted the same normative system, then we say that their normative system is evolutionarily stable. Let us illustrate the operation of our ENSS with an example. Consider a traffic scenario where agents are cars, and the coordination task is to ensure that cars reach their destinations as soon as possible without colliding. The actions available to the cars are "go" forward and "stop". Figure 3a depicts a junction at time t with three cars. Of these, cars 1 and 2 (circled in red) require coordination in order to avoid colliding. Figure 3b illustrates a collision between these cars at time t + 1 after both have performed action "go". The ENSS will detect such a coordination situation, and will model it as a one-shot game (labelled as Ga in Figure 4) with two roles: a car on the left (role 1) and a car on the right (role 2), both perceiving each other. The ENSS will create the following norms for Ga, which are listed in Table 3: – Norm n1 establishes no prohibitions, and hence a car is free to go forward (not giving way) when coming from either the left or the right (no matter the role it plays). – Norm n2 says that a car is prohibited to go forward when coming from the left (when playing role 1). In practice, n2 stands for a "give way to the right" norm. Analogously, norm n3 stands for a "give way to the left" norm. Synthesising Evolutionarily Stable Normative Systems 7 Fig. 2: Graphical representation of our evolutionary norm synthesis system (ENSS). It starts in (1), with an initial population of agents Pt; (2) agents interact for a given period of time, during which the ENSS identifies the coordination situations (games), and creates their corresponding norms; (3) the ENSS empirically computes norms' payoffs; (4) norms are replicated, growing in numbers proportional to their fitness; (5) a new population Pt+1 is generated in which the size of the set of agents adopting each norm is proportional to its frequency. Such new population is employed to perform a new iteration of the process. Fig. 3: (a) A junction at time t, with cars 1 and 2 playing a 2-role game; (b) At time t + 1, cars 1 and 2 collide after both perform action "go"; (c) At time t + 2, cars 3 and 4 play again the same game, this time by having norms to coordinate; (d) At time t + 3, cars 3 and 4 have avoided a collision. Also, car 6 plays a 1-role game in which it has to decide whether to go forward, or to stop to keep a security distance with car 5. – Norm n4 stands for a "give way always" norm, i.e., it prohibits a car to go forward once it is playing either role 1 or role 2. The ENSS will deliver each one of these norms to different cars. For instance, it will deliver norm n1 to 25% of the cars, and the same applies to norms n2, n3 and n4. Thereafter, the ENSS will detect once the cars play game Ga, and will monitor their outcome in order to compute the payoffs of each norm. For instance, at time t + 2 (Figure 3c) two new cars play Ga (car 3 plays role 1, and car 4 plays role 2). Suppose that both cars have and apply norm n3 (represented as a thought bubble containing n3). Thus, at time t + 3 (Figure 3d) car 3 goes forward and car 4 stops, avoiding collisions. The ENSS will monitor this positive outcome, increasing the payoff of norm n3 whenever two cars jointly use it to coordinate. Also, at time t + 3 the ENSS will detect a new one-role game played by car 6 (labelled as Gb), which can stop, keeping a security distance with car 5, or to go forward, possibly colliding with car 5 in case it stops. The ENSS will create new norms for this game, and will deliver them to the cars. In this way, each car will incrementally build a personal set of norms aimed to coordinate in different games. Population Pt1Game recognition2Norm replication4Population Pt+15Games base (GB)gamesgamesPayoff learning3gamestime t(a)car 2car1timet+1(b)time t+2(c)gamecar 1: gocar 2: gocar 3: gocar 3: gocar 4: goGacar 3car 2car 1car 3car 4car 3car 4Gacar 5time t+3(d)gamecar 3car 4car 5car 3: gocar 4: stopcar 5: gocar 6Gbn3n38 Javier Morales et al. role 2 (right car) stop go 1,0.7 0,0 0.7,1 0.7,0.7 role 1 (left car) go stop Norm role 1 Prohibitions role 2 n1 n2 n3 n4 - go - go - - go go Fig. 4: 2-role game played by cars 1 and 2 in Figure 3a. Table 2: Rewards for a given role in game Ga (depicted in Figure 4) given each pos- sible outcome of the game. Table 3: Different norms to co- ordinate any two cars that play game Ga in Figure 4. 3.2 Basic definitions and problem statement We model a multi-agent system (MAS) as a state transition system with a finite set of agents Ag, and a finite set of actions Ac available to these agents. Let S be the set of the states of the system. We consider a language L (with an entailment relation =) to describe what is true and false in each state. We adopt a synchronous model in which agents interact in some state, perform a collection of actions, and lead the MAS from its previous state to a new one. In each MAS state, agents may engage in strategic interactions in which they need co- ordination in order to avoid undesirable outcomes. Hereafter, we will refer to undesirable outcomes as conflicts. We will model coordination situations as one-shot games. Each game will have a context that describes the situation of each agent playing the game before acting. Each one of these agents will take on one role, which defines the part she plays in the game, and the actions she can perform. Hereafter, we will refer to an agent taking on role i as player i. Once the players of a game select a joint action, each player receives a reward that quantifies to what extent she has successfully avoided conflicts. Definition 1 (m-role game) An m-role game is a tuple G = (cid:104)ϕ, R, A,R(cid:105), where: – ϕ is an expression of L describing the starting conditions (context) of the game. – R = {1, . . . , m} is a set of m agent roles, one per agent playing the game. – A = (cid:104)A1, . . . , Am(cid:105) is an m-tuple of action sets available to each role, where Ai ⊆ Ac i=1 Ai → R≥0, – R = (cid:104)r1, . . . , rm(cid:105) is an m-tuple of reward functions of the form ri :(cid:81)m is the set of actions available to the agent taking on role i. each one returning a reward to player i after all players draw a joint action a.2 For example, we will define game Ga in Figure 4 as a 2-role game where role 1 repre- sents a car on the left, and role 2 a car on the right. Both roles have available actions "go" and "stop". This game will be a tuple Ga = (cid:104)ϕ, R, A,R(cid:105), where its context ϕ can be for- mally interpreted as: "there is a car playing role 1 that perceives a car playing role 2 on its right, and vice versa", R = {1, 2} is the set of roles, A = (cid:104){go, stop},{go, stop}(cid:105) is the set of action spaces of each role, and R contains the reward functions of Table 2. Specifi- cally, two cars that jointly go forward will collide and get reward 0 each. Once at least one of the cars stops, both will avoid colliding. In that case, a car will get reward 1 if it is able to progress (to move forward), and reward 0.7 otherwise. In general, in each state of the MAS the same m-role game can be simultaneously played by different groups of agents. We assume that each agent playing an m-role game in a given MAS state has available a function role that returns the role she plays in that game. 2 a = (cid:104)a1, . . . , am(cid:105) stands for a vector of actions, one for each player. Game Garole 1(left car)role 2(right car)Synthesising Evolutionarily Stable Normative Systems 9 Given an m-role game, a norm stands for a coordination strategy that specifies what an agent is prohibited to do when playing each possible role. For example, each norm in Table 3 establishes different prohibitions for any car playing either role 1 or 2 in game Ga. Formally, a norm is a (possibly empty) set of constraints that restricts the action space of each role of a game by prohibiting certain actions. Definition 2 (Norm) Given an m-role game G = (cid:104)ϕ, R, A,R(cid:105), a norm to coordinate the agents in G is a pair (cid:104)ψ, prh(cid:105), where: – ψ ∈ L is the precondition of the norm; and – prh : R → 2Ac is a function that returns the set of actions that an agent is prohibited to perform when taking on role i, where prh(i) ∈ 2Ai for all i ∈ R. Let G = (cid:104)ϕ, R, A,R(cid:105) be an m-role game, and n = (cid:104)ψ, prh(cid:105) a norm. We say that n applies in G if the precondition of n satisfies the context of G, namely if ϕ = ψ. Hereafter, we will refer as the set of norms that apply in a game G as the norm space of the game, and we will denote it by NG. For instance, the norm space of game Ga (Figure 4) is NGa = {n1, n2, n3, n4} (see Table 3). Agents in a MAS may play multiple, different m-role games. Henceforth, we shall de- note the set of games that agents can play as G = {G1, . . . , Gs}. A normative system is a set of G norms that provides an agent with the means to coordinate in each possible game in G. Following our example, each car will have one norm out of the norm space NGa to coordinate in Ga, one norm out of NGb to coordinate in Gb, and so on for each game. Definition 3 (Normative system) Let G be a set of m-role games. A normative system is a set of norms Ω such that for each game G ∈ G there is one and only one norm n ∈ Ω and n ∈ NG. First of all, each agent in a MAS agi ∈ Ag counts on her own normative system Ωi. Thus, in general we assume that a MAS is composed of a heterogeneous population whose agents may have different normative systems. Let Ag(cid:48) ⊆ Ag be a group of agents playing an m-role game G = (cid:104)ϕ, R, A,R(cid:105) at a given time t. Each agent counts on one and only one norm out of her normative system that applies in the game and prohibits her to perform some actions. We consider: – an injective function π : R → Ag that maps each role in R to one agent in Ag(cid:48), namely – a function η : Ag × G → NG1 ∪ . . . ∪ NGs that given an agent agi and a game G, tells to the agent enacting that role in G at time t; and us the norm in the normative system of the agent, Ωi, to apply in G. This allows us to define n = (cid:104)η(π(1), G), . . . , η(π(m), G)(cid:105) as the combination of norms that the normative systems of the agents in Ag(cid:48) prescribe them to apply in G at time t. Notice that η(π(i), G) stands for the norm for game G in the normative system of the agent playing role i. We assume that the agents always comply with the prohibitions prescribed by their norms. Therefore, based on the norms in n, the agents in the game will perform a tuple of actions a = (cid:104)a1, . . . , am(cid:105), where ai is an action that is not prohibited by norm η(π(i), G) for role i.3 After the agents perform a joint action a, each player i obtains a reward ri(a). 3 In principle, given a combination of norms n applicable to a group of agents, it is not possible to assume beforehand the joint action a that these agents will perform. However, we assume that the actions in a will comply with the prohibitions established by their respective norms in n. 10 Javier Morales et al. Let us illustrate these definitions with cars 3 and 4 in Figure 3c, which play game Ga at time t+2 by enacting roles 1 and 2, respectively. Say that these cars have normative systems Ω3 and Ω4, respectively, and that both normative systems have n3 as the applicable norm in game Ga. That is, η(3, G) = n3 and η(4, G) = n3. Thus, these cars will play with norm combination n = (cid:104)n3, n3(cid:105), which means that car 4 will be prohibited to go forward (and hence will stop). In practice, these cars will perform a joint action a = (cid:104)go, stop(cid:105). Then, the reward to each role i ∈ {1, 2} at time t + 3 can be computed as ri((cid:104)go, stop(cid:105)). Note therefore that the reward that an agent can expect obtain from a game depends on the norm that she uses to play the game, the role that she plays, and the norms that the other players use. Thus, at a given time an agent may get a high reward once she plays against agents with a particular combination of applicable norms, and at a different time she may get a low reward when playing against agents with a different norm combination. In practice, given an m-role game and the norms that apply in it, the agents will play an infinitely repeated one-shot game of norms against norms, namely a norms game, in which the norms used by the agents to play the game over time will lead them to obtain a history of rewards. Thus, a norms game will consist of an m-role game, the norm space of the game, and a memory that contains the history of rewards obtained by the agents once they repeatedly play the m-role game over time. Definition 4 (Norms game) A norms game is a tuple N G = (cid:104)G, NG, H, P(cid:105), where: – G = (cid:104)ϕ, R, A,R(cid:105) is an m-role game. – NG is the norm space of G, namely the set of norms that apply in G. – H = (cid:104)h0, . . . , hω(cid:105) is the memory of the game over a time window [t0, tω], where hj = (cid:104)nj, rj(cid:105) such that nj is the norm combination used by the agents to play G at time tj, and rj is the vector of rewards they obtained at time tj (one for each player). i=1 NG × H → R≥0, which return the expected payoff to player i based on the memory of the game H and a combination of norms n ∈ N Intuitively, the expected payoff of a combination of norms n tells us how successful that norm combination has been historically to coordinate the players of the game. Such a payoff is computed based on the rewards obtained by the players of the game within a time window. Further on, we provide an equation to compute the expected payoff in Section 3.3.2. – P = (cid:104)ρ1, . . . , ρm(cid:105) is an m-tuple of payoff functions of the form ρi :(cid:81)R R G applicable to all players. At this point we import from EGT the concept of fitness introduced in Section 2. Thus, given a norms game, the fitness of each one of its norms quantifies the average payoff that an agent can expect to obtain when using the norm to play the game by enacting different roles and by playing against agents with different norms. Formally: Definition 5 (Norm fitness) Given a norms game N G = (cid:104)G, NG, H, P(cid:105), the fitness of a norm n ∈ NG is represented as f (n, N G, t) ∈ R, where f stands for the norm fitness function at a particular point in time t ∈ N. Now we are ready to introduce the problem that we address in this paper. Let us assume a population of rational agents that will tend to adopt fitter norms. Given a norms game N G = (cid:104)G, NG, H, P(cid:105), our aim is to find a norm n ∈ NG such that, once it is used by all the agents to play G, there is no agent that can derive a greater fitness by using any alternative norm n(cid:48) ∈ NG. Then, the rational choice for all the agents will be complying with norm n. In terms of EGT, this amounts to saying that norm n is evolutionarily stable (Section 4.4), since no agent could be ever tempted to use alternative norms to play the game. Thus, given a norms game, our aim is to find a norm n applicable in the game such that n is evolutionarily Synthesising Evolutionarily Stable Normative Systems 11 stable. Likewise, given a collection of norms games, we aim to find a normative system that contains one evolutionarily stable norm for each norms game. Formally, our research problem is as follows. Definition 6 (Norm synthesis problem) Given a set of agents Ag and a set of norms games NG, our aim is to find a normative system Ω such that, from some time tu onwards, the fol- lowing conditions hold: 1. All agents adopt Ω. That is, Ωi = Ω for each agent agi ∈ Ag. 2. There is no norm in the normative system whose fitness is outperformed by that of an alternative norm. Namely, there is no norms game N G = (cid:104)G, NG, H, P(cid:105) ∈ NG and norm n ∈ Ω, n ∈ NG, such that f (n(cid:48), N G, t) > f (n, N G, t) for some alternative norm n(cid:48) ∈ NG and time t ≥ tu. 3.3 Formal model for evolutionary norm synthesis We now describe the tasks that our ENSS performs to synthesise a normative system that solves the norm synthesis problem in Definition 6. That is, game recognition (labelled as 2 in Figure 2), payoff learning (labelled as 3) and norm replication (labelled as 4). These tasks allow the ENSS to detect and abstract as games the coordination situations that the agents might encounter, and to carry out the evolutionary process whereby the agents can ultimately adopt an evolutionarily stable normative system. 3.3.1 Recognising new games from observation Game recognition is achieved by observation of the agents' activities. Agents interact in an environment for a given number of time steps, and the ENSS monitors their activities at regular time intervals. At each time step, the ENSS tries to detect new, untracked coordination situations that it abstracts as new m-role games. With this aim, we take inspiration from the work in [17,18,19], which performs automatic detection of coordination situations by detecting agent interactions that lead to undesirable outcomes (conflicts). Analogously, we consider that any type of m-agent interaction that may lead to conflicts is an m-role game. As in [17,18,19], we assume that conflicts can be detected at runtime, and that the agents involved in a conflict are the ones responsible for the conflict. Moreover, we assume that a conflict at time t is caused by the actions that the agents performed at time t − 1. At time t, the ENSS builds a new m-role game following the next steps: 1. Detect a new conflict at time t. Notice that detecting conflicts requires the use of domain- dependent knowledge to retrieve the conflicts at a given time. Examples of conflicts are: collisions in a traffic scenario. 2. Describe the situation involving the m conflicting agents. This amounts to generating an expression ϕ ∈ L that describes the situation involving these agents in the state prior to the occurrence of the conflict (i.e., at time t − 1). 3. Create a new m-role game to model coordination as G = (cid:104)ϕ, R, A,R(cid:105) with expression ϕ as its context, R as the set of roles played by the agents at time t − 1, A as the set of the actions available to these agents at time t− 1, and R as the set of reward functions.4 4 Note that assessing the rewards for a given game requires the use of domain-dependent knowledge. 12 Javier Morales et al. Then, if the new m-role game G does not exist in the Games Base (see Figure 2), the ENSS will add it in order to be able to detect when the agents play the game. We already illustrated this process in Section 3.2 by creating game Ga = (cid:104)ϕ, R, A,R(cid:105) from the interaction of cars 1 and 2 at time t (Figure 3a). Ga has ϕ ∈ L as its context, R = {1, 2}, A = {(cid:104)go, stop(cid:105),(cid:104)go, stop(cid:105)}, and R as its set of reward functions (Table 2). After creating a new m-role game G, the ENSS will create its norm space NG by: 1. identifying the actions performed by the m conflicting agents in the transition from the state prior to the conflict to the state containing the conflict, i.e., from time t − 1 to t. 2. creating norms prohibiting different roles to perform the conflicting actions in the game, each having ϕ as its precondition, and a combination of prohibitions as its postcondition. After that, the ENSS will create the corresponding norms game N G = (cid:104)G, NG, H, P(cid:105), with the new m-role game G, NG as its norm space, H as its tuple of history functions5, and P as its set of expected payoff functions. Next, the norms of NG are uniformly distributed among the agents in a MAS. This guarantees that the normative systems in an agent popu- lation are heterogeneous, namely the agents will play game N G by using different norms, and hence the ENSS will be able to evaluate which ones do better. Going back to the example of game Ga, the ENSS will now create its norm space NGa by first identifying action "go" as the action performed by the conflicting cars 1 and 2 during the transition from time t to time t + 1 in Figure 3. Then, it will create norms to prohibit to perform action "go" to: none of the roles (n1 in Table 3), role 1 (n2), role 2 (n3), and both roles (n4). The ENSS will now create and track the corresponding norms game N Ga = (cid:104)Ga, NGa , H, P(cid:105). Thereafter, it will deliver each norm to 25% of the agents. 3.3.2 Computing norms' payoffs empirically The ENSS continuously monitors the game play of the agents, detecting when they play each norms game, and keeping track of their rewards in the memory of the game. The ENSS will exploit this knowledge in order to approximate the expected payoffs of the game at a given point in time based on the following principles: 1. Those norms that have allowed the agents obtain high rewards in the past can be ex- pected to yield high rewards in future game plays. 2. The rewards obtained by the agents more recently in the past are more valuable and informative for the payoff computation than older rewards. Thus, the payoffs of a norms game are computed as follows. Let N G = (cid:104)G, NG, H, P(cid:105) be a norms game, and n a combination of norms applicable to a group of agents playing N G at a given time tω. First, we will retrieve from H a tuple (cid:104)r1, . . . , rk(cid:105) with the rewards obtained by the agents in the k times they played N G with norm combination n within a time window [t0, tω]. Then, we will compute the expected payoff to player i as the discounted average reward to role i within this time window: ρi(n, H) = k(cid:80) (cid:124) j=1 1 βk−j (cid:123)(cid:122) (cid:125) normalisation k(cid:88) (cid:124) j=1 i · βk−j (cid:125) (cid:123)(cid:122) rj discounted reward (8) 5 Initially, each history function hi ∈ H will return an empty sequence of rewards. Consequently, each empirical payoff function ρi ∈ P will return an undefined value. Synthesising Evolutionarily Stable Normative Systems 13 where rj i is the j-th reward obtained by player i in N G within a time window [t0, tω]; and β ∈ [0, 1] is a discount factor. Intuitively, the right part of equation 8 computes the weighted summation of the rewards to player i within a time window, where the j-th reward (an older reward) has a lower weight (is more discounted) than the (j + 1)-th reward (a more recent one). The left part of the equation normalises the payoff by dividing the weighted summation by the summation of weights. At this point we recall that our system assumes a non-deterministic setting in which the outcomes of a game, and hence its rewards, cannot be assumed a priori. Instead, the ENSS will assess the rewards for the players of a game by monitoring their outcomes a posteriori, i.e., once they have played the game. Thus, given an m-role game G = (cid:104)ϕ, R, A,R(cid:105) played by a group of agents at a given time t, the ENSS will approximate the reward function ri ∈ R i=1 Ai × N → R≥0, which returns the reward to the agent enacting role i in the game once all players perform a joint action a at time t + 1.6 for player i as an empirical reward function ri :(cid:81)R Back with the example of game Ga, say that at a given time t the cars have played this game three times so far. In the first two game plays, the cars played with norm combination n = (cid:104)n4, n4(cid:105) and performed a joint action a = (cid:104)stop, stop(cid:105), thus obtaining reward 0.7. In the third game play, the cars played with norm combination n = (cid:104)n1, n1(cid:105) and performed a joint action a = (cid:104)go, go(cid:105), thus getting reward 0. Hence, the memory of the game at time t is H = (cid:104)(cid:104)(cid:104)n4, n4(cid:105),(cid:104)0.7, 0.7(cid:105)(cid:105),(cid:104)(cid:104)n4, n4(cid:105),(cid:104)0.7, 0.7(cid:105)(cid:105),(cid:104)(cid:104)n1, n1(cid:105),(cid:104)0, 0(cid:105)(cid:105)(cid:105). Say that we want to compute the payoffs of norm combination (cid:104)n4, n4(cid:105). First, we will retrieve from H the sequence of rewards of (cid:104)n4, n4(cid:105) up to time t, that is, (cid:104)(cid:104)0.7, 0.7(cid:105),(cid:104)0.7, 0.7(cid:105)(cid:105). Note that k = 2, where k is the length of the history of rewards for (cid:104)n4, n4(cid:105). Let us consider a discount factor β = 0.9, which implies that the j-th reward of the sequence is 90% as valuable for the payoff computation as the (j + 1)-th one. We will compute the payoffs of this norm combination for each role i ∈ {1, 2} as: 0.91 + 0.90 ·(cid:104) 1 0.7 · 0.91 + 0.7 · 0.90(cid:105) ρi((cid:104)n4, n4(cid:105), H) = 3.3.3 Replicating norms = 0.7 As previously detailed, norm replication is the process of computing the fitness of each norm (Definition 5 in Section 3.2), and then making its frequency grow proportionally to its fitness. The ENSS computes a norm's fitness similarly to the way a strategy's fitness is computed in EGT (Section 2). Given a norms game, the fitness of a norm n will depend on: 1. the payoff that an agent can expect to obtain when using norm n to play the game against other agents with possibly different applicable norms in the game; and 2. the probability that the agent encounters these agents, which can be computed in terms of the frequencies of the norms applicable to these agents in the game. Intuitively, if an agent obtains a high payoff once she plays a game against agents with a highly frequent norm, then the agent will be very likely to encounter an agent that uses that norm to play that game, and hence to get a high fitness. Conversely, the same agent will very likely get a low fitness if she is highly likely to interact with agents against whom she always gets a low payoff. 6 We provide an example of empirical reward function for our traffic scenario in Section 4.2. 14 Javier Morales et al. As an example, let us consider that a car repeatedly plays game Ga in Figure 4 by using norm n1 in Table 3. According to this norm, the car will never give way. This car will yield a high payoff when playing against cars that have applicable norm n4, since this norm obliges a car to always give way. This occurs when the combination of norms used in the game is either (cid:104)n1, n4(cid:105) (with our car playing role 1), or (cid:104)n4, n1(cid:105) (with our car playing role 2). Conversely, this car will derive a low payoff when it interacts with cars that have n1 (since both will go forward and collide), namely when the combination of norms used in the game is (cid:104)n1, n1(cid:105). Now, say that the number of cars with norm n4 doubles the number of cars with norm n1. Then, our car will be twice as likely to play against cars that have n4, and hence to obtain a higher fitness. Given a norms game N G, we compute the fitness of a norm n at time t as the average payoff ρi to an agent once she uses norm n to play N G, for each role i and each combination of norms applicable to the players of N G. Formally: f (n, N G, t) = ρi(n, H) · p(n, t) (9) R(cid:88) (cid:88) i=1 n∈N R G ni=n where: R – N G is the set of all norm combinations that the agents playing the game can employ; – n is a norm combination and ni = n is the norm employed by the agent playing role i; – ρi(n, H) is the payoff to role i when the agents play with norm combination n, com- puted based on the game's memory H up to time t; and – p(n, t) is the joint frequency of the norms in n in the normative systems of the players. We compute the joint frequency of the norms in n in the normative systems of the players of N G at a given time t as: p(n, t) = F (n, t) = {ag ∈ Ag n ∈ Ωag} Ag (10) (cid:89) n∈n where F (n, t) ∈ [0, 1] is the frequency of norm n at time t, namely the proportion of agents whose normative systems contain n at time t. Next, those norms whose fitness is higher than average fitness will become more fre- quent, while those below average will decrease. This will be captured by a replication equa- tion that we compute as follows. Given a norms game N G, we update the frequency of a norm n ∈ NG as: F (n, t + 1) = F (n, t) + F (n, t) ·(cid:2)f (n, N G, t) − Θ(cid:3) (11) where Θ is the average fitness at time t of all the norms applicable in N G, computed as: Θ = f (n, N G, t) · F (n, t) (12) (cid:88) n∈NG Notice that equations 11 and 12 are the counterparts of 3-4 and 5 introduced in Section 2.1 to describe the replicator dynamics of EGT. As an example, let us compute the fitness of the norms in Table 3 for game Ga. For sim- plicity, let us consider that only norms n1 and n2 are available to the cars. At the outset, half of the cars have n1 in their normative systems, while the rest of cars have n2. Thus, it fol- lows that the cars can employ the following norm combinations to play the game: (cid:104)n1, n1(cid:105), Synthesising Evolutionarily Stable Normative Systems 15 role 2 (right car) n2 n1 0,0 0,0 0.7,1 0.7,1 role 1 (left car) n1 n2 Table 4: Payoffs computed for norms n1 and n2 after the agents have repeatedly played game Ga in Figure 4 for a sufficient amount of time. Once the cars use n1 to coordinate, they go forward and collide, getting payoff 0. Only when the car on the left (the one enacting role 1) applies norm n2 (hence stopping) will the cars avoid colliding. In that case, the left car gets payoff 0.7, and the right car gets payoff 1. (cid:104)n1, n2(cid:105), (cid:104)n2, n1(cid:105), and (cid:104)n2, n2(cid:105). The joint probability of each of these combinations is 0.25 (e.g., p((cid:104)n1, n1(cid:105), t) = F (n1, t) · F (n1, t) = 0.5 · 0.5 = 0.25). Also, let us consider that at time t our system computes the payoff matrix illustrated in Table 4 by means of equation 8 (Section 3.3.2) and based on a memory H of the norms game NGa. Then, we can compute the fitness of norm n1 by using equation 9 as follows: f (n1, Ga, t) = ρ1((cid:104)n1, n1(cid:105), H) · p((cid:104)n1, n1(cid:105), t) + ρ1((cid:104)n1, n2(cid:105), H) · p((cid:104)n1, n2(cid:105), t)+ ρ2((cid:104)n1, n1(cid:105), H) · p((cid:104)n1, n1(cid:105), t) + ρ2((cid:104)n2, n1(cid:105), H) · p((cid:104)n2, n1(cid:105), t) = 0 · 0.25 + 0 · 0.25 + 0 · 0.25 + 1 · 0.25 = 0.25 Analogously, we compute the fitness of norm n2 as follows: f (n2, Ga, t) = ρ1((cid:104)n2, n1(cid:105), H) · p((cid:104)n2, n1(cid:105), t) + ρ1((cid:104)n2, n2(cid:105), H) · p((cid:104)n2, n2(cid:105), t)+ ρ2((cid:104)n1, n2(cid:105), H) · p((cid:104)n2, n1(cid:105), t) + ρ2((cid:104)n2, n2(cid:105), H) · p((cid:104)n2, n2(cid:105), t) = 0.7 · 0.25 + 0.7 · 0.25 + 0 · 0.25 + 1 · 0.25 = 0.6 Note therefore that norm n2 is more than twice as fit as norm n1. Now, let us replicate both norms. We compute the average fitness of norms n1 and n2 using equation 12 as: Θ = f (n1, Ga) · F (n1) + f (n2, Ga) · F (n2) = 0.25 · 0.5 + 0.6 · 0.5 = 0.425 Since norm n2's fitness is larger than the average, its frequency in the next generation must increase, while that of n1 must decrease. Specifically: F (n1, t + 1) = F (n1, t) + F (n1, t) ·(cid:2)f (n1, Ga, t) − Θ(cid:3) = 0.5 + 0.5 ·(cid:2)0.25 − 0.425(cid:3) = 0.4125 F (n2, t + 1) = F (n2, t) + F (n2, t) ·(cid:2)f (n2, Ga, t) − Θ(cid:3) = 0.5 + 0.5 ·(cid:2)0.6 − 0.425(cid:3) = 0.5875 Hence, at time t + 1, approximately 59% of the agents will adopt n2 in their normative systems, which means that norm n2 will spread. The remaining 41% of the agents will adopt n1, and hence the presence of n1 in the agents' normative systems will shrink. 16 Javier Morales et al. 3.3.4 Evolutionarily stable normative systems At this point we are ready to provide the stability conditions for norms and normative sys- tems. Analogously to the definition of ESS (see Section 2.2), we say that a norm is evolu- tionarily stable if, once a majority of agents use it to coordinate in a game (that is, a majority of agents have the norm in their normative systems), there is no alternative norm applicable in the game with which the agents can derive a higher fitness. Consider a MAS such that the majority of its agents have norm n ∈ NG in their normative systems to employ when play- ing a norms game N G. Therefore, notice that the frequency of n in the agents' normative systems is close to 1, namely F (n, t) (cid:39) 1. We say that n is evolutionarily stable when it satisfies one of the following conditions: 1. It is a best response to itself. Once a whole group of agents play N G by having norm n applicable, no agent can get a higher payoff by using an alternative norm n(cid:48) (such that n (cid:54)= n(cid:48)) to play the game. 2. It is a best response to any alternative norm. Considering that condition 1 above does not hold, an agent that plays N G against a group of agents that have any alternative norm n(cid:48) ∈ NG (such that n (cid:54)= n(cid:48)) will derive a higher payoff by using norm n to play the game against these agents than by using norm n(cid:48). Next, we provide our formal definition of evolutionarily stable norm. Definition 7 (Evolutionarily stable norm) Let N G = (cid:104)G, NG, H, P(cid:105) be a norms game with a memory H at a given time t, and n ∈ NG a norm applicable in G. We say that n is evolutionarily stable at time t if, once a majority of agents have n in their normative systems at time t, namely once F (n, t) (cid:39) 1, for any alternative norm n(cid:48) ∈ NG, n (cid:54)= n(cid:48), one of the two following conditions holds: ρi((cid:104)n1, . . . , nm(cid:105), H) > ρi((cid:104)n1, . . . , n ρi((cid:104)n1, . . . , nm(cid:105), H) = ρi((cid:104)n1, . . . , n ρi((cid:104)n (cid:48) (cid:48) m(cid:105), H) < ρi((cid:104)n 1, . . . , ni, . . . , n (cid:48) 1, . . . , n (cid:48) i, . . . , nm(cid:105), H) for all i ∈ R (cid:48) i, . . . , nm(cid:105), H) and (cid:48) m(cid:105), H) for all i ∈ R (13) (14) where ni stands for the norm applicable to the agent taking on role i, and n(cid:48) norm applicable to the agent playing role i. i is an alternative Condition 13 means that n is a best response to itself, whereas condition 14 means that n is not necessarily a best response to itself, but it is a best response to any alternative norm n(cid:48). From this definition follows that an evolutionarily stable norm will be a best choice for coordination, since there will be no alternative norm that yields more expected fitness. Given a set of norms games, we say that a normative system is evolutionarily stable iff all its norms are evolutionarily stable. Formally: Definition 8 (Evolutionarily stable normative system) Let NG be a set of norms games and a MAS whose agents are Ag. We say that a normative system Ω is evolutionarily stable at time t iff: (i) for each norms game N G ∈ NG there is one and only one norm n ∈ Ω and n ∈ NG such that n is evolutionary stable from time t; and (ii) Ω is the normative system of all the agents in Ag, namely Ωi = Ω for all agi ∈ Ag. Then, if all the agents in a MAS count on the very same normative system, and this is evolutionary stable, no agent will be able to obtain a larger fitness by switching to any alter- native normative system. Therefore, complying with its norms will form a rational choice for the agents. Synthesising Evolutionarily Stable Normative Systems 17 Fig. 5: (a) Simulated traffic scenario with four entry points (labelled as in), and four exit points (labelled as out); (b) Graphical description of a 2-role game. Cells a, b, c and d describe the four positions in front of the car playing role 2 (the red car at cell e). Cells a, c, e and d describe the four positions in front of the car playing role 1 (the yellow car at cell b). 4 Empirical analysis and results In this section we empirically evaluate our approach in a simulated traffic scenario. We explore several dimensions. First, we analyse its convergence, showing that it manages to converge 100% of times to ESNSs that avoid collisions as far as cars give a sufficiently high importance to avoiding collisions, namely as far as they are sufficiently averse to colliding. We also test the adaptivity of our approach, namely its capability to adapt the normative systems it synthesises to the degree of collision aversion of the car population. Finally, we study the stability of the normative systems synthesised by our approach upon convergence. We demonstrate that, once all cars abide by an ESNS synthesised by our system, there is no alternative normative system that the agents can be tempted to switch to. 4.1 Empirical settings Our experiments consider a discrete simulator of a traffic scenario in which agents are au- tonomous cars, and the coordination task is to ensure that cars reach their destinations as soon as possible without colliding. Figure 5a illustrates an example of our scenario, com- posed of two orthogonal roads represented by a 7× 7 grid. At each tick, new cars may enter the scenario from four different entry points (labelled as "in"), and travel towards one of four exit points (labelled as "out"). Each car has a limited perception of the scenario and can perceive the four cells in front of it (one cell on its left, two consecutive cells in front, and one cell on its right). Each m-role game is described by means of the contents perceived by its players, namely by specifying the four cells in front of each car playing the game. Figure 5b graphically illustrates the description of a 2-role game played by two of the cars in Figure 5a. Specifically, cells a, b, c and d describe the four positions in front of the car playing role 2 (the one in cell e), and cells a, c, e and d describe the four positions in front of the car playing role 1 (the one in cell b). Each experiment consists of a set of simulations that start with a population of agents that have no norms to coordinate (that is, each agent agj ∈ Ag has an empty normative system Ωj = ∅). Simulations run in rounds of 200 ticks. In each round, cars interact in the ininininoutoutoutout(a)(b)abcderole 1(left car)role 2(right car)18 Javier Morales et al. junction and collisions occur as the simulation goes on7. Our ENSS continuously monitors the system, and captures these coordination situations as m-role games, creating norms that the cars incorporate to their normative systems (see Section 3.3.1). Over time, the ENSS computes the payoffs and the fitnesses of each norm, and evolves norms based on their fitness. We consider that the system has converged whenever the frequency of each norm has not changed during the last 30 rounds (I = 30). Upon convergence, we consider that a norm n is evolutionarily stable if all the agents have adopted it (that is, if F (n, t) = 1). Moreover, if all the cars have adopted the same set of norms and these are evolutionarily stable, then we say that they have converged to an evolutionarily stable normative system. 4.2 Convergence analysis We first analyse the capability of our approach to synthesise an ESNS that successfully coordinates the cars in avoiding collisions as far as they are sufficiently willing to avoid collisions (namely, as far as they are sufficiently averse to colliding). With this aim, we run 1,000 simulations that consider the following empirical reward function (see Section 3.3.2):  0 ri(a, t) = if player i collides at time t 0.7 if player i avoids collisions but cannot move forward at time t 1 if player i avoids collisions and can move forward at time t where a is a joint action performed by the players of an m-role game G at a given time t. Thus, a car gets the worst possible reward (reward 0) once it plays a game at time t − 1 and collides at time t. The best possible reward (reward 1) it gets once avoids collisions and can go forward (hence not delaying). Finally, a car gets a less positive reward (reward 0.7) when it has to stop in order to not collide (which, on the other hand, is detrimental to the goal of reaching its destination as soon as possible)8. Note that the two rewards for not colliding are significantly higher than the reward for colliding. Thus, cars will give a higher importance to avoiding collisions at the expense of travelling time. In other words, we say that the cars will be highly averse to colliding. Finally, we consider a discount factor of 0.8 (β = 0.8) to compute empirical payoffs in Equation 8 (Section 3.3.2).9 Out of 1,000 simulations, the ENSS takes an average of 54 rounds to converge. During that time, it generates 64 different games that can be grouped into the four categories illus- trated in Figure 6. The first category (Figure 6a), which we call single-stop games (SSG), stands for 2-role games in which the best strategy to avoid collisions and to delay as little as possible is that one of the cars stops, giving way to the other. Two examples of SSG are the one illustrated in Figure 6a, and the one depicted in Figure 5b, which is very similar to the former one but also considers a third and a fourth car in cells a and d. In general, any variation in cells a, c and d of a 2-role game is considered as a different game10. The second category, which we call double-stop games (DSG), stands for 2-role games in which both cars need to stop in order to avoid collisions (at the expense of extra travel 7 Whenever two or more cars collide, they remain in the scenario for 5 ticks until they are removed. With this aim we aim to simulate the time that the emergency services require to move away collided cars. 8 Note that with this reward function the system will learn payoff matrices that are similar to the one considered in our example game of Table 2 in Section 3. 9 We also performed simulations with discount factors that ranged from 1 to 0.1, obtaining similar results. 10 This is due to the assumed non-determinism of the MAS, whereby our system cannot assume that two similar situations will stand for similar games. Synthesising Evolutionarily Stable Normative Systems 19 Fig. 6: Different game types detected in our simulations: (a) a 2-role single-stop game (SSG), in which the best strategy is that one of the cars stops, giving way to the other; (b) a 2-role double-stop game (DSG), in which both cars have to stop in order to not collide; (c) a 1-role prevention game (PG), in which a car (the one at cell e) can stop for one tick, thus keeping a security distance with the car in front (the one at cell c), or to go forward, assuming the risk of colliding in case the car in front stops; (d) a 1-role traffic-jam game (TJG), in which a car (the one at cell e) has to stop in a traffic jam in order to not collide with the car in front. time). Figure 6b shows an example of DSG, in which two cars are waiting for a collision to be moved away. The third category (Figure 6c), called prevention games (PG), stands for 1-role games in which a car can go forward, assuming a collision risk in case the car in front stops, or to stop for one time step in order to keep a security distance. The fourth category (Figure 6d) we call traffic-jam games (TJG). They are similar to prevention games, but in this case a car is stuck in a jam and it can decide whether to stop, doing the queue, or to go forward, colliding with the car in front. Overall, the ENSS detects 12 different single-stop games, 7 double-stop games, 40 pre- vention games, and 5 different traffic jam games. Table 5 illustrates the six possible types of strategies to coordinate cars in each type of game, along with the average number of simu- lations that converged to norms that prescribed these strategies. The first four strategies are aimed to regulate 2-role games (SSGs and DSGs)11. They state that a car has to: "never give way" when playing either role 1 or 2 (no prohibitions imposed); "give way to the right", i.e., it is prohibited to go when playing role 1 ({(cid:104)1, go(cid:105)}); "give way to the left" ({(cid:104)2, go(cid:105)}); and "give way always" ({(cid:104)1, go(cid:105),(cid:104)2, go(cid:105)}). The last two strategies regulate 1-role games (PGs and TJGs). Strategy "go" ({}) says that a car is free to go forward, and strategy "stop" ({(cid:104)1, go(cid:105)}) says that a car is prohibited to go (either to keep a security distance with the car in front, or to do a queue). As a matter of fact, strategy "stop" is the only one that can avoid 100% of collisions in both PGs and TJGs. On average, cars converge to adopting an evolutionarily stable norm that avoids colli- sions in 100% of SSGs and DSGs. Specifically, in SSGs cars adopt 49% of times a norm to "give way to the right", and the remaining 51% of times they adopt a "give way to the left" norm. As for DSGs, cars adopt a norm prescribing a "give way always" strategy 100% of times. In PGs, cars adopt 90% of times a norm prescribing a "stop" strategy in order to keep a security distance, and the remaining 10% of times they adopt the "go" norm (hence assuming the risk of colliding). This happens because in PGs, the car in front (e.g., the car in cell e in Figure 6c) does not always stop, and hence the collision risk of proceeding is lower than 100%. Since the reward for going forward and not colliding (reward 1) is higher than that of stopping and not colliding (reward 0.7), sometimes cars prefer to assume the collision risk and adopting a "go" norm in order to save travel time. Conversely, in TJGs cars adopt 11 For instance, norms n1, n2, n3, n4 in Table 3 (Section 3.3) prescribe these strategies to regulate Ga. role 1(left car)role 1(car behind)(a)(b)(c)(d)Single stop game (SSG)Double stop game (DSG)Prevention game (PG)Traffic jam game (TJG)role 2(right car)role 1(left car)role 2(right car)role 1(car behind)20 SSG DSG PG TJG Javier Morales et al. "never give way" {} 0% 0% - - "give way to the right" {(cid:104)1, go(cid:105)} 49% 0% - - Norms "give way to the left" {(cid:104)2, go(cid:105)} 51% 0% - - "give way always" {(cid:104)1, go(cid:105), (cid:104)2, go(cid:105)} 0% 100% - - "go" {} - - 90% 0% "stop" {(cid:104)1, go(cid:105)} - - 10% 100% Table 5: Different types of norms to coordinate the cars in each type of game, along with the average number of times that cars converged to adopting each type of norm. The first four norms regulate SSGs and DSGs (2-role games). The last two norms regulate PGs and TJGs (1-role games). Fig. 7: Evolutionary dynamics of norm adoption in SSGs (a) and DSGs (b). Each square represents the possible frequency distributions of the first four norms in Table 5, which establish that a car has to: "never give way", "give way to the right", "give way to the left", and "give way always". Arrows represent the gradient of norm adoption for each norm distribution, i.e., the most likely trajectory in terms of norm adoption that a population with a given norm distribution will follow. Figure (a) shows the dynamics of SSGs, in which any norms prescribing strategies "give way to the right" or "give way to the left" are evolutionarily stable. Also, there is an evolutionarily stable state with a polymorphic population whose 50% of cars "never give way", and the remaining 50% "give way always". This state can be reached when cars do not have available norms to give way to either the left or the right. Figure (b) shows the dynamics of DSGs, in which the evolutionarily stable norms are those prescribing strategy "give way always". Also, there are two evolutionarily stable states in which cars only have available either norms to "give way to the left" or norms to "give way to the right". a norm prescribing "stop" 100% of times, since in those games the risk of colliding once going forward is 100%. Figure 7 shows the evolutionary dynamics of norm adoption for SSGs (Figure 7a) and DSGs (Figure 7b). Each square represents the possible frequency distributions of norms prescribing strategies to "never give way", "give way to the right", "give way to the left" and "give way always". For instance, the top-left corner represents a population in which 100% of cars adopt a "never give way" norm, and the middle point of the square represents a population in which the four norms are 25% frequent. Arrows represent the gradient of norm adoption for each norm distribution, i.e., the most likely trajectory in terms of norm adoption that a population with a given norm distribution will follow. "never give way""give way always""give way to the right""give way tothe left" (a)(b)"never give way""give way always""give way to the right""give way tothe left" Synthesising Evolutionarily Stable Normative Systems 21 Outcome Collides Stops & avoids collisions Goes & avoids collisions r0 i 0 0 1 r1 i 0 0.1 1 r2 i 0 0.2 1 r3 i 0 0.3 1 r4 i 0 0.4 1 r5 i 0 0.5 1 r6 i 0 0.6 1 r7 i 0 0.7 1 r8 i 0 0.8 1 r9 i 0 0.9 1 r10 i 0 1 1 Table 6: Empirical reward functions to model populations with different degrees of collision aversion. The lower rewards (e.g., r0 i ) represent populations with lower aversion to colliding. The higher rewards (e.g., r9 i , r10 i ) represent populations with higher aversion to colliding. i , r1 In SSGs, norms to "give way to the right" or to "give way to the left" are the only evolutionarily stable ones. Both norms are attractor points of the norm evolution process (indicated with big black dots). If the mass of cars giving way to the right is bigger than the mass of cars giving way to the left, then the whole population will tend to give way to the right in order to synchronise. There is also an evolutionarily stable state in which 50% of cars adopt a "never give way" norm, which the remaining 50% of cars compensate adopting a "give way always" norm. This state can be reached whenever cars do not have available norms to give way to a side (which could invade the population because they are fitter). As for DSGs, norms to "give way always" are the only evolutionarily stable ones. Thus, no matter what the initial norm distribution is, as long as at least one car adopts a "give way always" norm, its fitness will be higher than that of any other car, and the whole population will eventually adopt its norm. 4.3 Adaptivity analysis Next, we analyse the capability of our ENSS to adapt norm synthesis to the characteristics of the population to regulate. With this aim, we run simulations with populations that have different degrees of aversion to colliding. We model these populations by considering a collection of empirical reward functions depicted in Table 6. Each function returns 0 once a car collides at time t, and 1 once a car goes forward without colliding at time t. These functions differ in the reward given to the cars once they stop to avoid collisions, which balances their "hurry" to get to their destinations with their willingness to avoid collisions. If this reward is low (e.g., r0 i ), then the cars prefer not to stop in order to not being delayed, even if it implies a collision risk. Then, we say that cars will have a low collision aversion degree. As this reward increases, cars will have a lower aversion to stopping, which can be interpreted as a higher aversion to colliding. We run 1,000 simulations for each reward function with a 0.8 discount factor (β = 0.8). Figure 8 shows averaged results of all simulations. The x-axis depicts the different empirical reward functions (collision aversion degrees), and the y-axis shows: – the number of rounds that the ENSS requires to converge. – the average frequency with which cars optimally converge in 2-role games. That is, the frequency with which they converge to a norm like "give way to the right" or a norm like "give way to the left" in SSGs, and to a norm like "give way always" in DSGs. – the average frequency with which cars optimally converge 1-role games. That is, the frequency with which they converge to "stop" in PGs and TJGs (which are the only strategies that allow to avoid 100% of collisions). – the collision avoidance rate during the last round of the simulation (once the simulation has converged and the cars have adopted an ESNS). 22 Javier Morales et al. Fig. 8: Averaged results of 1,000 simulations with a 0.8 discount factor (β = 0.8) and different degrees of collision aversion. The x-axis represents different reward functions, which allow to model the different collision aversion degrees. For instance, r0 rep- resents a population with total collision aversion. The y-axis shows: (1) the number of rounds the simulations required to converge; (2) the frequency with which cars converge optimally in 2-role games; (3) the frequency with which cars converge optimally in 1-role games; and (4) the collision avoidance rate during the last round of the simulation (once agents had converged to an ESNS). i represents a population with null collision aversion, and r10 i With null collision aversion (r0 i ), simulations take a high number of rounds to converge (286 rounds). This happens because the rewards for colliding and for stopping are equal, and hence the fitness of the norms that cause collisions (i.e., those prescribing "never give way") and those that prohibit to go in order to avoid collisions (e.g., "give way to the right") are similar. Consequently, cars take a long time to decide which norm to adopt. Upon conver- gence, cars adopt ESNSs containing "never give way" norms, which avoid 0% of collisions. As the collision aversion increases, simulations take less number of rounds to converge, and cars adopt norms that prohibit to go more frequently. For low collision aversions (r1 i to r3 i ), simulations still take a high number of rounds (281, 265 and 189 rounds, respectively), but cars adopt ESNSs that avoid up to 80% of collisions. Specifically, cars converge opti- mally up to 73% of times in 2-role games, and up to 55% of times in 1-role games. The rea- son that this frequency is slightly lower in 1-role games is because in prevention games (PG) cars not always collide once they choose to go forward. Hence, cars occasionally converge to norm "go", which cannot fully avoid collisions. For middle and high collision aversion degrees (r4 i ), the number of rounds necessary to converge decreases significantly. The best results are given by functions r7 i , with which cars optimally converge in the 100% of 2-role games – but still, they optimally converge up to 90% of 1-role games, hence avoiding up to 93% of collisions. i and r8 i to r9 With total collision aversion (r10 i ), the number of rounds necessary to converge increases again. This happens because the reward for stopping and not colliding and the reward for going forward and not colliding are equal. Hence, the fitness of all the norms that avoid collisions – either by prohibiting one role to go, or by prohibiting both roles to go – are similar. In consequence, cars need extra time to decide which norm to adopt. Upon conver- gence, cars adopt ESNSs containing only "give way always" norms to coordinate in SSGs and DSGs (hence converging optimally for DSGs, but not for SSGs). As a result, cars re- 0 50 100 150 200 250 300r0ir1ir2ir3ir4ir5ir6ir7ir8ir9ir10iReward functionsnumber of rounds to converge2-role games optimal convergence rate1-role games optimal convergence ratecollision avoidance degreeSynthesising Evolutionarily Stable Normative Systems 23 Fig. 9: Competition of an ESNS Ω∗ (represented with id 1000) against mutant normative systems in one simulation. The x-axis illustrates the different rounds of the simulation. The y-axis illustrates the id's of the different normative systems generated during the simulation. Black dots represent the creation of mutant normative systems with a certain id at a given round. The red line illustrates the id of the most frequent normative system. main stopped indefinitely and 100% of collisions are avoided. It turns out that cars are so afraid of colliding that they do not mind to stay still indefinitely in order to avoid collisions. 4.4 Stability analysis Finally, we analyse the stability of the normative systems synthesised by our approach upon convergence. With this aim, we perform 100 simulations that start with a population in which 100% of agents abide by an ESNS of those synthesised in the experiment of Section 4.2, which we will call Ω∗. Each simulation lasts 400 rounds, and each round lasts 200 ticks. In each round, the system creates random agents that abide by a normative system with alternative norms for each possible game. Then, we let agents interact. At the end of each round, the system performs norm replication. As considered in the literature in EGT [29], we consider that our normative system is an ESNS if none of its norms can be invaded by any alternative norm. That is, if in every single round, the fitness of the mutant norms is lower than the fitness of the norms in Ω∗ (and hence, the invader norms cannot grow in frequency). Thus, if our normative system is an ESNS, the agents will end up adopting it at the end of every single simulation. In 100% of simulations, the cars ultimately adopted normative system Ω∗. Figure 9 illustrates the dynamics of one of these simulations. The x-axis shows the different rounds of the simulation, and the y-axis depicts the id's of the normative systems created over time. Black dots represent mutant normative systems created in each round (with which Ω∗ has to compete), and the red line indicates the id of the most frequent normative system. For the sake of clarity, we represent Ω∗ as the normative system with id 1,000. After 200 rounds, the simulation created 2,500 different mutant normative systems. Upon round 400, normative system Ω∗ remained stable most of the time. In punctual rounds, the simulation generated a high number of mutant normative systems, making the frequency of Ω∗ to go below stability. But, after a few rounds, Ω∗ replicated and became again the most frequent 0 500 1000 1500 2000 2500 0 50 100 150 200 250 300 350 400Normative system id'sRoundsmost frequent normative system24 Javier Morales et al. normative system. Upon round 400, the cars converged to adopting Ω∗, thus demonstrating that Ω∗ is a best choice for the agents. 5 Related work Broadly speaking, research on norm synthesis can be classified into two main strands of work: off-line design, and on-line synthesis. Pioneered by Shoham and Tennenholtz [27], off- line design aims at designing the norms that will coordinate the agents before they start to do their work [27,12,32,9]. This approach is arguably less flexible, since norms are typically hard-wired into the agents' behaviours and cannot be adapted over time. Alternatively, on-line synthesis studies how norms can come to exist while the agents interact at runtime. Most on-line approaches focus on investigating how norms emerge from agent societies through an iterated process whereby agents tend to adopt best-performing strategies [2,26,34,11,33,1]. Alternatively, recent work by Morales et. al. approached on- line synthesis by employing designated agents that observe agents' interactions at runtime and generate norms aimed at resolving conflict situations [17,18,19]. Later on, Mashayekhi et. al. [16] extended this work by proposing a hybrid mechanism in which norms are syn- thesised by designated agents, and the agent society iteratively selects and spread best- performing norms, ultimately adopting most successful ones. The closest approach to our work in the literature is that of norm emergence, and partic- ularly the study of norm evolution and stability. One of the pioneer works on this approach is the one by Axelrod [2,3], which studies how norms evolve and emerge as stable patterns of behaviour of agent societies. Axelrod considers a game-theoretic setting in which agents repeatedly play a single two-player game by employing different strategies. The strategies that allow the agents to achieve better personal results prosper and spread. A (stable) norm is said to have emerged once a majority of agents abide by the same strategy that is sustained over time. Subsequently, many researchers have studied norm emergence by employing the game theoretic approach. In [23], Rajiv Sethi extended the work of Axelrod and studied how so- cial norms of vengeance and cooperation emerge within agent societies. With this aim, Sethi incorporated the solution concept of evolutionarily stable strategy (ESS) and the principle of replicator dynamics from evolutionary game theory (EGT) [29]. Again, this work considers that the agents play a single two-player game. Shoham and Tennenholtz [28] introduced a framework for the emergence of social conventions as points of (Nash) equilibria in stochas- tic settings. They introduced a natural strategy-selection rule whereby the agents eventually converge to rationally acceptable social conventions. Later, Sen and Airiau proposed in [22] a social learning model whereby agents can learn their policies and norms can emerge over repeated interactions with multiple agents in two- player games. Many works have considered this model to study further criteria that affect to the emergence of norms. Of these, the closest to our work is perhaps the one by Sugawara et. al. [30,31], in which conflict situations are characterised as Markov games, and a model is introduced to evolve norms that successfully coordinate the agents in these games. More recent work (such as that by Santos et. al. [21]) studies how cooperation norms can emerge once the agents can explore alternative strategies, i.e. they have arbitrary explo- ration rates. They show that cooperation emergence depends on both the exploration rate of the agents and the underlying norms at work. Similarly, Soham et. al. [10] introduce an EGT-based model to study how norms change in agent societies in terms of the need for coordination and the agents' exploration rate. They show that societies with high needs for Synthesising Evolutionarily Stable Normative Systems 25 coordination tend to lower exploration rates and higher norm compliance, while societies with lower coordination needs lead to higher exploration rates. Also, Lorini et. al. [15] in- troduce a model for the evolution and emergence of fairness norms in relation to the degree of sensitivity (internalisation) of the agents to these norms. They show that, in the long term, the higher the sensitivity of the agents to norms, the more beneficial for the social welfare. From Axelrod [2,3] to Lorini [15], our approach is different to all the aforementioned works for several reasons. First, most previous works consider that the agents play a sin- gle game whose payoffs are known beforehand. Unlike them, our work considers a non- deterministic setting in which the games (i.e., the conflict situations) that the agents might engage in are initially unknown, likewise the payoffs of these games. We provide a frame- work to perform runtime detection of the multiple games that the agents might play, and to learn their payoffs based on the rewards to the agents once they repeatedly play each game over time. Our framework allows to evolve norms based on their capability to coordinate agents in different games, and leads agents to converge to normative systems that maximise their fitness. Moreover, to the best of our knowledge, our framework is the first one in in- troducing the analytical concept of evolutionarily stable normative system (ESNS), which allows to assess the combinations of norms that the agents are likely to adopt. 6 Conclusions In this work we introduced a framework for the synthesis of evolutionarily stable norma- tive systems (ESNS) for non-deterministic settings. Our framework allows to synthesise sets of norms that populations of rational agents are likely to adhere by. With this aim, our framework carries out an evolutionary game theoretic process inspired in evolutionary game theory (EGT), whereby the agents tend to adopt the norms that are fittest to coordinate them in strategic situations. Our framework assumes no previous knowledge about the coordina- tion situations that the agents can engage in, neither about their potential outcomes. Instead, it learns these by letting the agents interact, detecting the situations in which they need co- ordination, and modelling them as games of which it iteratively learns their payoffs. Norms that are more useful to coordinate the agents in these games prosper and spread, and are ultimately adopted by the agents. The outputs of this evolutionary process are sets of norms that, once adopted by an entire agent population, no agent can benefit from switching to alternative norms (and hence, we say that they are evolutionarily stable). We provided evidence of the quality and the relevance of our approach through an em- pirical evaluation in a simulated traffic scenario. We showed that our framework allows cars to converge 100% of times to normative systems that successfully avoid car collisions as far as they are sufficiently willing to avoid collisions. We illustrated the capability of our ap- proach to adapt norm synthesis to the preferences of the agent population, showing that the agents will tend to adopt different types of normative systems as their preferences change. In this sense, we believe that our framework provides a valuable decision support tool for policy makers. It can be used to make predictions about the type of normative systems that a given population will end up adopting, and to assess whether a particular population will accept a given normative system. Note that our model for norm evolution is described in terms of the individual goals of the agents. Thus, the norms that will be more likely to spread are those that are useful to the agents from their point of view. As previously mentioned, this can be very useful in order to predict which norms a given population will ultimately adopt. Nevertheless, our framework cannot be employed to synthesise norms that achieve a global goal that is not 26 Javier Morales et al. aligned with the individual goals of the agents. For instance, cars will only adopt norms that avoid collisions as far as they do not want to collide. Therefore, the sensible next step is to extend our model in order to be able to synthesise evolutionarily stable norms that accomplish a coordination task even when coordination is not in the interests of the agents. For instance, to synthesise norms that avoid collisions even when cars do not mind to collide. With this aim, we plan to include a deterrence mechanism in our model by incorporating sanctions in norms so that coordination can emerge even when the agents' goals are not aligned with the global system goals. References 1. S. Airiau, S. Sen, and D. Villatoro. Emergence of conventions through social learning. Autonomous Agents and Multi-Agent Systems, 28(5):779–804, 2014. 2. R. Axelrod. An evolutionary approach to norms. American political science review, 80(04):1095–1111, 3. R. M. Axelrod. The complexity of cooperation: Agent-based models of competition and collaboration. 1986. Princeton University Press, 1997. 4. O. H. Azar. What sustains social norms and how they evolve?: The case of tipping. Journal of Economic Behavior & Organization, 54(1):49–64, 2004. 5. C. Bicchieri and A. Chavez. Behaving as expected: Public information and fairness norms. Journal of Behavioral Decision Making, 23(2):161–178, 2010. 6. C. Bicchieri and E. Xiao. Do the right thing: but only if others do so. Journal of Behavioral Decision Making, 22(2):191–208, 2009. 7. J. Bjornerstedt and J. W. Weibull. Nash equilibrium and evolution by imitation. Working Paper Series 407, Research Institute of Industrial Economics, 1994. 8. G. Boella, L. Van Der Torre, and H. Verhagen. Introduction to normative multiagent systems. Computa- tional & Mathematical Organization Theory, 12(2-3):71–79, 2006. 9. D. Corapi, A. Russo, M. de Vos, J. Padget, and K. Satoh. Normative design using inductive learning. Theory and Practice of Logic Programming, 11(4-5):783–799, 007 2011. 10. S. De, D. S. Nau, and M. J. Gelfand. Understanding norm change: An evolutionary game-theoretic approach. In Proceedings of the 16th Conference on Autonomous Agents and MultiAgent Systems, AA- MAS '17, pages 1433–1441, Richland, SC, 2017. International Foundation for Autonomous Agents and Multiagent Systems. 11. J. Delgado. Emergence of social conventions in complex networks. Artificial intelligence, 141(1-2):171– 185, 2002. 12. D. Fitoussi and M. Tennenholtz. Choosing social laws for multi-agent systems: Minimality and simplic- ity. Artificial Intelligence, 119(1-2):61–101, 2000. 13. G. Governatori and A. Rotolo. Norm compliance in business process modeling. In International Work- shop on Rules and Rule Markup Languages for the Semantic Web, pages 194–209. Springer, 2010. 14. K. Iwai. Schumpeterian dynamics: An evolutionary model of innovation and imitation. Journal of Economic Behavior & Organization, 5(2):159–190, 1984. 15. E. Lorini and R. Muhlenbernd. The long-term benefits of following fairness norms: A game-theoretic analysis. In International Conference on Principles and Practice of Multi-Agent Systems, pages 301– 318. Springer, 2015. 16. M. Mashayekhi, H. Du, G. F. List, and M. P. Singh. Silk: A simulation study of regulating open normative In Proceedings of the Twenty-Fifth International Joint Conference on Artificial multiagent systems. Intelligence, IJCAI'16, pages 373–379. AAAI Press, 2016. 17. J. Morales, M. Lopez-Sanchez, J. A. Rodriguez-Aguilar, M. Wooldridge, and W. Vasconcelos. Auto- mated synthesis of normative systems. In Proceedings of the AAMAS'13, pages 483–490. IFAAMAS, 2013. 18. J. Morales, M. Lopez-Sanchez, J. A. Rodriguez-Aguilar, M. Wooldridge, and W. Vasconcelos. Minimal- ity and simplicity in the on-line automated synthesis of normative systems. In AAMAS '14: Proceedings of the 13th international conference on autonomous agents and multiagent systems, pages 109–116, Richland, SC, 2014. International Foundation for Autonomous Agents and Multiagent Systems (IFAA- MAS). 19. J. Morales, M. Lopez-Sanchez, J. A. Rodriguez-Aguilar, M. Wooldridge, and W. Vasconcelos. Synthe- sising liberal normative systems. In AAMAS '15: Proceedings of the 14th international conference on autonomous agents and multiagent systems, pages 433–441. International Foundation for Autonomous Agents and Multiagent Systems (IFAAMAS), 04/05/2015 2015. Synthesising Evolutionarily Stable Normative Systems 27 20. A. Rapoport and A. M. Chammah. Prisoner's dilemma: A study in conflict and cooperation, volume 165. University of Michigan press, 1965. 21. F. P. Santos, J. M. Pacheco, and F. C. Santos. Evolution of cooperation under indirect reciprocity and arbitrary exploration rates. Scientific Reports, 6, 2016. 22. S. Sen and S. Airiau. Emergence of norms through social learning. In Proceedings of the 20th Interna- tional Joint Conference on Artifical Intelligence, IJCAI'07, pages 1507–1512, San Francisco, CA, USA, 2007. Morgan Kaufmann Publishers Inc. 23. R. Sethi. Evolutionary stability and social norms. Journal of economic behavior & organization, 29(1):113–140, 1996. 24. R. Sethi and E. Somanathan. Norm compliance and strong reciprocity. Moral sentiments and material interests: The foundations of cooperation in economic life, pages 229–250, 2005. 25. Y. Shoham and K. Leyton-Brown. Multiagent Systems: Algorithmic, Game-Theoretic, and Logical Foun- dations. Cambridge University Press, New York, 2009. 26. Y. Shoham and M. Tennenholtz. Emergent conventions in multi-agent systems: Initial experimental results and observations (preliminary report). In B. Nebel, C. Rich, and W. R. Swartout, editors, KR, pages 225–231. Morgan Kaufmann, 1992. 27. Y. Shoham and M. Tennenholtz. On social laws for artificial agent societies: off-line design. Artificial Intelligence, 73(1-2):231–252, February 1995. 28. Y. Shoham and M. Tennenholtz. On the emergence of social conventions: modeling, analysis, and sim- ulations. Artificial Intelligence, 94(1):139–166, 1997. 29. J. M. Smith. Evolution and the Theory of Games, pages 202–215. Springer US, Boston, MA, 1988. 30. T. Sugawara. Emergence and stability of social conventions in conflict situations. In Proceedings of the Twenty-Second international joint conference on Artificial Intelligence-Volume Volume One, pages 371–378. AAAI Press, 2011. 31. T. Sugawara. Emergence of conventions for efficiently resolving conflicts in complex networks. In Proceedings of the 2014 IEEE/WIC/ACM International Joint Conferences on Web Intelligence (WI) and Intelligent Agent Technologies (IAT) - Volume 03, WI-IAT '14, pages 222–229, Washington, DC, USA, 2014. IEEE Computer Society. 32. W. van Der Hoek, M. Roberts, and M. Wooldridge. Social laws in alternating time: Effectiveness, feasibility, and synthesis. Synthese, 156(1):1–19, 2007. 33. D. Villatoro. Self-organization in decentralized agent societies through social norms. In The 10th In- ternational Conference on Autonomous Agents and Multiagent Systems-Volume 3, pages 1373–1374. International Foundation for Autonomous Agents and Multiagent Systems, 2011. 34. A. Walker and M. Wooldridge. Understanding the emergence of conventions in multi-agent systems. In International Conference on Multiagent Systems, ICMAS'95, pages 384–389, 1995.
1303.2789	1	1303	2013-03-12T07:00:04	A Cooperative Q-learning Approach for Real-time Power Allocation in Femtocell Networks	[ "cs.MA", "cs.LG" ]	In this paper, we address the problem of distributed interference management of cognitive femtocells that share the same frequency range with macrocells (primary user) using distributed multi-agent Q-learning. We formulate and solve three problems representing three different Q-learning algorithms: namely, centralized, distributed and partially distributed power control using Q-learning (CPC-Q, DPC-Q and PDPC-Q). CPCQ, although not of practical interest, characterizes the global optimum. Each of DPC-Q and PDPC-Q works in two different learning paradigms: Independent (IL) and Cooperative (CL). The former is considered the simplest form for applying Qlearning in multi-agent scenarios, where all the femtocells learn independently. The latter is the proposed scheme in which femtocells share partial information during the learning process in order to strike a balance between practical relevance and performance. In terms of performance, the simulation results showed that the CL paradigm outperforms the IL paradigm and achieves an aggregate femtocells capacity that is very close to the optimal one. For the practical relevance issue, we evaluate the robustness and scalability of DPC-Q, in real time, by deploying new femtocells in the system during the learning process, where we showed that DPC-Q in the CL paradigm is scalable to large number of femtocells and more robust to the network dynamics compared to the IL paradigm	cs.MA	cs	A Cooperative Q-learning Approach for Real-time Power Allocation in Femtocell Networks Hussein Saad∗, Amr Mohamed† and Tamer ElBatt∗ ∗Wireless Intelligence Network Center (WINC), Nile University, Cairo, Egypt. [email protected] [email protected] †Computer Science and Engineering Department Qatar University, P.O. Box 2713, Doha, Qatar. [email protected] Abstract—In this paper, we address the problem of distributed interference management of cognitive femtocells that share the same frequency range with macrocells (primary user) using distributed multi-agent Q-learning. We formulate and solve three problems representing three different Q-learning algorithms: namely, centralized, distributed and partially distributed power control using Q-learning (CPC-Q, DPC-Q and PDPC-Q). CPC- Q, although not of practical interest, characterizes the global optimum. Each of DPC-Q and PDPC-Q works in two different learning paradigms: Independent (IL) and Cooperative (CL). The former is considered the simplest form for applying Q- learning in multi-agent scenarios, where all the femtocells learn independently. The latter is the proposed scheme in which femtocells share partial information during the learning process in order to strike a balance between practical relevance and performance. In terms of performance, the simulation results showed that the CL paradigm outperforms the IL paradigm and achieves an aggregate femtocells capacity that is very close to the optimal one. For the practical relevance issue, we evaluate the robustness and scalability of DPC-Q, in real time, by deploying new femtocells in the system during the learning process, where we showed that DPC-Q in the CL paradigm is scalable to large number of femtocells and more robust to the network dynamics compared to the IL paradigm. I. INTRODUCTION Femtocells have been recently proposed as a promising solution to the indoor coverage problem. Although femtocells offer significant benefits to both the operator and the user, several challenges have to be solved to fully reap these benefits. One of the most daunting challenges is their interfer- ence on macro-users and other femtocells [1], [2]. Typically, femtocells are installed by the end user and hence, their number and positions are random and unknown to the network operator a priori. Adding to this the typical dynamics of the wireless environment, a centralized approach to handle the interference problem can not be feasible which, in turn, calls for a distributed interference management strategies. Based on these observations, in this paper, we focus on closed access femtocells [3] working in the same bandwidth with macrocells (i.e. cognitive femtocells), where the femto- cells will be the secondary users who try to perform power control to maximize their own performance while maintain the macrocell capacity at certain level. In order to handle the interference generated by the femtocells on the macrocell users, we will use a distributed reinforcement learning [4] technique called multi-agent Q-learning [5] and [6]. In our context, a prior model of the environment cannot be achieved due to 1) the unplanned placement of the femtocells, 2) the typical dynamics of the wireless enshrinement. In such context, Q-Learning offers significant advantages to achieve optimal decision policies through realtime learning of the environment [7]. In the literature, Q-learning has been used several times to perform power allocation in femtocell networks. In [8], authors used independent learning (IL) Q-learning to perform power allocation in order to control the interference generated by the femtocells on the macrocell user. In [7], authors introduced a new concept called docitive femtocells where a new femtocell can fasten its learning process by learning the policies acquired by the already deployed femtocells, instead of learning from scratch. The policies are shared by Q-table exchange between the femtocells. However, after the Q-tables are exchanged, all the femtocells take their actions (powers) independently, which may generate an oscillating behavior in the system. In [9], we developed a distributed power allocation algorithm called dis- tributed power control using Q-learning (DPC-Q). In DPC-Q, two different learning paradigms were proposed: independent learning (IL) and cooperative learning (CL). It was shown that both paradigms achieves convergence. Moreover, the CL paradigm outperforms the IL one through achieving higher aggregate femtocells capacity and better fairness (in terms of capacity) among the learning femtocells. However, in [9] we did not evaluate the performance of DPC-Q against the networks dynamics, specially after conver- gence. Also, we did not have any benchmarking algorithm to compare the performance of DPC-Q to. Thus, the contribution of this paper can be summed up as follows: • we propose two new Q-learning based power allocation algorithms: namely, centralized power control using Q- learning (CPC-Q) and partially distributed power con- trol using Q-learning (PDPC-Q). CPC-Q is used for benchmarking purposes, where a central controller, which has all the information about the system (channel gains of all femtocells, system noise, · · · ), is responsible for calculating the optimal powers that the femtocells should use. PDPC-Q, which is an agent based power control algorithm, is proposed as: 1) it gives the operator the flexibility to work on a global base (e.g. aggregate femtocell capacity instead of subcarrier based femtocell capacity as in DPC-Q), 2) it makes DPC-Q comparable to CPC-Q. indicates the channel gain between FBS n′ transmitting on subcarrier k and the femto user associated to FBS n. • we evaluate the robustness and scalability of DPC-Q, in both IL and CL paradigms, against two of the dynamics that typically exist in the wireless environment: namely, the random activity of femtocells (when new femtocells are deployed in the system during the learning process) and the density of the femtocells in the macrocell cov- erage area (the number of femtocells that are interfering on the macro users). • we compare our proposed DPC-Q in both IL and CL paradigms to the idea of docitive femtocells presented in [7]. the system model The rest of this paper is organized as follows. In section II, is described. Section III presents a brief background about multi-agent Q-learning. In section IV, the proposed Q-learning based power allocation algorithms are presented. The simulation scenario and the results are discussed in section V. Finally the conclusions are drawn in section VI. II. SYSTEM MODEL We consider a wireless network composed of one macro cell (with one single transmit and receive antenna Macro Base Station (MBS)) that coexists with Nf femtocells, each with one single transmit and receive antenna Femto Base Station (FBS). The Nf femtocells are placed indoors within the macrocell coverage area. Both the MBS and the FBSs’ transmit over the same K subcarriers where orthogonal down- link transmission is assumed. Um and Uf macro and femto users are located randomly inside the macro and femto cells respectively. Femtocells within the same range can share partial information during the learning process to enhance their performance. p(k) o and p(k) n n ≤ P f PK max respectively, where PK denote the transmission powers of the MBS and FBS n on subcarrier k respectively. Moreover, the max- imum transmission powers for the MBS and any FBS n are max and max and P f P m k=1 p(k) The system performance is analyzed in terms of the capacity measured in (bits/sec/Hz). The capacity achieved by the MBS at its associated user on subcarrier k is: oo p(k) h(k) n=1 h(k) no p(k) o n + σ2 o = log2(1 + k=1 p(k) o ≤ P m max. C(k) (1) ) where h(k) oo indicates the channel gain between the trans- mitting MBS and its associated user on subcarrier k; h(k) no indicates the channel gain between FBS n transmitting on subcarrier k and the macro user. Finally σ2 indicates the noise power. The capacity achieved by FBS n at its associated user on subcarrier k is: PNf n nn p(k) h(k) n′np(k) n′=1,n′6=n h(k) n′ + h(k) on p(k) o + σ2 ) (2) C(k) n = log2(1 + PNf where h(k) nn indicates the channel gain between FBS n transmitting on subcarrier k and its associated user; h(k) n′n III. MULTI-AGENT Q-LEARNING The scenario of distributed cognitive femtocells can be mathematically formulated using stochastic games [10], where the learning process of each femtocell is described by a task defined by the quintuple {N, S, A, P, R(s, ~a)}, where: • N = {1, 2, · · · , Nf } is the set of agents (i.e. femtocells). • S = {s1, s2, · · · , sm} is the set of possible states that each agent can occupy, where m is the number of possible states. • A = {a1, a2, · · · , al} is the set of possible actions that each agent can perform for each task, where l is the number of possible actions. • P is the probabilistic transition function that defines the probability that an agent transits from one state to another, given the joint action performed by all agents. • R(s, ~a) is the reward function that determines the reward fed back to an agent n by the environment when the joint action ~a is performed in state s ∈ S. In the distributed cognitive femtocells scenario, P can not be deduced due to the dynamics of the wireless environment. Thus, one of the most famous techniques that calculates optimal policies without any prior model of the environment is Q-learning . Q-learning assigns each task of each agent a Q-table whose entries are known as Q-values Q(sm, al), for each state sm ∈ S and action al ∈ A. Thus, the dimension of this table is m × l. The Q-value Q(sm, al) is defined to be the expected discounted reward over an infinite time when action al is performed in state sm, and an optimal policy is followed thereafter [5]. The learning process of each agent n at time t can be described as follows: 1) the agent senses m ∈ S, 2) the environment and observes its current state sn randomly with based on sn probability ǫ or according to: an l = arg maxa∈A Qt m, a) n(sm, a) is the row of the Q- with probability 1 − ǫ, where Qt table of agent n that corresponds to state sn m at time t, and ǫ is an exploration parameter (a random number) that guarantees that all the state-action pairs of the Q-table is visited at least once, 3) the environment makes a transition to a new state m′ ∈ S and the agent receives a reward rt m, ~a) due sn to this transition, 4) the Q-value is updated using equation 3 and the process is repeated. m, the agent selects its action an l n = R(sn n(sn Qt+1 n (sn m, an l ) :=(1 − α)Qt n(sn n + γ max a′ ∈A m, an Qt l )+ n(sn α(rt ′ )) (3) m′ , a where α is called the learning rate and 0 ≤ γ ≤ 1 is the discount factor that determines how much effect the future rewards have on the decisions at each moment. It should be noticed that the reward rt n depends on the joint action ~a of all agents not on the individual action al. This is the main difference between the multi-agent scenario described here and the single-agent one (when Nf = 1). In the single-agent case, one of the conditions needed to guarantee that the Q- values converges to the optimal ones is that: the reward of the agent must be dependent only on its individual actions (i.e. the reward function is stationary for each state-action pair) [5], [11]. However, for the multi-agent scenario, the reward function is not stationary from the agent point of view, since it now depends on the actions of other agents. Thus, the convergence proof used for the single-agent case can not be used in the multi-agent one. IV. POWER ALLOCATION USING Q-LEARNING In this section, the three proposed Q-learning based power allocation algorithms will be presented: A. Distributed Power Control Using Q-learning (DPC-Q) DPC-Q is a distributed algorithm where multiple agents (i.e: femtocells) aim at learning a sub-optimal decision policy (i.e: power allocation) by repeatedly interacting with the envi- ronment. The DPC-Q algorithm is proposed in two different learning paradigms: • Independent learning (IL): In this paradigm, each agent learns independently from other agents (i.e: ignores other agents’ actions and considers other agents as part of the environment). Although, this may lead to oscillations and convergence problems, the IL paradigm showed good results in many applications [8]. • Cooperative learning (CL): In this paradigm, each agent shares a portion of its Q-table with all other cooperat- ing agents1, aiming at enhancing the femtocells’ perfor- mance. CL is performed as follows: each agent shares the row of its Q-table that corresponds to its current state with all other cooperating agents (i.e. femtocells in the same range). Then, each agent n selects its action an l according to the following equation: an l = arg max a∈A ( X1≤n′≤Nf Qn′(sn′ , a)) (4) The main idea behind this strategy is explained in details in [9]. In terms of overhead, if the number of femtocells is Nf , then the total overhead needed is Nf .(Nf − 1) messages (each of size l) per unit time (i.e. the overhead is quadratic in the number of cooperating femtocells). DPC-Q is an agent and subcarrier based algorithm (i.e. the capacities, states, actions, reward functions are defined for each agent over each subcarrier) [9]: • Agents: F BSn, ∀1 ≤ n ≤ Nf • States: At time t for femtocell n on subcarrier k, the state is defined as: sn,k t ∈ {0, 1} indicates the level of interference measured at the macro- user on subcarrier k at time t: t } where I k t = {I k t , Pn I k t =(1, C(k) 0, C(k) o < Γo o ≥ Γo (5) where Γo is the target capacity determining the QoS per- formance of the macrocell. We assume that the macrocell 1We assume that the shared row of the Q-table is put in the control bits of the packets transmitted between the femtocells. The details of the exact protocol lie out of the scope of this paper. o to all FBSs through the backhaul reports the value of C(k) connection. Pn t defines the power levels used to quantize the total power FBS n is using for transmission at time t: t ≤ P f max (6) Pn t =  1, (P f 0, PK 2, PK k=0 pn,k t < (P f max − A2) ≤PK k=0 pn,k t > P f max max − A1) k=0 pn,k where A1 and A2 are arbitrary selected thresholds (sev- eral values for A1 and A2 as well as more power levels were tried through the simulations and the performance gain between these values was marginal). • Actions: The action here is scalar, where the set of actions available for each FBS is defined as the set of possible powers that a FBS can use for transmission on each subcarrier. In the simulations, a range from −20 to max dBm with step of 2 dBm is used. P f • Reward Functions: The reward fed back to agent n on subcarrier k at time t is defined as: rn,k t =(e−(C (k) −2, o −Γo)2 − e−C (k) k=0 pn,k k=0 pn,k t ≤ P f t > P f max n , PK PK max (7) The rationale behind this reward function is that each femtocell will aim at maximizing its own capacity while: 1) maintaining the capacity of the macrocell around the target capacity Γo (convergence is assumed to be within a range of ±1 bits/sec/Hz from Γo), 2) not exceeding the allowed P f This reward function was compared to the reward func- tion defined in [9]: max. , PK PK rn,k t =(e−(C (k) −1, o −Γo)2 k=0 pn,k k=0 pn,k t ≤ P f t > P f max max (8) where it was shown that both reward functions maintain the capacity of the macrocell within the convergence range. However, reward function 7 was able to achieve higher aggregate femtocell capacity. In this paper, we show another advantage for reward function 7, which is: it learns (explores or reacts to network dynamics) better than reward function 8 even when the exploration parameter ǫ is not used. This mainly depends on the initial value of the Q-values. In this paper, we initially set all the Q-values to zero. Thus, when ǫ is not used, using reward function 8 will always feed the agent back with a max is not exceeded). Thus, positive reward (given that P f if initially agent n was in state sn,k on subcarrier k and took action pn,k , the Q-value of this action Qn(sn,k, pn,k) will be updated using a positive valued reward, thus this Q-value will increase with time, and agent n will keep using the same action forever (since the action is chosen according to the maximum Q-value). Thus, using ǫ with reward function 8 is a must to have better exploration behavior. On the other hand, using reward function 7 may feed the agent back with positive or negative valued rewards (e−(C (k) n ). Thus, given the same initial conditions, agent n could o −Γo)2 could be smaller than e−C (k) t t receive a negative valued reward after taking action pn,k , leading to the decrease of its Q-value with time. Once the Q-value decreases below zero, the agent will take another action whose Q-value is greater than the decreased one. Thus, reward function 7 learns (explores) better than reward function 8. t In this paper, we also evaluate the robustness and scalability of DPC-Q, in both IL and CL paradigms. We believe that the CL paradigm is much more robust and scalable against the network dynamics compared to the IL paradigm. The reason is that after sharing the row of the Q-table, each femtocell will know the states that all other cooperating femtocells are occupying, and since a state at a certain moment can be defined as: how the agent sees the environment at that moment, each femtocell can implicitly know 1) how all other femtocells can react to the network dynamics, 2) what actions other femtocells are going to take. However, if the femtocells took their actions independently (i.e. IL paradigm), even after knowing the states of each other, oscillating behaviors that may not reach convergence may be generated. One way to overcome this problem is to force the femtocells to make use of the information shared while taking their actions (i.e. taking the actions cooperatively: equation 4). This could decrease the oscillations in the system, making the femtocells more robust towards the increase of the number of deployed femtocells, and towards the sudden effect caused by any new deployed femtocell. B. Partially Distributed Power Control Using Q-learning (PDPC-Q) PDPC-Q is a partial distributed algorithm, where it is a multi-agent algorithm but only agent dependent (i.e. the states, actions, reward functions are defined for each agent over all subcarriers). As DPC-Q, PDPC-Q works in both IL and CL paradigms. The agents, states, actions and reward functions used for the PDPC-Q algorithm are defined as follows: • Agents: F BSn, ∀1 ≤ n ≤ Nf • States: At time t the state is defined as: st = {It} where It ∈ {0, 1} indicates the level of interference measured at the macro-user over all subcarriers at time t: It =(1, Co < βo k=1 C(k) 0, Co ≥ βo (9) o where Co = PK is the aggregate macrocell capacity and βo is the target aggregate macrocell capacity. • Actions: For FBS n, the set of actions is defined to be a set of vectors where each vector represents the powers FBS n is using on all subcarriers. • Reward Functions: Reward function 7 can be redefined as: − e−Cn (10) t = e−(Co−βo)2 rn k=1 C(k) where Cn =PK n is the aggregate capacity of FBS n. Note that since PDPC-Q is not subcarrier based, a power vector in which P f max is exceeded will never be assigned for any FBS. Thus, there is no need to put a TAXONOMY OF THE PROPOSED ALGORITHMS. TABLE I Complexity Reaction to network dynamics Scalability DPC-Q/IL action is scalar DPC-Q/CL action is scalar Inefficient & non-robust Efficient & robust Inefficient at large Nf Efficient at large Nf Speed of Convergence Medium convergence Fast convergence Overhead None f − Nf N 2 messages each of size A CPC-Q A grows exponentially in Nf and K in K - PDPC-Q A grows exponentially Infeasible at large Nf Slow convergence since A is huge Huge CL is more efficient & robust than IL CL is more scalable than IL CL is faster than IL CL has larger overhead than IL negative reward here as in DPC-Q case. The same goes for CPC-Q. C. Centralized Power Control Using Q-learning (CPC-Q) CPC-Q is a centralized power control algorithm used to evaluate the performance of our proposed DPC-Q algorithm. CPC-Q can be regarded as the single-agent version of the DPC-Q, and hence, its convergence to the optimal Q-values and thus optimal powers is guaranteed. However, using a centralized controller is not feasible in terms of overhead in multi-agent scenarios. Thus, CPC-Q works only for small scale problems. The agent, states, actions and reward functions used for CPC-Q are defined as follows: • Agents: A centralized controller. • States: The same as PDPC-Q. • Actions: For the central controller, the set of actions is defined to be a set of matrices where each matrix represents the powers of all femtocells over all subcarri- ers. However, the size of this set grows exponentially with both the number of femtocells and the number of subcarriers. Thus, forming the matrices (all possible actions) from a large set of powers such as the one used in DPC-Q will be infeasible2. • Reward Functions: Since CPC-Q is global, reward function 7 can be redefined as: rt = e−(Co−βo)2 − e−Cf emto where Cf emto is defined as Cf emto =PNf (11) k=1 C(k) n . Finally, for the rest of the paper, reward functions 7, 11 and 10 will be referred to as R1, while reward function 8 will be referred to as R0. The three proposed algorithms are compared qualitatively in table I. n=1PK V. PERFORMANCE EVALUATION A. Simulation Scenario We consider a wireless network consisting of one macrocell serving Um = 1 macro user underlaid with Nf femtocells. Each femtocell serves Uf = 1 femto-user, which is randomly located in the femtocell coverage area. All of the macro 2In the simulations, the set of powers used to form the matrices and the vectors in CPC-Q and PDPC-Q respectively is: {0, 6, 12} dBm. and femto cells share the same frequency band composed of K subcarriers, where orthogonal downlink transmission is assumed. In the simulations, K will change according to the algorithm used: for DPC-Q, K = 6, while for both CPC-Q and PDPC-Q, K = 3. The channel gain between any transmitter i and any receiver j on subcarrier k is assumed to be path-loss dominated and is given by: ij = d(−P L) h(k) (12) ij where dij is the physical distance between transmitter i and receiver j, and P L is the path loss exponent. In the simulations P L = 2 is used. The distances are calculated according to the following assumptions: 1) The maximum distance between the MBS and its associated user is set to 1000 meters, 2) The maximum distance between the MBS and a femto-user is set to 800 meters, 3) The maximum distance between a FBS and its associated user is set to 80 meters, 4)The maximum distance between a FBS and another femtocell’s user is set to 300 meters, 5) The maximum distance between a FBS and the macro-user is set to 800 meters. We used MatLab on a cluster computing facility with 300 cores to simulate such scenario, where in the simulations we set the noise power σ2 to 10−7, the maximum transmission the maximum power of the macrocell P m max to 15 dBm, each transmission power of each femtocell P f of the power levels A1 and A2 is set to 5 dBm, the learning rate α to 0.5, the discounted rate γ to 0.9 and the random number ǫ to 0.1 [7] and [9]. max to 43 dBm, B. Numerical Results Figure 1(a) shows the aggregate femtocells capacity (as a function of the number of femtocells) using CPC-Q and PDPC-Q with R1 in both IL and CL paradigms. It can be observed that CL is much better than IL, where from the figure it can be shown that the aggregate capacity gain of CPC-Q over PDPC-Q in case of CL is marginal. Since CPC-Q is considered the single agent version of DPC-Q, it should converge to the global optimal values. This is shown in the figure at small number of femtocells (Nf = 1 and 2). The optimal values are calculated using exhaustive search over all possible actions, where the optimal value is defined to be the maximum aggregate capacity the system can achieve while maintaining the capacity of the macrocell in the convergence range (±1 bits/sec/Hz from βo). However, starting from Nf = 3, CPC-Q begins to be infeasible since the size of the possible actions set A becomes very large (at Nf = 5: A = 320, 000, 0). So, besides the computational problems, the condition of visiting all state-action pair becomes infeasible. Thus, getting the optimal value is also not feasible (that’s why we stopped CPC-Q at Nf = 4). Note also that we stopped the exhaustive search at Nf = 5 due to complexity and memory problems, while PDPC-Q is shown at Nf = 6 and 7 just to illustrate the continuity of our algorithm. Figures 2 and 3 show the robustness of the proposed DPC- Q algorithm. In these figures, we started with Nf = 5, then we added a new femtocell after every 4000 iterations3to reach ) z H / c e s / s t i b ( y t i c a p a c l l t e c o m e F d e a g e r g g A t 3 2.5 2 1.5 1 0.5 0 Exhaustive Search PDPC−Q_R1_IL PDPC−Q_R1_CL CPC−Q_R1 1 2 3 4 5 6 7 Number of Femtocells (a) Aggregate femtocells capacity versus the number of femtocells. ) z H / c e s / s t i b ( y t i c a p a C s l l e c o t m e F e t a g e r g g A 2 1.9 1.8 1.7 1.6 1.5 1.4 1.3 1.2 1.1 4 3.5 3 2.5 2 1.5 1 0.5 0 0 Exhaustive Search pdpc(cid:237)Q_R1_IL_scratch pdpc(cid:237)Q_R1_IL_share pdpc(cid:237)Q_R1_CL_scratch pdpc(cid:237)Q_R1_CL_share 1345 1350 1355 1360 1365 1370 1375 500 Q(cid:237)iterations 1000 1500 (b) Aggregate femtocells capacity versus learning iterations. Fig. 1. Aggregate femtocells capacity using CPC-Q and PDPC-Q with R1 in both IL and CL paradigms. Nf = 29 at the 96000th iteration. Finally, we add another femtocell at the 99000th iteration. The figures show how DPC- Q using the CL paradigm is more robust to the deployment of new femtocells compared to the IL paradigm. Moreover, in these figures we compare the performance of DPC-Q to the docitive idea presented in [7]. We investigated two cases: 1) the already deployed femtocells share their Q-tables with the new femtocells when they first join the system (suffixed with share on the figure), 2) the new deployed femtocells starts with a zero initialized Q-tables (suffixed with scratch on the figure). Figure 2 shows the macrocell convergence on a certain subcarrier using DPC-Q with R1 in both IL and CL paradigms, where it can be observed that the CL paradigm maintains the macrocell capacity within the range of conver- gence (6 ± 1 bits/sec/Hz) and reacts well to the effect of the new deployed femtocells, without the need to have a learning phase again every time a new femtocell is deployed, which is a very interesting observation. It can also be observed that our proposed CL paradigm converges to the same value regardless the already deployed femtocells shared its Q-tables with the new ones or not. So, sharing could be ignored, thus decreasing the overall overhead. On the other hand, the IL paradigm showed a very bad reaction to the network dynamics, where 1) convergence is not attained (i.e. an oscillating behavior is generated), 2) as Nf increases, IL paradigm may push the macrocell capacity out of the convergence range when the network becomes more dense. Thus, CL is more scalable than IL. However, it can be noticed that the docitive idea is useful in the IL paradigm, where sharing the Q-tables of the already deployed femtocells with the new ones is much better (in terms of the value that the macrocell capacity oscillates around) than beginning with zero-initialized (scratch) Q-tables. In terms of speed of convergence, it can be noticed that, although the learning process may need large number of iterations initially, CL decreases the dynamics of the learning process, and hence, making it faster. This can be noticed from the figure, where ) z H / c e s / s t i b ( y t i c a p a c l l e c o r c a M 12 11 10 9 8 7 6 5 4 3 2 6.2 6.1 6 5.9 5.8 5.7 5.6 5.5 5.4 5.3 640 660 680 700 720 740 760 DPC−Q_R1_IL_share DPC−Q_R1_CL_share DPC−Q_R1_IL_scratch DPC−Q_R1_CL_scratch 0 100 200 300 400 500 600 700 800 900 1000 Q_iterations Fig. 2. Convergence of the macrocell capacity over the Q-iterations on a certain subcarrier where Nf was initially 5, then incremented until Nf = 30. ) z H / c e s / s t i b ( y t i c a p a c s l l t e c o m e F e a g e r g g A t 2 1.8 1.6 1.4 1.2 1 6 5 4 3 2 1 0 0 640 660 680 700 720 740 760 DPC−Q_R1_IL_share DPC−Q_R1_CL_share DPC−Q_R1_IL_scratch DPC−Q_R1_CL_scratch 100 200 300 400 500 600 700 800 900 1000 Q_iterations Fig. 3. Aggregate femtocells capacity over the Q-iterations where Nf was initially 5, then incremented until Nf = 30. CL converged almost at the 300th iteration, which is much earlier than the IL paradigm. Also, after the deployment of each new femtocell, CL took less than 10 iterations only - around 0.01 seconds - to re-achieve convergence. Figure 3 shows the aggregate femtocells capacity over the learning iterations. It can be noticed that using the CL paradigm, the aggregate capacity increases as more femtocells are deployed in the network, while in the IL paradigms, since convergence is already not maintained, the aggregate capacity behavior has a sporadic behavior, which indicates clearly that IL is not efficient to react to the network dynamics. However, in the CL paradigm from the 64000th to the 72000th iteration(640th to 720th according to figure’s scale), it can be noticed that the aggregate capacity decreases. The reason is that as more femtocells are being deployed, the network becomes very dense and since using the CL paradigm makes the cooperating femtocells use the same powers, this may force the macrocell capacity to violate the range of convergence. Thus, all the femtocells will have to decrease the power used to maintain again the macrocell capacity within the range of convergence leading to the decrease of their aggregate capacity. Note that at the 64000th and 72000th iterations, ǫ is already removed, which proves that R1 learns well even when ǫ is removed. Finally, in order to compare the aggregate capacity the CL paradigm achieves, after the incremental deployment of femtocells, to the ideal value, we used the small scale problem again. This is shown in figure 1(b), where we started with Nf = 2 and added an extra femtocell at the 8000th, 12000th and 13500th iterations4. Again, it can be observed that CL achieves aggregate capacity that is very close to the optimal one while the IL paradigm is far from it. 3In figures 2 and 3, ǫ is removed at the 50000th iteration and the figures were drawn with step = 100 in order to achieve better resolution. 4In figure 1(b), ǫ is removed at the 12000th iteration and the figure was drawn with step = 10 in order to achieve better resolution. VI. CONCLUSION In this paper, three Q-learning based power allocation algo- rithms for cognitive femtocells scenario are presented: namely, DPC-Q, CPC-Q and PDPC-Q. Although DPC-Q was pre- sented in previous work, in this paper it is extended, in both of its learning paradigms: IL and CL, to evaluate its performance, robustness and scalability. In terms of performance, DPC-Q is extended to PDPC-Q and then compared to CPC-Q, where the simulations showed that the CL paradigm outperforms the IL and achieves aggregate femtocell capacity that is very close the optimum one. In terms of robustness, the CL paradigm was found to be much more robust against the deployment of new femtocells during the learning process, where the results showed that the CL paradigm outperforms the IL paradigm in: 1) maintaining convergence, 2) learning better (i.e. reacting better to the network dynamics), especially when a suitable reward function such as the one defined in the simulations is used, 3) converging to the target capacity regardless the old femtocells share their experience (i.e. Q-tables) with the new deployed ones or not and 4) speeding up the convergence. Finally, in terms of scalability, CL paradigm reacted better to the network dynamics and maintained convergence, even when the number of the femtocells is large . ACKNOWLEDGMENT This work is supported by the Qatar Telecom (Qtel) Grant No.QUEX-Qtel-09/10-10. REFERENCES [1] V. Chandrasekhar, J. Andrews, and A. Gatherer, “Femtocell networks: a survey,” Communications Magazine, IEEE, vol. 46, September 2008. [2] S. Saunders, S. Carlaw, A. Giustina et al., Femtocells: Opportunities and Challenges for Business and Technology. Great Britain: John Wiley and Sons Ltd, 2009. [3] P. Xia, V. Chandrasekhar, and J. G. Andrews, “Open vs closed access femtocells in the uplink,” CoRR, vol. abs/1002.2964, 2010. [4] R. S. Sutton and A. G. Barto, Reinforcement learning: an introduction. Cambridge MA, MIT press, 1998. [5] J. C. H. Watkins and P. Dayan, “Technical note Q-learning,” Journal of Machine Learning, vol. 8, 1992. [6] J. R. Kok and N. Vlassis, “Collaborative multiagent reinforcement learning by payoff propagation,” J. Mach. Learn. Res., vol. 7, December 2006. [7] A. Galindo-Serrano, L. Giupponi, and M. Dohler, “Cognition and doci- tion in OFDMA-based femtocell networks,” in Proceeding of the IEEE Global Telecommunications Conference (GLOBECOM), may 2010. [8] A. Galindo-Serrano and L. Giupponi, “Distributed Q-learning for inter- ference control in OFDMA-based femtocell networks,” in Proceedings of the 71st IEEE Vehicular Technology Conference, 2010. [9] H. Saad, A. Mohamed, and T. ElBatt, “Distributed cooperative Q- learning for power allocation in cognitive femtocell networks,” in Proceedings of the IEEE 76th Vehicular Technology Conference, Sep. 2012. [10] A. Burkov and B. Chaib-draa, “Labeled initialized adaptive play q- learning for stochastic games,” in Proceedings of the AAMAS’07 Work- shop on Adaptive and Learning Agents (ALAg’07), May 2007. [11] F. S. Melo, “Convergence of Q-learning: A simple proof,” Institute Of Systems and Robotics, Tech. Rep., 2001.
1104.1905	2	1104	2011-08-12T08:14:41	A simulation of the Neolithic transition in Western Eurasia	[ "cs.MA", "q-bio.PE" ]	Farming and herding were introduced to Europe from the Near East and Anatolia; there are, however, considerable arguments about the mechanisms of this transition. Were it people who moved and outplaced the indigenous hunter- gatherer groups or admixed with them? Or was it just material and information that moved-the Neolithic Package-consisting of domesticated plants and animals and the knowledge of its use? The latter process is commonly referred to as cultural diffusion and the former as demic diffusion. Despite continuous and partly combined efforts by archaeologists, anthropologists, linguists, paleontologists and geneticists a final resolution of the debate has not yet been reached. In the present contribution we interpret results from the Global Land Use and technological Evolution Simulator (GLUES), a mathematical model for regional sociocultural development embedded in the western Eurasian geoenvironmental context during the Holocene. We demonstrate that the model is able to realistically hindcast the expansion speed and the inhomogeneous space-time evolution of the transition to agropastoralism in Europe. GLUES, in contrast to models that do not resolve endogenous sociocultural dynamics, also describes and explains how and why the Neolithic advanced in stages. In the model analysis, we uncouple the mechanisms of migration and information exchange. We find that (1) an indigenous form of agropastoralism could well have arisen in certain Mediterranean landscapes, but not in Northern and Central Europe, where it depended on imported technology and material, (2) both demic diffusion by migration or cultural diffusion by trade may explain the western European transition equally well, (3) [...]	cs.MA	cs	A simulation of the Neolithic transition in Western Eurasia Carsten Lemmen1,∗, Detlef Gronenbornb, Kai W. Wirtz1 aHelmholtz-Zentrum Geesthacht, Institut fur Kustenforschung, Max-Planck Strasse 1, 21501 Geesthacht, Germany bRomisch-Germanisches Zentralmuseum, Ernst-Ludwig Platz 2, 55116 Mainz, Germany 1 1 0 2 g u A 2 1 ] A M . s c [ 2 v 5 0 9 1 . 4 0 1 1 : v i X r a Abstract Farming and herding were introduced to Europe from the Near East and Anatolia; there are, however, considerable arguments about the mechanisms of this transition. Were it the people who moved and either outplaced or admixed with the indigenous hunter-gatherer groups? Or was it material and information that moved-the Neolithic Package- consisting of domesticated plants and animals and the knowledge of their use? The latter process is commonly referred to as cultural diffusion and the former as demic diffusion. Despite continuous and partly combined efforts by archaeologists, anthropologists, linguists, palaeontologists and geneticists, a final resolution of the debate has not yet been reached. In the present contribution we interpret results from the Global Land Use and technological Evolution Simulator (GLUES). This mathematical model simulates regional sociocultural development embedded in the geoenvironmental context during the Holocene. We demonstrate that the model is able to realistically hindcast the expansion speed and the inhomogeneous space-time evolution of the transition to agropastoralism in western Eurasia. In contrast to models that do not resolve endogenous sociocultural dynamics, our model describes and explains how and why the Neolithic advanced in stages. We uncouple the mechanisms of migration and information exchange and also of migration and the spread of agropastoralism. We find that (1) an indigenous form of agropastoralism could well have arisen in certain Mediterranean landscapes but not in northern and central Europe, where it depended on imported technology and material; (2) both demic diffusion by migration and cultural diffusion by trade may explain the western European transition equally well; (3) migrating farmers apparently contribute less than local adopters to the establishment of agropastoralism. Our study thus underlines the importance of adoption of introduced technologies and economies by resident foragers. Keywords: Europe, Linearbandkeramik, cultural diffusion, demic diffusion, agriculture, adaptation, migration, modelling 1. Introduction 1.1. Archaeology technological The transition to agropastoralism in western Eurasia between 10 000 and 3000 cal BC was associated with cultural, enormous and sociopolitical changes. Growing crops and herding animals have profoundly changed and continue to change global human history (e.g. Roth, 1887; Westropp, 1872; Diamond, 2002; Mithen, 2004; Barker, 2006; Ruddiman, 2006; Kaplan et al., 2010; Kutzbach et al., 2010). These changes may be viewed positively as a trajectory of progress in the way it had been seen by nineteenth and early twentieth century evolutionists (Westropp, 1872; Childe, 1936), or it may be seen as a road to perdition as it was, for instance, considered by J. Diamond (1997): " . . . a catastrophe from which we have never recovered. With agriculture came the gross social and sexual inequality." However the interpretation, the transition to agropastoralism, often termed Neolithisation, constitutes a major period of change in the history of humankind. ∗Tel +49 4152 87-2013, Fax -2020 Email address: [email protected] (Carsten Lemmen) Neolithisation is believed to have begun during the early Holocene in the Fertile Crescent, a mountainous region between the Levantine coast and the Zagros ridge (Flan- nery, 1973). Archaeobotanical, archaeozoological and ar- chaeogenetic work has demonstrated that all food crops and animals-except the dog-have their origins in and around the Fertile Crescent as a single founder region. The assemblage making up the Neolithic Package includes wheat, barley, rye, lentils, peas (Willcox, 2005), and cattle, sheep, goat and pigs (Luikart et al., 2001; Edwards et al., 2007; Larson et al., 2007; Zeder, 2008). While tendencies towards sedentism and storage of wild plants may already be interpreted from the Natufian data (Boyd, 2006), intensive cultivation and domestica- tion of both plants and animals gradually began during the Younger Dryas and only fully developed during the early Holocene (Zeder, 2008; Willcox et al., 2009). European agropastoralism is allochthonous and its most likely ori- gins are in the Fertile Crescent, where farming and herding began-still in mixture with a broad spectrum of foraging practices-during the tenth millennium cal BC (Flannery, 1973; Kuijt and Goring-Morris, 2002). Preprint submitted to Journal of Archaeological Sciences 31st May 2018 The wider expansion of agropastoralism started around 8500 cal BC, approximately 1000 years after the first ap- pearance of domesticated cereals in the Levant. The first clear evidence for colonist farmers was found on Cyprus (Peltenburg et al., 2000; Colledge et al., 2004; Willcox, 2005); the expansion ended after 4000 cal BC, when Neo- lithic sites emerged on the British isles and throughout northern Europe (Sheridan, 2007; Whittle, 2007). Details of the intermediate region specific accounts of transitions have been collected, for example, by Price (2000), Whittle (2007), and Gronenborn and Petrasch (2010), including the prominent sixth millennium linear pottery cultures of central Europe (LBK, e.g. Luning and Stehli, 1994) and the funnel beaker culture of the northern European plains (TRB, after 4500 cal BC, Midgeley, 1992). Not only did agropastoralism spread to the northwest from the Near East centres but also eastward as far as the Indus valley (Fuller, 2006). The question as to why agropastoral life style spread has been recently connected to environmental variations and conflict resolution: Dolukhanov (1973), Weninger et al. (2009) and Gronenborn (2009a) suggested an emergence and spread of farming as a result of climate induced crises periods, during which it may have become necessary for groups to fission, i.e. to move from one location to the other to escape conflicts. Two contrasting concepts on the mechanism of the spread of farming across western Eurasia have existed side by side. One suggests the introduction of the new agropastoral technologies through movements of people- migrations of any form; the other suggests a techno- logy shift through indigenous adaptations and inven- tions fostered by culture contact-information dispersal of any form. Zvelebil (1998) discriminates seven spreading modes, for example elite exchange or leap-frog colonisa- tion, as combinations or intermediate forms of the two opposite spread mechanisms. The acculturation or cultural diffusion model corpus has, in the more recent past, been applied by a num- ber of post-processual archaeologists-more typically for the British isles but also for the continent (Hodder, 1990; Thomas, 1991). It may go back to a critique by, for ex- ample, Zvelebil and Zvelebil (1988) of a migrationist model proposed by Renfrew (1987) and later by others (e.g. Bell- wood, 2005). In a way connected to these models are those where farming or animal husbandry were seen as regional developments within Europe. Such indigenist scenarios have been proposed for southern France, northwest Africa or Greece (e.g. Gedd`es, 1980; Courtin and Erroux, 1974; Winiger, 1998; Theocharis and Bokonyi, 1973). Opposing this position is the one of migration, where the new technology and cultigens arrive from Anatolia and the Fertile Crescent into Europe through migrating people. This migrationist position goes back to 1925, when V. G. Childe noticed a gradient in the spatiotem- poral distribution of ceramics from western Eurasia em- anating from the Fertile Crescent northwest into Europe. In a later publication, Childe (1942) suggested that popu- lation pressure in the source region was the driver of this outmigration. His position was supported by Ammerman and Cavalli-Sforza (1971, 1973), who formulated the 'wave of advance model', which was based on the concept of de- mic diffusion. Today, scholars from a number of disciplines favour migrationist models both for people and for culti- gens (Sokal et al., 1991; Richards, 2003; Pinhasi et al., 2005; Edwards et al., 2007; Bramanti et al., 2009; Balar- esque et al., 2010; Haak et al., 2010). Archaeology, partic- ularly continental European archaeology, sees evidence for more complex scenarios of migrations and local accultur- ation notably in western central Europe and France (e.g. Jeunesse, 2000; Bentley et al., 2002; Gronenborn, 2007b), but acknowledges that long-distance contacts across west- ern Eurasia did exist during the mid-Holocene and should, at least partly, have been maintained by the migration of people. It is yet unclear exactly when migrations began and what the relative importance of acculturation and movement of people was. It may not be ruled out that the large-scale population replacements around the Neo- lithic began with the onset of the sixth millennium cal BC (Gronenborn, 2007a, 2011). 1.2. Mathematical Models The spatiotemporal structure of the advance of farm- ing in Europe was first and very coarsely quantified by Edmonson (1961), who estimated the speed of the agro- pastoral transmission frontier at 1.9 km a−1. Later, Clark (1965) and Ammerman and Cavalli-Sforza (1971, 1973) based their analysis on radiocarbon dates (in three areas, at 53 sites, 103 sites, respectively) to calculate the velocity of the appearance of farming practice in Europe along a southeast-northwest gradient. All three studies found an approximately linear relationship between temporal and spatial distance of European Neolithic sites to four Near Eastern sites, with a slope of approximately 1 km a−1. Pin- hasi et al. (2005) confirmed this finding on a more exten- ded data set of 765 sites; they calculated a spreading rate between 0.8 and 1.3 km a−1; even when shortest-path dis- tances are considered (longer than great circle distances because of the detour necessary from the Near East to Anatolia and from central Europe to Iberia), a similar rate (0.6–1.1 km a−1) is found. A data set of 477 sites, including boreal European sites, was used by Davison et al. (2006, 2007, 2009) who simulate for a noninteracting agropas- toral subsistence style Neolithisation and its speed with a reaction-diffusion model. They too arrive at a mean speed of 1 km a−1 into Europe. Ackland et al. (2007) simulated the spread of farming by including a 'hitchhiking' advantageous trait in their reaction-diffusion model. This new lifestyle addition could either reflect the immigration of farmers, or a dynamical conversion of foragers into farmers upon contact. Their model predicts that Neolithic farmers outplaced the indi- genous population up to a line approximately connecting today's Venice–Prague–Warsaw–Moscow. To the north 2 In the following section, we shortly introduce the GLUES model and the radiocarbon site data which are used for validation; a full description of the algorithms used in GLUES can be found in the supplementary on- line material (SOM). The spatiotemporal pattern of the emergence and advancement of agropastoralism in west- ern Eurasia is reconstructed and analysed in detail. This is achieved through a model-data comparison for ten focus regions along a southeast-northwest trajectory, from the Levant to north Germany; model-based expansion rates are put into the context of prior estimates from radiocar- bon dates. A major part of our discussion concentrates on the discrimination of migration versus trade and the max- imum contribution of immigrants to emerging agropastoral communities in Europe. 2. Material and Methods GLUES mathematically resolves the dynamics of local human populations' density and characteristic sociocul- tural traits in the context of a changing biogeographical en- vironment. A local sociocultural coevolution is described by changes in mean population density, technology, share of agropastoral activities, and economic diversity, within a simulation region of approximately country-size extent (Figure 1). Each local population utilises its regional natural resources, which are described by vegetation pro- ductivity and climatic constraints. Each local population interacts with its geographical neighbours via trade and migration. The conceptual model is outlined below, for details on the algorithms used and the mathematical im- plementation we refer the reader to Wirtz and Lemmen (2003), summarised in the SOM. 2.1. Characteristic traits For pre-industrial human societies, we define three char- acteristic traits: 1. Technology is a trait which describes the efficiency of food procurement-related to both foraging and farming-and improvements in health care. In partic- ular, technology as a model describes the availability of tools, weapons, and transport or storage facilities. It aggregates over various relevant characteristics of early societies and also represents social aspects re- lated to work organisation and knowledge manage- ment. It quantifies improved efficiency of subsistence, which is often connected to social and technological modifications that run in parallel. An example is the technical and societal skill of writing as a means for cultural storage and administration, with the latter acting as a organisational lubricant for food procure- ment and its optimal allocation in space and among social groups. 2. A second model variable represents the share of farm- ing and herding activities, encompassing both animal Figure 1: Map of Europe with simulation region boundaries (solid lines) in the Global Land Use and technological Evolution Simulator (GLUES). Dots represent 631 radiocarbon dated Neolithic sites from Pinhasi et al. (2005). Our discussion focusses on a southeast to northwest transect along the highlighted regions A to J. and west of this line, the so-called converts adopted the new lifestyle by cultural diffusion. Available evidence, in part originating from isotopic and genetic studies, points to a discontinuous expansion se- quence for western Eurasia (Guilaine, 2001; Gronenborn, 2009a; Bocquet-Appel et al., 2009; Schier, 2009) during which short dynamic phases of long distance rapid expan- sions were followed by periods of stand-still with local or regional colonisation. Discontinuities in the Neolithic ad- vance, however, have not been hindcasted by the aforemen- tioned models. One possible reason is that in the frame- works provided by Ackland et al. (e.g. 2007) or Davison et al. (2006, 2009) important aspects are missing. These may comprise more detailed descriptions of the resources needed and used by the people, the influence of the local biogeographic suitability for farming or herding, and of temporal variation in resource availability. Their mod- els do not simulate any endogenous cultural, technical or agrarian development. All these factors may in principle accelerate or slow down the process of Neolithisation and lead to a more complex spatiotemporal pattern than may be predicted by simple reaction-diffusion models. In this study, we employ the Global Land Use and tech- nological Evolution Simulator (GLUES), which resolves local innovation, migration and cultural diffusion of traits (Wirtz and Lemmen, 2003; Lemmen, 2009; Lemmen and Wirtz, 2010). Although GLUES has been developed for the global domain, we restrict our analysis in this study to western Eurasia, where radiocarbon dates from Neolithic sites are abundant and of high quality, and where the is- sue of migration versus cultural diffusion is most intens- ively debated. The model is chosen because it allows us to differentiate between exchange (i.e. information exchange as cultural diffusion) and migration (demic diffusion) as important vectors of the expansion of agriculture. 3 ABCDEFGHIJhusbandry and plant cultivation. It describes the al- location of energy, time, or manpower to agropastoral- ism with respect to the total food sector. We define a local population as Neolithic when this share is larger than the share of foragers-regardless of its techno- logy, economic diversity, or population density. 3. Economic diversity resolves the number of different agropastoral economies available to a regional popu- lation. This trait is closely tied to regional vegeta- tion resources and climate constraints. We do not, however, attribute specific plants and animals to each economy. As an example, a value of four would be ob- tained when (1) domestic pigs and (2) goats and the growing of (3) barley and (4) wheat were present in a given population. A larger economic diversity offering different niches for agricultural or pastoral practices enhances the reliability of subsistence and the efficacy in exploiting heterogeneous landscapes. The temporal change of each of these characteristic traits follows the direction of increased benefit for suc- cess (i.e. growth) of its associated population; this concept had been derived for genetic traits in the works of Fisher (1930), and was recently more stringently formulated by Metz and colleagues (Metz et al., 1992; Dieckmann and Law, 1996; Kisdi, 2010) as adaptive dynamics (AD). In AD, the population averaged value of a trait changes at a rate which is proportional to the gradient of the fitness function evaluated at the mean trait value. The AD ap- proach was extended to functional traits of ecological com- munities (Wirtz and Eckhardt, 1996; Merico et al., 2009; Smith et al., 2011), and was first applied to cultural traits of human communities by Lemmen (2001) and Wirtz and Lemmen (2003). 2.2. Local resources Each simulation region is defined by a largely homo- geneous vegetation productivity (measured as net primary productivity, NPP), resulting in an average size of 130 · 103 km2 (Figure 1). We reconstruct past distributions of NPP with a global climate model coupled to a vegeta- tion module. Climber-2 (Claussen et al., 1999) temperat- ure and precipitation anomalies from the IIASA climato- logical data base (International Institute for Applied Sys- tems Analysis, Leemans and Cramer, 1991) are converted to NPP according to the climate constraints on NPP from Lieth (1975); we do not use soil maps to constrain veget- ation productivity. From NPP, both the regional utility of natural food re- sources and the number of potential domesticates are de- rived. According to Braidwood and Braidwood's (1949; 1950) hilly flanks hypothesis, potential domesticates were most abundant in open woodlands at low to intermediate NPP. The number of potential domesticates furthermore depends on a continental aggregation to account for the area-biodiversity relationship (e.g. Begon et al., 1993). 2.3. Exchange of information and people between regions Information exchange and migration are vectors of the spread of technology, economic diversity, and farming practice from the founding centres to adjacent simulation regions. We discriminate the diffusion of traits without involving resettlement of people (cultural diffusion by in- formation exchange), and the diffusion of traits via mi- gration (demic diffusion). In GLUES, both mechanisms are driven by differences in influence between neighbour- ing local populations. We assume that information travels two orders of mag- nitude faster than people. Exchange networks extend over distances of up to 1000 km, in the later Mesolithic and Neolithic (Mauvilly et al., 2008; Gronenborn, 1999); these networks were crossed many times during the active time-say ten years-of a Neolithic trader. Within this time span, a migration model like the one by Ammerman and Cavalli-Sforza (1973), would allow for an advance of only ten km. This parameterisation leads to diffusivities for migration on the order of ten km2 a−1, a value which is comparable to the diffusivities employed by other model studies of demic diffusion (e.g. Davison et al., 2006; Ack- land et al., 2007; Patterson et al., 2010). 2.4. Reference data and simulated time scale Our reference data set is the comprehensive data col- lection of 765 sites by Pinhasi et al. (2005). These au- thors used site data provided by the United Kingdom Ar- chaeology Data Service, the Central Anatolian Neolithic e-Workshop (CANeW), the radiocarbon CONTEXT data- base, and the Radiokarbondaten Online (RADON) data- base. In their compilation, they included only sites with small dating uncertainty (< 200 a); they report dates as calibrated calendar years before present (relative to 1950) based on calibration of original 14C measurements with CalPal 2004. This data set was created by the Pinhasi et al. to provide a high quantity of dates and good spa- tial coverage at the expense of chronologic uncertainties, which could have been avoided, if for example only AMS- dated (accelerator mass spectroscopy) samples had been used; only few of the 765 sites, however, have been AMS dated. For our purpose, this data set with many (possibly uncertain) dates represents the expansion of agropastoral- ism at a satisfactory level of detail. Future simulations at a refined spatial scale would benefit from a data set with better chronologic control, where local and regional events are presented in higher resolution and where the region- ally patchy nature of the expansion of agropastoralism is better represented. From this data set, we choose for comparison those 631 sites which are located in the spatial model domain (10◦W–42◦E and 31◦N–57◦N) and the period of interest (8000–3500 cal BC). For each site, we use the age range computed from the reported calibrated radiocarbon age and the reported standard deviation. 4 For the mathematical model, we introduce the age scale 'simulated time BC' (sim BC) to distinguish between em- pirically determined age models and the model time scale. Ideally, sim BC should be numerically equal to cal BC. We set up the eight global model parameters such that the simulation is able to hindcast an accurate timing and location of the early farming centres Fertile Crescent, northern China, and Mesoamerica (Smith, 1997), and a reasonable global pattern of the subsequent Neolithisation. The simulation is started at 9500 sim BC. All of the 685 biogeographically defined regions (including 71 in western Eurasia) are initially set with farming activity at 4% and established agropastoral communities at 0.25, what repres- ents a low density Mesolithic technology population and a broad spectrum foraging lifestyle with low unintentional farming activity. The latter is assumed to represent early animal harvesting, selective seed gathering, and the active use of fire. 3. Results In the GLUES simulation, farming originates in the Le- vant (focus region A, cmp. Figure 1) around 7000 sim BC and penetrates into Europe in a northwest direction. By 3500 sim BC all of continental Europe has converted to farming as the predominant subsistence style. This emer- gence of farming in western Asia and Europe is shown as a series of snapshots in Figure 2 (see SOM for an animated version with finer temporal resolution). 3.1. Expansion of agropastoralism The initial development progresses slowly and at a low level. It begins during the first century of the seventh millennium sim BC in a region encompassing today's Le- banon, coastal Syria and a small part of the adjacent coastal Anatolia. In the 67th century sim BC, northern Greece converts to agropastoral subsistence with rapid extension into the central Balkan. Over the next four hundred years, these agropastoral nuclei spread out fur- ther, encompassing the whole of Greece and the south- ern Balkan, and the coast of Anatolia by the 63rd cen- tury sim BC. A rapid expansion of agropastoralism occurs between 6200 and 6000 sim BC, transforming the entire Balkan re- gion and Anatolia. By 5750 sim BC, the new subsistence mode has reached the northwestern and the easternmost coasts of the Black Sea. In the 57th century sim BC, in- dependent agropastoralism arises in north Africa in the region around the Strait of Gibraltar, and it emerges on the Italian peninsula. The 55th century sim BC sees a rapid expansion of farm- ing and herding into the area of the central LBK, and its spread into the south coast of the Iberian peninsula. By the 54th century, the LBK has expanded west- and east- wards and covers a vast stretch of land from southern Ger- many to the Ukraine; this central–eastern European area Figure 2: The spread of agropastoralism in western Eurasia from 7500 sim BC (top left, then downwards) until 3500 sim BC, hindcas- ted with the Global Land Use and technological Evolution Simulator. A finer temporal resolution (50 a time step) animation of this evolu- tion is available in the SOM. intensifies agropastoral activity without notable expansion until 5100 sim BC. Around 5000 sim BC, forager societies on the north coast of the Black Sea, in north Africa and on the Iberian penin- sula have converted to predominantly agropastoralism. At 4750 sim BC, the Neolithic package reaches the Baltic Sea at the Oder river mouth; this coastal agropastoralism ex- pands eastward until 4500 sim BC and resembles the rise of the eastern TRB culture. By this time, farming and herding have-in the model-reached the south coast of France and the north coast of Portugal. An agropastoral area resembling the western TRB ap- pears by 4400 sim BC, also in southern Germany the new life style becomes dominant. The later half of the fifth millennium sees a slow expansion towards the northeast of Europe, and the gap closure in central and northern France. After 4000 sim BC, agropastoralism reaches the British isles. 5 Figure 3: Timing of the transition to agropastoralism in Western Eurasia. The simulated transition (background pastel shading) is contrasted with the radiocarbon ages of Neolithic sites from Pinhasi et al. (2005, solid colour triangles). The lower right inset image shows the transition for a scenario without migration or exchange, i.e, it shows the propensity of regions to endogenously develop agropastoralism. 3.2. Timing of agropastoralism in model and data A summary description of the timing of agropastoralism between 7500 and 3500 sim BC is illustrated by Figure 3; shown alongside are the median radiocarbon dates of Neo- lithic sites within this period from the data compilation by Pinhasi et al. (2005). From this time-integrated perspect- ive, the simulated centres of agropastoralism in the Fertile Crescent, in northern Greece and at the Strait of Gibral- tar are evident, as well as the southeast to northwest tem- poral gradient of the Neolithic transition. The model-data comparison shows many good matches between radiocar- bon dates and simulated transition dates. We can clearly see, however, the spatial scale difference between simula- tion region and site data. The spatial distribution of ra- diocarbon dated sites has good coverage along the transect from the Levant to northwestern Europe discussed below, it provides few or no information on eastern Europe, on the Iberian peninsula, and in north Africa. To assess the quality of the simulated onset of agropas- toralism, we compare in Figure 4 the change in fractional agropastoralism to the radiocarbon site statistics for ten focus regions A–H (a transect from the the Levant A to north Germany H) and radiocarbon dated sites within the region, or within 200 km distance of the region centre for small regions. We also indicate by colour selected cultural attributions. Seventeen sites within or near region A are dated between 8000 and 5500 BC, of which the most frequent cultural attribution is Pre Pottery Neolithic (PPNX, 9 sites) and Pottery Neolithic (6 sites). The most fre- quent century is the 68th cal BC (4 sites); the simulated change in agropastoral activity is greatest in the 70th cen- tury sim BC. Ten sites are found in or near region B, most of which are assigned to the Pottery Neolithic (6 sites). All sites date to before 6000 cal BC, with a maximum around 6400 cal BC; the largest simulated change to farming occurs in this region and in region C, around 6300 sim BC. Near or in C, 15 sites cover a wide temporal range from 6500 to 4400 cal BC. The site statistic within or around region D is poor with only six sites, which date to 5800–5000 cal BC. The timing of the largest simulated change is 6100 sim BC; this simulated transformation resembles the occurrence of the Koros culture. Like region D, most sites near region E are attributed to Koros; the second most frequent cultural complex in re- gion E is the LBK, which is also the dominant attribution at sites around regions F to I. In regions H to J, the site his- togram is bimodal with the latter peak assigned to funnel beaker sites. Many site dates near region E fall within the period 6000–4800 cal BC, whereas the simulated change is greatest at 6100 sim BC; around region F, the simulated transition occurs around 5800 sim BC. The most frequent date is the 52nd century cal BC, with a large range of 1500 years. Radiocarbon dates for region G range from 6600 to 3500 cal BC; a maximum occurs between the 53rd and 47th century, which is concurrent with the largest sim- ulated change at 5200 sim BC. Seven LBK sites around re- gion H are dated to 5600–5000 cal BC, coterminous with 6 TimingofNeolithictransitionshading:simulationregionssymbols:radiocarbondatedsites35004000450050005500600065007000calendaryearsBCsimulationyearsBCNoexchange,nomigrationscenarioFigure 5: Agricultural onset timing versus distance from an assumed archaeological centre near Beirut (Lebanon) in the GLUES simula- tion and in the data set by Pinhasi et al. (2005). The correlation of distance and timing in the simulated data for n = 66 regions is r2 = .39 (large circles, significant at p = .01), the average slope is 0.81 km a−1. For n = 631 sites, the correlation is r2 = .61 (blue dots, significant at p = .01), the slope is 0.72 km a−1. The solid lines show for each distance interval of 500 km the probability of the earliest and farthest occurrence (p = 0.05 percentile) of agropastoralism in the site dates (blue) and the simulation (red). radiocarbon dates between 5800 and 4600 cal BC (incl. 9 LBK sites) and the latter dates from 4200 cal BC (incl. 10 TRB sites). 3.3. Time lag–distance relationships The average speed of the expansion of agropastoralism can be estimated from the time lag–distance relationship relative to an assumed founding centre of agropastoral- ism (e.g. Ammerman and Cavalli-Sforza, 1973). Figure 5 shows this relationship for all regions and radiocarbon dated sites within the model domain. Here, the assumed agropastoral centre is near today's Beirut, which lies in the middle of focus region A. The (great circle) distances and time differences to this assumed centre extend over 4000 km and 4000 a, respectively. Time lag and distance from the sites are highly cor- related (r2 = .61) but also indicate a stair-case like distribution around a linear regression line with slope 0.72 km a−1(cmp. Guilaine, 2001; Gronenborn, 2009a; Schier, 2009). This means that in the data collection the spread of Neolithisation is slower in spatial proximity of the founder region than is predicted by a linear correla- tion; between 6000 and 4500 cal BC, however, the graph of the majority of sites lies above the regression line, which Figure 4: Timing of the transition to farming in GLUES and cal- ibrated radiocarbon date statistics for the ten focus regions (map inset). In each panel A–I (ordered from southeast to northwest, bottom to top), the change in the fraction of agropastoralism from GLUES (grey shading) is contrasted to the number of radiocarbon dates compiled by Pinhasi et al. (2005) (bars, including 1σ uncer- tainties of the age determination). Colour indicates selected cul- tural attributions within the site record to pre-pottery Neolithic (of any kind, PPNX ), Koros, Linear pottery culture (LBK), and fun- nel beaker culture (TRB). To account for the scale difference, the simulated transition was broadened in time by a convolution with a normal distribution with mean zero and standard deviation propor- tional to region area and diffusion length. the simulated shift at 5500 sim BC. For region I, with mostly LBK-attributed sites and radiocarbon dates (5200– 4400 cal BC), the simulated subsistence change culminates around 4600 sim BC. The respective model transition for region J appears at 4400 sim BC. Here, the site histogram can be divided into two modes, where the first encompasses 7 −8000−7500−7000−6500−6000−5500−5000−4500−4000−350000.21J−8000−7500−7000−6500−6000−5500−5000−4500−4000−350000.21simulatedtransitionsimBC−8000−7500−7000−6500−6000−5500−5000−4500−4000−350000.21radiocarbondistributioncalBC−8000−7500−7000−6500−6000−5500−5000−4500−4000−350000.21I1-2-3-4-5-#sitesTRBsites−8000−7500−7000−6500−6000−5500−5000−4500−4000−350000.21H−8000−7500−7000−6500−6000−5500−5000−4500−4000−350000.21GLBKsites−8000−7500−7000−6500−6000−5500−5000−4500−4000−350000.21F−8000−7500−7000−6500−6000−5500−5000−4500−4000−350000.21E−8000−7500−7000−6500−6000−5500−5000−4500−4000−350000.21DKorossites−8000−7500−7000−6500−6000−5500−5000−4500−4000−350000.21C−8000−7500−7000−6500−6000−5500−5000−4500−4000−350000.21B−8000−7500−7000−6500−6000−5500−5000−4500−4000−350000.21 10oW 0o 10oE 20oE 30oE 40oE 35oN 40oN 45oN 50oN 55oN 10oW 0o 10oE 20oE 30oE 40oE 35oN 40oN 45oN 50oN 55oN ABCDEFGHIJA4000BC4500BC5000BC5500BC6000BC6500BC7000BC7500BCPPNXsites350040004500500055006000650070007500800005001000150020002500300035004000Distance from Levante (km)Time (year BC) RadiocarbonsitesSimulatedregionalshifts·Sites/regionsLinearregression5%timingpercentileindicates a more rapid wave of advance from the Balkan towards central Europe. Lag and distance for GLUES- simulated regions are also correlated to a marked degree (r2 = .40) and are similarly scattered around the regres- sion line. Of the ten focus regions, regions A and B develop more slowly than expected from the regression and regions E–G develop agriculture faster than the linear regression. The average speed for the expansion of agropastoralism from the Levant into Europe calculated from the model is 0.81 km a−1. 4. Discussion The Global Land Use and technological Evolution Simu- lator is able to hindcast a realistic spatiotemporal pattern of the introduction of farming and herding into Europe between 8000 and 3500 sim BC. The simulated expansion speed of agropastoralism compares well to a large dataset of radiocarbon dated Neolithic sites; the inhomogeneous spatial distribution of Neolithisation is reproduced. 4.1. spatiotemporal onset and expansion of agriculture The differences we observe between simulated timing and the radiocarbon age of sites within a simulation re- gion (Figures 3 and 4) are less than 1000 a for almost all sites, for the majority of sites less than 500 a; only a hand- ful of sites show differences greater than 1000 a. These differences are similar in magnitude to those obtained by Davison et al. (2007) between their numerical model and radiocarbon dated sites in Europe. At this scale of model uncertainty, the radiocarbon dating uncertainty of indi- vidual sites (< 200 a) can be neglected: the mismatch between the onset definitions in the data (presence of a Neolithic site, Figure 4) and in the model (50 % agropas- toral activity), as well as the spatial scale mismatch (local site data versus country-size simulation region, Figure 3) introduce larger temporal differences. To overcome the spatial scale problem, Zimmermann (2004) argued for a landscape approach to archaeology, whereby a multitude of local sites are used to infer the archaeological context at the regional scale or larger. The landscape approach can only succeed, however, if many sites within a region are excavated, as was the case for the lignite mining area of the Aldenhovener Platte studied by Luning and Stehli (1994) and Zimmermann (2004). From a model perspect- ive, more studies on methodologies to scale up the site (or many sites) information to the landscape are highly desirable. The scale difference illuminates the resolution limits of our model: GLUES resolves societal dynamics in larger environmental contexts rather than the history at individual sites. We find a marked correlation between the timing of first agropastoralism and the distance from a founding centre in the model (r2 = .40), and an average speed of agropastor- alism in western Eurasia of 0.81 km a−1. Using radiocar- bon data, a marked or high correlation was also found by 8 Gkiasta et al. (2003, r2 = 0.53, n = 510), Ammerman and Cavalli-Sforza (1971, r2 = 0.79, n = 103), and Pinhasi et al. (2005, r2 = 0.64, n = 765). Differences between these empirical results can be attributed to the number of sites under consideration, to the location of the assumed found- ing centre, to site selection, or to the consideration of the shortest land route versus great circle distance (Pinhasi et al., 2005). For calibrated dates, Pinhasi et al. calculate a speed range of 0.6–1.3 km a−1 when these differences are taken into account. The validity of comparing the onset of agropastoralism between simulated regions (with large areal extent) and (local) radiocarbon dated sites, despite the different scales, is supported by our simulation in two ways: (1) the marked correlation obtained between lag and distance of first agropastoralism to an assumed founding centre; (2) a calculated speed of 0.81 that agrees closely with other published estimates. Our simulations do not take coastal expansions into ac- count, which would seem a major model deficiency at first glance. The independent Moroccan model centre, how- ever, acts for the Iberian peninsula similar as an explicit fast migration process (like leapfrogging, on the order of 20 km a−1, Zilhao, 1993; Zilhao, 2000) along the Mediter- ranean coast and islands. GLUES does not currently ac- count for rivers: we attribute the late transition of north- ern France in the simulation to a missing pathway from the Mediterranean coast through the Rhone valley. Indeed, Davison et al. (2006) found in their model a significant role of waterways in the Neolithisation of Europe, whereas our results only indirectly (through the definition of regions by homogeneous vegetation) include river basins; GLUES performs well despite the lack of explicit river pathways like the Rhine or Danube valleys for all regions of Europe except central France. In the simulation, as well as in the data, the expansion of farming occurs in stages with periods of rapid spread followed by periods of local intensifications. The rapid Neolithisation from Greece to the central Balkan in the 67th century sim BC is followed by a 400 a period of rel- ative stagnation. A very similar pattern is hindcasted for the LBK-like Neolithisation in the 55th and 54th century cal BC, and for the relative stagnation before the onset of a TRB-like Neolithic further north. Several regions exhibit a slow conversion to agropastoralism: according to Bellwood and Oxenham's (2008) classification of zones in Neolithic Europe, France (see discussion of rivers above) would rep- resent rather a friction than a spread zone. Another fric- tion zone, where the Neolithic is introduced gradually, ex- ists in the northern European lowlands, where empirical data supports the simulated late arrival of farming (Midge- ley, 1992; Zvelebil, 2006; Hartz et al., 2007). Our simulations predict a second and eastern expansion path around the Black Sea, which was archaeologically suggested by Kotova (2003, 2009). From a model per- spective, Davison et al. (2007, 2009) suggested that to find a suitable simulation consistent with the radiocarbon dated site context, one needed an additional early wave of advance emanating around 8200 cal BC from an East- ern European centre; agropastoral expansion would then proceed via the steppe corridor and be responsible for the eastern version of the Neolithic in Europe. This steppe cor- ridor is archaeologically visible from the spread of pottery across Eurasia and the expansion of farming from east- ern Anatolia into the Caucasus (Dolukhanov et al., 2005; Gronenborn, 2009b; Kotova, 2003, 2009). In the reaction- diffusion migration model by Ackland et al. (2007), a cir- cular expansion from a single Mesopotamian centre is sim- ulated; the geographical bottlenecks between the Mediter- ranean and the Black Sea, and between the Black and the Caspian sea act as secondary wave centres and thus the ar- rival of the farming wave in Europe "appears to come from two sources: north and south of the Black Sea" (Ackland et al., p. 8715). GLUES exhibits the same behaviour, with a southern and eastern path around the Black Sea based on a single Mesopotamian source area; for central Europe, however, our model suggests that the secondary centre is rather located in northern Greece and the central Balkan. Archaeologically, the land route bottleneck may have not been too important for the Mediterranean coast, where the many islands provided a fast sea route from the Levant to Cyprus, Greece, and as far as Portugal (Peltenburg et al., 2000; Theocharis and Bokonyi, 1973; Arias, 1999; Zilhao, 2000). A separate non-Eurasian independent centre of agropas- toralism is simulated by GLUES in the Maghreb. From there, agropastoralism enters Europe via the Strait of Gibraltar around 5500 sim BC (cmp. Manen et al., 2007). Archaeological records to show this are sparse; it is clear, however, that the strait had been in use as a migra- tion path long since pre-Neolithic times, which is evid- ent in gene pool analyses (e.g. Currat et al., 2010). Not only people but also domesticates crossed the strait in prehistory. This was verified in a study by Anderung et al. (2005), who found mitochondrial DNA of Bronze age (1800 cal BC) Iberian cattle, of which a significant number possessed African haplotypes. 4.2. Demic or cultural diffusion The relative contribution of demic versus diffusive pro- cesses can be calculated by following the streams of mi- gration and trade in the model. We find that exchange processes contribute much less to local Neolithisation than adoption does. In Figure 6, we show this for the fraction of immigrant agropastoralists: for some (mostly Mediter- ranean) regions, immigrants are unimportant; for most re- gions, immigrants constitute one fourth of the agropastoral community; for a few northern and alpine regions, immig- rants dominate. We assessed the model sensitivity to different config- urations for the speed of both information exchange and migration. We find that (1) the model does not show sens- itivity to either of these two parameters for a wide range of values, and that (2) the substitution fraction of agropastor- alists with immigrants generally remained around 25% for Figure 6: Simulated fraction of immigrants in the agropastoralist population of each region at the time when the transition is loc- ally 90% complete. Green colour indicates mainly local adoption by resident foragers, yellow to red a major contribution of immigrants. most regions; local invention and adoption of ideas dom- inate the Neolithisation, irrespective of whether people or information moved. Is demic diffusion a sufficient explanation for the European Neolithic? This has been suggested by many authors (Ammerman and Cavalli-Sforza, 1971, 1973; Sokal et al., 1991; Richards, 2003; Edwards et al., 2007; Bramanti et al., 2009; Balaresque et al., 2010; Haak et al., 2010). We can confirm this finding from a model simulation where cultural diffusion was deactivated. Demic processes alone can reproduce the timing and lag-distance relationship seen in the radiocarbon data. But is demic diffusion necessary for explaining the ra- diocarbon record? Pinhasi et al. (2005) answered this question positively and pointed out that, at present, no working model existed that could explain the European Neolithic without demic diffusion. Alike the demic diffusion-only experiment, we set up a simulation experi- ment where only information was allowed to diffuse, and where migration was inhibited: we could successfully re- produce the spatiotemporal emergence of the Neolithic in Europe with purely cultural diffusion processes. To the question of necessity of demic diffusion for the ex- planation of the radiocarbon record, our answer is no. This shows that the recently published evidence for ma- jor population transfers around the time of the Neolith- isation process (Haak et al., 2010) may have to be func- tionally disconnected from the spread of agropastoralism: apparently, people did move at greater scales during the sixth and fifth millennium but these movements were not triggered by the spread of farming. It may entirely be pos- sible that early-yet hypothetical-migrations already oc- curred before and during the seventh millennium and were 9 10oW 0o 10oE 20oE 30oE 40oE 35oN 40oN 45oN 50oN 55oN 10oW 0o 10oE 20oE 30oE 40oE 35oN 40oN 45oN 50oN 55oN 10oW 0o 10oE 20oE 30oE 40oE 35oN 40oN 45oN 50oN 55oN 10oW 0o 10oE 20oE 30oE 40oE 35oN 40oN 45oN 50oN 55oN 0.10.20.30.40.50.60.70.80.91FHPercentageofimmigrantfarmersandherders 10oW 0o 10oE 20oE 30oE 40oE 35oN 40oN 45oN 50oN 55oN 10oW 0o 10oE 20oE 30oE 40oE 35oN 40oN 45oN 50oN 55oN 10oW 0o 10oE 20oE 30oE 40oE 35oN 40oN 45oN 50oN 55oN 10oW 0o 10oE 20oE 30oE 40oE 35oN 40oN 45oN 50oN 55oN 0.10.20.30.40.50.60.70.80.910%25%50%75%100%Region F Region H agropastoralism agropastoralism Region H technology migr:adopt migr:adopt migr:exch:adopt demic mixed cultural 54:46 22:78 0:100 85:15 41:59 0:100 22:0:78 6:13:79 0:21:79 Table 1: Contributions (in percent) to local agropastoralism and technology from three different sources (1) demic diffusion (labelled migr), (2) cultural diffusion (exch), and (3) local adoption and in- vention (adopt) for three different model configurations with demic only, mixed (our standard configuration discussed in Figures 2–6), and cultural only diffusion. Simulation results are shown for model regions corresponding to today's Hungary (focus region F) and south- ern Poland (focus region H). undertaken by hunter-gatherers or mixed hunter-gatherer- horticulturalists originating from Anatolia. These immig- rants then gradually pushed the original Mesolithic hunter- gatherer population of Europe towards the continental margins. Later, this migratory stream was complemen- ted with farmers from Anatolia who then interacted with those hunter-gatherers who had arrived earlier (Gronen- born, 2011). This scenario would explain the archaeo- genetic evidence for migrations as well as the archaeolo- gical evidence of interactions by disconnecting the spread of farming from the mid-Holocene migratory processes. The insensitivity of the simulation results to the absence of either trade or migration processes prevents us from constraining the parameters for these processes quantitat- ively. Even more, it tells us that from the phenology (the timing of agropastoralism) we cannot infer which of the two processes was responsible, or to which degree. For the interpretation of radiocarbon dates of sites with an attri- bution to farming subsistence, one cannot find out whether demic or cultural diffusion was responsible for the appar- ent distribution in space and time of these sites. Or, put differently, the question on demic or cultural vectors may be not the most critical one for understanding the Neo- lithisation of Europe as a whole. What was the contribution of local adoption and inven- tion? From Figure 6 it is evident that for most regions conversion of resident foragers to farmers played a larger role than immigration. We quantitatively examined the relative importance of different sources (demic diffusion, cultural diffusion and local adoption or invention) to local Neolithisation in different model configurations (Table 1). Even in a scenario where demic diffusion is the only active process, migration as a source does not explain 100% of the agropastoralists in any focus region along the transect A–H but at most 85% (in region H), less in regions closer to the Mediterranean coast (e.g. 54% in region F). The local source (adoption and innovation) for an exchanged commodity like technology is in all configurations and fo- cus regions more important (70–90%) than migration or exchange. 4.3. Independent agropastoralism Was regional exchange (via migration or trade) neces- sary at all for the onset of agropastoralism everywhere in western Europe? In Figure 3 (lower right panel) we show the result of a simulation where both exchange processes were suppressed, thus endogenous transitions to agropas- toralism become visible. The timing of the onset of agro- pastoralism around the Mediterranean Sea exhibits-next to the Levantine, Greek, and Moroccan centres which also appear in the reference simulation-many centres of hypo- thetical independent agropastoralism. This independent agropastoralism is solely predicted on the basis of suit- able environmental conditions (open vegetation type, not too cold) and internal development of sociocultural traits and demography; it corresponds to indigenist scenarios proposed in the older literature, for example for south- ern France (Gedd`es, 1980; Courtin and Erroux, 1974) or Greece (Theocharis and Bokonyi, 1973; Winiger, 1998). The indigenous agropastoralism hypothesis is, however, currently disregarded in archaeology because the genetic evidence points to the Near East as the centre for all Neolithic cultigens and nearly all domestic animals. For northern Europe, GLUES does not simulate the emergence of agropastoralism without the contribution of migration; these regions critically depended on the introduction of technologies and economies through the actual movement of people, commodities, and information. 4.4. Model comparison and outlook In addition to the prior approaches to simulating the European Neolithic (Davison et al., 2006, 2007, 2009; Ack- land et al., 2007) which use geographic and topographic constraints for describing environmental heterogeneity, our model considers vegetation. Vegetation production is dir- ectly coupled to the carrying capacity and it determines the economic potential of a given environment. Already many regions, mainly around the Mediterranean Sea, have a high propensity for developing independent agricultures based on the palaeoecological background (Figure 3, lower right panel). We couple the diffusion rates of traits and migration of people not only to the background geography but also to the (evolving) technology. With these as- sumptions we can realistically reproduce the spatiotem- poral pattern of Neolithisation in greater detail than was done for the front speed of Neolithisation by Davison et al. (2006, 2007) and Ackland et al. (2007). Ackland et al. (2007) and more recently Patterson et al. (2010, for the Indian transition to agropastoralism) use the concept of converts to describe resident foragers which have converted to agropastoralism. We have shown that these converts may have played a larger role in the European Neolithisation than immigrant farmers. Our model provides additional insight into the processes re- sponsible for local adoption-often, a small share of in- troduced technology is sufficient to spark local invention 10 and trigger the transition. Alternatively, a few immig- rant farmers and the technologies and economic possibil- ities they carry along may suffice to stimulate the local transition. Our regional prediction for western Eurasia emerges in the context of a global simulation: not only is the subcon- tinental prediction embedded in the larger spatial scale, but every local transition occurs within the temporal con- text of preceding predominant foraging subsistence with continuous innovation and succeeding intensification peri- ods. While we have shown that the model realistically reproduced the European Neolithisation, where archeolo- gical data is plenty and most reliable, the model's spati- otemporal consistency gives us confidence to draw conclu- sions about regions outside Europe in further studies. With the expected availability of more reliable palaeo- climate and palaeovegetation reconstructions from both models and data (Kutzbach et al., 2010; Gaillard et al., 2010), we expect to refine the large-scale biogeographic context of cultural evolution and the impact of local en- vironmental disturbances (Wirtz et al., 2010). We will then be able to assess better the degree to which the en- vironment determined the potential transition to farm- ing. This potential should, however, be interpreted in G. Ackland et al. (2007)'s way as providing a "historical null hypothesis. Its predictions can be taken as requir- ing no special explanation, and its failures can be taken as evidence of rare events that had significant and long- lived consequences". Numerical modelling of culture as a (natural) ecosystem may help to isolate the significant and non-deterministic events and concentrate our histor- ical interpretation on those events where culture was most emancipated from the environment. 5. Conclusion We presented a spatially explicit mathematical model of the Neolithisation of western Eurasia from 8000 BC to 3500 BC. Our model incorporates endogenous sociotech- nological dynamics, where culture is represented by the adaptation of characteristic population traits (technology, fraction of farmers, and economic diversity) and their in- teraction with demographics. The study resolved the spa- tial expansion of Neolithic culture via indigenous develop- ment, migration and information exchange and reproduced the chronology of agropastoral onset observed in field data across western Eurasia, particularly reproducing and ex- plaining the discontinuous speed of the 'wave of advance'. Our results encourage us to rethink possible indigenous centres along the northeastern shore of the Mediterranean: these might have not been able to develop since they were overrun by Near Eastern populations. Alternatively, the evidence for independent agropastoralism may have gotten lost in the admixture with the Fertile Crescent Neolithic package. According to our simulations, a north African contribution to the European Neolithic should equally not be discounted. The assessment of the relative importance cultural dif- fusion and demic diffusion in the model shows that either of these processes can explain the spatiotemporal pattern of agropastoral onset in Europe equally well. The phen- ology of the spatiotemporal pattern of agropastoral onset cannot discriminate the underlying process. Furthermore, even if only migration was considered and the diffusion of traits occurred only via immigrants, the prevalence of immigrant farmers in any of the emerging agropastoral re- gions was much less than the prevalence of foragers who adopted the agropastoral life style. To the long-lasting dis- pute between cultural and demic diffusionists our novel in- terpretation offers a balanced explanation of predominant adoption despite migration. Whether the adopting pop- ulation, however, did not also ultimately originate from Anatolia needs to be investigated by further archaeogen- etic studies. Acknowledgments We acknowledge the financial support for C.L. re- ceived from the Dutch Agency for Environmental Planning (Milieu- en Natuurplanbureau, De Bilt, The Netherlands, made possible by H. de Vries) and the German National Science Foundation (DFG priority project Interdynamik 1266). We thank the two anonymous reviewers for their comments and J. Jago for improving the language of the manuscript. References Ackland, G., Signitzer, M., Stratford, K., Cohen, M., 2007. Cultural hitchhiking on the wave of advance of beneficial technologies. Pro- ceedings of the National Academy of Sciences 104 (21), 8714. Ammerman, A., Cavalli-Sforza, L., 1971. Measuring the rate of spread of early farming in Europe. Man 6 (4), 674–688. Ammerman, A., Cavalli-Sforza, L., 1973. A population model for the diffusion of early farming in Europe. In: Renfrew, C. (Ed.), The explanation of cultural change. Duckworth, London, pp. 343–57. Anderung, C., Bouwman, A., Persson, P., Carretero, J., Ortega, A., Elburg, R., Smith, C., Arsuaga, J., Ellegren, H., Gotherstrom, A., 2005. Prehistoric contacts over the Straits of Gibraltar indicated by genetic analysis of Iberian Bronze Age cattle. Proceedings of the National Academy of Sciences 102 (24), 8431. Arias, P., 1999. The origins of the Neolithic along the Atlantic coast of continental Europe: A survey. Journal of World Prehistory 13 (4), 403–464. Balaresque, P., Bowden, G., Adams, S., Leung, H., King, T., Rosser, Z., Goodwin, J., Moisan, J., Richard, C., Millward, A., et al., 2010. A predominantly Neolithic origin for European paternal lin- eages. Public Library of Science Biology 8 (1), e1000285. Barker, G., 2006. The Agricultural Revolution in Prehistory. Why did Foragers become Farmers? Oxford University Press, Oxford, United Kingdom. Begon, M., Harper, J. L., Townsend, C. R., 1993. Ecology- Individual, Populations and Communities, 2nd Edition. Blackwell Scientific Publications, Oxford. Bellwood, P., 2005. First Farmers: The Origins of Agricultural Soci- eties. Blackwell Publishers. Bellwood, P., Oxenham, M., 2008. The expansions of farming so- cieties and the role of the Neolithic Demographic Transition. In: Bocquet-Appel, J., Bar-Yosef, O. (Eds.), The Neolithic Demo- graphic Transition and its consequences. Springer, pp. 13–34. 11 Bentley, A., Price, T., Luning, J., Gronenborn, D., Wahl, J., Ful- lagar, P., 2002. Prehistoric migration in Europe: strontium iso- tope analysis of early Neolithic skeletons. Current Anthropology 43 (5), 799–804. Bocquet-Appel, J.-P., Naji, S., van der Linden, M., Kozlowski, J. K., 2009. Detection of diffusion and contact zones of early farming in Europe from the space-time distribution of 14C dates. Journal of Archaeological Science 36, 807–820. Boyd, B., 2006. On 'sedentism' in the Later Epipalaeolithic L., Choyke, A. M., Coqueugniot, E., Dohle, H.-J., Goldner, H., Hartz, S., Helmer, D., Herzig, B., Hongo, H., Mashkour, M., Ozdogan, M., Pucher, E., Roth, G., Scahde-Lindig, S., Schmolcke, U., Schulting, R. J., Stephan, E., Uerpmann, H.-P., Voros, I., Voytek, B., Bradley, D. G., Burger, J., 2007. Mitochondrial DNA analysis shows a Near Eastern Neolithic origin for domestic cattle and no indication of domestication of European aurochs. Philo- sophical Transactions of the Royal Society of London Series B 274, 1377–1385. (Natufian) Levant. World Archaeology 38 (2), 164–178. Fisher, R. A., 1930. The Genetical Theory of Natural Selection. Braidwood, L., Braidwood, R., 1949. On the treatment of the prehis- toric near Eastern materials in Steward's "cultural causality and law". American Antiquity 51 (4), 665–669. Dover, New York. Flannery, K. V., 1973. The origins of agriculture. Annual Review of Anthropology 2 (1), 271–310. Braidwood, R., Braidwood, L., 1950. Jarmo: A village of early farm- Fuller, D. Q., 2006. Agricultural origins and frontiers in south Asia: ers in Iraq. Antiquity 24 (96), 189–195. Bramanti, B., Thomas, M., Haak, W., Unterlaender, M., Jores, P., Tambets, K., Antanaitis-Jacobs, I., Haidle, M., Jankauskas, R., Kind, C., et al., 2009. Genetic discontinuity between local hunter- gatherers and central Europe's first farmers. Science 326 (5949), 137. Childe, V. G., 1925. The Dawn of European civilization. Kegan Paul, Trench, Trubner, London. Childe, V. G., 1936. Man makes himself. Watts, London. Childe, V. G., 1942. What happened in history. Pelican/Penguin, Harmondsworth. Clark, J. D., 1965. Radiocarbon dating and the spread of farming economy. Antiquity 39, 45–48. Claussen, M., Kubatzki, C., Brovkin, V., Ganopolski, A., Hoelzmann, P., Pachur, H., 1999. Simulation of an abrupt change in Saharan vegetation at the end of the mid-Holocene. Geophysical Research Letters 24, 2037–2040. Colledge, S., Conolly, J., Shennan, S., 2004. Archaeobotanical evid- ence for the spread of farming in the eastern Mediterranean. Cur- rent Anthropology 45 (suppl), S35–S58. Courtin, J., Erroux, J., 1974. Aper¸cu sur l'agriculture pr´ehistorique dans le sud-est de la France. Bulletin de la Soci´et´e Pr´ehistorique Fran¸caise 71 (1), 321–334. Currat, M., Poloni, E., Sanchez-Mazas, A., 2010. Human genetic differentiation across the Strait of Gibraltar. BMC Evolutionary Biology 10 (1), 237. Davison, K., Dolukhanov, P., Sarson, G., Shukurov, A., 2006. The role of waterways in the spread of the Neolithic. Journal of Ar- chaeological Science 33 (5), 641–652. Davison, K., Dolukhanov, P., Sarson, G., Shukurov, A., Zaitseva, G., 2009. Multiple sources of the European Neolithic: Mathematical modelling constrained by radiocarbon dates. Quaternary Interna- tional 203 (1-2), 10–18. Davison, K., Dolukhanov, P. M., Sarson, G., Shukurov, A., Zaitseva, G. I., 2007. A pan-European model of the Neolithic. Documenta Praehistorica XXXIV, 641–652. Diamond, J., 1997. The worst mistake in the history of the human race. Discover May, 64–66. Diamond, J., Aug. 2002. Evolution, consequences and future of plant and animal domestication. Nature 418, 700–707. Dieckmann, U., Law, R., 1996. The dynamical theory of coevolu- tion: a derivation from stochastic ecological processes. Journal of Mathematical Biology 34, 579–612, adap. Dolukhanov, P., 1973. The Neolithisation of Europe: a chronological and ecological approach. In: Renfrew, C. (Ed.), The explanation of cultural change. Duckworth, Gloucester, United Kingdom, pp. 329–342. Dolukhanov, P., Shukurov, A., Gronenborn, D., Sokoloff, D., Timofeev, V., Zaitseva, G., 2005. The chronology of Neolithic dis- persal in central and eastern Europe. Journal of Archaeological Science 32 (10), 1441–1458. Edmonson, M., 1961. Neolithic diffusion rates. Current Anthropology 2, 71–102. Edwards, C. J., Bollongino, R., Scheu, A., Chamberlain, A., Tresset, A., Vigne, J.-D., Baird, J. F., Larson, G., Ho, S. Y. W., Heupink, T. H., Shapiro, B., Freeann, A. R., Thomas, M. G., Arbogast, R., Arndt, B., Bartosiewicz, L., Benecke, N., Budja, M., Chaix, A working synthesis. Journal of World Prehistory 20 (1), 1–86. Gaillard, M.-J., Sugita, S., Mazier, F., Trondman, A.-K., Brostrom, A., Hickler, T., Kaplan, J. O., Kjellstrom, E., Kokfelt, U., Kunes, P., Lemmen, C., Miller, P., Olofsson, J., Poska, A., Rundgren, M., Smith, B., Strandberg, G., Fyfe, R., Nielsen, A. B., Alenius, T., Balakauskas, L., Barnekow, L., Birks, H. J. B., Bjune, A., Bjorkman, L., Giesecke, T., Hjelle, K., Kalnina, L., Kangur, M., van der Knaap, W. O., Koff, T., Lageras, P., Lata(cid:32)lowa, M., Ley- det, M., Lechterbeck, J., Lindbladh, M., Odgaard, B., Peglar, S., Segerstrom, U., von Stedingk, H., Seppa, H., 2010. Holocene land- cover reconstructions for studies on land cover-climate feedbacks. Climate of the Past 6 (4), 483–499. Gedd`es, D., 1980. De la chasse au troupeau en M´editerran´ee Occi- dentale. les d´ebuts de l'´elevage dans le bassin de l'Aude. Bulletin de la Soci´et´e pr´ehistorique fran¸caise. 81 (10-12), 370–378. Gkiasta, M., Russell, T., Shennan, S., Steele, J., 2003. Neolithic transition in Europe: the radiocarbon record revisited. Antiquity 77 (295), 45–62. Gronenborn, D., 1999. A variation on a basic theme: The transition to farming in southern central Europe. Journal of World Prehis- tory 13 (2), 123–210. Gronenborn, D., 2007a. Beyond the models: Neolithisation in central Europe. In: Whittle and Cummings (2007), pp. 73–98, 73–98. Gronenborn, D., 2007b. Climate change and socio-political crises: some cases from Neolithic central Europe. In: Pollard, T., Banks, I. (Eds.), War and sacrifice studies in the archaeology of conflict. Koninklijke Brill NV, Leiden, pp. 13–32. Gronenborn, D., 2009a. Climate fluctuations and trajectories to com- plexity in the Neolithic: towards a theory. Documenta Praehistor- ica 36, 97–110. Gronenborn, D., 2009b. Transregional culture contacts and the Neo- lithization process in Northern Central Europe. In: Jordan, P., Zvelebil, M. (Eds.), Ceramics before farming: the dispersal of pottery among prehistoric Eurasian hunter-gatherers. Left Coast Press (Walnut Creek), pp. 527–550. Gronenborn, D., 2011. Neolithic western temperate and central Europe. In: Cummings, V., Jordan, P., Zvelebil, M. (Eds.), Ox- ford Handbook of the Archaeology and Anthropology of Hunter Gatherers. Vol. in press. Oxford University Press. Gronenborn, D., Petrasch, J. (Eds.), 2010. The Spread of the Neo- lithic to Central Europe. Vol. 4 of RGZM Tagungen. Romisch- Germanisches Zentralmuseum, Mainz, Germany. Guilaine, J., 2001. La diffusion de l'agriculture en Europe: une hy- poth`ese arythmique. Zephyrus 53 (4), 267–272. Haak, W., Balanovsky, O., Sanchez, J. J., Koshel, S., Za- porozhchenko, V., Adler, C. J., Der Sarkissian, C. S. I., Brandt, G., Schwarz, C., Nicklisch, N., Dresely, V., Fritsch, B., Bal- anovska, E., Villems, R., Meller, H., Alt, K. W., Cooper, A., the Genographic Consortium, 2010. Ancient DNA from European early Neolithic farmers reveals their Near Eastern affinities. Public Library of Science Biology 8 (11), e1000536. Hartz, S., Lubke, H., Terberger, T., 2007. From fish and seal to sheep and cattle: new research into the process of Neolithisation in northern Germany. Proceedings of the British Academy 144, 567–594. Hodder, I., 1990. The Domestication of Europe: Structure and Con- tingency in Neolithic Europe. Basil Blackwell, Oxford. 12 Jeunesse, C., 2000. Les composantes autochtone et danubienne en Europe centrale et occidentale entre 5500 et 4000 av. J.-C.: contacts, transferts, acculturations. In: Les derniers chasseurs- cueilleurs d'Europe occidentale (13000-5500 av. J.-C.): actes du colloque international de Besan¸con. pp. 361–78. Kaplan, J., Krumhardt, K., Ellis, E., Ruddiman, W., Lemmen, C., Klein Goldewijk, K., 2010. Holocene carbon emissions as a result of anthropogenic land cover change. The Holocene, prepublished Dec 30, 2010, doi: 10.1177/0123456789123456. Kisdi, ´Eva and Geritz, S. A. H., 2010. Adaptive dynamics: a frame- work to model evolution in the ecological theatre. Journal of Math- ematical Biology 61, 165–169. Kotova, N. S., 2003. Neolithization in Ukraine. British Archaeological Reports Ltd. Kotova, N. S., 2009. The Neolithization of northern Black Sea area in the context of climate changes. Documenta Praehistorica XXXVI, 159–174. Kuijt, I., Goring-Morris, N., 2002. Foraging, farming, and social com- plexity in the pre-pottery Neolithic of the southern Levant: A re- view and synthesis. Journal of World Prehistory 16 (4), 361–440. Kutzbach, J., Ruddiman, W., Vavrus, S., Philippon, G., 2010. Cli- mate model simulation of anthropogenic influence on greenhouse- induced climate change (early agriculture to modern): the role of ocean feedbacks. Climatic Change 99 (3), 351–381. Larson, G., Albarella, U., Dobney, K., Rowley-Conwy, P., Schibler, J., Tresset, A., Vigne, J. D., Edwards, C. J., Schlumbaum, A., Dinu, A., Bala¸csescu, A., Dolman, G., Tagliacozzo, A., Mana- seyran, N., Miracle, P., Van Wijngaarden-Bakker, L., Masseti, M., Bradley, D. G., Cooper, A., 2007. Ancient DNA, pig domest- ication, and the spread of the Neolithic into Europe. Proceedings of the National Academy of Sciences 104 (39), 15276. Leemans, R., Cramer, W. P., Nov. 1991. The IIASA database for mean monthly values of temperature, precipitation and cloudiness of a global terrestrial grid. Research report, International Institute of Applied Systems Analyses, Laxenburg. Lemmen, C., 2001. Understanding the regional rise of civilizations by means of a dynamic model. Master thesis, Carl von Ossietzky University, Oldenburg. Lemmen, C., 2009. World distribution of land cover changes during pre-and protohistoric times and estimation of induced carbon re- leases. G´eomorphologie: relief, processus, environnement 4, 303– 312. Lemmen, C., Wirtz, K. W., 2010. Socio-technological revolutions and migration waves re-examining early world history with a math- ematical model. In: Gronenborn and Petrasch (2010), Vol. 4 of RGZM Tagungen, pp. 25–38. Lieth, H., 1975. Modeling the primary productivity of the world. Primary Productivity of the Biosphere 14, 237–263. Luikart, G., Gielly, L., Excoffier, L., Vigne, J.-D., Bouvet, J., Taber- let, P., 2001. Multiple maternal origins and weak phylogeographic structure in domestic goats. Proceedings of the National Academy of Sciences 98 (10), 5927–5932. Luning, J., Stehli, P., 1994. Die Bandkeramik im Merzbachtal auf der Aldenhovener Platte. Rheinland-Verlag. Manen, C., Marchand, G., Carvalho, A. F., 2007. Le N´eolithique ancien de la p´eninsule Ib´erique: vers une nouvelle ´evaluation du mirage Africain? In: ´Evin, J. (Ed.), XXVIe Congr`es pr´ehistorique de France, Congr`es du centenaire de la Soci´et´e pr´ehistorique fran¸caise, Avignon, 21–25 Sep 2004. Vol. 3 of Aux conceptions d'aujourd'hui. Soci´et´e Pr´ehistorique Fran¸caise, Paris, France, pp. 133–151. Mauvilly, M., Jeunesse, C., Doppler, T., 2008. Ein Tonstempel aus der spatmesolithischen Fundstelle Arconciel/La Souche (Kanton Freiburg/Schweiz). Quartar 55, 151–157. Merico, A., Bruggeman, J., Wirtz, K., 2009. A trait-based approach for downscaling complexity in plankton ecosystem models. Ecolo- gical Modelling 220 (21), 3001–3010. Metz, J., Nisbet, R., Geritz, S., 1992. How should we define fitness for general ecological scenarios. Trends in Ecology and Evolution 7, 198–202. Midgeley, M. S., Nov. 1992. TRB culture: the first farmers of the North European Plain. Edinburgh University Press. Mithen, S. J., 2004. After the Ice: A Global Human History, 20,000- 5000 BC. Weidenfels & Nicolson, London, United Kingdom. Patterson, M., Sarson, G., Sarson, H., Shukurov, A., 2010. Modelling the Neolithic transition in a heterogeneous environment. Journal of Archaeological Science 37 (11). Peltenburg, E., Colledge, S., Croft, P., Jackson, A., McCartney, C., Murray, M. A., 2000. Agro-pastoralist colonization of Cyprus in the 10th millennium BP: initial assessments. Antiquity 74, 844– 853. Pinhasi, R., Fort, J., Ammerman, A., 2005. Tracing the origin and spread of agriculture in Europe. PloS Biology 3 (12), 410. Price, T. D. (Ed.), 2000. Europe's First Farmers. Cambridge Univer- sity Press. Renfrew, C., 1987. Archaeology and language. the puzzle of indo- European origins. Current Anthropology 29, 437–441. Richards, M., 2003. The Neolithic transition in Europe: archaeolo- gical models and genetic evidence. Documenta Praehistorica 30, 159–68. Roth, H. L., 1887. On the origin of agriculture. The Journal of the Anthropological Institute of Great Britain and Ireland 16, 102– 136. Ruddiman, W., 2006. The early anthropogenic hypothesis- challenges and responses. Geophysical Research Abstracts 8, 01749. Schier, W., 2009. Extensiver Brandfeldbau und die Ausbreitung der Neolithischen Wirtschaftsweise in Mitteleuropa und Sudskand- inavien am Ende des 5. Jahrtausends v. Chr. Praehistorische Zeits- chrift 84 (1), 15–43. Sheridan, J., 2007. Scottish beaker dates: the good, the bad and the ugly. In: From Stonehenge to the Baltic: Living with Cultural Diversity in the Third Millennium BC. British Archaeological Re- ports Ltd, pp. 91–123. Smith, B., 1997. The initial domestication of Cucurbita pepo in the Americas 10, 000 years ago. Science 276 (5314), 932–934. Smith, S. L., Pahlow, M., Merico, A., Wirtz, K. W., 2011. Optimality as a unifying concept for planktonic organisms and their ecology. Limnology and Oceanography submitted. Sokal, R., Oden, N., Wilson, C., 1991. Genetic evidence for the spread of agriculture in Europe by demic diffusion. Nature 351 (6322), 143–145. Theocharis, D. R., Bokonyi, S., 1973. Neolithic Greece. National Bank of Greece Athens. Thomas, J., 1991. Rethinking the Neolithic. Cambridge University Press, Cambridge. Weninger, B., Clare, L., Rohling, E. J., Bar-Yosef, O., Bohner, U., Budja, M., Bundschuh, M., Feurdean, A., Gebel, H.-G., Joris, O., Linstadter, J., Mayewski, P., Muhlenbruch, T., Reingruber, A., Rollefson, G., Schyle, D., Thissen, L., Todorova, H., Zielhofer, C., 2009. The impact of rapid climate change on prehistoric societies during the Holocene in the eastern Mediterranean. Documenta Praehistorica 902 (4-5), 551–583. Westropp, H. M., 1872. Pre-historic phases: or, Introductory essays on pre-historic archaeology, 1st Edition. Bell & Daldy. Whittle, A., 2007. The temporality of transformation: dating the early development of the southern British Neolithic. In: Whittle and Cummings (2007), Vol. 144 of Proceedings of the British Adademy, pp. 377–398. Whittle, A., Cummings, V., 2007. Going Over: the Mesolithic- Neolithic transition in North-West Europe. Vol. 144 of Proceed- ings of the British Adademy. Oxford University Press. Willcox, G., 2005. The distribution, natural habitats and availabil- ity of wild cereals in relation to their domestication in the near east: multiple events, multiple centres. Vegetation History and Archaeobotany 14 (4), 534–541. Willcox, G., Buxo, R., Herveux, L., 2009. Late Pleistocene and early Holocene climate and the beginnings of cultivation in northern Syria. Holocene 1 (1), 151–158. Winiger, J., 1998. Ethnoarchaologische Studien zum Neolithikum Sudwesteuropas. British Archaeological Reports. Wirtz, K., Lemmen, C., 8 2003. A global dynamic model for the 13 Neolithic transition. Climatic Change 59 (3), 333–367. Wirtz, K., Lohmann, G., Bernhardt, K., Lemmen, C., 2010. Mid- Holocene regional reorganization of climate variability: Analyses of proxy data in the frequency domain. Paleogeography Palaeocli- matology Palaeoecology 298 (3–4). Wirtz, K. W., Eckhardt, B., 1996. Effective variables in ecosystem models with an application to phytoplankton succession. Ecolo- gical Modelling 92, 33–54. Zeder, M., 2008. Domestication and early agriculture in the Mediter- ranean basin: Origins, diffusion, and impact. Proceedings of the National Academy of Sciences 105 (33), 11597. Zilhao, J., 1993. The spread of agro-pastoral economies across Medi- terranean Europe: a view from the Far-West. Journal of Mediter- ranean Archaeology 6, 5–63. Zilhao, J., 2000. From the Mesolithic to the Neolithic in the Iberian peninsula. In: Price (2000), pp. 144–182. Zimmermann, A., 2004. Landschaftsarchaologie II. Uberlegungen zu Prinzipien einer Landschaftsarchaologie. Vol. 85 of Bericht der Romisch- Germanischen Kommission. Philipp von Zabern, Mainz. Zvelebil, M., 1998. What's in a name: the Mesolithic, the Neolithic, and social change at the Mesolithic-Neolithic transition. In: Ed- monds, M., Richards, C. (Eds.), Understanding the Neolithic of north-western Europe. Cruithne Press, pp. 1–36. Zvelebil, M., 2006. Mobility, contact, and exchange in the Baltic Sea basin 6000–2000 BC. Journal of Anthropology and Archaeology 25 (2), 178–192. Zvelebil, M., Zvelebil, K. V., 1988. Agricultural transition and Indo- European dispersals. Antiquity 62, 574–583. 14
1810.06913	1	1810	2018-10-16T10:20:06	How to share a cake with a secret agent	[ "cs.MA", "cs.GT" ]	In this note we study a problem of fair division in the absence of full information. We give an algorithm which solves the following problem: n $\ge$ 2 persons want to cut a cake into n shares so that each person will get at least 1/n of the cake for his or her own measure, furthermore the preferences of one person are secret. How can we construct such shares? Our algorithm is a slight modification of the Even-Paz algorithm and allows to give a connected part to each agent. Moreover, the number of cuts used during the algorithm is optimal: O (n log(n)) .	cs.MA	cs	HOW TO SHARE A CAKE WITH A SECRET AGENT GUILLAUME CH`EZE Abstract. In this note we study a problem of fair division in the absence of full information. We give an algorithm which solves the following problem: n ≥ 2 persons want to cut a cake into n shares so that each person will get at least 1/n of the cake for his or her own measure, furthermore the preferences of one person are secret. How can we construct such shares? Our algorithm is a slight modification of the Even-Paz algorithm and allows to give a connected part to each agent. Moreover, the number of cuts used during the algorithm is optimal: O(cid:0)n log(n)(cid:1). Introduction Fair division is an old problem. The following situation already appears in the Bible. Two persons (in the Bible Abraham and Lot) want to share a land. Fur- thermore, this division must be fair. This means that each agent think that he or she has obtained at least 1/2 of the land for his or her own point of view. The following protocol, called "Cut and Choose", is then used: The first player cuts the land into two pieces with equal values for him or her. The second player chooses one of the two pieces. With this strategy each player get a connected piece with a value bigger than 1/2 for his or her point of view. This protocol can also be used to divide a cake or an heterogeneous good as time or computer memory between two agents. How can we generalize this protocol to n agents? Several answers exist. In order to describe them we need to precise some points. We consider an heterogeneous good, for example: a cake, represented by the interval X = [0, 1] and n agents with different points of view. We associate to each agent a non-atomic probability measure µi on the interval X = [0; 1]. These measures represent the utility functions of the agent. This means that if [a, b] ⊂ X is a part of the cake then µi([a, b]) is the value associated by the i-th agent to this part of the cake. As µi are probability measures, we have µi(X) = 1 for all i. A division of X is a partition X = ⊔jXj where each Xj is given to one agent. Thus there exists a permutation σ ∈ Sn such that µi is associated to Xσ(i). A division is simple when each Xi is an interval. Guillaume Ch`eze: Institut de Math´ematiques de Toulouse, UMR 5219, Universit´e de Toulouse ; CNRS, UPS IMT, F-31062 Toulouse Cedex 9, France E-mail address: [email protected] . Date: October 17, 2018. Key words and phrases. fair division; cake cutting algorithm; partial information. 1 2 G. CH `EZE Several notions of fair division exists. We say that a division is proportional when µi(Xσ(i)) ≥ 1/n. We say that a division is envy-free when for i 6= j, we have µi(Xσ(i)) ≥ µi(Xσ(j)). The problem of fair division (theoretical existence of fair division and construc- tion of algorithms) has been studied in several papers [14, 8, 10, 9, 4, 12, 2], and books about this topic, see e.g. [13, 5, 11, 3]. These results appear in the math- ematics, economics, political science, artificial intelligence and computer science literature. Recently, the cake cutting problem has been studied intensively by com- puter scientists for solving resource allocation problems in multi agents systems, see e.g. [7]. In this note we are going to study proportional fair divisions. This topic has been [13]. In order to describe very studied. Several algorithms already exist, see e.g. algorithms we thus need a model of computation. There exist two main classes of cake cutting algorithms: discrete and continuous protocols (also called moving knife methods). Here, we study only discrete algorithms. These kinds of algorithms can be described thanks to the classical model introduced by Robertson and Webb and formalized by Woeginger and Sgall in [16]. In this model we suppose that a mediator interacts with the agents. The mediator asks two type of queries: either cutting a piece with a given value, or evaluating a given piece. More precisely, the two type of queries allowed are: (1) evali(x, y): Ask agent i to evaluate the interval [x, y]. This means compute µi([x, y]). (2) cuti(x, α): Asks agent i to cut a piece of cake [x, y] such that µi([x, y]) = α. This means: for given x and α, solve µi([x, y]) = α. We remark that in the "Cut and Choose" algorithm only these two kinds of queries are used. In the Robertson-Webb model the mediator can adapt the queries from the previ- ous answers given by the players. In this model, the complexity counts the finite number of queries necessary to get a fair division. For a rigorous description of this model we can consult: [16] An optimal algorithm for proportional fair division has been given by Even and Paz in [10]. When there are only two agents this algorithm corresponds to "Cut and Choose". When there are n ≥ 3 agents, this algorithm uses a recursive strategy and it is sometimes called "Divide and Conquer". Some properties of this approach are studied in [6]. However, it seems that one property has never been studied. Indeed, in the "Cut and Choose" algorithm the second agent do not give his or her preference. A par- tition X = X1 ⊔ X2 is given and the second player choose X1 or X2. The measure µ2 is not used for the construction of the partition. Thus, with two agents, even if one player do not participate to the construction of the partition, we can get a proportional division. In this note, we are going to show that when there are n + 1 agents we can construct a proportional fair division with connected portions even if the measure of one agent is unknown. We call this agent the secret agent. HOW TO SHARE A CAKE WITH A SECRET AGENT 3 One application suggested by the article [1] is the following: During a birthday party with n guests and one host, a cake is divided into n + 1 pieces before it is presented to the birthday boy or girl, but he or she gets any portion. In this article, we give an algorithm which returns a partition assuring 1/(n + 1) of the cake to each persons (the n guests and the host) for his or her own measure. More precisely we are going to prove: Theorem 1. Consider n + 1 players and X = [0, 1]. We denote by Ej the set {1, 2, . . . , n, n + 1} \ {j}. In the Roberston-Webb model, there exists an algorithm using only queries with the first n players and giving a fair division X = ⊔n+1 j=1 Xj such that: • each Xj is an interval, • for all j ∈ {1, 2, . . . , n, n + 1} there exists a bijection σj from {1, . . . , n} to Ej such that for all i ∈ {1, . . . , n} µi(cid:0)Xσj (i)(cid:1) ≥ µi(X) n + 1 . Furthermore, this algorithm uses at most O(cid:0)n log(n)(cid:1) queries. This theorem says that the first n players can construct a partition X = ⊔n+1 j=1 Xj without asking any queries to the n + 1-th agent. Then, this last agent (the secret agent) choose first a portion Xj. Therefore, this agent can choose a portion Xj such that µn+1(Xj) ≥ µn+1(Xi) for i 6= j and then µn+1(Xj) ≥ 1/(n + 1). The second part of the theorem says that the remaining portions Xi can be allocated to the other agents in such a way that each agent obtained at least 1/(n + 1) for his or her own measure. This theorem asserts that we can obtain a proportional fair division in the absence of full information. Indeed, no query is asked to the n+1-th player. The preferences of this player are secret. We do not know at the end of the division what is the value of µn+1(Xj). This is the reason why we call this last agent a secret agent. The "Divide and Conquer" algorithm presented by Even and Paz in [10] ask queries to all agents and then cannot be used to prove Theorem 1. Indeed, if there are 4 agents this algorithm ask to each agent the query cuti(0, 1/2). This gives four cut-points ci. We can suppose without loss of generality that c1 ≤ c2 ≤ c3 ≤ c4. Then the first and the second agent apply the "Cut and Choose" algorithm on [0, c2] and the third and fourth agent do the same on [c2, 1]. The first step of this algorithm prevent its use with a secret agent. However, we will see in the next section that this approach can be modified and used with a secret agent. Other protocols can be modified and used with a secret agent. However, the com- plexity of these algorithms will be bigger than O(cid:0)n log(n)(cid:1). The benefit of the modification of the "Divide and Conquer" protocol is the following: it leads to an optimal algorithm. Indeed, it has been proved in [16, 9] that a proportional fair division needs at least O(cid:0)n log(n)(cid:1) queries in the Robertson-Webb model. We remark that our proportional fair division method cannot be generalized with more than one secret agent. Indeed, suppose that we have n + 2 agents and 2 agents are secret agents, i.e. we cannot ask queries to them, they do not par- ticipate to the construction of the division. Suppose also that the n agents give 4 G. CH `EZE a partition X = ⊔n+2 j=1 Xj. If the two secret agents have measure µn+1, µn+2 such that µn+1(X1) = µn+2(X1) = 1, then it is impossible to obtain a proportional fair division for these n + 2 agents with this partition. We can also remark that we cannot obtain the same result for envy-free divi- sion. Indeed, Stromquist in [15] has proved that there exist no algorithm in the Roberston-Web model giving a simple (connected) and envy-free fair division for n ≥ 3 players. However, it has been proved in [1] that simple and envy-free divisions theoretically exist even if the preferences of one person are secret. This means that this kind of partition exists but cannot be obtained with an algorithm in the Robertson-Webb model. In the next section, we give a slightly modified version of the "Divide and Con- quer" algorithm in order to prove Theorem 1. 1. A modified "Divide and Conquer" algorithm DC secret Inputs: X = [a, b], a list l = [µ1, . . . , µn] of n players. Outputs: A partition of X. (1) If n = 1 then c1 := cut1(a, 1/2). Return(X = [a, c1] ∪ [c1, b]). (2) If n > 1 is odd then For i from 1 to n do ci := cuti(a, µi(X)/2) End For. Sort the set {c1, . . . , cn} in order to have: ci1 ≤ ci2 ≤ · · · ≤ cin . Set XL := [a, c n+1 Return(cid:0)DC secret(XL, l1)⊔ DC secret(XR, l2)(cid:1). , b], l1 := [µi1 , . . . , µi n+1 ], XR = [c n+1 −1 2 2 2 ], l2 := [µi n+1 , . . . , µin ]. +1 2 (3) If n > 1 is even then For i from 1 to n do ci := cuti(cid:16)a, µi(X) n+1 (cid:17), End For. Compute ci0 := mini=1,...,n(ci). Set XL := [a, ci0] and l′ := l \ [µi0 ]. Return(cid:0)XL⊔ DC secret([ci0 , b], l′)(cid:1). Proof. We are going to prove by induction Theorem 1. We consider the following claim: Hn: The algorithm DC secret applied with n measures returns a partition satisfying the conclusion of Theorem 1. HOW TO SHARE A CAKE WITH A SECRET AGENT 5 For n = 1, H1 is true. Indeed, in this case the algorithm DC secret gives the same partition as the "Cut and Choose" algorithm. Now, we suppose that Hk is true for k ≤ n − 1. If n is even then the algorithm returns X = XL ⊔ DC secret(X ′, l′), where X ′ = X \XL and l′ is a list of n−1 measures. By construction XL is an interval. Without loss of generality we can suppose that l′ = [µ2, . . . , µn]. As Hn−1 is true we get: X = XL ⊔n+1 j=2 Xj, where Xj are intervals and for all j ∈ {2, . . . , n, n+ 1} there exists a bijection σj from {2, . . . , n} to {2, . . . , n+ 1} \ {j} such that for all i in {2, . . . , n} we have µi(cid:0)Xσj (i)(cid:1) ≥ µi(X ′) n . Furthermore, for i ≥ 2, we have µi(XL) ≤ µi(X) Thus n+1 and µi(X ′) = µi(X) − µi(XL). µi(X ′) ≥ It follows n n + 1 µi(X). µi(Xσj (i)) ≥ µi(X) n + 1 . Moreover, by construction, we have µ1(XL) = µ1(X)/(n + 1). This proves the theorem when the secret agent chooses a portion Xj and j ≥ 2. It remains thus to prove the theorem when the secret agent chooses the portion XL. However, in this situation the agent with associated measure µ1 plays the role of the secret agent in DC secret(X \ XL, l′) and we get the desired result. If n > 1 is odd then the algorithm returns X = DC secret(XL, l1)⊔DC secret(XR, l2). Without loss of generality we can suppose that l1 = [µ1, . . . , µ n+1 [µ n+1 By our induction hypothesis we get: 2 +1, . . . , µn]. 2 −1] and l2 = X = n+1 2 G j=1 Xj n+1 G j= n+1 2 +1 Xj. We set XL = ⊔ n+1 2 j=1 Xj and XR = ⊔n+1 j= n+1 2 +1 Xj. Suppose that the secret agent chooses a portion Xj where j ≤ situation is similar). As H n+1 {1, . . . , n+1 is true, there exists a bijection σ1 from the set {1, . . . , n+1 2 } \ {j} such that for all i in {1, . . . , n+1 2 − 1} we have 2 n + 1 2 , (the other 2 − 1} to µi(Xσ1(i)) ≥ µi(XL) n+1 2 . 6 G. CH `EZE By construction, we have for all i ∈ {1, . . . , n+1 gives, 2 − 1}, µi(XL) ≥ µi(X)/2. This µi(Xσ1(i)) ≥ µi(X) n + 1 , for all i ∈ n1, . . . , n + 1 2 − 1o. In the same way, we consider the sub-cake XR, players with associated mea- as a secret agent. By 2 +1, . . . , µn and the agent with measure µ n+1 sures µ n+1 construction, this last agent is such that 2 µ n+1 2 (XR) = µ n+1 2 (X)/2. Furthermore, XR is divided in (n + 1)/2 parts, thus there exists a portion Xk such that We deduce µ n+1 2 (Xk) ≥ µ n+1 (XR) 2 n+1 . 2 µ n+1 2 (Xk) ≥ µ n+1 2 (X) n + 1 . As before, as H n+1 2 + 1, . . . , n} to { n+1 is true, we get: there exists a bijection σ2 from the set 2 + 1, . . . , n + 1} \ {k}, such that for i ∈ { n+1 2 + 1, . . . , n} 2 { n+1 we have µi(Xσ2(i)) ≥ µi(X) n + 1 . From the bijection σ1 and σ2 we can construct a bijection σ giving the conclusion of the theorem. The complexity study is classical and is the same as the one done for the usual "Divide and Conquer" algorithm: the number of steps is in O(cid:0) log(n)(cid:1) and the number of queries in each step is in O(n). This gives the desired complexity. (cid:3) References [1] Megumi Asada, Florian Frick, Vivek Pisharody, Maxwell Polevy, David Stoner, Ling Hei Tsang, and Zoe Wellner. Fair division and generalizations of Sperner- and KKM-type results. SIAM J. Discrete Math., 32:591 -- 610, 2018. [2] H. Aziz and S. Mackenzie. A discrete and bounded envy-free cake cutting protocol for any number of agents. In IEEE 57th Annual Symposium on Foun- dations of Computer Science, FOCS 2016, 9-11 October 2016, pages 416 -- 427, 2016. [3] J. Barbanel. The geometry of efficient fair division. Cambridge University Press, 2005. [4] S. Brams and A. Taylor. An envy-free cake division protocol. The American Mathematical Monthly, 102(1):9 -- 18, 1995. [5] S. Brams and A. Taylor. Fair division - from cake-cutting to dispute resolution. Cambridge University Press, 1996. [6] Steven J. Brams, Michael A. Jones, and Christian Klamler. Divide-and- conquer: A proportional, minimal-envy cake-cutting algorithm. SIAM Review, 53:291 -- 307, 2011. HOW TO SHARE A CAKE WITH A SECRET AGENT 7 [7] Y. Chevaleyre, P. Dunne, U. Endriss, J. Lang, M. Lemaıtre, N. Maudet, J. Pad- get, S. Phelps, J. Rodr´ıguez-Aguilar, and P. Sousa. Issues in multiagent re- source allocation. INFORMATICA, 30:3 -- 31, 2006. [8] L.E. Dubins and E. H. Spanier. How to cut a cake fairly. The American Mathematical Monthly, 68(1):1 -- 17, 1961. [9] J. Edmonds and K. Pruhs. Cake cutting really is not a piece of cake. ACM Trans. Algorithms, 7(4):51, 2011. [10] S. Even and A. Paz. A note on cake cutting. Discrete Applied Mathematics, 7(3):285 -- 296, 1984. [11] A. Procaccia. Cake cutting algorithms. In F. Brandt, V. Conitzer, U. En- driss, J. Lang, and A. D. Procaccia, editors, Handbook of Computational Social Choice, chapter 13. Cambridge University Press, 2016. [12] J. Robertson and W. Webb. Near exact and envy-free cake division. Ars Combinatoria, 45:97 -- 108, 1997. [13] J. Robertson and W. Webb. Cake-cutting algorithms - be fair if you can. A K Peters, 1998. [14] H. Steinhaus. The problem of fair division. Econometrica, 16(1):101 -- 104, January 1948. [15] Walter Stromquist. Envy-free cake divisions cannot be found by finite proto- cols. Electr. J. Comb., 15(1), 2008. [16] Gerhard J. Woeginger and Jir´ı Sgall. On the complexity of cake cutting. Discrete Optimization, 4(2):213 -- 220, 2007.
1303.3827	1	1303	2013-03-15T16:53:29	Towards a serious games evacuation simulator	[ "cs.MA", "cs.CY" ]	The evacuation of complex buildings is a challenge under any circumstances. Fire drills are a way of training and validating evacuation plans. However, sometimes these plans are not taken seriously by their participants. It is also difficult to have the financial and time resources required. In this scenario, serious games can be used as a tool for training, planning and evaluating emergency plans. In this paper a prototype of a serious games evacuation simulator is presented. To make the environment as realistic as possible, 3D models were made using Blender and loaded onto Unity3D, a popular game engine. This framework provided us with the appropriate simulation environment. Some experiences were made and results show that this tool has potential for practitioners and planners to use it for training building occupants.	cs.MA	cs	TOWARDS A SERIOUS GAMES EVACUATION SIMULATOR João Ribeiro1, João Emílio Almeida1†, Rosaldo J. F. Rossetti1†, António Coelho1‡, António Leça Coelho2 1Department of Informatics Engineering †LIACC – Laboratory of Artificial Intelligence and Computer Science ‡INESC TEC – INESC Technology and Science Faculty of Engineering, University of Porto Rua Roberto Frias, S/N, 4200-465, Porto, Portugal {joao.pedro.ribeiro, joao.emilio.almeida, rossetti, acoelho}@fe.up.pt 2LNEC – National Laboratory of Civil Engineering Av. Brasil, 101, 1700-066, Lisboa, Portugal [email protected] KEYWORDS Evacuation simulation, fire drill, modelling and simulation, serious games. ABSTRACT The evacuation of complex buildings is a challenge under any circumstances. Fire drills are a way of training and validating evacuation plans. However, sometimes these plans are not taken seriously by their participants. It is also difficult to have the financial and time resources required. In this scenario, serious games can be used as a tool for training, planning and evaluating emergency plans. In this paper a prototype of a serious games evacuation simulator is presented. To make the environment as realistic as possible, 3D models were made using Blender and loaded onto Unity3D, a popular game engine. This framework provided us with the appropriate simulation environment. Some experiences were made and results show that this tool has potential for practitioners and planners to use it for training building occupants. INTRODUCTION The problem of evacuation from large facilities during an emergency or disaster has been addressed by researchers and practitioners in recent years. Real-world fire drills lack the realistic atmosphere of the emergency situation. Typically, the scenario is set up with the help of fire consultants and experts in the field, and the evacuation procedures follow some predefined rules and participants are expected to proceed accordingly. In this paper, Serious Games (SG) are proposed as a means to overcome such drawbacks, since immersion into the emergency scenario artificially created using computer videogames is easier to accomplish. Also , the commitment of players, due to the excitement of using computer digital games, is expected to achieve better results than the traditional approaches. In this paper the concept of serious games is used to build an evacuation simulator as an attempt to address some of the issues that were identified in real-world fire drills. It is our intention to improve the way people participate their such experiments enhancing in experience in many different ways. We have adapted and customised the environment of a game engine, in this case Unity3D, to support simulation features that enabled users to be tracked and assessed while playing. To test our approach and demonstrate its feasibility, we have carried out preliminary experiences with our prototype, in which subjects using the game environment were asked to evacuate a building in the case of fire. The remaining part of this paper is organised as follows. We start by briefly presenting some related concepts that concern this project, such as pedestrian simulation and serious games. We then discuss on applying serious games to evacuation training, following the presentation and formalisation of our problem. We propose the approach implemented in this paper and suggest a preliminary experiment using our prototype. Some results are also discussed, after which we finally draw some conclusions and give clues of some further steps in this research. BACKGROUND AND RELATED WORK Pedestrian simulators There are three main reasons for developing pedestrian computer simulations: i) to test scientific theories and hypotheses; ii) to assess design strategies; iii) to recreate the phenomena about which we want to theorize (Pan et al., 2007). Pedestrian flow management demands the correct representation of both the collective as well as the individual (Hoogendoorn et al., 2004). Timmermans et al. (2009) argue that understanding the pedestrian decision-making and movement is of critical importance to develop valid pedestrian models. According to (Teknomo, 2002), pedestrian studies can be divided in two phases, namely data collection and focus on former the analysis. Whereas data characteristics such as speed, movement and path- planning, the latter is instead related to understanding how pedestrians behave. Predicting the movement of crowds (macroscopic level) or individual pedestrian actions (microscopic level) is the main goal of pedestrian simulation. For level, the macroscopic hydraulic or gas models are used (Santos and Aguirre, 2004). Microscopic models are based on behavioural approaches, in which entities are described individually (Castle et al., 2007). Traditional models, however, are mainly tested and validated through direct observations, techniques based on photography, as well as time-lapse films (Coelho, 1997; Helbing et al., 2001; Qingge et al., 2007) and also by stated preferences questionnaires (Cordeiro et al, 2011). In such models to verify certain is possible it phenomena such as herding or flocking that happen due to people following other individuals instinctively. However, in conditions of low visibility or little knowledge of the surroundings this can provoke flocks of wandering people, contributing to the panic and confusion of the whole group, which is also a social reaction rather to be avoided if possible (Reynolds, 1987). Kuligowski proposes a model to mimic the human behavioural process during evacuation from buildings. Social science studies are needed to develop these theories, which could then yield more realistic results leading to safer and more efficient building design (Kuligowski, 2008, 2011). Although many approaches exist to virtually simulate the behaviour of crowds with varying levels of realism, three models seem to be the most used (Heïgeas et al., 2003; Santos and Aguirre, 2004; Pelechano et al., 2007; Pretto, 2011). Cellular Automata Models (Neumann, 1966, Beyer et al., 1985) treat individuals as separate objects in an area divided into the so-called cells. Forces-based Models use mathematical formulaes to calculate the position variations of individual elements through the application of forces (its most explored subtypes consider Magnetic Forces and Social Forces). Finally, in Artificial Intelligence (AI) based Models, the decisions are made by individuals that compose the crowd on an autonomous basis. This sort of structure very much resembles a society of several interacting entities and has inspired much research in the Social Sciences (Kuligowski et al., 2010; Almeida et al., 2011). The Serious Games Concept Serious Games has gained a great prominence in the Digital Games field within the last decade, using appealing software with high-definition graphics and state-of-the-art gaming technology. It presents a great potential of application in a wide range of domains, naturally including social simulation. Contrary to the primary purpose of entertainment in traditional digital games, SG are designed for the purpose of solving a problem. Athough they are indeed expected to be entertaining, their main purpose is rather serious with respect to the outcomes reflected in changes to the player behavior (Frey et al., 2007; McGonigal, 2011). According to (Hays, 2005), a game is an artificially constructed, competitive activity with a specific goal, a set of rules and constraints that is located in a specific context. Serious Games refer to video games whose application is focused on supporting activities such as education, training, health, advertising, or social change. A few benefits from combining them with other training activities include (Freitas, 2006): the learners’ motivation is higher; completion rates are higher; possibility of accepting new learners; possibility of creating collaborative activities; learn through doing and acquiring experience. Other aspects that draw video game players’ attention are fantasy elements, challenging situations and the ability to keep them curious about the outcomes of their possible actions (Kirriemuir et al., 2004). Serious Games can be classified in five categories: Edutainment, Advergaming, Edumarket Games, Political Games and Training and Simulation Games (Alvarez et al., 2007). Bearing in mind the aforementioned characteristics of SG-based frameworks, we expect to contribute to the creation of the next-generation pedestrian simulators. A SERIOUS GAMES EVACUATION SIMULATOR The Serious Games Evacuation Simulator proposed in this research is based on the Unity3D game engine, that was selected due to its characteristics, among them: i) powerful graphical interface that allows visual object placement and property changing during runtime (especially useful to rapidly create new scenarios from existing models and assets and quick tweaking of script variables); ii) the ability to develop code in JavaScript, C# or Boo; iii) simple project deployment for multiple platforms without additional configuration, including for instance the Web (which makes it possible to run the game on a Web browser). Detailed characteristics of the implemented environment are presented below. Combining Simulation and Serious Games By starting the application the user gains control over the player character. Its aim will be to evacuate the building in the shortest time possible. The User Interface displays time, which starts the elapsed counting as soon as the player presses the “start fire simulation” key (illustrated in Figure 1). Figures 1: Gameplay example The game genre – First Person Shooter First Person Shooters (FPS) are characterised by placing players in a 3D virtual world which is seen through the eyes of an avatar. This attempts to recreate the experience of the user being physically there and exploring their surroundings. The controls for this game follow the common standards for the FPS genre, using a combination of keyboard and mouse to move the player around the environment. The complete action mapping is as follows:  Mouse movement - camera control, i.e. where the player is looking at;  W - move forward;  S - move backwards;  A - move to the left;  D - move to the right;  Space bar - jump;  O - start fire simulation. Game scenarios Figure 2 DEI plan and 3D representation The environment to support various is prepared scenarios modelled in 3D. For the trial described in this paper, a single simulation scenario was considered. It takes place Informatics Engineering in FEUP’s Department (DEI). A model of the FEUP campus was used, focusing only on one of the buildings where our a virtual located. As is laboratory research representation of the outside already existed, it was only necessary to create its interiors. This task was handled in Blender and used the official plans in order to recreate it as real as possible in terms of topology, dimensions, scale and proportions. Images of the plans and the 3D model are presented in Figure 2. The player starts in a predefined room and, upon starting the evacuation event, a fire appears in a random room and the alarm sounds. At this very moment the timer starts. The player must then traverse the building in order to go to the outside as quickly as possible, choosing from one of the two possible exits. Several emergency signs are in place in order to help the player identify the nearest exit. Challenges, Rules and Scoring Systems The main challenge involved in the evacuation of a building comes from identifying the exact location of the nearest exit and how to get there. Also to consider is that computer-controlled agents are present and trying to evacuate the building at the same time, possibly clogging the passage and delaying the player. After starting, fire keeps spreading to adjacent areas in small intervals of time; as fire is not surmountable, this can eventually constitute another obstacle and forces the player to look for a different exit route. At the current stage, the score given to a player is solely based on the time taken to evacuate the building – meaning that the lower the score, the better. Whether the player picked the nearest exit or not is inherently reflected in the time taken to reach the outside. Model calibration Calibration is an important issue to assure the validity of the model. For this purpose, three different paths were considered, named P1, P2 and P3. One in a straight line (P1), two involving taking sets of stairs. Of these latter two, one involved taking the nearest exit from the building (P2) and another the farthest one (P3). These paths were measured using the AutoCAD plans for the building. The comparison was made between data collected from real evacuations and from the game. The real times were measured with a stopwatch while traversing the paths, whereas for the game times the clock in the interface was used. It is also worth noticing that the adult profile of (1.5 m/s) was used in the game. Subject’s speed values were calculated from the measured distance and time taken. Error values were calculated according to the equation: (1) The values for distances and times are registered in Table 1. Table 1: Model Validation Distance (m) Real Time (s) Subject’s Speed (m/s) Game Time (s) Error (%) P1 24 17.53 1.34 15.86 9.53 P2 31 21.50 1.44 19.28 10.33 P3 72 55.91 1.29 48.08 14.00 One aspect to notice is that subject’s speed is consistent at around 1.3 and 1.4 m/s. Thus, subject’s times would always be longer than the ones registered in the game, as the player moves at 1.5 m/s. It is also worth considering that the error is higher for routes involving stairs. This was also expected as the player’s speed does not decrease when taking stairs, which in turn is verified in reality. EXPERIMENTAL SETUP Description All subjects were divided according to their previous knowledge of the building. Each had to try to reach a safe exit to the outside as quickly as possible. Besides the time, it was expected that players would select the nearest emergency exit, just outside of the laboratory set as starting point, instead of using the normal way which is longer. Users were tested individually so not to spoil the experience to each other regarding details of their chosen routes. Tests were also performed only once in order to capture first reactions to the game experience and its controls. Population Sample A total of 30 subjects were selected as sample to test the developed prototype. These testers can be classified according to the following parameters:  Regular video game player - Yes or No;  Familiar with the building -Yes or No. An attempt was made to equalise these variables, as well as age and gender, so as to receive as many different experiences as possible and maintain a balance among categories. The distribution is shown in Table 2. Table 2: User Times – Results by Categories Previous Knowledge of the Building Yes No Regular Video Game Player No Yes 6 8 5 11 Test Setup Each subject could play only once. Some time was given to the user so as to get acquainted with the keyboard and mouse controls. Players with no previous knowledge of the building were taken to the lab where they had to escape from. The purpose was to show, like a regular visitor (for instance a student or foreign professor) the normal way, from the building entrance, up to the first floor and end of the corridor, where the laboratory is located. After the siren signed, the player was instructed to leave the building following the emergency signs leading to the nearest exit. Preliminary results Intuitively it was expected that all subjects would selected the nearest exit available. However, some of the players misbehaved according to these expectations and chose the longer way out. These testers can be classified according to the following parameters , whose distribution is shown in Table 3. Table 3: Experiment Results Was the nearest exit chosen? Previous knowledge of the building Y N 11 2 No previous knowledge of the building 6 11 From the analysis of Table 3, it is possible to conclude that users with previous knowledge of the building were aware of the emergency exit and used it. Nevertheless, 2 of them (aprox. 15% - 2 out of 13) missed it and used the longer way out. The remaining players, only 6 out of 17 (aprox. 35%) chose to exit using the emergency way, whilst the remaining 11 of that group followed the same way they were shown initially to get to the starting point of this experience. CONCLUSION AND FUTURE WORK This work explores the concept of serious games as an important asset to aid and improve traditional fire drills. The contribution of this work can then be considered two-fold. First we extended a popular game engine to implement a pedestrian simulator to study evacuation dynamics. Second, our an approach provided appropriate environment to test with and influence behaviour of egresses of a building in hazardous situations, such as fires. It also addresses the common notion that people tend to leave buildings using the same way they use to get into it, unless they are told otherwise. This was highlighted by the experiment in which approximately 65% of players without previous knowledge of the building missed the emergency exit and signage, following the longer but more intuitive path to exit the building. It is important to bear in mind that this framework does neither completely replace nor avoid the need for in -site drills to train people for emergency situations, such as with the prospect of fire in an office building or school. Nonetheless, game environments can be very attractive in many different ways, and have proven to be an invaluable tool for training. Additionally, this approach is built upon the potential of such a concept to ease and improve the understanding of human behaviour in such situations, as subjects are monitored during their playing the game and some performance measures are logged to be further analysed later on. We have implemented our prototype on the basis of a popular game engine, namely Unity3D, which provided us with a customisable framework and allowed us to feature environment with virtual game the characteristics of a serious game platform. We invited some subjects to use the game and collected some preliminary results that demonstrated the viability of the approach. We have then conceived a methodology which is both instrumental as an aid to train people and an invaluable instrument to help practitioners and scientists to better understand group behaviour and the social phenomenon in a vast range of circumstances. The very next steps in this research include the improvement of the prototype featuring it with tools for rapidly setting up simulation environments from CAD blueprints of buildings. We also intend to include other performance measures to study individual and social behaviour in circumstances other the hazardous scenarios. Ultimately, this framework is also expected to be used as an imperative decision support tool, providing necessary and additional into insights evacuation plans, building layouts, and other design criteria to enhance places where people usually gather and interact rather socially, such as shopping malls, stadiums, airports, and so on. ACKNOWLEDGMENT This project has been partially supported by FCT (Fundação para a Ciência e a Tecnologia), the Portuguese Agency for R&D, under grant SFRH/BD/72946/2010. REFERENCES Almeida, J. E., Rosseti, R., and Coelho, A. L. 2011. “Crowd Simulation Modeling Applied to Emergency and Evacuation Simulations using Multi-Agent Systems”. In Proceedings of the 6th Doctoral Symposium on Informatics Engineering, DSIE’11, Porto. Alvarez, J., Rampnoux, O., Jessel, J.P., and Methel, G. 2007. “Serious Game: Just a Question of Posture? “. In Artificial & Ambient Inteligence, AISB’07, pages 420–423. Beyer, W.; Sellers, P.; Waterman, M. 1985 . “Stalinslaw m. Ulam’s contributions to theoretical theory”. Letter in Mathematical Physics 10:231-242. Castle, Christian J. E. 2007. Guidelines for Assessing Pedestrian Evacuation Software Applications. Centre for Advanced Spatial Analysis - UCL University College London. http://eprints.ucl.ac.uk/3471/1/3471.pdf. Coelho, L. 1997. Modelação de Evacuação de Edifícios Sujeitos à Acção de um Incêndio (in Portuguese), Ph.D. Dissertation, FEUP-LNEC, Lisboa. Cordeiro, E., Coelho, A. L., Rossetti, R. J. F., and Almeida, J. a. E. 2011. “Human Behavior under Fire Situations – A case–study in the Portuguese Society”. In the Proceedings of Advanced Research Workshop: Evacuation and Human Behavior in Emergency Situations, pages 63–80, Santander, Spain. GIDAI. Universidad de Cantabria. Freitas, S. (2006). Using Games and Simulations for Supporting Learning. Learning, Media and Technology, 31(4):343– 358. Frey, A., Hartig, J., Ketzel, A., Zinkernagel, A., and Moosbrugger, H. 2007. “The Use of Virtual Environments Based on a Modification of the Computer Game Quake III ArenaR in Psychological Experimenting”. Computers in Human Behavior, 23(4):2026–2039. Hays, R. 2005. “The Effectiveness of Instructional Games: a Literature Review and Discussion”. Technical report, Naval Air Warfare Center Training Systems Division Orlando, FL. Heïgeas, L., Luciani, A., Thollot, J., and Castagn´e, N. 2003. “A Physically-Based Particle Model of Emergent Crowd Behaviors”. In Graphicon. D. Helbing, I. Farkas, P. Molnar, T. Vicsek. 2001. “Simulating of Pedestrian Crowds in Normal and Evacuation Situations”. M.Schreckenberg, Sharma(ed.) S.D. Pedestrian and Evacuation Dynamics. Springer Verlag Berlin and Heidelberg, pp. 21-58. Hoogendoorn, S.P., and P.H.L. Bovy. 2004. “Pedestrian route- choice and activity scheduling theory and models”. Transportation Research Part B: Methodological 38 (2) (February): 169-190. Kirriemuir, J. and McFarlane, A. 2004. “Literature Review in Games and Learning”. Technical report, Futurelab. Kuligowski, E. D. 2008 “Modeling Human Behavior during Building Fires”. NIST Technical Note 1619. Kuligowski, E. D., Peacock, R., and Hoskins, B. L. 2010. “A Review of Building Evacuation Models, 2nd Edition”. NIST Technical Note 1680. Kuligowski, E. D. 2011. “Predicting Human Behavior During Fires.” Fire Technology (November 13). http://www.springerlink.com/index/10.1007/s10694-011- 0245-6 McGonigal, J. 2011. Reality Is Broken: Why Games Make Us Better and How They Can Change the World , volume 22. The Penguin Press HC. Neumann, V. 1966. Theory of self-reproducing automata. Champaign IL: University of Illinois Press. Pan, X., Han, C. S., Dauber, K., & Law, K. H. 2007. “A multi- agent based framework for the simulation of human and social behaviors during emergency evacuations”. Ai & Society 22 (2) (June 29) , 113-132. Pelechano, N., Allbeck, J., and Badler, N. 2007. “Controlling Individual Agents in High-density Crowd Simulation”. In ACM SIGGRAPH/Eurographics Proceedings of Symposium on Computer Animation, pages 99–108. Eurographics Association. Pretto, C. O. 2011. Desenvolvimento de um Simulador de Pedestres (in Portuguese). PhD thesis, Universidade Federal do Rio Grande do Sul. Qingge, J., Can G. 2007. “Simulating Crowd Evacuation with a Leader-Follower Model”. IJCSES International Journal of Computer Sciences and Engineering Systems, Vol.1, No.4, October 2007. Reynolds, C. 1987. “Flocks, Herds and Schools: A Distributed Behavioral Model”. ACM SIGGRAPH Computer Graphics, 21(4):25–34. Santos, G. and Aguirre, B. E. 2004. A critical review of emergency evacuation simulation models. Critical Review, (1032):25–50. 2002. Microscopic Pedestrian Flow Teknomo, K. Characteristics: Development of an Image Processing Data Collection and Simulation Model. Ph.D. Thesis. Tohoku University, Japan. Timmermans, H. 2009. Pedestrian Behavior: Models, Data Collection and Applications. Emerald Group Publishing Limited. AUTHOR BIOGRAPHIES JOAO PEDRO RIBEIRO concluded his MSc in Informatics and Computing Engineering in 2012, from Faculty of Engineering, University of Porto, Portugal. He specialised in Digital Games development and Artificial Intelligence, combining the concepts of multi- agent systems and serious games. He can be reached by e-mail at: [email protected]. in JOAO EMILIO ALMEIDA holds a BSc Informatics (1988), and MSc in Fire Safety Engineering (2008). He is currently reading for a PhD in Informatics Engineering at the Faculty of Engineering, University of Porto, Portugal, and a researcher at LIACC. He has co- authored many fire safety projects for complex buildings such as schools, hospitals and commercial centres. His areas of interest include Serious Games, Artificial Intelligence, and multi-agent systems; more specifically he is interested in validation methodologies for pedestrian and social simulation models. His e-mail is [email protected]. ROSALDO ROSSETTI is an Assistant Professor with the Department of Informatics Engineering at the University of Porto, Portugal. He is also a Research Fellow in the Laboratory of Artificial Intelligence and Computer Science (LIACC) at the same University. Dr. Rossetti is a member of the Board of Governors of IEEE Intelligent Transportation Systems Society (IEEE ITSS) and a co-chair of the Technical Activities sub - committee on Artificial Transportation Systems and Simulation of IEEE ITSS. His areas of interest include Artificial Intelligence and agent-based modelling and simulation for the analysis and engineering of complex systems and optimisation. His e-mail is [email protected] and his Web page can be found at http://www.fe.up.pt.com/~rossetti/. ANTONIO COELHO was born in 1971, in Porto, Portugal, and is currently an Assistant Professor at the Informatics Engineering Department of the Faculty of Engineering, University of Porto, where he teaches in the areas of Computer Graphics, Programming and Digital Games. He is also a Research Fellow at INESC TEC (INESC Technology and Science). His e-mail is [email protected]. A. LEÇA COELHO holds both the Electrotechnical and Civil Engineering degrees, as well as a Master’s and PhD in Civil Engineering. He is currently a Principal Researcher with Habilitation at LNEC. His areas of interest include fire safety and risk analysis. He can be reached by e-mail at [email protected].
1807.05283	3	1807	2018-10-11T10:15:18	When Are Two Gossips the Same? Types of Communication in Epistemic Gossip Protocols	[ "cs.MA" ]	We provide an in-depth study of the knowledge-theoretic aspects of communication in so-called gossip protocols. Pairs of agents communicate by means of calls in order to spread information---so-called secrets---within the group. Depending on the nature of such calls knowledge spreads in different ways within the group. Systematizing existing literature, we identify 18 different types of communication, and model their epistemic effects through corresponding indistinguishability relations. We then provide a classification of these relations and show its usefulness for an epistemic analysis in presence of different communication types. Finally, we explain how to formalise the assumption that the agents have common knowledge of a distributed epistemic gossip protocol.	cs.MA	cs	When Are Two Gossips the Same? Types of Communication in Epistemic Gossip Protocols⋆ Krzysztof R. Apt1,2, Davide Grossi3, and Wiebe van der Hoek4 1 CWI, Amsterdam, The Netherlands 2 MIMUW, University of Warsaw, Warsaw, Poland 3 Bernoulli Institute, University of Groningen, Groningen, The Netherlands 4 Department of Computer Science, University of Liverpool, Liverpool, U.K. Abstract. We provide an in-depth study of the knowledge-theoretic aspects of communication in so-called gossip protocols. Pairs of agents communicate by means of calls in order to spread information -- so-called secrets -- within the group. Depending on the nature of such calls knowl- edge spreads in different ways within the group. Systematizing existing literature, we identify 18 different types of communication, and model their epistemic effects through corresponding indistinguishability rela- tions. We then provide a classification of these relations and show its usefulness for an epistemic analysis in presence of different communica- tion types. Finally, we explain how to formalise the assumption that the agents have common knowledge of a distributed epistemic gossip proto- col. 1 Introduction In the gossip problem [31,6] a number of agents, each one knowing a piece of information (a secret ) unknown to the others, communicate by one-to-one in- teractions (e.g., telephone calls). The result of each call is that the two agents involved in it learn all secrets the other agent knows at the time of the call. The problem consists in finding a sequence of calls which disseminates all the secrets among the agents in the group. It sparked a large literature in the 70s and 80s [31,6,16,8,29], typically on establishing -- in the above and other variants of the problem -- the minimum number of calls to achieve dissemination of all the secrets. This number has been proven to be 2n − 4, where n, the number of agents, is at least 4. The gossip problem constitutes an excellent toy problem to study informa- tion dissemination in distributed environments. A vast literature on distributed protocols has taken up the problem and analyzed it together with a wealth of variations including different communication primitives (e.g., broadcasting instead of one-to-one calls), as well as communication structures (networks), ⋆ A shorter version of this paper appears in the Proceedings of the 22nd International Conference on Logic for Programming Artificial Intelligence and Reasoning (LPAR- 22). faulty communication channels [9], and probabilistic information transmission, where the spreading of gossips is used to model the spread of an epidemic [5,28]. Surveys are [13,22,19,23]. Background The present paper investigates a knowledge-based approach to the gossip problem in a multi-agent system. Agents perform calls following individ- ual epistemic protocols they run in a distributed fashion. These protocols tell the agents which calls to execute depending on what they know, or do not know, about the information state of the agents in the group. We call the resulting dis- tributed programs epistemic gossip protocols, or gossip protocols, for short. Such protocols were introduced and studied in [4,1]. 'Distributed' means that each agent acts autonomously, and 'epistemic' means that the gossip protocols refer to the agents' knowledge. The reliance of these protocols on epistemic properties makes them examples of so-called knowledge-based protocols, as studied in the context of distributed systems [27,25,18,11]. Besides the aforementioned [4,1], a number of papers have recently focused on epistemic gossip protocols. In [21] gossip protocols were studied that aim at achieving higher-order shared knowledge, for example knowledge of level 2 which stipulates that everybody knows that everybody knows all secrets. In particular, a protocol is presented and proved correct that achieves in (k + 1)(n − 2) steps shared knowledge of level k. Further, in [10] gossip protocols were studied as an instance of multi-agent epistemic planning that is subsequently translated into the classical planning language PDDL. More recently, [33] presented a study of dynamic gossip protocols in which the calls allow the agents not only to share the secrets but also to share the communication channels (that is, who can call whom). In turn, [2] studied the computational complexity of distributed epistemic gossip protocols, while [3] showed that implementability, partial cor- rectness, termination, and fair termination of these protocols is decidable. More broadly, the paper positions itself within the long-standing tradition of analysis of distributed systems from the perspective of epistemic logic [12,26]. Such a perspective has led in [27,25,11] to a useful level of abstraction allow- ing one to study a number of topics in distributed computing from the knowl- edge theoretic perspective, in particular protocols for the sequence transmission problem (for instance the alternating bit protocol) in [18], coordination [17], and secure communication [7], to mention some. The characteristic feature of these programs is that they use tests for knowledge. Contributions The form of communication underpinning the epistemic gossip problem may vary from work to work, and the above papers sometimes make different assumptions on the nature of communication upon which the consid- ered protocols are based. Little attention has been devoted to a systematic anal- ysis, with the notable exception of [15], which singled out some of the key in- formational assumptions on calls -- specifically observability, synchronicity and asynchronicity assumptions -- and systematically studied the effects of such as- sumptions on the aforementioned 2n − 4 call-length bound. 2 It is our claim that research on epistemic gossip protocols can at this point benefit from a systematisation of the key possible assumptions that a modeler can make on the type of communication (call) underpinning such protocols. From an epistemic logic point of view, each call type induces a specific notion of knowledge. The comparison of the resulting definitions of knowledge is of obvious importance for the study of epistemic aspects of communication. By 'type of communication' we mean the way in which communication takes place and may be observed, and to focus on it we disregard the type of infor- mation exchanged (in particular, whether higher order knowledge, or commu- nication links may be exchanged -- matters we do not address), or the type of information the agents have initially at their disposal (e.g., whether it is common knowledge what the number of agents is). More specifically, here are the features we focus on. First of all, a call between two agents takes place in the presence of other agents. What these other agents become aware of after the call is one natural parameter. We call it privacy. The second parameter, that we call direction, clarifies in which direction the informa- tion flows. Here we focus on three possibilities: they exchange all information, one agent passes all information to the other one, or one agent acquires all infor- mation available to the other one. The final parameter of a call is what we call observance. It determines whether the agent(s) affected by the call learn what information was held by the other agent prior to the call. By a call type we mean a combination of these three parameters. What the agents know after a call, or more generally a sequence of calls, depends on the assumed call type. This yields in total 18 possibilities. The paper provides a framework in which we model these possibilities in a unified way. This allows us to provide in Theorem 1 a complete classification of the resulting indistin- guishability relations. This in turn makes it possible to clarify in Propositions 2 and 3 the effect of a call type on the truth of the considered formulas. Addition- ally, we provide in Proposition 4 a natural proposal on how to incorporate into this framework an assumption that the agents have common knowledge of the underlying protocol. Paper outline Section 2 introduces gossip protocols by example, and identifies the features of calls we will focus on. Section 3 introduces the syntax and se- mantics of a simple epistemic language to study communication and its effects in gossip protocols, together with some motivating examples. Crucially the se- mantics introduced is parametrised by the indistinguishability relations which, for each call type, identify the call sequences that the agents cannot distinguish. These equivalence relations are systematically introduced and defined in Section 4, and then compared in terms of their relative informativeness in Section 5. The proposed systematisation is then applied in Section 6: first, to deliver general results on the analysis of how knowledge depends on the assumed call types (Section 6.1); second, to offer a natural approach to the problem of modelling common knowledge of protocols in the epistemic gossip setting (Section 6.2). Finally, Section 7 summarises our results and charts several directions for future research. 3 2 Knowledge-Based Gossip We start by recalling the notion of a gossip protocol, moving then to introduce the formal set-up of the paper. 2.1 Gossip protocols Gossip protocols aim at sharing knowledge between agents in a pre-described way. This is the paradigmatic setup: Six friends each know a secret. They can call each other by phone. In each call they exchange all the secrets they know. How many calls are needed for everyone to know all secrets? Let us generalise this to the case of n ≥ 2 agents and focus on protocols that are correct (in the sense that they spread all secrets). If n = 2, the two agents a and b need to make only one phone call, which we denote by ab ('a calls b'). For n = 3, the call sequence ab, bc, ca will do. Let us look at a protocol for n ≥ 4 agents. Protocol 1 Choose four from the set of agents Ag, say a, b, c, d, and one of those four, say a. First, a makes n − 4 calls to each agent in Ag \ {a, b, c, d}. Then, the calls ab, cd, ac, bd are made. Finally a makes another call to each agent from Ag \ {a, b, c, d}. This adds up to (n − 4) + 4 + (n − 4) = 2n − 4 calls. For n = 6 we get a call sequence ae, af , ab, cd, ac, bd, ae, af of 8 calls. All agents are then familiar with all secrets. It was shown that less than 2n − 4 calls is insufficient to distribute all secrets [31]. The above protocol assumes that the agents can coordinate their actions before making the calls. But often such coordination is not possible. Suppose some students of a given cohort receive an unexpected invitation for a party. The members of the cohort may be curious to find out about who received an invitation, in which case they will resort to phone calls based on the knowledge, or better, ignorance, they have about the secrets (in this context: extended invitations) of others. Since in such a distributed protocol several agents may decide to initiate a call at the same time, we assume the presence of an arbiter who breaks the ties in such cases. Let us now consider such an epistemic protocol. Protocol 2 (Hear my secret) Any agent a calls agent b if a does not know that b is familiar with a's secret. This protocol has been proven in [1] to terminate and be correct, under specific assumptions on the type of communication taking place during each call. In this paper we aim at providing a systematic presentation of such assumptions and at an analysis of their logical interdependencies. 4 Throughout the paper we assume a fixed finite set Ag of at least three agents. We further assume that each agent holds exactly one secret and that the secrets are pairwise different. We denote by S the set of all secrets, the secret of agent a by A, the secret of agent b by B, and so on. A secret can be any piece of data, for instance birthday, salary or social security number. Furthermore, we assume that each secret carries information identifying the agent to whom this secret belongs. So once agent b learns secret A she knows that this is the secret of agent a. 2.2 Calls In the context of gossip protocols calls constitute the sole form of knowledge acquisition the agents have at their disposal. Each call concerns two agents, the caller (a, below) and the callee (b, below). We call a the partner of b in the call, and vice versa. Any agent c different from a and b is called an outsider. We study the following properties of calls: -- privacy , which is concerned with what the outsiders note about the call, -- direction, which clarifies the direction of the information flow in the call, -- observance, which clarifies, when an agent a is informed by b, whether a sees b's secrets before adding them to her own set, or only sees the result of the fusion of the two sets of secrets. More specifically, we distinguish three privacy degrees of a call where agent a calls b: -- : every agent c 6= a, b notes that a calls b, -- : every agent c 6= a, b notes that some call takes place, though not between whom, -- : no agent c 6= a, b notes that a call is taking place. Intuitively, these degrees can be ordered as <p <p , with meaning no privacy at all, denoting full privacy. Conversely, from the perspective of the agents not involved in the is the most informative, while a call with the call, a call with the privacy level privacy level ensuring anonymity of the caller and callee, and is the most opaque. We distinguish three direction types, in short directions, of a call: -- push, written as ⊲. As a result of the call the callee learns all the secrets held by the caller. -- pull , written as ⊳. As a result of the call the caller learns all the secrets held by the callee. -- push-pull , written as ♦. As a result of the call the caller and the callee learn each other's secrets. Depending on the direction of a call between a and b, one or both agents can learn directly new information thanks to it. We say that these are the agents 5 affected in the call. More formally, an agent a is affected by a call c if c is one of the following forms: a♦b, b♦a, b ⊲ a, or a ⊳ b. Intuitively, a is affected by the call if it can affect the set of secrets a is familiar with. This brings us to two possible levels of observance of a call: -- α: During the call the affected agent(s) add the secrets of their partner to their own secrets, and only after that, inspect the result.5 -- β: During the call the affected agent(s) inspect the secrets of their partner before adding them to their own secrets. Intuitively, the observance level α is less informative for an affected agent than β, because in the latter case she also learns which secrets were known to the other agent before adding them to the secrets she is familiar with. Let -- P = { , , }, -- D = {♦, ⊳, ⊲}, -- O = {α, β}. Each call between agents a and b is of the shape ab τ , where τ = (p, d, o) ∈ P × D × O is called its type. So we defined in total 18 call types. To clarify their effect on communication we will elaborate on some representative call types in Examples 3 -- 5. The types ( , ♦, β) and ( , ♦, β) were studied in [4] while the types ( , ♦, α), ( , ⊲, α), and ( , ⊳, α), were analyzed in [1]. For a type τ like ( , ♦, β), we define τ (p) = , τ (d) = ♦ and τ (o) = β. Often, the call type (or parts of it) is (are) clear from the context, and we omit it (them). In our examples, at the level of calls, we often only explicitly mention the direction type. Given a call between a and b we shall sometimes write it simply as ab for the direction type ♦, a ⊲ b for the direction type ⊲ and a ⊳ b for the direction type ⊳. 3 Language and Semantics In this section we introduce a modal language for epistemic gossip and its formal semantics. 3.1 Modal language We are interested in determining agents' knowledge after a sequence of calls took place. To this end we use the following modal language L for epistemic logic: φ ::= FaS ¬φ φ ∧ φ φ ∨ φ Kaφ, 5 This mode is akin to the caller and callee interacting through a third party, who first collects the caller's and callee's secrets separately, and then shares their union with the affected agent(s). We are indebted to R. Ramezanian for this observation. 6 where a ∈ Ag and S ∈ S. In what follows we refer to the elements φ of L as epistemic formulas, or in short, just formulas. We read FaS as 'agent a is familiar with the secret S' (or 'S belongs to the set of secrets a has learned') and Kaφ as 'agent a knows that formula φ is true'. So L is an epistemic language with the atomic formulas of the form FaS. The above language was introduced in [1]. It is a modification of the language introduced in [4]. Example 1. Consider the statement that agent a is familiar with all the secrets. This can be expressed as the formula FaB ^ b∈Ag that we subsequently abbreviate to Expa ("a is an expert"). Here and elsewhere for simplicity we refer in the conjunction limits only to agents and not to their secrets. This convention allows us to write more complex statements, for instance that each agent is familiar only with her own secret. This can be expressed as the formula ^ a∈Ag (FaA ∧ ^ b∈Ag,b6=a ¬FaB). (1) Finally, consider the statement that for each agent, say a, it is not the case that a is an expert and each other agent is familiar with at most her own secret and that of a. This can be expressed as the formula ¬(Expa ∧ ^ a∈Ag ^ b,c∈Ag, {a,b,c}=3 ¬FbC). (cid:3) Next, we clarify the use of the knowledge operators. In the presented reason- ing we assume that the agents have the knowledge of the underlying call type. In all cases we assume that the initial situation is the one in which every agent is only familiar with her own secret, that is, we assume (1) to be true for each agent before any communication takes place. The examples provide intuitions about how agents' knowledge is influenced by the types of calls underpinning their communications. Such intuitions will then be formalised in Section 3.2. Example 2. Initially, each agent is familiar with her secret and each agent knows this fact. Additionally, she does not know that any other agent is familiar with a secret different from her own. This can be expressed by means of the formula ^ a∈Ag ( ^ b∈Ag KaFbB ∧ ^ b,c∈Ag,a6=b,b6=c ¬KaFbC) that holds initially, for all call types. (cid:3) 7 Example 3. Suppose there are four agents, a, b, c and d. Consider the call type is ( , ♦, α). Assume the call sequence ab, bc. Let us reason from the perspective of agent d. Because of the assumed privacy level, after the first call, ab, agent d knows that both agents a and b are familiar with A and B. This can be expressed as the formula Kd(FaA ∧ FaB ∧ FbA ∧ FbB). This then implies that after the second call, bc, agent d also knows that both b and c are familiar with A, B, and C. Agent's d factual knowledge after the second call can be expressed as the formula Kdφ, where φ = FaA ∧ FaB ∧ ^ i∈{b,c}, j∈{a,b,c} FiJ. In fact, because of the assumed privacy level , how the knowledge evolves during communication is completely transparent to all agents. Hence after both calls everybody knows φ, i.e., Kaφ. ^ a∈Ag An analogous argument applies for the call type ( , ♦, β). Suppose now that the privacy level is . Then we cannot conclude the formula Kdφ after the second call, since agent d only knows then that two calls took place, but not between which pairs of agents. In fact, in this case we can only conclude (note that the same call can be made twice): Kd( _ (FaB ∧ FbA)). a,b∈Ag\{d},a6=b Finally, if the privacy level is , then d is not aware of the calls ab and bc. She considers it possible that a, b, c are already familiar with all secrets except her own, but also considers it possible that all other agents only know their own secret. As she has not yet been involved in any call, she knows that they are not familiar with D. So after the call sequence ab, bc agent's d knowledge can be expressed as Kd( ^ e∈Ag\{d} (FeE ∧ ¬FeD)). Example 4. Suppose there are three agents, a, b and c. Consider the two call types ( , ♦, o), where o ∈ O, and assume the call sequence ac, bc, ab. After it the agents a and b (and c too) are familiar with all the secrets, which can be expressed as the formula φ = Expa ∧ Expb, and both know this fact, which can be expressed as Kaφ ∧ Kbφ. (cid:3) 8 If the observance of the calls is β, agent a also learns that prior to the call ab agent b was familiar with a's secret, i.e., with A. This allows a to conclude that agent b was involved in a call with c and hence agent c is familiar with B. We can express this as KaFcB. Contrast the above with the situation when the observance is α. Although again after the considered call sequence both agents a and b are familiar with all the secrets, now agent a cannot conclude that agents b and c communicated. Hence agent a does not know whether agent c is familiar with B, i.e., the formula KaFcB is not true. In both cases agent c (who also is an expert) does not know that agents a and b communicated, so she does not know that they are experts. In other words, the formula Kcφ is not true. This changes when the privacy degree is , i.e., in that case the formula Kcφ is true. Moreover, because there are three agents, the same conclusion holds when the privacy degree is . However, the last conclusion does not hold anymore when there are more than three agents. (cid:3) Example 5. Assume the same call sequence as in the previous example but sup- pose that the call parameters are now ( , ⊳, o), where o ∈ O. So we consider now the call sequence c = a ⊳ c, b ⊳ c, a ⊳ b. Because of the assumed privacy level, after this call sequence agent a knows that agent b learned the secret C and agent c knows that agent a learned the secret B, i.e., the following holds after c KaFbC ∧ KcFaB. Suppose now the privacy degree is and the observance is β. Then we only have KaFbC as agent a cannot distinguish c from a ⊳ c, c ⊳ b, a ⊳ b. Clearly, KcFaB does not hold after c as agent c cannot distinguish c from a ⊳ c, c ⊳ b, b ⊳ a. Finally, if the privacy degree is hold after c either. then for the same reason KcFaB does not (cid:3) We conclude that what the agents know after a call sequence crucially de- pends on the parameters of the calls. Further, the precise effect of a single call on the agents' knowledge is very subtle, both for the agents involved in it and for the outsiders. 3.2 Semantics We provide now a formal semantics for the modal language L. It is parameterized by a call type τ . Gossip situations and calls First we recall the following crucial notions intro- duced in [1]. A gossip situation is a sequence s = (Qa)a∈Ag, where Qa ⊆ S for each agent a. Intuitively, Qa is the set of secrets agent a is familiar with in the situation s. Given a gossip situation s = (Qa)a∈Ag, we denote Qa by sa. The 9 initial gossip situation is the one in which each Qa equals {A} and is denoted by i (for "initial"). The initial gossip situation reflects the fact that initially each agent is familiar only with her own secret. Each call transforms the current gossip situation by possibly modifying the set of secrets the agents involved in the call are familiar with. The definition depends solely on the direction of the call. Definition 1. The application of a call c to a gossip situation s is defined as follows, where s := (Qa)a∈Ag: c = ab c(s) = (Q′ c = a ⊲ b c(s) = (Q′ a)a∈Ag, where Q′ a = Q′ a)a∈Ag, where Q′ b = Qa ∪ Qb, Q′ b = Qa ∪ Qb, Q′ c = Qc, for c 6= a, b. a = Qa, Q′ c = Qc, for c 6= a, b. c = a ⊳ b c(s) = (Q′ a)a∈Ag, where Q′ a = Qa∪Qb, Q′ b = Qb, Q′ c = Qc, for c 6= a, b. This definition captures the meaning of the direction type: for ab the secrets are shared between the caller and callee , for a ⊲ b they are pushed from the caller to the callee, and for a ⊳ b they are retrieved by the caller from the callee. Note that (a♦b)(s) = (b♦a)(s) and (a ⊲ b)(s) = (b ⊳ a)(s), as expected. In turn, the privacy degree of a call captures what outsiders of the call learn from it and the observance level determines informally what caller and callee can learn about each other's calling history. The meaning of these two parameters will be determined by means of the appropriate equivalence relations between call sequences. A call sequence is a finite sequence of calls, all of the same call type. The empty sequence is denoted by ǫ. We use c to denote a call sequence and Cτ to denote the set of all call sequences of call type τ . Given the call sequence c and a call c, c.c denotes the sequence obtained by appending c with c. The result of applying a call sequence c to a situation s is defined by induction using Definition 1, as follows [Base] ǫ(s) := s, [Step] c.c(s) := c(c(s)). Note that this definition does not depend on the privacy degree and observance of the calls. Example 6. Let Ag be {a, b, c}. We use the following concise notation for gossip situations. Sets of secrets will be written down as lists. E.g., the set {A, B, C} will be written as ABC. Gossip situations will be written down as lists of lists of secrets separated by dots. E.g., i = A.B.C and the gossip situation ({A, B}, {A, B}, {C}) will be written as AB.AB.C. So, (ab)(A.B.C) = AB.AB.C, (ab, ca)(A.B.C) = ABC.AB.ABC and (ab, ca, ab)(A.B.C) = ABC.ABC.ABC. (cid:3) 10 Truth of formulas We illustrated in Examples 3 -- 5 that each call has an effect on the knowledge of the agents. After a sequence of calls took place the agents may be uncertain about the current gossip situation because they do not know which call sequence actually took place. This leads to appropriate indistinguishability relations that allow us to reason about the knowledge of the agents. This is in a nutshell the basis of the approach to epistemic gossip protocols put forth in [1], and upon which we build here. To clarify matters consider the situation analyzed in Example 4. We noticed there that depending on the assumed observance level the knowledge of agent a differs. This has to do with the call sequences the agent considers possible. If the call type is ( , ♦, α) agent a cannot distinguish between the call sequences ac, ab and ac, bc, ab. Indeed, after both sequences she is familiar with all the secrets but she cannot determine whether agents b and c communicated. From her perspective both call sequences are possible, that is, she cannot distinguish between them. In contrast, if the call type is ( , ♦, β) agent a can distinguish between these two call sequences, which has in turn an effect on her knowledge. In general, to determine what agents know after a call sequence we need then to consider an appropriate equivalence relation between the call sequences. Let c and d be two call sequences of call type τ and a an agent. The statement c ∼τ a d informally says that agent a cannot distinguish between c and d. The definition of ∼τ a crucially depends on the call type τ and is provided in the next subsection. Here we assume that it is given and proceed to define the truth of the formulas of the language L with respect to a gossip model (for a given set of agents Ag) Mτ = (Cτ , {∼τ a}a∈Ag) and a call sequence c as follows: Definition 2. Let Mτ be a gossip model for a call type τ and a set of agents Ag, and let c ∈ Cτ . The truth relation for language L is inductively defined as follows (with Boolean connectives omitted): (Mτ , c) = FaS iff S ∈ c(i)a, (Mτ , c) = Kaφ iff ∀d ∈ Cτ such that c ∼τ a d, (Mτ , d) = φ. Since the gossip model is clear from the context, we will from now on write c =τ φ for (Mτ , c) = φ. We also write Mτ = φ (φ is valid in Mτ ) if for all c ∈ Cτ we have Mτ , c = φ. So the formula FaS is true after a sequence of calls c whenever agent a is familiar with the secret S in the gossip situation generated by c applied to the initial gossip situation i. The knowledge operator Ka is interpreted as is cus- tomary in the multimodal S5n logic (see, e.g., [26,32]), so using the equivalence relation ∼τ a. It is important to notice that to determine the truth of a propositional for- mula (so in particular to determine which secrets an agent is familiar with) only the direction parameter of the type of the calls is used. In contrast, to determine the truth of formulas involving the knowledge operator all three parameters of the call type are needed, through the definition of the ∼τ a relations, to which we turn next. 11 4 Indistinguishability of Call Sequences Below we say that an agent a is involved in a call c, and write a ∈ c, if a is one of the two agents involved in it, i.e., if it is either a caller or a callee in c. So agent a is involved but not affected (a notion introduced in Section 2) by a call c if c = a ⊲ b or c = b ⊳ a for some agent b. 4.1 The ∼ τ a relations a⊆ Cτ × Cτ in two steps. First we define the auxiliary relation ≈τ For every call type τ and agent a we define the indistinguishability relation ∼τ a (Definition 3). Intuitively, the expression c ≈τ a d can be interpreted as "from the point of view of a, if c is an (epistemically) possible call sequence, so is d, and vice versa". Then, we define ∼τ a as the least equivalence relation that contains ≈τ a. Definition 3. Let a ∈ Ag and fix a type τ . The relation ≈τ subset of Cτ × Cτ satisfying the following conditions: a is the smallest [Base] ǫ ≈τ [Step] Suppose that c ≈τ a ǫ. a d and let c and d be calls. [Step-outτ] if Out τ [Step-inτ] if In τ a(c, d) then Concl τ a(c, d, c) then Concl τ a(c, d, c, d), a(c, d, c), where the used relations are defined in Table 1. (b is there the partner of a in the call c.) The definition of ≈τ a captures the complex effect of each of the three pa- rameters of a call type on the knowledge of an agent. Let us discuss it now in detail. The Base condition is clear. Consider now the Step-outτ clause which refers to Table 1, top. Suppose that c ≈τ a d. Consider first the privacy type . According to its informal description the condition a 6∈ c means that agent a is not involved in the call c but knows who calls whom. The conclusion c.c ≈τ a d.c then coincides with this intuition. Consider now the privacy type . The conditions a 6∈ c and a 6∈ d mean that agent a is not involved in the calls c and d, thus according to the informal description of she cannot distinguish between these two calls. This explains the conclusion c.c ≈τ a d.d. Note that this conclusion is not justified for the privacy because if c 6= d then agent a can distinguish between these two calls, so type a fortiori between the call sequences c.c and d.d. Finally, consider the privacy type . According to its informal description, the condition a 6∈ c means that agent a is not aware of the call c. This justifies the conclusions c.c ≈τ a d and c ≈τ Next, consider the Step-inτ clause. It spells the conditions that allow one to extend the ≈τ a relation in case agent a is involved in the last call, c. Table 1, middle, formalises the intuition that when agent a is not affected by the call c, then we can conclude that c.c ≈τ a d.c. a d.c. 12 Agent a is not involved in the last call: τ (p) Out τ a(c, d) Concl τ a(c, d, c, d) a 6∈ c a 6∈ c, a 6∈ d a 6∈ c a d.c c.c ≈τ c.c ≈τ a d.d a d, c ≈τ c.c ≈τ a d.c Agent a is involved in but not affected by the last call: In τ a(c, d, c) Concl τ a(c, d, c) c ∈ {a ⊲ b, b ⊳ a} c.c ≈τ a d.c Agent a is involved in and affected by the last call: τ (o) In τ a(c, d, c) Concl τ a(c, d, c) α c ∈ {a♦b, b♦a, b ⊲ a, a ⊳ b}, c.c ≈τ a d.c c.c(i)a = d.c(i)a β c ∈ {a♦b, b♦a, b ⊲ a, a ⊳ b}, c.c ≈τ a d.c c(i)b = d(i)b Table 1. Defining indistinguishability of call sequences Table 1, bottom, focuses on the remaining case. Consider first the observance α. According to its informal description, affected agents incorporate the secrets of their partner with their own secrets and then inspect the result. So we check what secrets agent a is familiar with after the call sequences c and d are both extended by c. If these sets are equal, then we can conclude that c.c ≈τ a d.c. In the case the observance is β, the informal description stipulates that the agent inspects the set of secrets of the call partner before incorporating them with their own secrets. So we compare these sets of secrets after, respectively, the call sequences c and d took place. If these sets are equal, then we conclude that c.c ≈τ a d.c. This explains why in this case a reference to agent b is made in In τ a(c, d, c). 4.2 Examples and a useful observation Example 7. We first illustrate Table 1, top, by analyzing situations in which the considered agent is not involved in the last call. Assume four agents, a, b, c and d. Suppose that the privacy of τ is . We have ab, bc 6∼τ a ab, cd, because ab, bc 6≈τ a ab, cd as bc 6= cd and bc 6= dc. So we fail to apply Table 1, top, first row and the transitive reflexive closure does not give us that either. On the other hand, if the privacy of τ is a ab, cd, as a 6∈ bc, a 6∈ cd and ab ∼τ ab, bc ≈τ the other hand, ab, bc 6∼τ apply, as the lengths of the compared sequences are different. a ab, cd, because a ab (Table 1, top, second row). On a ab, cd, bc as now the clause in the second row fails to , we have ab, bc ∼τ 13 Finally, if the privacy of τ is a ab, cd for the same reason as in the previous paragraph, but we now also have ab, bc ∼τ a ab, cd, bc, because ab, bc ≈τ a ab and hence by Table 1, top, third row, applied three times, first ab ∼τ a ab, cd, then ab, bc ∼τ a ab, cd, and finally ab, bc ∼τ (cid:3) a ab, cd, bc. Indeed, we have ab ∼τ , we of course also have ab, bc ∼τ a ab, cd, bc. Example 8. To illustrate Table 1, middle, consider the same four agents and sequence d ⊲ c, b ⊲ c, and . Then d ⊲ c, b ⊲ c ∼τ b c ⊲ d, b ⊲ c, because agent b is involved in the second call but not affected (Table 1, middle), and d ⊲ c ∼τ b c ⊲ d, because b 6∈ d ⊲ c and b 6∈ c ⊲ d (Table 1, top, second row). (cid:3) Example 9. Now consider Table 1, bottom. The difference between observances α and β is seen in Example 5. For the observancy α we have that a ⊳ c, b ⊳ c, a ⊳ b ∼τ a a ⊳ c, c ⊳ b, a ⊳ b, because agent a is afterwards familiar with the same secrets on the lefthand side and the righthand side, namely A, B, C (Table 1, bottom, first row). On the other hand, for the observancy β we get a ⊳ c, b ⊳ c, a ⊳ b 6∼τ a a ⊳ c, c ⊳ b, a ⊳ b, because a ⊳ c, b ⊳ c 6∼τ b a ⊳ c, c ⊳ b (note that this concerns indistin- guishability for agent b, not a); here the second row of Table 1, bottom, applies. c c ⊲ d, b ⊲ c, because c is involved but not affected in call c ⊲ d, while it is involved and affected in call d ⊲ c. Observe that after d ⊲ c, b ⊲ c agent c is familiar with the secrets B, C, D, whereas after c ⊲ d, b ⊲ c agent c is only familar with B, C. (cid:3) As a final example, we have that d ⊲ c, b ⊲ c 6∼τ Let us focus now on some properties of the ∼τ a equivalence relations. Note 1. For all agents a and call types τ ∼τ a= (≈τ a)∗, where ∗ is the transitive, reflexive closure operation on binary relations. Proof. A straightforward proof by induction show that each ≈τ metric. This implies the claim. a relation is sym- (cid:3) The following observation will be needed later. Proposition 1. For all call types τ if c ∼τ a d, then c(i)a = d(i)a. Proof. By Note 1 it is sufficient to prove the conclusion under the assumption that c ≈τ a d. We proceed by induction on the sum k of the lengths c + d of both se- quences. If k = 0, then c = d = ǫ, so the claim holds. Suppose the claim holds for all pairs of sequences such that the sum of their lengths is < k and that k > 0, c + d = k and c ≈τ a is the smallest relation satisfying the Base and Step conditions of Definition 3. Let c be the last call of c or of d if c is empty. a d. By definition ≈τ If agent a is not involved in c, then four cases arise, depending on the form of c and d. We consider one representative case, when c is of the form c′.c, where c′ ≈τ a d. Then by the assumption about c and the induction hypothesis c(i)a = c′.c(i)a = c′(i)a = d(i)a. 14 If agent a is involved in but not affected by the last call, then c is of the form a d′. Then by the form of c c′.c, d is of the form d′.c, c ∈ {a ⊲ b, b ⊳ a} and c′ ≈τ and the induction hypothesis c(i)a = c′.c(i)a = c′(i)a = d′(i)a = d′.c(i)a = d(i)a. Finally, if agent a is involved in and affected by the last call, then c is of the form c′.c, d is of the form d′.c, c ∈ {a♦b, b♦a, b ⊲ a, a ⊳ b} and c′ ≈τ a d′. If τ (o) = α, then by assumption c′.c(i)a = d′.c(i)a, i.e., c(i)a = d(i)a. If τ (o) = β, then by assumption c′(i)b = d′(i)b. Also, by the induction hypothesis c′(i)a = d′(i)a, so by the form of c c(i)a = c′.c(i)a = c′(i)a ∪ c′(i)b = d′(i)a ∪ d′(i)b = d′.c(i)a = d(i)a. (cid:3) Corollary 1. For all call types τ , agents a, b and call sequences c c =τ KaFaB iff c =τ FaB. Proof. By Proposition 1 and the definition of truth of KaFaB and FaB. ⊓⊔ 5 Classification of the ∼ τ a Relations We introduced in the previous section 18 equivalence relations ∼τ a, each parame- trised by an agent a. The uniform presentation makes it possible to compare these relations by means of a classification, which we now provide. First, let us introduce some notation. Given two call types τ1 and τ2 we abbreviate the statement ∀a ∈ Ag, ∼τ1 a to τ1 ⊂ τ2 and similarly for τ1 ⊆ τ2 and τ1 = τ2. Such statements presuppose that we systematically change the types of all calls in the considered call sequences. a ⊂∼τ2 The following theorem provides the announced classification. It clarifies in total 153 (= 18·17 2 ) relationships between the equivalence relations. Theorem 1. The ∼τ a equivalence relations form preorders presented in Figures 1 and 2. An arrow → from τ1 to τ2 stands here for τ1 ⊂ τ2, ( , d, o) for the set of six call types with the privacy degree that are all equal, and ( , ♦, o) for the set {( , ♦, α), ( , ♦, β)}. The remainder of this section is devoted to the proof of Theorem 1. Below we say that the call types τ1 and τ2 are incomparable when neither τ1 ⊆ τ2 nor τ2 ⊆ τ1 holds. The proofs concerning the incomparability that are established below also hold for a stronger definition, namely that τ1 and τ2 are incomparable when for all agents a neither ∼τ1 a holds. This way Figures 1 and 2 can be alternatively interpreted as preorders on the ∼τ a equivalence relations, for any agent a, where an arrow → from τ1 to τ2 stands then for ∼τ1 a nor ∼τ2 a ⊆ ∼τ1 a ⊆ ∼τ2 a ⊂∼τ2 a . We first establish the claimed equalities between the call types. 15 ( , ⊲, α) ( , ♦, α) ( , ⊳, α) ( , ⊲, β) ( , ⊲, α) ( , ♦, β) ( , ⊳, β) ( , ⊳, α) ( , ⊲, β) ( , ♦, o) ( , d, o) ( , ⊳, β) Fig. 1. Classification of the ∼τ a relations when Ag = 3. ( , ⊲, α) ( , ♦, α) ( , ⊳, α) ( , ⊲, β) ( , ⊲, α) ( , ♦, β) ( , ♦, α) ( , ⊳, β) ( , ⊳, α) ( , ⊲, β) ( , ♦, β) ( , d, o) ( , ⊳, β) Fig. 2. Classification of the ∼τ a relations when Ag > 3. Lemma 1. (i) Suppose that τ (p) = . Then each ∼τ (ii) Suppose that τ1(p) = τ2(p) = . Then τ1 = τ2. (iii) If Ag = 3 then ( , ♦, β) = ( , ♦, α). a is the identity relation. a d. By definition ≈τ Proof. (i) By Note 1 it is sufficient to prove that c ≈τ a d implies c = d. We proceed by induction on the sum k of the lengths c + d of both sequences. If k = 0, then c = d = ǫ, so the claim holds. Suppose the claim holds for all pairs of sequences such that the sum of their lengths is < k and that k > 0, c + d = k and c ≈τ a is the smallest relation satisfying the Base and Step con- ditions of Definition 3. So, since τ (p) = , by the Step condition c is of the a d′. By the induction hypothesis form c′.c and d is of the form d′.c, where c′ ≈τ c′ = d′, so c = d. (ii) By (i). (iii) Suppose Ag = {a, b, c}. Take τ ∈ {( , ♦, β), ( , ♦, α)}. Then by Definition 3 c ∼τ a d iff c and d differ only in some of the calls a is not involved in. Because there are exactly 3 agents, each such call must be b♦c or c♦b and both have the same effect independently of the type of observance. (cid:3) 16 Next we establish the claimed strict inclusions. Below the unspecified param- eters are implicitly universally qualified. For example, ( , d, o) ⊂ ( , d, o) is an abbreviation for the statement ∀a ∈ Ag ∀d ∈ D ∀o ∈ O ∼( ,d,o) a ⊂ ∼( ,d,o) a . Lemma 2. (i) ( , d, o) ⊂ ( , d, o). (ii) ( , d, o) ⊂ ( , d, o). (iii) If Ag > 3 or d 6= ♦ then ( , d, β) ⊂ ( , d, α). (iv) If Ag = 3, d 6= ♦ and o1, o2 ∈ O, then ( , d, o1) ⊂ ( , ♦, o2). (v) ( , d, β) ⊂ ( , d, α). Proof. First we establish the ⊆ inclusions. (i) and (ii) These are direct consequences of Definition 3. (iii) We prove that ( , d, β) ⊆ ( , d, α) always holds. Let τ1 = ( , d, β) and τ2 = ( , d, α). Fix an agent a. a d implies c ≈τ2 By Note 1 it is sufficient to prove that c ≈τ1 a d. We proceed by induction on the sum k of the lengths c + d of both sequences. If k = 0, then c = d = ǫ, so the claim holds. Suppose the claim holds for all pairs of sequences such that the sum of their lengths is < k and that k > 0, c + d = k and c ≈τ1 a is the smallest relation satisfying the Base and Step conditions of Definition 3. So, since τ1(p) = , by the Step condition c is of the form c′.c and d is of the form d′.d, where c′ ≈τ1 a d′. By the induction hypothesis c′ ≈τ2 a d. By definition ≈τ1 a d′. Three cases arise that reflect the case analysis in Definition 3, where b is the partner of a in the call c: (a) a 6∈ c, a 6∈ d. Then c′ ≈τ2 a d′ implies c′.c ≈τ2 a d′.d. (b) c ∈ {a ⊲ b, b ⊳ a}. Then c′.c ≈τ1 d′.d. a d′.d implies c = d and consequently c′ ≈τ2 a d′ implies c′.c ≈τ2 a (c) c ∈ {a♦b, b♦a, b ⊲ a, a ⊳ b}. Then c′.c ≈τ1 by Proposition 1 c′(i)a = d′(i)a, so a d′.d implies c = d and c′(i)b = d′(i)b, because τ1(o) = β. Also c′.c(i)a = c′(i)a ∪ c′(i)b = d′(i)a ∪ d′(i)b = d′.c(i)a. Hence c′ ≈τ2 a d′ implies c′.c ≈τ2 a d′.d, because τ2(o) = α. (iv) Let τ1 = ( , d, o1) and τ2 = ( , ♦, o2). Fix an agent a. By Note 1 it suffices to prove that c ≈τ1 a d implies c ≈τ2 a d. Two cases arise. (a) d = ⊲. Because there are only 3 agents, by Definition 3 if c ≈τ1 a d then c and d differ only in some of the calls a is not involved in. So then c and d, when 17 interpreted under τ2, differ only in some of the calls between b and c, which are b♦c or c♦b, and both have the same effect independently of the type of observance. So c ∼τ1 a d implies c ∼τ2 a d. (b) d = ⊳. The argument is the same as in (a). (v) The proof is analogous to the one given in (iii) and is omitted. We prove now that the inclusions are strict. Suppose that a, b, c ∈ Ag are different agents. (i) Note that for τ1 = ( , d, o) and τ2 = ( , d, o) we have bc ∼τ2 a bc 6∼τ1 (ii) Note that for τ1 = ( , d, o) and τ2 = ( , d, o) we have bc ∼τ2 a cb. a ǫ, while bc 6∼τ1 a ǫ. cb, while (iii) Let τ1 = ( , d, β) and τ2 = ( , d, α). Assume first that Ag > 3. Suppose a, b, c, d ∈ Ag are different agents. Three cases arise. (a) d = ♦. Then a♦b, a♦c, b♦c, a♦b ∼τ2 a♦b, a♦c, c♦d, a♦b. a a♦b, a♦c, c♦d, a♦b, while a♦b, a♦c, b♦c, a♦b 6∼τ1 a (b) d = ⊲. Then c⊲a, b⊲c, b⊲a ∼τ2 a c⊲a, c⊲b, b⊲a, while c⊲a, b⊲c, b⊲a 6∼τ1 a c⊲a, c⊲b, b⊲a. (c) d = ⊳. a ⊳ c, c ⊳ b, a ⊳ b ∼τ2 a a ⊳ c, b ⊳ c, a ⊳ b, while a ⊳ c, c ⊳ b, a ⊳ b 6∼τ1 a a ⊳ c, b ⊳ c, a ⊳ b. Assume now that d 6= ♦. Then the desired conclusion is established in (b) and (c), as the examples used there involve only three agents. (iv) Let τ1 = ( , d, o1) and τ2 = ( , ♦, o2). Assume Ag = {a, b, c}. Two cases arise. (a) d = ⊲. Note that c♦b, c♦a ∼τ2 a b♦c, c♦a, while c ⊲ b, c ⊲ a 6∼τ1 a b ⊲ c, c ⊲ a. (b) d = ⊳. Note that b♦c, a♦c ∼τ2 a c♦b, a♦c, while b ⊳ c, a ⊳ c 6∼τ1 a c ⊳ b, a ⊳ c. (v) Let τ1 = ( , d, α) and τ2 = ( , d, β). Three cases arise. (a) d = ♦. Note that a♦c, a♦b ∼τ1 a a♦c, b♦c, a♦b, while a♦c, a♦b 6∼τ2 a a♦c, b♦c, a♦b. (b) d = ⊲. Note that c ⊲ a, b ⊲ a ∼τ1 a c ⊲ a, c ⊲ b, b ⊲ a, while c ⊲ a, b ⊲ a 6∼τ2 a c ⊲ a, c ⊲ b, b ⊲ a. (c) d = ⊳. Note that a ⊳ c, a ⊳ b ∼τ1 a a ⊳ c, b ⊳ c, a ⊳ b, while a ⊳ c, a ⊳ b 6∼τ2 a a ⊳ c, b ⊳ c, a ⊳ b. (cid:3) 18 As a side remark note that Lemmas 1(ii), (iii) and 2(iii), (v) imply that (p, d, β) ⊆ (p, d, α). Finally, we establish the claims concerning incomparability of the types. Lemma 3. Let d, d1, d2 ∈ D and o1, o2 ∈ O. (i) Suppose that Ag > 3 or ♦ 6∈ {d1, d2}, and d1 6= d2. Then ( , d1, o1) and ( , d2, o2) are incomparable. (ii) Suppose that d1 6= d2. Then ( , d1, o1) and ( , d2, o2) are incomparable. (iii) Suppose that Ag = 3 and d 6= ♦. Then ( , ♦, α) and ( , d, α) are incompa- rable. (iv) Suppose that Ag > 3 or ♦ 6∈ {d1, d2}, and d1 6= d2. Then ( , d1, β) and ( , d2, α) are incomparable. (v) Suppose that Ag > 3 or d 6= ♦. Then ( , d, α) and ( , d, β) are incompa- rable. Proof. Suppose that a, b, c ∈ Ag are different agents. (i) Let τ1 = ( , d1, o1) and τ2 = ( , d2, o2). Assume first that Ag > 3. Suppose a, b, c, d ∈ Ag are different agents. For each pair of distinct direction types we exhibit appropriate call sequences. In each case the conclusions do not depend on the observance level. (a) d1 = ⊳ and d2 = ♦. Then b ⊳ c, c ⊳ a ∼τ1 Further, b♦c, a♦c ∼τ2 a b ⊳ d, c ⊳ a, while b♦c, c♦a 6∼τ2 a c♦b, a♦c, while b ⊳ c, a ⊳ c 6∼τ1 a b♦d, c♦a. a c ⊳ b, a ⊳ c. (b) d1 = ⊲ and d2 = ♦. Then c ⊲ b, a ⊲ c ∼τ1 Further, c♦b, c♦a ∼τ2 a d ⊲ b, a ⊲ c, while c♦b, a♦c 6∼τ2 a b♦c, c♦a, while c ⊲ b, c ⊲ a 6∼τ1 a d♦b, a♦c. a b ⊲ c, c ⊲ a. (c) d1 = ⊲ and d2 = ⊳. Then b ⊲ c, a ⊲ c ∼τ1 Further, c ⊳ b, c ⊳ a ∼τ2 a c ⊲ b, a ⊲ c, while b ⊳ c, a ⊳ c 6∼τ2 a c ⊳ b, a ⊳ c. a b ⊳ c, c ⊳ a, while c ⊲ b, c ⊲ a 6∼τ1 a b ⊲ c, c ⊲ a. Assume now that ♦ 6∈ {d1, d2}. Then the desired conclusion is established in (c), as both examples used there involve only three agents. (ii) Let τ1 = ( , d1, o1) and τ2 = ( , d2, o2). We proceed by the same case analysis as in the proof of (i). (a) d1 = ⊳ and d2 = ♦. Then b ⊳ c, c ⊳ a ∼τ1 (b) d1 = ⊲ and d2 = ♦. Then c ⊲ b, a ⊲ c ∼τ1 (c) d1 = ⊲ and d2 = ⊳. a c ⊳ a, while b♦c, c♦a 6∼τ2 a c♦a. a a ⊲ c, while c♦b, a♦c 6∼τ2 a a♦c. Then both examples used in the proof of item (c) in (i) apply here, as well. To prove that c ∼τ2 a d does not imply c ∼τ1 a d we can use the same examples as in the proof of (i). (iii) Let τ1 = ( , ♦, α) and τ2 = ( , d, α). We distinguish two cases. 19 (a) d = ⊲. Then c♦b, c♦a ∼τ1 a b♦c, c♦a, while c ⊲ b, c ⊲ a 6∼τ2 a b ⊲ c, c ⊲ a. (b) d = ⊳. Then b♦c, a♦c ∼τ1 a c♦b, a♦c, while b ⊳ c, a ⊳ c 6∼τ2 a c ⊳ b, a ⊳ c. Next, note that for all d 6= ♦ we have bc ∼τ2 (iv) Let τ1 = ( , d1, β) and τ2 = ( , d2, α). a ǫ, while bc 6∼τ1 a ǫ. Assume first that Ag > 3. Suppose a, b, c, d ∈ Ag are different agents. For each pair of distinct direction types we exhibit appropriate call sequences. (a) d1 = ⊲, d2 = ⊳. Then c ⊲ b, a ⊲ b ∼τ1 a b ⊲ c, a ⊲ b, while c ⊳ b, a ⊳ b 6∼τ2 a b ⊳ c, a ⊳ b. (b) d1 = ⊲, d2 = ♦. Then c ⊲ b, a ⊲ c ∼τ1 a d ⊲ b, a ⊲ c, while c♦b, a♦c 6∼τ2 a d♦b, a♦c. (c) d1 = ⊳, d2 = ⊲. Then b ⊳ c, b ⊳ a ∼τ1 a c ⊳ b, b ⊳ a, while b ⊲ c, b ⊲ a 6∼τ2 a c ⊲ b, b ⊲ a. (d) d1 = ⊳, d2 = ♦. Then b ⊳ c, c ⊳ a ∼τ1 a b ⊳ d, c ⊳ a, while b♦c, c♦a 6∼τ2 a b♦d, c♦a. (e) d1 = ♦, d2 = ⊲. Then c♦b, c♦a ∼τ1 a b♦c, c♦a, while c ⊲ b, c ⊲ a 6∼τ2 a b ⊲ c, c ⊲ a. (f) d1 = ♦, d2 = ⊳. Then b♦c, a♦c ∼τ1 a c♦b, a♦c, while b ⊳ c, a ⊳ c 6∼τ2 a c ⊳ b, a ⊳ c. Assume now that ♦ 6∈ {d1, d2}. Then the desired conclusion is established in (a) and (c), as both examples used there involve only three agents. Next, note that for all direction types bc ∼τ2 a ǫ, while bc 6∼τ1 a ǫ. (v) Let τ1 = ( , d, α) and τ2 = ( , d, β). Assume first that Ag > 3. Suppose a, b, c, d ∈ Ag are different agents. (a) d = ♦. Then a♦b, a♦c, b♦c, a♦b ∼τ1 a♦b, a♦c, c♦d, a♦b. a a♦b, a♦c, c♦d, a♦b, while a♦b, a♦c, b♦c, a♦b 6∼τ2 a (b) d = ⊲. Then c⊲a, b⊲c, b⊲a ∼τ1 a c⊲a, c⊲b, b⊲a, while c⊲a, b⊲c, b⊲a 6∼τ2 a c⊲a, c⊲b, b⊲a. (c) d = ⊳. Then a ⊳ c, c ⊳ b, a ⊳ b ∼τ1 a a ⊳ c, b ⊳ c, a ⊳ b, while a ⊳ c, c ⊳ b, a ⊳ b 6∼τ2 a a ⊳ c, b ⊳ c, a ⊳ b. Assume now that d 6= ♦. Then the desired conclusion is established in (b) and (c) as both examples used there involve only three agents. Finally, note that for all direction types bc ∼τ2 a ǫ, while bc 6∼τ1 a ǫ. (cid:3) The above Lemmas imply the classification of the ∼τ a relations given in The- orem 1 and visualized in Figures 1 and 2. Indeed, the equalities (represented as sets) are established in Lemma 1, the strict inclusions (that correspond to the arrows) are established in Lemma 2, and Lemma 3 implies that no further strict inclusions (i.e., arrows) are present. For example, there is no arrow in Fig- ure 2 between two different diamond shaped subgraphs that correspond to the direction types ⊲, ♦, and ⊳ because by Lemma 3(iv) for d1 6= d2 the call types ( , d1, β) and ( , d2, α) are incomparable. 20 6 Applications of the Classification The section shows how the above systematisation of ∼τ a relations, through the standard epistemic logic semantics of Definition 2, enables general insights into the epistemic effects of call sequences and offers a natural handle on how to model assumptions to the effect that agents have common knowledge of the protocol in use. 6.1 Epistemic effects of communication types The above classification is useful in order to draw general epistemic consequences in presence of different communication types. Below we will be using two frag- ments of L: -- L+ -- L+ 1 , consisting of the literals FaS and ¬FaS, ∧, ∨ and Ka, 2 , consisting of the atomic formulas FaS, ∧, ∨ and Ka. Proposition 2. Consider two call types τ1 and τ2 such that τ1(d) = τ2(d). (i) For all literals ψ and all c, c =τ2 ψ =⇒ c =τ1 ψ. (ii) If τ1 ⊆ τ2 then for all formulas φ ∈ L+ 1 and all c, c =τ2 φ =⇒ c =τ1 φ. Proof. (i) By assumption τ1(d) = τ2(d), so both occurrences of c refer to identical call sequences. Hence for all atomic formulas FaS and all c, c =τ2 FaS iff c =τ1 FaS. (ii) We proceed by induction on the structure of φ. The only case that requires explanation is when φ is of the form Kaψ. Suppose that c =τ2 Kaψ. To prove c =τ1 Kaψ take a call sequence d such that c ∼τ1 a d. By assumption τ1 ⊆ τ2, a d and so d =τ2 ψ. By the induction hypothesis d =τ1 ψ, so by hence c ∼τ2 definition c =τ1 Kaψ. (cid:3) It is easy to construct examples showing that the implication in (ii) does not hold for all formulas. For instance, for τ1 = ( , ♦, α) and τ2 = ( , ♦, α) we have τ1 ⊆ τ2 by Theorem 1 and bc =τ2 ¬KaFbC but not bc =τ1 ¬KaFbC. We finally compare knowledge for call types with different direction types. Then claim (i) in the above Proposition does not hold anymore. Indeed, for τ1 and τ2 such that τ1(d) = ♦ and τ2(d) = ⊲ we have ab =τ2 ¬FaB but not ab =τ1 ¬FaB. However, the following weaker claim does hold. Proposition 3. Consider two call types τ1 and τ2 such that τ1(d) = ♦. (i) For all atomic formulas ψ and all c, c =τ2 ψ =⇒ c =τ1 ψ. (ii) If τ1 ⊆ τ2 then for all formulas φ ∈ L+ 2 and all c, c =τ2 φ =⇒ c =τ1 φ. 21 Proof. By Proposition 2 we can assume that τ2(d) 6= ♦. (i) We use induction on the length c of c. Assume that τ2(d) = ⊲. If c = 0 then c = ǫ and ǫ =τ2 FcD iff D = C iff ǫ =τ1 FcD. Now suppose the claim is proven for c and consider c.ab. For any agent c 6= b, we have by Definition 1 c.a ⊲ b =τ2 FcD iff c =τ2 FcD, which implies by the induction hypothesis c =τ1 FcD, and hence c.a♦b =τ1 FcD. For agent b, we have c.a ⊲ b =τ2 FbD iff (c =τ2 FaD or c =τ2 FbD) and c.a♦b =τ1 FbD iff (c =τ1 FaD or c =τ1 FbD), so the claim for b holds by the induction hypothesis, as well. The proof for τ2(d) = ⊳ is analogous and omitted. (ii) The claim follows by (i) and the argument used in the proof of Proposition 2. (cid:3) Proposition 2 holds for example for τ1 = ( , ♦, β) and τ2 = ( , ♦, α), since by Theorem 1 ( , ♦, β) ⊂ ( , ♦, β) ⊂ ( , ♦, β) ⊂ ( , ♦, α). In turn, Proposition 3 holds for example for τ1 = ( , ♦, β) and τ2 = ( , ⊲, α), since by Theorem 1 ( , ♦, β) = ( , ⊲, β) ⊂ ( , ⊲, β) ⊂ ( , ⊲, β) ⊂ ( , ⊲, α). In particular, for both pairs of τ1 and τ2 for all call sequences c, c =τ2 KaFbC implies c =τ1 KaFbC. Informally, under τ1 the agents are then more informed about the knowledge of other agents than under τ2. Further, note that by Theorem 1 if τ1(d) = ♦ 6= τ2(d) then τ1 ⊆ τ2 iff τ1(p) = , so under the assumption τ1(d) = ♦ 6= τ2(d) the second claim of Proposition 3 can be rewritten as If τ1(p) = then for all formulas φ ∈ L+ 2 and all c, c =τ2 φ =⇒ c =τ1 φ. This implication (under the assumption τ1(d) = ♦ 6= τ2(d)) does not hold for the other two privacy types because of the following instructive counterexample. Example 10. Assume Ag = {a, b, c}. Suppose τ1 = (p, ♦, α), τ2 = (p, ⊲, α), where p 6= , and c = ac, cb, ba. We claim that then c =τ2 KaKcFbC but not c =τ1 KaKcFbC. (i) p = . Then c =τ2 KaKcFbC. The reason is that the only call sequence ∼τ2 a equiv- alent to c is c itself. Indeed, if c ∼τ2 a d, then d has to be of the form ac, c, ba, where a 6∈ c, and by the first entry in Table 1, bottom, also c(i)a = (ac, c, ba)(i)a has to hold. But c(i)a = {A, B, C}, which implies that c = cb. Thus c =τ2 KaKcFbC iff c =τ2 KcFbC and the latter is easy to check. However, c 6=τ1 KaKcFbC since c ∼τ1 (ii) p = . a ac, de, ba and ac, de, ba 6=τ1 KcFbC. The reasoning is now a bit more involved. To show that c =τ2 KaKcFbC a d. Then d is of the form d1, ac, d2, ba, d3, take a call sequence d such that c ∼τ2 22 where agent a is not involved in any call from d1, d2, d3. Moreover c(i)a = d(i)a holds, as well. But, as already noted in (i), c(i)a = {A, B, C}, which implies that one of the calls in d1 or d2 is cb. Now, for any call sequence d in which the call cb appears we have d =τ2 c d′ then the call cb appears in d′, as well, and hence a ac, de, ba and KcFbC. Indeed, if d ∼τ2 d′ =τ2 FbC. However, c 6=τ1 KaKcFbC since, as in (i), c ∼τ1 ac, de, ba 6=τ1 KcFbC. Analogous examples can be constructed for the pull calls. (cid:3) This example shows that for the privacy degrees the push calls may convey more knowledge than the push-pull calls, even though the former ones result in less informative communication. The same is the case for the pull calls. and 6.2 Common knowledge of protocols When reasoning about specific protocols it is necessary to limit the set of con- sidered call sequences to those that are 'legal' for it. When the agents form a graph given in advance one can simply limit the set of considered call sequences by allowing only syntactically legal calls. This affects the definition of seman- tics and can be of importance when reasoning about the correctness of specific protocols. For example, in [1] a specific protocol for a directed ring is proved correct (Protocol R2 on page 61, for 3 or 4 agents) by allowing for each agent a only the calls between her and her successor a ⊕ 1, and using the fact that the formula KaFa⊕1A ⊖ 1 → FaA ⊖ 1 is then true. Here A ⊖ 1 is the secret of the predecessor of agent a, so this formula states that if agent a knows that her successor is familiar with the secret A ⊖ 1 of her predecessor then agent a is familiar with the secret A ⊖ 1. A more challenging task is to incorporate into the framework an assumption that the agents have common knowledge of the underlying protocol.6 Example 11. Consider Protocol 2 (Hear my Secret) from Section 2 with the direction type ♦. Recall that in this protocol an agent a can call agent b if ¬KaFba is true after the current call sequence. So each pair of agents can communicate at most once. Assume now four agents a, b, c, d. Then the call sequence ab, bc, bd is compliant with the protocol independently on the assumptions about the privacy degree and observance. Let us analyse the situation after this call sequence took place. Assume first the privacy degree . Then agent c knows which calls took place and hence knows that after the third call agent d is familiar with her secret, C. So after these three calls agent c cannot call agent d anymore. The situation changes when the privacy degree is . Through the second call agent c learns the secret A, so she knows that the first call was ab or ba. Agent 6 This issue was identified as an open problem for epistemic gossip in [1]. The same issue manifests itself in other knowledge-based asynchronous protocols, such as the one investigated recently in [24]. 23 c is not involved in the third call, but by the assumed privacy degree she still knows that a third call has taken place. Assume now that the agents have common knowledge of the protocol. So agent c knows that each pair of agents can communicate at most once. Hence she can conclude that d must be involved in the third call and consequently that the third call was between agent d and agent a or b. Agent c therefore now knows that after the third call agent d is familiar with at least 3 secrets: A, B, D if the call was with agent a or A, B, C, D if the call was with agent b. But agent c cannot anymore conclude that agent d is familiar with her secret, C, and consequently can call d. Suppose now that the privacy degree is still but the call sequence is ab, bc, cd, bd. Consider now agent a. After the fourth call she knows that after the call ab three calls took place between the agents b, c, d. Further, she knows that each pair of agents can communicate at most once. So agent a concludes that each pair of agents from {b, c, d} communicated precisely once. In particu- lar both agents c and d communicated with agent b and hence both of them are familiar with the secret A. So after these four calls agent a cannot call anymore any agent. Finally, consider the privacy degree and suppose the call sequence is ab, bc, bd or ab, bc, cd, bd. Then agent a does not know whether any calls took place after the call ab. In particular she cannot conclude that any of the agents c and d are familiar with her secret and hence can call either c or d. (cid:3) To discuss the matters further let us be more precise about the syntax of the protocols. An epistemic gossip protocol (in short a protocol) consists of the union of Ag sets of instructions, one set for each agent. Each instruction is of the form if φ then execute call c, in symbols φ → c, where φ is a Boolean combination of formulas of the form Kaψ, where a is the caller in the call c. The formula φ is referred to as an epistemic guard . Such instructions are executed iteratively, where at each time one instruction is selected (at random, or based on some fairness considerations) whose guard is true after the call sequence executed so far.7 We therefore view a protocol P as a set of instructions φ → c. For example, the instructions composing Protocol 2, are of the form ¬KaFbA → ab for all agents a and b. That is, if a does not know whether b is not familiar with her secret, a calls b. To justify the restriction on the syntax of the epistemic guards note the following observation. 7 This simple rendering of protocols suffices for the purposes of this section. More sophisticated formalizations of epistemic gossip protocols have been provided in [4,1]. 24 Note 2. Consider a call type τ such that τ (p) = . Then for all agents a, b, c and all call sequences c and formulas φ c =τ Kaφ iff c.bc =τ Kaφ. Consequently, the same equivalence holds for all formulas that are Boolean com- binations of formulas of the form Kaφ, so in particular for all epistemic guards used in the instructions for agent a. Proof. By Definition 3 if the privacy type of τ is the claim. then c ∼τ a c.bc, which implies ⊓⊔ This note states that the calls in which agent a is not involved have no effect on the truth of the epistemic guards used in the instructions for agent a. If we allowed in the epistemic guards for agent a as conjuncts formulas not prefixed by Ka, this natural and desired property would not hold anymore. Indeed, assume the privacy type and consider the protocol for three agents, a, b, c, in which the only instructions are ¬FbA ∧ FbC → ab for agent a and ¬FbC → bc for agent b. Then initially only the call bc can be performed. After it, the call ab can be performed upon which the protocol terminates. In other words, the call bc, of which agent a is not aware, affects the truth of its epistemic guard, which contradicts the idea behind the privacy type . For the privacy type this restriction on the syntax of the epistemic guards is not needed as then all formulas are equivalent to the propositional ones. Note 3. Consider a call type τ such that τ (p) = . Then for all agents a and all formulas φ and call sequences c c =τ Kaφ iff c =τ φ. Proof. This is a direct consequence of the fact that when the privacy type of τ ⊓⊔ is then by Lemma 1(i) each relation ∼τ a is the identity. Let us return now to the matter of common knowledge of a protocol. In Defi- nition 3 the τ -dependent indistinguishability relations are constructed assuming that any call is possible after any call sequence. This builds in the resulting gos- sip models Mτ = (Cτ , {∼τ a}a∈Ag) the assumption that agents may consider any call sequence possible in principle, including calls that are not legal if we assume that the agents have common knowledge of the protocol in use. Specifically, given a gossip model Mτ = (Cτ , {∼τ a}a∈Ag) and a protocol P we P ⊆ Cτ of P (cf. [1]) as the set of call sequences define the computation tree Cτ inductively defined as follows: [Base] ǫ ∈ Cτ P , [Step] If c ∈ Cτ So Cτ P is a (possibly infinite) set of finite call sequences that is iteratively ob- tained by performing a 'legal' call (according to protocol P ) from a 'legal' (ac- cording to protocol P ) call sequence. We refer to such legal call sequences as P -compliant.8 P and c =τ φ then c.c ∈ Cτ P , where φ → c ∈ P . 8 We call Cτ P a tree since its elements can be arranged in an obvious way in (a possibly infinite, but finitely branching) tree. 25 Note however, that when building such a computation tree, the epistemic guard φ is evaluated with respect to the underlying gossip model Mτ , which may well include call sequences that are not P -compliant. So in order to restrict the domain of the gossip model to only P -compliant sequences, the epistemic guards of the protocol need to be evaluated, and to do that one needs in turn a gossip model, which contains only P -compliant sequences. This circularity is not problematic for the call types involving privacy degrees and , as the ∼τ a relations then link only sequences of equal length, allowing therefore for call sequences and these equivalence relations to be inductively constructed in parallel. That is however not the case for the call types involving privacy degree , as then call sequences of any length may be indistinguishable from the actual call sequence. We propose here a solution to the above issue, showing how under a natural assumption on the syntax of the epistemic guards one can construct, also for the privacy degree , a gossip model which consists only of call sequences that are compliant with a given protocol P . Fix till the end of the section an arbitrary call type τ . First, we introduce the definition of semantics relativised to a set X ⊆ Cτ of call sequences. Let Mτ a relation is restricted to X × X, and let c ∈ X. Then the definition of semantics is the same as before with the except of the formulas of the form Kaφ: a}a∈Ag), where each ∼τ X = (X, {∼τ (Mτ X , c) = Kaφ iff ∀d ∈ X such that c ∼τ a d, (Mτ X , d) = φ. Fix now a protocol P and a set X ⊆ Cτ . We define the relativised computa- (P,X) obtained by replacing the above Base and Step tion tree of P as the set Cτ conditions by [Base] ǫ ∈ Cτ (P,X), [Step] If c ∈ X ∩Cτ (P,X) and (Mτ X , c) = φ then c.c ∈ Cτ (P,X), where φ → c ∈ P , and refer to each call sequence from Cτ (P,X) as (P, X)-compliant. We now limit the syntax of epistemic guards as follows. A formula Kaφ is an abbreviation for ¬Ka¬φ and L denotes the existential fragment of L, consisting of only literals, ∨, ∧, and Ka. The following lemma clarifies the introduction of the language L. Lemma 4. If X ⊆ Y ⊆ Cτ then for all formulas φ ∈ L and all c ∈ X, (Mτ X , c) = φ =⇒ (Mτ Y , c) = φ. Proof. The only case that requires explanation is when φ is of the form Kaψ. Suppose that (Mτ X , d) = ψ. By the induction hypothesis (Mτ Y , c) = φ. ⊓⊔ X , c) = φ. Then for some d ∈ X such that c ∼τ Y , d) = ψ, so by definition (Mτ a d, (Mτ Define next an operator ρP : 2Cτ → 2Cτ by ρP (X) = X ∩ Cτ (P,X). 26 That is, ρP removes from a given set X of call sequences those that are not (P, X)-compliant. What we are after is a set from which no sequences would be removed, so a fixpoint of ρP . Proposition 4. Suppose the epistemic guards of a protocol P are all from L. Then there exists an X ⊆ Cτ such that X = ρP (X). Proof. Suppose that X ⊆ Y and c ∈ X ∩ Cτ induction on the length of c. If c = ǫ, then c ∈ Cτ (P,X). We prove that c ∈ Cτ (P,Y ) by (P,Y ) by the Base condition. Otherwise, by the Step condition c is of the form c′.c, where c′ ∈ X ∩Cτ (P,X), X , c′) = φ, and φ → c ∈ P . By the induction hypothesis (P,Y ). Further, c′ ∈ Y and by Lemma 4 (Mτ Y , c′) = φ, so c ∈ Cτ (P,Y ). and for some φ ∈ L, (Mτ c′ ∈ Cτ It follows that ρP is a monotonic function, that is, X ⊆ Y implies ρP (X) ⊆ ρP (Y ). By the Knaster-Tarski theorem of [30] ρP has therefore fixpoints, includ- ing a largest and a smallest one. (cid:3) Intuitively, when the domain X ⊆ Cτ of a gossip model is a fixpoint of ρP , then the restriction of the definition of the indistinguishability relations ∼τ a to such a domain has the effect that the call sequences considered possible by the agents coincide with the call sequences generated by the protocol. Such gossip models incorporate then the assumption that there is common knowledge among the agents about the protocol in use. Furthermore, by the Knaster-Tarski theorem one can construct the largest fixpoint of ρP by iteratively applying ρP to Cτ . Such fixpoint νρP is the most natural domain for a gossip model that realises the assumption of common knowl- edge of the protocol, with the (P, νρP )-compliant call sequences viewed as the P -compliant ones. When the privacy degree is such a gossip model has a very simple structure, namely (Cτ P , {∼τ a}a∈Ag). Corollary 2. Consider a protocol P and a call type τ such that τ (p) = . Then νρP = Cτ P . Proof. Note that we always have ρP (Cτ ) = Cτ (P,Cτ P ) by induction on the length of the call sequences. We only need to consider the induction step. So consider some c.c ∈ Cτ P and c =τ φ, where φ → c ∈ P , and by the induction hypothesis c ∈ Cτ P . By definition c ∈ Cτ P ). P . We now show that Cτ P ⊆ Cτ (P,Cτ Let φ′ be obtained from φ by removing all occurrences of Ka for all agents a. By Note 3 relativised to an arbitrary X ⊆ Cτ such that c ∈ X we have c =τ φ iff c =τ φ′ iff (Mτ , c) = φ and hence by definition c.c ∈ Cτ P ) = Cτ X , c) = φ′ iff (Mτ X , c) = φ. So in particular (Mτ P ). (P,Cτ P ∩ Cτ Consequently ρP (Cτ P and hence Cτ Cτ P P ) = Cτ (P,Cτ P is the largest ⊓⊔ fixpoint of ρP . The syntactic restriction on the epistemic guards used in Proposition 4 is clearly satisfied by Protocol 2 as its guards can be rewritten as Ki¬Fj I. The 27 same is the case for all protocols studied in [1] since by Corollary 1 for all call types and all agents a and b the formulas KaFaB and FaB are equivalent and consequently each formula FaB can be replaced by ¬ Ka¬FaB. 7 Conclusions We provided an in-depth study of 18 different types of communication relevant for epistemic gossip protocols and modelled their epistemic effects in a uniform way through different indistinguishability relations. This led us to establish a precise map of the relative informativeness of these types of communication (Theorem 1). In turn, this result allowed us to prove general results concerning the epistemic effects of call sequences under different communication regimes (Propositions 2 and 3) and to advance a natural proposal on how to model and analyse agents' common knowledge of gossip protocols (Proposition 4), a still under-investigated issue in the literature. Several natural directions for future research present themselves. We men- tion three of them. The first question concerns the axiomatisation of the modal language L introduced in Section 3. This problem is parametrised by the un- derlying indistinguishability relations introduced in Section 4. For example, by Note 3 the equivalence φ ↔ Kaφ holds for the privacy type but not for the other two. Actually, even the axiomatization of the FaS formulas is not straightforward, as it has to take into account the nature of the communication. Indeed, consider the following formula, where a 6= b: (cid:16)FbA ∧ ^ i6=a,b ¬FiA(cid:17) → FaB. It states that if agent b is the only agent (different from a) familiar with the secret of a, then agent a is familiar with the secret of b. A more general version is: (cid:16) _ i∈X FiA ∧ ^ i6∈X∪{a} ¬FiA(cid:17) → _ FaI, i∈X where a 6∈ X. Intuitively it states that if somebody from a group X, to which a does not belong, is familiar with her secret and nobody from outside of the group X (except a) is familiar with this secret, then agent a is familiar with a secret of somebody from the group X. Clearly, both formulas are valid for the ♦ direction type. In general such an axiomatisation project could be carried out at several levels (cf. [14]): by considering FiS formulas as primitive, as we did in this paper; or analysing them as "knowing whether" formulas (in epistemic logic notation, KiS∨Ki¬S) as in [4]. Whether the latter level of analysis can be easily reconciled with the one proposed in this paper is an interesting open problem. 28 The second question addresses the problem of decidability of the 18 defini- tions of truth we introduced. In the terminology of this paper [3] established for the call type ( , ♦, α) that the semantics and the definition of truth are both decidable for the formulas without nested modalities. It would be interesting to establish analogous results for the remaining call types, ideally by providing a single, uniform proof that generalises the arguments of [3]. The final question concerns the robustness of our analysis, and specifically of the relationships identified in Theorem 1, with respect to modes of gossip that involve the transfer of higher-order epistemic information as introduced and studied in [20,21]. Intuitively, we would expect this type of higher-order epistemic communication to have an impact on the effects of the asymmetric communication types ⊲ and ⊳ and for the full privacy degree. Finally, one could envisage other aspects of a call not considered in this framework. For example in [4] yet another notion of privacy was considered, according to which given a call ab every agent c 6= a, b noted that at most one call took place. Then for agent c the call sequences ǫ and ab are equivalent but ǫ and ab, ab are not. Another possibility could be to consider a notion of privacy that is intermediate between and , according to which the caller is anonymous but the callee not. Then for agent c the call sequences ab and ad are equivalent but ab and bd are not. Acknowledgments We thank Hans van Ditmarsch for most useful and extensive discussions on the subject of this paper. We are also grateful to anonymous referees of this and earlier versions of this paper for helpful comments. The first author was partially supported by the NCN grant nr 2014/13/B/ST6/01807. References 1. K. R. Apt, D. Grossi, and W. van der Hoek. Epistemic protocols for distributed gossiping. In Proceedings Fifteenth Conference on Theoretical Aspects of Rational- ity and Knowledge (TARK 2015), volume 215, pages 51 -- 56. EPTCS, 2016. 2. K. R. Apt and D. Wojtczak. On the computational complexity of gossip protocols. In Proceedings of IJCAI 2017, pages 765 -- 771, 2017. 3. K. R. Apt and D. Wojtczak. Verification of distributed epistemic gossip protocols. Journal of Artificial Intelligence Research, 62:101 -- 132, 2018. 4. M. Attamah, H. van Ditmarsch, D. Grossi, and W. van der Hoek. Knowledge and gossip. In Proceedings of ECAI'14, pages 21 -- 26. IOS Press, 2014. 5. N. Bailey. The Mathematical Theory of Epidemics. Griffen Press, 1957. 6. B. Baker and R. Shostak. Gossips and telephones. Discrete Mathematics, 2:197 -- 193, 1972. 7. O. Bataineh and R. van der Meyden. Abstraction for epistemic model checking of dining-cryptographers based protocols. In Proceedings of TARK'11, 2011. 8. R. Bumby. A problem with telephones. SIAM Journal of Algorithms and Discrete Methods, 2:13 -- 18, 1981. 29 9. B. Chlebus and D. Kowalski. Robust gossiping with an application to consensus. Journal of Computer and System Sciences, 72:1262 -- 1281, 2006. 10. M. Cooper, A. Herzig, F. Maffre, F. Maris, and P. R´egnier. Simple epistemic planning: Generalised gossiping. In Proceedings of ECAI 2016, pages 1563 -- 1564, 2016. 11. R. Fagin, J. Halpern, Y. Moses, and M. Vardi. Knowledge-based programs. Dis- tributed Computing, 10:199 -- 225, 1997. 12. Ronald Fagin, Joseph Y. Halpern, Yoram Moses, and Moshe Y. Vardi. Reasoning about knowledge. The MIT Press, Cambridge, 1995. 13. P. Fraigniaud and E. Lazard. Methods and problems of communication in usual networks. Discrete Applied Mathematics, 53:79 -- 133, 1994. 14. M. Gattinger. New Directions in Model Checking Dynamic Epistemic Logic. PhD thesis, ILLC, 2018. 15. van Ditmarsch H., Grossi D., Herzig A., van der Hoek W., and Kuijer L. Parameters for epistemic gossip problems. In Proceedings of LOFT'16, 2016. 16. A. Hajnal, E. C. Milner, and E. Szemeredi. A cure for the telephone disease. Canadian Mathematical Bulletin, 15:447 -- 450, 1972. 17. J. Halpern and Y. Moses. Knowledge and common knowledge in a distributed environment. Journal of the ACM, 37(3):549 -- 587, 1990. 18. J. Halpern and L. Zuck. A little knowledge goes a long way: Knowledge-based derivations and correctness proofs for a family of protocols. Journal of the ACM, 39(3):449 -- 478, 1992. 19. S. M. Hedetniemi, S. T. Hedetniemi, and A. L. Liestman. A survey of gossiping and broadcasting in communication networks. Networks, 18(4):319 -- 349, 1988. 20. A. Herzig and F. Maffre. How to share knowledge by gossiping. In Proceedings of EUMAS/AT, pages 249 -- 263, 2015. 21. A. Herzig and F. Maffre. How to share knowledge by gossiping. AI Communica- tions, 30(1):1 -- 17, 2017. 22. J. Hromkovic, R. Klasing, B. Monien, and R. Peine. Dissemination of informa- tion in interconnection networks (broadcasting and gossiping). In Combinatorial Network Theory, pages 125 -- 212. Kluwer, 1996. 23. J. Hromkovic, R. Klasing, A. Pelc, P. Ruzicka, and W. Unger. Dissemination of In- formation in Communication Networks: Broadcasting, Gossiping, Leader Election, and Fault-Tolerance. Springer, 2005. 24. S. Knight, B. Maubert, and F. Scharzentruber. Reasoning about knowledge and messages in asynchronous multi-agent systems. Mathematical Structures in Com- puter Science, pages 1 -- 42, 2017. 25. R. Kurki-Suonio. Towards programming with knowledge expressions. In Proceed- ings of POPL'86, pages 140 -- 149, 1986. 26. J.-J. Ch. Meyer and W. van der Hoek. Epistemic Logic for AI and Computer Sci- ence, volume 41 of Cambridge Tracts in Theoretical Computer Science. Cambridge University Press, 1995. 27. R. Parikh and R. Ramanujam. Distributed processing and the logic of knowledge. In Logic of Programs, LNCS 193, pages 256 -- 268. Springer, 1985. 28. A. Procaccia, Y. Bachrach, and J. Rosenschein. Gossip-based aggregation of trust in decentralized reputation systems. In Proceedings of IJCAI'07, pages 1470 -- 1475, 2007. 29. ´A. Seress. Quick gossiping without duplicate transmissions. Graphs and Combi- natorics, 2:363 -- 383, 1986. 30. A. Tarski. A lattice-theoretical fixpoint theorem and its applications. Pacific Journal of Mathematics, 5:285 -- 309, 1955. 30 31. R. Tijdeman. On a telephone problem. Nieuw Archief voor Wiskunde, 3(XIX):188 -- 192, 1971. 32. H. van Ditmarsch, B. Kooi, and W. van der Hoek. Dynamic Epistemic Logic, volume 337 of Synthese Library Series. Springer, 2007. 33. H. van Ditmarsch, J. van Eijck, P. Pardo, R. Ramezanian, and F. Scharzentruber. Epistemic protocols for dynamic gossip. Journal of Applied Logic, 20:1 -- 31, 2017. 31
1804.06311	1	1804	2018-04-17T15:10:44	Leveraging Statistical Multi-Agent Online Planning with Emergent Value Function Approximation	[ "cs.MA" ]	Making decisions is a great challenge in distributed autonomous environments due to enormous state spaces and uncertainty. Many online planning algorithms rely on statistical sampling to avoid searching the whole state space, while still being able to make acceptable decisions. However, planning often has to be performed under strict computational constraints making online planning in multi-agent systems highly limited, which could lead to poor system performance, especially in stochastic domains. In this paper, we propose Emergent Value function Approximation for Distributed Environments (EVADE), an approach to integrate global experience into multi-agent online planning in stochastic domains to consider global effects during local planning. For this purpose, a value function is approximated online based on the emergent system behaviour by using methods of reinforcement learning. We empirically evaluated EVADE with two statistical multi-agent online planning algorithms in a highly complex and stochastic smart factory environment, where multiple agents need to process various items at a shared set of machines. Our experiments show that EVADE can effectively improve the performance of multi-agent online planning while offering efficiency w.r.t. the breadth and depth of the planning process.	cs.MA	cs	Leveraging Statistical Multi-Agent Online Planning with Emergent Value Function Approximation Thomy Phan, Lenz Belzner, Thomas Gabor and Kyrill Schmid Institute of Informatics LMU Munich {thomy.phan, belzner, thomas.gabor, kyrill.schmid}@ifi.lmu.de 8 1 0 2 r p A 7 1 ] A M . s c [ 1 v 1 1 3 6 0 . 4 0 8 1 : v i X r a Abstract Making decisions is a great challenge in distributed autonomous environments due to enormous state spaces and uncertainty. Many online planning algo- rithms rely on statistical sampling to avoid search- ing the whole state space, while still being able to make acceptable decisions. However, plan- ning often has to be performed under strict com- putational constraints making online planning in multi-agent systems highly limited, which could lead to poor system performance, especially in stochastic domains. In this paper, we propose Emergent Value function Approximation for Dis- tributed Environments (EVADE), an approach to in- tegrate global experience into multi-agent online planning in stochastic domains to consider global effects during local planning. For this purpose, a value function is approximated online based on the emergent system behaviour by using methods of reinforcement learning. We empirically evalu- ated EVADE with two statistical multi-agent on- line planning algorithms in a highly complex and stochastic smart factory environment, where mul- tiple agents need to process various items at a shared set of machines. Our experiments show that EVADE can effectively improve the performance of multi-agent online planning while offering effi- ciency w.r.t. the breadth and depth of the planning process. 1 Introduction Decision making in complex and stochastic domains has been a major challenge in artificial intelligence for many decades due to intractable state spaces and uncertainty. Sta- tistical approaches based on Monte-Carlo methods have be- come popular for planning under uncertainty by guiding the search for policies to more promising regions in the search space [Kocsis and Szepesv´ari, 2006; Silver and Veness, 2010; Weinstein and Littman, 2013; Amato and Oliehoek, 2015; Belzner et al., 2015; Silver et al., 2016; Claes et al., 2017]. These methods can be combined with online planning to adapt to unexpected changes in the environment by inter- leaving planning and execution of actions [Silver and Ve- ness, 2010; Amato and Oliehoek, 2015; Belzner et al., 2015; Silver et al., 2016; Claes et al., 2017]. However, online planning often has to meet strict real- time constraints limiting the planning process to local search. This makes the consideration of possible global effects dif- ficult, which could lead to suboptimal policies, especially in stochastic domains. The problem is further intensi- fied in multi-agent systems (MAS), where the search space grows exponentially w.r.t. the dimension and the number of agents, which is known as the curse of dimensionality [Boutilier, 1996; Amato and Oliehoek, 2015; Oliehoek and Amato, 2016]. Furthermore, one has to cope with the co- ordination of individual actions of all agents to avoid po- tential conflicts or suboptimal behaviour [Boutilier, 1996; Bus¸oniu et al., 2010]. Many multi-agent planning approaches assume the avail- ability of a pre-computed value function of a more simpli- fied model of the actual environment to consider possible global effects in the local planning process, which can be exploited to prune the search space or to further refine the policy [Emery-Montemerlo et al., 2004; Szer et al., 2005; Oliehoek et al., 2008b; Spaan et al., 2011]. This might be insufficient for highly complex and uncertain domains, where the dynamics cannot be sufficiently specified before- hand [Belzner et al., 2015]. Depending on the domain com- plexity, pre-computing such a value function might be even computationally infeasible [Boutilier, 1996; Silver and Ve- ness, 2010]. Thus, an adaptive and model-free approach is desirable for learning a value function at system runtime in MAS. Recently, approaches to combine online planning and rein- forcement learning (RL) have become popular to play games with high complexity like Go and Hex [Silver et al., 2016; Silver et al., 2017; Anthony et al., 2017]. A tree search algo- rithm is used for planning, which is guided by a value func- tion approximated with RL. These approaches were shown to outperform plain planning and RL, even achieving super- human level performance in Go without any prior knowledge about the game beyond its rules [Silver et al., 2017]. So far, these approaches have only been applied to deterministic do- mains with only one agent. In this paper, we propose Emergent Value function Approx- imation for Distributed Environments (EVADE), an approach to integrate global experience into multi-agent online plan- ning in stochastic domains. For this purpose, a value func- tion is approximated online based on the emergent system behaviour by using methods of RL. With that value function, global effects can be considered during local planning to im- prove the performance and efficiency of existing multi-agent online planning algorithms. We also introduce a smart factory environment, where multiple agents need to process various items with different tasks at a shared set of machines in an au- tomated and self-organizing way. Given a sufficient number of agents and stochasticity w.r.t. the outcome of actions and the behaviour of agents, we show that our environment has a significantly higher branching factor than the game of Go [Silver et al., 2016]. We empirically evaluate the effective- ness of EVADE in this stochastic and complex domain based on two existing multi-agent planning algorithms [Oliehoek et al., 2008a; Belzner and Gabor, 2017a]. The rest of the paper is organized as follows. Section 2 provides some background about decision making in gen- eral. Section 3 discusses related work. Section 4 describes EVADE for enhancing multi-agent planning algorithms. Sec- tion 5 presents and discusses experimental results achieved by two statistical multi-agent planning algorithms enhanced with EVADE in our smart factory environment. Finally, sec- tion 6 concludes and outlines a possible direction for future work. 2 Background 2.1 Markov Decision Processes We formulate our problem as multi-agent Markov Decision Process (MMDP) assuming a fully cooperative setting, where all agents share the same common goal [Boutilier, 1996; Oliehoek and Amato, 2016]. this work only focuses on fully observable problems as modeled in [Boutilier, 1996; Tampuu et al., 2017; Claes et al., 2017]. For simplicity, Although more realistic models exist for describing large- scale MAS like Dec-MDPs or Dec-POMDPs [Oliehoek and Amato, 2016], the focus of this work is just to evaluate the possible performance and efficiency gain based on integrating global experience into the multi-agent online planning pro- cess. An extension of our approach to partially observable models is left for future work. MDP Decision-making problems with discrete time steps and a sin- gle agent can be formulated as Markov Decision Process (MDP) [Howard, 1961; Boutilier, 1996; Puterman, 2014]. An MDP is defined by a tuple M = (cid:104)S,A,P,R(cid:105), where S is a (finite) set of states, A is the (finite) set of ac- tions, P(st+1st, at) is the transition probability function and R(st, at) is the scalar reward function. In this work, it is al- ways assumed that st, st+1 ∈ S, at ∈ A, rt = R(st, at), where st+1 is reached after executing at in st at time step t. Π is the policy space and Π is the number of all possible policies. The goal is to find a policy π : S → A with π ∈ Π, which maximizes the (discounted) return Gt at state st for a horizon h−1(cid:88) k=0 h: γk · R(st+k, at+k) Gt = (1) where γ ∈ [0, 1] is the discount factor. If γ < 1, then present rewards are weighted more than future rewards. A policy π can be evaluated with a state value function V π = Eπ[Gtst], which is defined by the expected return at state st [Bellman, 1957; Howard, 1961; Boutilier, 1996]. π is optimal if V π(st) ≥ V π(cid:48) (st) for all st ∈ S and all policies π(cid:48) ∈ Π. The optimal value function, which is the value function for any optimal policy π∗, is denoted as V ∗ and defined by [Bellman, 1957; Boutilier, 1996]: V ∗(st) = maxat∈A(cid:8)rt + γ P (s(cid:48)st, at) · V ∗(s(cid:48))(cid:9) (2) (cid:88) s(cid:48)∈S Multi-Agent MDP An MMDP is defined by a tuple M = (cid:104)D,S,A,P,R(cid:105), where D = {1, ..., n} is a (finite) set of agents and A = A1 × ... × An is the (finite) set of joint actions. S, P and R are defined analogously to an MDP, given joint actions in- stead of atomic actions [Boutilier, 1996]. The goal is to find a joint policy π = (cid:104)π1, ..., πn(cid:105), which maximizes the return Gt of eq. 1. πi is the individual policy of agent i ∈ D. Given n agents in the MMDP, the number of i=1 Πi. If all agents share the same individual policy space Πi, then Π = Πin. possible joint policies is defined by Π = (cid:81)n Similarly to MDPs, a value function V π can be used to evaluate the joint policy π. 2.2 Planning Planning searches for a policy, given a generative model M, which represents the actual environment M. M pro- vides an approximation for P and R of the underlying MDP or MMDP [Boutilier, 1996; Weinstein and Littman, 2013; Belzner et al., 2015]. We assume that M perfectly models the environment such that M = M. Global planning meth- ods search the whole state space to find π∗ or V ∗. An ex- ample is value iteration, which computes the optimal value function V ∗ by iteratively updating value estimates for each state according to eq. 2 [Bellman, 1957; Howard, 1961; Boutilier, 1996]. Local planning methods only regard the current state and possible future states within a horizon of h to find a local policy πlocal [Weinstein and Littman, 2013; Belzner et al., 2015]. An example for local planning is given in fig. 1a for a problem with a branching factor of two and a planning horizon of h = 2. The nodes in the search tree represent states and the links represent actions. In this paper, we only focus on local planning methods for online planning, where planning and execution of actions are performed alternately at each time step, given a fixed com- putation budget nbudget [Silver and Veness, 2010; Amato and Oliehoek, 2015; Belzner et al., 2015; Silver et al., 2016; Claes et al., 2017]. Local planning can be performed via closed-loop or open- loop search. Closed-loop search corresponds to a tree search, where a search tree is constructed and traversed guided by Sutton and Barto, 1998]. The experience samples are ob- tained from interaction between the agent and the environ- ment. 3 Related Work Hybrid Planning Some hybrid approaches to combine of- fline and online planning in partially observable domains were introduced in [Paquet et al., 2006; Ross et al., 2007]. 1 is In the offline planning phase, a value function VM DP computed based on a fully observable model of the actual environment by using variants of value iteration. VM DP is used to enhance online planning to search for a policy un- der the consideration of possible global effects. It was shown that VM DP provides an upper bound to V ∗ of the actual en- vironment [Cassandra and Kaelbling, 2016; Oliehoek et al., 2008b]. This can be exploited to prune the search space without loosing optimality of the solutions found. Many multi-agent planning algorithms use similar methods to enhance planning with such a pre-computed value function VM DP [Emery- Montemerlo et al., 2004; Szer et al., 2005; Oliehoek et al., 2008b; Spaan et al., 2011]. In our approach, V ∗ is approximated online based on ac- tual experience without requiring a model. A generative model is only used for online planning to find a joint pol- icy. We intend to apply our approach to highly complex and stochastic domains, where an offline computation is not fea- sible, since any change in the model would require the re- computation of VM DP . Online Planning and Deep RL AlphaGo is a program in- troduced in [Silver et al., 2016], which is able to play Go at a super-human level. It recently defeated the currently best hu- man Go players in various tournaments [Silver et al., 2016; Silver et al., 2017]. AlphaGo uses MCTS for online planning and deep neural networks, which approximate π∗ and V ∗ to guide the tree search. With this approach, AlphaGo is able to develop extremely complex strategies within given time con- straints. MCTS-based planning combined with an approximation of V ∗ was shown to improve the performance of plain on- line planning or RL in complex and deterministic games like Go and Hex [Silver et al., 2016; Silver et al., 2017; Anthony et al., 2017]. The idea of these approaches is based on the human mind, which is able to think ahead into the future, while guiding the thoughts with intuition learned from experience. In the context of artificial intelligence, on- line planning represents the future thinking, while deep RL represents the integration of strong intuition [Evans, 1984; Kahneman, 2003; Anthony et al., 2017]. Our approach extends these ideas to environments with multiple agents. We also focus on stochastic domains, where the outcome of actions and the behaviour of agents are not deterministic. 1The action value function Q(st, at) is often used instead of the state value function V (st). We limit our scope to the computation of V (st), however. (a) local planning local planning with value (b) function Figure 1: Illustration of local planning with a horizon of h = 2. The nodes in the search tree represent states and the links represent ac- tions. The red path represents a sampled plan. The dashed gray links mark unreachable paths. (a) plain local planning. (b) local planning with a value function to consider global effects in the unreachable subtree. an action selection strategy πtree [Perez Liebana et al., 2015; Belzner and Gabor, 2017b]. The nodes of the tree repre- sent states and the links represent actions. The state values V πtree (st) are computed recursively according to eq. 1 start- ing from the leaves of the search tree. Monte Carlo Tree Search (MCTS) is a popular closed-loop planning approach, which is applied to very large and complex domains [Chaslot, 2010; Kocsis and Szepesv´ari, 2006; Silver and Veness, 2010; Silver et al., 2016; Silver et al., 2017]. MCTS can also be adapted to multi-agent planning [Amato and Oliehoek, 2015; Claes et al., 2017]. Open-loop planning searches for action sequences or plans of length h [Bubeck and Munos, 2010; Weinstein and Littman, 2013; Perez Liebana et al., 2015; Belzner and Gabor, 2017b]. These plans are typically sam- pled from a sequence of distributions Φ1, ..., Φh and simu- lated in M. The resulting rewards are accumulated accord- ing to eq. 1 and used to update the distributions. Open-loop planning does not store any information about intermediate states, thus enabling efficient planning in large-scale domains [Weinstein and Littman, 2013; Perez Liebana et al., 2015]. An approach to open-loop planning in MAS is proposed in [Belzner and Gabor, 2017a]. 2.3 Reinforcement Learning Reinforcement Learning (RL) corresponds to a policy search for an unknown environment M. In general, an agent knows the state and action space S and A but it does not know the effect of executing at ∈ A in st ∈ S [Boutilier, 1996; Sutton and Barto, 1998]. Model-based RL methods learn a model M ≈ M by approximating P and R [Boutilier, 1996; Sutton and Barto, 1998; Hester and Stone, 2013]. M can be used for planning to find a policy. In this paper, we focus on model- free RL to approximate V ∗ based on experience samples et = (st, at, st+1, rt) and a parametrized function approxi- mator Vθ with parameters θ without learning a model M [Sut- ton and Barto, 1998]. A policy π can be derived by maximiz- ing Vθ such that π(st) = argmaxat∈A( Qθ(st, at)), where st+1∈S P(st+1st, at) Vθ(st+1) is the approximated action value function [Boutilier, 1996; Qθ(st, at) = R(st, at) + γ(cid:80) Distributed Value Function Approximation In this paper, we focus on centralized learning of V ∗, where all agents share the same parameters θ similarly to [Foerster et al., 2016; Tan, 1997]. Unlike previous work on multi-agent RL, we do not use the approximated value function to directly derive a policy. Instead, we use it to guide online planning in MAS. Besides, there exist approaches to approximate the value function asynchronously and in parallel [Nair et al., 2015; Mnih et al., 2016]. In that case, multiple agents act indepen- dently of each other in different instances of the same domain. They share experience with each other in order to update the same value function approximation Vθ in parallel to acceler- ate the learning process. Our approach approximates V ∗ based on the global expe- rience of multiple agents, which act in the same environment. Our approximation Vθ is not meant to improve the perfor- mance of individual agents but to improve the behaviour of the MAS as a whole. 4 EVADE We now describe Emergent Value function Approximation for Distributed Environments (EVADE) for leveraging statistical multi-agent online planning with a value function, which is approximated online at system runtime. EVADE is a frame- work for combining multi-agent online planning and RL to further improve the performance in MAS. 4.1 Combining Online Planning and RL Given a perfect generative model M = M, online planning can be used for decision making with high quality and accu- racy w.r.t. the expected return. However, due to computa- tional constraints, online planning is unable to make looka- heads for arbitrarily long horizons, which would be required for highly complex tasks that require much more time steps to solve than the actually feasible horizon as sketched in fig. 1a. In contrast, model-free RL with a parametrized function ap- proximator Vθ allows for potentially infinite future prediction but has approximation erros due to the compressing nature of Vθ. By combining online planning and RL, a decision maker can benefit from both advantages [Silver et al., 2016; Sil- ver et al., 2017; Anthony et al., 2017]. The limited looka- head of planning can be enhanced with Vθ as shown in fig. 1b. Online planning can plan accurately for h initial time steps, which are weighted more than the outcome estimate Vθ(st+h), given a discount factor of γ < 1. The discount can also neglect possible approximation errors of Vθ. Espe- cially in highly complex and stochastic domains with multi- ple agents, we believe that the integration of a value function approximation could improve the performance of otherwise limited multi-agent online planning. 4.2 Multi-Agent Planning with Experience We focus on online settings, where there is an alternating planning and learning step for each time step t. In the plan- ning step, the system searches for a joint policy πlocal, which maximizes Gt,EVADE: Gt,EVADE = Gt + γh Vθ(st+h) (3) Gt,EVADE extends Gt from eq. 1 with Vθ(st+h) as the pro- vided global outcome estimate to enhance local planning with a limited horizon of h as sketched in fig. 1b. The planning step can be implemented with an arbitrary multi-agent plan- ning algorithm, depending on the concrete problem. After the planning step, all agents execute the joint action at = πlocal(st) causing a state transition from st to st+1 with a reward signal rt. This emergent result is stored as experi- ence sample et = (st, at, st+1, rt) in an experience buffer D. A sequence of experience samples e1, ..., eT is called episode of length T . In the subsequent learning step, a parametrized function approximator Vθ is used to minimize the one-step temporal difference (TD) error of all samples et in D w.r.t. θ. The TD error for et is defined by [Sutton, 1988; Sutton and Barto, 1998]: δt = Vθ(st) − (rt + γ Vθ(st+1)) (4) It should be noted that the approximation only depends on the experience samples et ∈ D and does not require a model like hybrid planning approaches explained in section 3. The up- dated value function Vθ can then be used for the next planning step at t + 1. The complete formulation of multi-agent online planning with EVADE is given in algorithm 1, where T is the length of an episode, M is the generative model used for planning, n is the number of agents in the MAS, h is the planning horizon, nbudget is the computation budget and Vθ is the value function approximator. The parameter M ASP lan can be an arbitrary multi-agent planning algorithm for searching a joint policy πlocal by maximizing Gt,EVADE. Given that the computation budget nbudget is fixed and the time to update Vθ at each time step is constant2, EVADE is suitable for online planning and learning in real-time MAS. Algorithm 1 Multi-agent online planning with EVADE 1: procedure EV ADE(M ASP lan, M , n, h, nbudget, Vθ) 2: 3: 4: 5: Initialize θ of Vθ Observe s1 for t = 1, T do Find πlocal using M ASP lan(st, M , n, h, nbudget, Vθ) 6: 7: 8: 9: Execute at = πlocal(st) Observe reward rt and new state st+1 Store new experience et = (st, at, st+1, rt) in D Refine θ to minimize the TD error δt for all et ∈ D 2In practice, θ is updated w.r.t. experience batches of constant size, which are sampled from D [Mnih et al., 2013; Mnih et al., 2015]. 4.3 Architecture We focus on centralized learning, since we believe that V ∗ can be approximated faster if all agents share the same pa- rameters θ [Tan, 1997; Foerster et al., 2016]. Online plan- ning can be performed in a centralized or decentralized way by using a concrete MAS planning algorithm. In both cases, each planner uses the common value function approximation Vθ to search for πlocal by maximizing Gt,EVADE. A conceptual overview of the EVADE architecture is shown in fig. 2. Com- pletely decentralized architectures, where all agents plan and learn independently of each other, are not considered here and left for future work. processed by the machines marked as green pentagons before going to the machines marked as blue rectangles. Note that i can choose between two different machines for processing its requests ai,1 = 9 and ai,2 = 3, which are rendered as light green pentagons or light blue rectangles. In the presence of multiple agents, coordination is required to choose an appro- priate machine, while avoiding conflicts with other agents. (a) Centralized planning (b) Decentralized planning Figure 2: Illustration of the possible MAS planning architectures for EVADE. The planners get global feedback from a value function, which is approximated in a centralized way. The red dashed arrow between the planners in fig. 2b represents a coordination mechanism for decentralized planning. Decentralized planning approaches require an explicit co- ordination mechanism to avoid convergence to suboptimal joint policies as shown in fig. 2b and in [Boutilier, 1996; Bus¸oniu et al., 2010]. This could be done by using a con- sensus mechanism to synchronize on time or on a common seed value to generate the same random numbers when sam- pling plans [Emery-Montemerlo et al., 2004]. Agents could also exchange observations, experience, plans or policies via communication [Tan, 1997; Wu et al., 2009]. Another way is to predict other agents' actions by using a policy func- tion similarly to [Silver et al., 2016] or by maintaining a belief about other agents' behaviour [Bus¸oniu et al., 2010; Oliehoek and Amato, 2016]. 5 Experiments 5.1 Evaluation Environment Description We implemented a smart factory environment to evaluate multi-agent online planning with EVADE. Our smart fac- tory consists of a 5 × 5 grid of machines with 15 different machine types as shown in fig. 3a. Each item is carried by one agent i and needs to get processed at various ma- chines according to its randomly assigned processing tasks tasksi = [{ai,1, bi,1},{ai,2, bi,2}], where each task ai,j,bi,j is contained in a bucket. While tasks in the same bucket can be processed in any order, buckets themselves have to be pro- cessed in a specific order. Fig. 3b shows an example for an agent i with tasksi = [{9, 12},{3, 10}]. It first needs to get (a) machine grid (b) an agent and its tasks Figure 3: Illustration of the smart factory setup used in the experi- ments. (a) the 5 × 5 grid of machines. The numbers in each grid cell denote the machine type. (b) an agent i (red circle) in the fac- tory with tasksi = [{9, 12},{3, 10}]. It should get processed at the green pentagonal machines first before going to the blue rectangular machines. All agents have a random initial position and can move along the machine grid or enqueue at their current position represented by a machine. Each machine can process exactly one item per time step with a cost of 0.25 but fails with a probability of 0.1 to do so. Enqueued agents are unable to perform any actions. If a task is processed, it is removed from its bucket. If a bucket is empty, it is removed from the item's tasks list. An item is complete if its tasks list is empty. The goal is to complete as many items as possible within 50 time steps, while avoiding any conflicts or enqueuing at wrong ma- chines. MMDP Formulation The smart factory environment can be modeled as MMDP M = (cid:104)D,S,A,P,R(cid:105). D is the set of n agents with Dactive ∩ Dcomplete = ∅ and D = Dactive ∪ Dcomplete. Dactive is the set of agents with incomplete items and Dcomplete is the set of agents with complete items. S is a set of sys- tem states described by the individual state variables of all agents, items and machines. A is the set of joint actions. Each agent i ∈ D has the same individual action space Ai enabling it to move north, south, west or east, to enqueue at its current machine m = posi or to do nothing. Any at- tempt to move across the grid boundaries is treated the same as "do nothing". P is the transition probability function. R is the scalar reward function. R at time step t is defined by R(st, at) = scoret+1 − scoret, where scoret is the immedi- ate evaluation function for the system state: where taskst = (cid:80) scoret = Dcomplete − taskst − costt − tpent (5) c is the total number of currently unprocessed tasks, costt is the total sum of pro- cessing costs for each machine after processing an enqueued (cid:80) i∈Dactive c∈tasksi item and tpent = tpent−1 +(cid:80) i∈Dactive penalty is the total sum of time penalities with penalty = 0.1 for all incomplete items at time step t. Processing tasks and completing items increases scoret. Otherwise, scoret decreases for each in- complete item or enqueuing at a wrong machine. Complexity Depending on the number of agents n, the number of possi- ble joint actions is A = Ain = 6n. The machine failure probability of 0.1 increases the branching factor of the prob- lem even more. Given a planning horizon of h, the number of possible joint plans is defined by: (6) πlocal = Πlocal,in = (Aih)n = Aih·n = 6h·n We tested EVADE in settings with 4 and 8 agents. In the 4- agent case, there exist 64 ≈ 1300 possible joint actions. In the 8-agent case, there exist 68 ≈ 1.68·106 possible joint actions. In our stochastic smart factory setup, where machines can fail with a probability of 0.1 and where agents are not acting in a deterministic way, the environment has a significantly higher branching factor than the game of Go, which has a branching factor of 250 [Silver et al., 2016]. 5.2 Methods Online Open-Loop Planning Due to the stochasticity and high complexity of our environ- ment, we focus on open-loop planning because we think that current state-of-the-art algorithms based on closed-loop plan- ning would not scale very well in our case [Perez Liebana et al., 2015; Amato and Oliehoek, 2015]. Also, we do not aim for optimal planning, since our goal is to enhance existing local planning algorithms, which might even perform subop- timal in the first place. The individual policy πi for each agent i is implemented as a stack or sequence of multi-armed bandits (MAB) of length h as proposed in [Belzner and Gabor, 2017b]. Each MAB Φt = P (atDat) represents a distribution, where Dat is a buffer of size 10 for storing local returns, which are observed when selecting arm at ∈ A. Each buffer Dat is implemented in a sliding window fashion to consider only most recent ob- servations to adapt to the non-stationary joint behaviour of all agents during the planning step. Thompson Sampling is implemented as concrete MAB al- gorithm because of its effectiveness and robustness for mak- ing decisions under uncertainty [Thompson, 1933; Chapelle and Li, 2011; Belzner and Gabor, 2017b]. The imple- mentation is adopted from [Honda and Takemura, 2014; Bai et al., 2014], where the return values in Dat for each arm at are assumed to be normally distributed. To optimize πi, a plan of h actions is sampled from the MAB stack. The plan is evaluated in a simulation by us- ing a generative model M. The resulting rewards are ac- cumulated to local returns according to eq. 3 and used to update the corresponding MABs of the MAB stack. This procedure is repeated (cid:98) nbudget h (cid:99) times. Afterwards, the action at = argmaxa1∈A{Da1} is selected from the MAB Φ1 for execution in the actual environment, where Da1 is the mean of all local returns currently stored in Da1. Multi-Agent Planning We implemented two multi-agent planning algorithms to evaluate the performance achieved by using EVADE. All al- gorithms enhanced with EVADE were compared with their non-enhanced counterparts w.r.t. performance and efficiency. fξ(π) = (cid:81)n Direct Cross Entropy (DICE) method for policy search in distributed models DICE is a centralized planning algo- rithm proposed in [Oliehoek et al., 2008a] and uses stochas- tic optimization to search joint policies, which are optimal or close to optimal. In DICE a multivariate distribution i=1 fξi(πi) is maintained to sample candidate joint policies π. These candidates are evaluated in a simula- tion with a global model M. The Nb best candidates are used to update fξ. This procedure is repeated until convergence is reached or nbudget has run out. Our implementation of DICE uses n MAB stacks representing fξ(π) to sample joint plans of length h, which are simulated in M. The resulting local returns are used to update all MAB stacks. Distributed Online Open-Loop Planning (DOOLP) DOOLP is a decentralized version of DICE proposed in [Belzner and Gabor, 2017a], where each agent is controlled by an individual planner with an individual model Mi = M for simulation-based planning. At every time step each agent i iteratively optimizes its policy πi by first sampling a plan and then querying the sampled plans of its neighbours to construct a joint plan. The joint plan is simulated in Mi and the simulation result is used to update the individual policy πi of agent i. The individual MAB stacks are assumed to be private for each agent i. Due to the stochasticity of the environment described in section 1 and 5.1, the planners can have different simulation outcomes leading to different updates to the individual MAB stacks. As a decentralized approach, DOOLP requires an ex- plicit coordination mechanism to avoid suboptimal joint poli- cies (see section 4.3 and fig. 2b). We implemented a communication-based coordination mechanism, where each planner communicates its sampled plans to all other planners, while keeping its actual MAB stack private. Value Function Approximation We used a deep convolutional neural network as Vθ to ap- proximate the value function V ∗. The weights of the neu- Vθ was trained with TD ral network are denoted as θ. learning by using methods of deep RL [Mnih et al., 2013; Mnih et al., 2015]. An experience buffer D was implemented to uniformly sample minibatches to perform stochastic gra- dient descent on. D was initialized with 5000 experience samples generated from running smart factory episodes us- ing multi-agent planning without EVADE. An additional target network Vθ− was used to generate TD regression targets for Vθ (see eq. 4) to stabilize the training [Mnih et al., 2015]. All hyperparameters used for training Vθ are listed in table 1. The factory state is encoded as a stack of 5 × 5 feature planes, where each plane represents the spatial distribution of hyperparameter update rule for optimization learning rate discount factor γ minibatch size replay memory size target network update frequency C value ADAM 0.001 0.95 64 10000 5000 Table 1: Hyperparameters for the value network Vθ. in the first hundred episodes. The average score increases slowly afterwards or stagnates as shown in the 8-agent case in fig. 4c and 4d. There are no significant differences between the enhanced versions with nbudget ∈ {384, 512}. Planning with a budget of nbudget = 192 leads to worse performance than the corresponding enhanced variants with a larger bud- get. machines or agents w.r.t. some aspect. An informal descrip- tion of all feature planes is given in table 2. The input to Vθ is a 5 × 5 × 35 matrix stack consisting of 35 matrices. The first hidden layer convolves 128 filters of size 5× 5 with stride 1. The next three hidden layer convolve 128 filters of size 3 × 3 with stride 1. The fifth hidden layer convolves one filter of size 1 × 1 with stride 1. The sixth hidden layer is a fully connected layer with 256 units. The output layer is a fully connected with a single linear unit. All hidden layers use exponential linear unit (ELU) activation as proposed in [Clevert et al., 2015]. The architecture of Vθ was inspired by the value network of [Silver et al., 2016]. 5.3 Results Various experiments with 4- and 8-agent settings were con- ducted to study the effectiveness and efficiency achieved by the multi-agent online planning algorithms from section 5.2 with EVADE. An episode is reset after T = 50 time steps or when all items are complete such that Dactive = ∅. A run consists of 300 episodes and is repeated 100 times. Multi-agent online planning with EVADE searches for a joint policy πlocal by maximizing Gt,EVADE with a value function approximation Vθ (see eq. 3). All baselines perform planning without EVADE by maximizing Gt instead (see eq. 1). The performance of multi-agent online planning is evalu- ated with the value of score50 at the end of each episode (see eq. 5) and the item completion rate Rcompletion at the end of the 300th episode, which is defined by: Rcompletion = Dcomplete ∪ Dactive (7) with 0 ≤ Rcompletion ≤ 1. If all items are complete within 50 time steps, then Rcompletion = 1. If no item is complete within 50 time steps, then Rcompletion = 0. All baselines were run 500 times to determine the average of score50 and Rcompletion. Efficiency w.r.t. Computation Budget The effect of EVADE w.r.t. the breadth of the policy search was evaluated. The experiments for each algorithm were run with different budgets nbudget ∈ {192, 384, 512}3 and a fixed horizon of h = 4. The baselines represented by the corre- sponding non-enhanced planning algorithms had a computa- tion budget of nbudget = 512. Fig. 4 shows the average progress of score50. In all cases, the EVADE enhanced versions outperform their correspond- ing baselines. There is a relatively large performance gain 3We also experimented with nbudget = 256 but there was no sig- nificant difference to planning with nbudget = 384. Dcomplete D = Dcomplete (a) DICE (4 agents) (b) DOOLP (4 agents) (c) DICE (8 agents) (d) DOOLP (8 agents) Figure 4: Average progress of score50 of 100 runs shown as running mean over 5 episodes for different computation budgets nbudget ∈ {192, 384, 512} and a horizon of h = 4. All baselines have a com- putation budget of nbudget = 512. Shaded areas show the 95% con- fidence interval. The average completion rates Rcompletion at the end of the 300th episode of all experiments are listed in table 3. In the 4-agent case, the completion rates of the baselines are about 63%, while the rates achieved by the EVADE enhanced ver- sions range from 86 to 92%. In the 8-agent case, the com- pletion rates of the baslines are about 54%, while the rates achieved by the EVADE enhanced versions range from 65 to 78%. EVADE enhanced planning with nbudget ∈ {384, 512} always tends to achieve a higher completion rate than using a budget of nbudget = 192. Efficiency w.r.t. Horizon Next the effect of EVADE w.r.t. the depth of the policy search was evaluated. The experiments for each algorithm were run with different horizon lengths h ∈ {2, 4, 6} and a fixed com- putation budget of nbudget = 384. The baselines represented by the corresponding non-enhanced planning algorithms had a horizon of h = 6. The planning horizon h influences the reachability of ma- chines in each simulation step as shown in fig. 5. In this example, the agent can only reach about one fifth of the grid when planning with h = 2 (see fig. 5b), while it can theoret- ically reach almost any machine when planning with h = 6 Feature Machine type Agent state 1 4 Tasks (1st bucket) Tasks (2nd bucket) 15 15 Table 2: Description of all feature planes as input for Vθ. # Planes Description The type of each machine as a value between 0 and 14 (see fig. 3a) The number of agents standing at machines whose types are (not) contained in their current tasks and whether they are enqueued or not. Spatial distribution of agents containing a particular machine type in their first bucket of tasks for each available machine type. Same as "Tasks (1st bucket)" but for the second bucket of tasks. Table 3: Average rate of complete items Rcompletion at the end of the 300th episode of all experiments within a 95% confidence interval. Planning was performed with different computation budgets nbudget and a horizon of h = 4. algorithm (# agents) DICE (4 agents) DOOLP (4 agents) DICE (8 agents) DOOLP (8 agents) baseline (nbudget = 512) 62.5 ± 2.1% 63.7 ± 2.1% 55.2 ± 1.5% 53.9 ± 1.4% EVADE (nbudget = 192) 86.8 ± 3.6% 88.5 ± 3.0% 65.0 ± 3.4% 65.8 ± 3.7% EVADE (nbudget = 384) 89.3 ± 3.0% 91.3 ± 2.9% 73.1 ± 3.3% 72.8 ± 3.6% EVADE (nbudget = 512) 91.8 ± 3.0% 91.0 ± 3.4% 77.5 ± 3.2% 73.0 ± 3.5% (see fig. 5d). a longer horizon after about one hundred episodes as shown in fig. 6c and 6d. This phenomenon will be discussed in the next section. (a) an agent and its tasks (b) horizon of h = 2 (a) DICE (4 agents) (b) DOOLP (4 agents) (c) horizon of h = 4 (d) horizon of h = 6 Figure 5: Reachability of machines for an agent (red circle) in a simulation step depending on the planning horizon h. Gray grid cells mark unreachable machines. (a) The example from fig. 3b. (b), (c) and (d) Reachable machines within the dashed red boundaries, given resp. horizons of h. (c) DICE (8 agents) (d) DOOLP (8 agents) Figure 6: Average progress of score50 of 100 runs shown as running mean over 5 episodes for different horizons h ∈ {2, 4, 6} and a computation budget of nbudget = 384. All baselines have a horizon of h = 6. Shaded areas show the 95% confidence interval. Fig. 6 shows the average progress of score50. Planning with a horizon of h = 2 always had the worst initial average performance but the largest performance gain in the first hun- dred episodes, while planning with a horizon of h = 6 had the best initial average performance but the smallest perfor- mance gain. In the 8-agent case, planning with EVADE and a horizon of h = 2 even outperforms the planning variants with The average completion rates Rcompletion at the end of the 300th episode of all experiments are listed in table 4. In the 4-agent case, the completion rate of the baselines are about 70%, while the rate achieved by the EVADE enhanced ver- sions range from about 82 to 92%. In the 8-agent case, the completion rates of the baslines are about 59%, while the rate achieved by the EVADE enhanced versions range from about 66 to 77%. Increasing the horizon from 2 to 6 in the 4-agent case tends to slightly increase Rcompletion, while in the 8-agent case it leads to a decrease of Rcompletion. 5.4 Discussion Our experiments show that statistical multi-agent online plan- ning can be effectively improved with EVADE, even when us- ing a smaller computation budget nbudget than planning with- out any value function. However, nbudget must not be too small, since statistical online planning algorithms always re- quire a minimum of computation to reach promising states with higher probability. This is shown in the experimental settings with nbudget = 192 in fig. 4 and table 3. In the smart factory environment, planning with a sufficient horizon length is crucial to find joint policies with high qual- ity as shown in fig. 5 and table 3 and 4 regarding the perfor- mance of the baselines. If a needed machine is unreachable in the simulation, it cannot be considered in the local planning process, thus possibly leading to poor solutions. In our ex- periments, the value function approximation could improve multi-agent planning with horizons which were too short to consider the entire factory. If the discount factor is γ < 1, then the value function influences planning with short horizons more than planning with a long horizon (see eq. 3). In our experiments, plan- ning with a horizon of h = 2 was able to keep up with plan- ning variants with a longer horizon, even outperforming them in the 8-agent case, given an equal computation budget of nbudget = 384. These are strong indications that our approach offers planning efficiency w.r.t. the breadth and the depth of the policy search after a sufficient learning phase. The performance stagnation in the 8-agent case after hun- dred episodes can be explained with the enormous policy space to be searched and the relatively small computation budget nbudget. This also explains the rather poor performance of online planning with a horizon of h = 6 compared to vari- ants with h = 2 or h = 4 as shown in fig. 6c and 6d. Given nbudget = 384, the former only performs (cid:98) nbudget h (cid:99) = 64 sim- ulations per time step, while searching a much larger policy space (πlocal > 1037) than the latter (πlocal < 1025) accord- ing to eq. 6. When using the value function approximation Vθ, more simulations should lead to high quality results with a higher accuracy. Thus, a larger performance gain can be expected when increasing nbudget. 6 Conclusion & Future Work In this paper, we presented EVADE, an approach to effec- tively improve the performance of statistical multi-agent on- line planning in stochastic domains by integrating global experience. For this purpose, a value function is approxi- mated online based on the emergent system behaviour by us- ing model-free RL. By considering global outcome estimates with that value function during the planning step, multi-agent online planning with EVADE is able to overcome the limita- tion of local planning as sketched in fig. 1. We also introduced a smart factory environment, where multiple agents need to process various items with different tasks at a shared set of machines in an automated and self- organizing way. Unlike domains used in [Silver et al., 2016; Silver et al., 2017; Anthony et al., 2017], our environment can have multiple agents, is stochastic and has a higher branching factor, given a sufficient number of agents. EVADE was successfully tested with two existing statisti- cal multi-agent planning algorithms in this highly complex and stochastic domain. EVADE offers planning efficiency w.r.t. the depth and the breadth of the joint policy search re- quiring less computational effort to find solutions with higher quality compared to multi-agent planning without any value function. For now, EVADE has only been applied to fully observable settings. Decentralized partially observable problems can of- ten be decomposed into smaller subproblems, which are fully observable themselves. This is common in distributed envi- ronments, where agents can sense and communicate with all neighbours within their range. EVADE could be directly ap- plied to those subproblems. As a possible direction for future work, EVADE could be extended to partially observable do- mains without any problem decomposition. References [Amato and Oliehoek, 2015] Christopher and Frans A Oliehoek. Scalable planning and learning for the Twenty- multiagent pomdps. Ninth AAAI Conference on Artificial Intelligence, pages 1995–2002. AAAI Press, 2015. In Proceedings of Amato [Anthony et al., 2017] Thomas Anthony, Zheng Tian, and David Barber. Thinking fast and slow with deep learn- In Advances in Neural Information ing and tree search. Processing Systems, pages 5366–5376, 2017. [Bai et al., 2014] Aijun Bai, Feng Wu, Zongzhang Zhang, and Xiaoping Chen. Thompson sampling based monte- carlo planning in pomdps. In Proceedings of the Twenty- Fourth International Conferenc on International Confer- ence on Automated Planning and Scheduling, pages 29– 37. AAAI Press, 2014. [Bellman, 1957] Richard Bellman. Dynamic Programming. Princeton University Press, Princeton, NJ, USA, 1 edition, 1957. [Belzner and Gabor, 2017a] Lenz Belzner and Thomas Ga- bor. Scalable multiagent coordination with distributed on- line open loop planning. arXiv preprint arXiv:1702.07544, 2017. [Belzner and Gabor, 2017b] Lenz Belzner and Thomas Ga- In Proceedings of the bor. Stacked thompson bandits. 3rd International Workshop on Software Engineering for Smart Cyber-Physical Systems, pages 18–21. IEEE Press, 2017. [Belzner et al., 2015] Lenz Belzner, Rolf Hennicker, and Martin Wirsing. Onplan: A framework for simulation- In International Workshop on based online planning. Formal Aspects of Component Software, pages 1–30. Springer, 2015. Table 4: Average rate of complete items Rcompletion at the end of the 300th episode of all experiments within a 95% confidence interval. Planning was performed with different horizons h and a computation budget of nbudget = 384. algorithm (# agents) DICE (4 agents) DOOLP (4 agents) DICE (8 agents) DOOLP (8 agents) baseline (h = 6) 69.7 ± 2.0% 71.6 ± 2.0% 58.3 ± 1.5% 60.0 ± 1.5% EVADE (h = 2) 87.0 ± 3.2% 82.3 ± 4.0% 77.0 ± 3.8% 72.9 ± 3.5% EVADE (h = 4) 89.3 ± 3.0% 91.3 ± 2.9% 73.1 ± 3.3% 72.8 ± 3.6% EVADE (h = 6) 90.8 ± 3.2% 88.5 ± 3.4% 66.1 ± 3.1% 67.6 ± 2.9% [Boutilier, 1996] Craig Boutilier. Planning, learning and co- ordination in multiagent decision processes. In Proceed- ings of the 6th conference on Theoretical aspects of ratio- nality and knowledge, pages 195–210. Morgan Kaufmann Publishers Inc., 1996. [Bubeck and Munos, 2010] S Bubeck and R Munos. Open loop optimistic planning. In Conference on Learning The- ory, 2010. [Bus¸oniu et al., 2010] Lucian Bus¸oniu, Robert Babuska, and Bart De Schutter. Multi-agent reinforcement learning: An overview. In Innovations in multi-agent systems and applications-1, pages 183–221. Springer, 2010. [Cassandra and Kaelbling, 2016] Anthony R Cassandra and Leslie Pack Kaelbling. Learning policies for partially ob- In Machine Learn- servable environments: Scaling up. ing Proceedings 1995: Proceedings of the Twelfth Inter- national Conference on Machine Learning, Tahoe City, California, July 9-12 1995, page 362. Morgan Kaufmann, 2016. [Chapelle and Li, 2011] Olivier Chapelle and Lihong Li. An empirical evaluation of thompson sampling. In Advances in neural information processing systems, pages 2249– 2257, 2011. [Chaslot, 2010] Guillaume Chaslot. Monte-carlo tree search. Maastricht: Universiteit Maastricht, 2010. [Claes et al., 2017] Daniel Claes, Frans Oliehoek, Hendrik Baier, and Karl Tuyls. Decentralised online planning for multi-robot warehouse commissioning. In Proceedings of the 16th Conference on Autonomous Agents and MultiA- gent Systems, pages 492–500. International Foundation for Autonomous Agents and Multiagent Systems, 2017. [Clevert et al., 2015] Djork-Arn´e Clevert, Thomas Un- terthiner, and Sepp Hochreiter. Fast and accurate deep network learning by exponential linear units (elus). CoRR, abs/1511.07289, 2015. [Emery-Montemerlo et al., 2004] Rosemary Emery- Montemerlo, Geoff Gordon, Jeff Schneider, and Sebastian Thrun. Approximate solutions for partially observable stochastic games with common payoffs. In Proceedings of the Third International Joint Conference on Autonomous Agents and Multiagent Systems-Volume 1, pages 136–143. IEEE Computer Society, 2004. [Evans, 1984] Jonathan St BT Evans. Heuristic and analytic processes in reasoning. British Journal of Psychology, 75(4):451–468, 1984. [Foerster et al., 2016] Jakob Foerster, Yannis M Assael, Nando de Freitas, and Shimon Whiteson. Learning to communicate with deep multi-agent reinforcement learn- ing. In Advances in Neural Information Processing Sys- tems, pages 2137–2145, 2016. [Hester and Stone, 2013] Todd Hester and Peter Stone. Tex- real-time sample-efficient reinforcement learning plore: for robots. Machine learning, 90(3):385–429, 2013. [Honda and Takemura, 2014] Junya Honda and Akimichi Takemura. Optimality of thompson sampling for gaussian In Artificial Intelligence and bandits depends on priors. Statistics, pages 375–383, 2014. [Howard, 1961] Ronald A. Howard. Dynamic Programming and Markov Processes. The MIT Press, 1961. [Kahneman, 2003] Daniel Kahneman. Maps of bounded rationality: Psychology for behavioral economics. The American economic review, 93(5):1449–1475, 2003. [Kocsis and Szepesv´ari, 2006] Levente Kocsis and Csaba Szepesv´ari. Bandit based monte-carlo planning. In ECML, volume 6, pages 282–293. Springer, 2006. [Mnih et al., 2013] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller. Playing atari with deep In NIPS Deep Learning Work- reinforcement learning. shop. 2013. [Mnih et al., 2015] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Belle- mare, Alex Graves, Martin Riedmiller, Andreas K Fidje- land, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. Nature, 518(7540):529–533, 2015. [Mnih et al., 2016] Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asyn- chronous methods for deep reinforcement learning. In International Conference on Machine Learning, pages 1928–1937, 2016. [Nair et al., 2015] Arun Nair, Praveen Srinivasan, Sam Blackwell, Cagdas Alcicek, Rory Fearon, Alessandro De Maria, Vedavyas Panneershelvam, Mustafa Suleyman, Charles Beattie, Stig Petersen, et al. Massively parallel methods for deep reinforcement learning. arXiv preprint arXiv:1507.04296, 2015. [Oliehoek and Amato, 2016] Frans A Oliehoek and Christo- pher Amato. A concise introduction to decentralized POMDPs. Springer, 2016. for solving decentralized pomdps. In 21st Conference on Uncertainty in Artificial Intelligence-UAI'2005, 2005. [Tampuu et al., 2017] Ardi Tampuu, Tambet Matiisen, Do- rian Kodelja, Ilya Kuzovkin, Kristjan Korjus, Juhan Aru, Jaan Aru, and Raul Vicente. Multiagent cooperation and competition with deep reinforcement learning. PloS one, 12(4):e0172395, 2017. [Tan, 1997] Ming Tan. Multi-agent reinforcement learning: independent vs. cooperative agents. In Readings in agents, pages 487–494. Morgan Kaufmann Publishers Inc., 1997. [Thompson, 1933] William R Thompson. On the likelihood that one unknown probability exceeds another in view of the evidence of two samples. Biometrika, 25(3/4):285– 294, 1933. [Weinstein and Littman, 2013] Ari Weinstein and Michael L Littman. Open-loop planning in large-scale stochastic do- mains. In AAAI, 2013. [Wu et al., 2009] Feng Wu, Shlomo Zilberstein, and Xiaop- ing Chen. Multi-agent online planning with communi- cation. In Nineteenth International Conference on Auto- mated Planning and Scheduling, 2009. [Oliehoek et al., 2008a] Frans A Oliehoek, Julian FP Kooij, and Nikos Vlassis. The cross-entropy method for policy search in decentralized pomdps. Informatica, 32(4), 2008. [Oliehoek et al., 2008b] Frans A Oliehoek, Matthijs TJ Spaan, and Nikos Vlassis. Optimal and approximate q- value functions for decentralized pomdps. Journal of Ar- tificial Intelligence Research, 32:289–353, 2008. [Paquet et al., 2006] S´ebastien Paquet, Brahim Chaib-draa, In Pro- and St´ephane Ross. Hybrid pomdp algorithms. ceedings of The Workshop on Multi-Agent Sequential De- cision Making in Uncertain Domains (MSDM-06), pages 133–147, 2006. [Perez Liebana et al., 2015] Diego Perez Liebana, Jens Dieskau, Martin Hunermund, Sanaz Mostaghim, and Simon Lucas. Open loop search for general video game playing. In Proceedings of the 2015 Annual Conference on Genetic and Evolutionary Computation, pages 337–344. ACM, 2015. [Puterman, 2014] Martin L Puterman. Markov decision pro- cesses: discrete stochastic dynamic programming. John Wiley & Sons, 2014. [Ross et al., 2007] St´ephane Ross, Brahim Chaib-Draa, et al. Aems: An anytime online search algorithm for approxi- mate policy refinement in large pomdps. In IJCAI, pages 2592–2598, 2007. [Silver and Veness, 2010] David Silver and Joel Veness. In Advances in Monte-carlo planning in large pomdps. neural information processing systems, pages 2164–2172, 2010. [Silver et al., 2016] David Silver, Aja Huang, Chris J Maddi- son, Arthur Guez, Laurent Sifre, George Van Den Driess- che, Julian Schrittwieser, Ioannis Antonoglou, Veda Pan- neershelvam, Marc Lanctot, et al. Mastering the game of go with deep neural networks and tree search. Nature, 529(7587):484–489, 2016. [Silver et al., 2017] David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. Mastering the game of go without human knowledge. Nature, 550(7676):354–359, 2017. [Spaan et al., 2011] Matthijs TJ Spaan, Frans A Oliehoek, and Christopher Amato. Scaling up optimal heuristic search in dec-pomdps via incremental expansion. In Pro- ceedings of the Twenty-Second international joint con- ference on Artificial Intelligence-Volume Volume Three, pages 2027–2032. AAAI Press, 2011. [Sutton and Barto, 1998] Richard S Sutton and Andrew G Introduction to reinforcement learning, volume Barto. 135. MIT Press Cambridge, 1998. [Sutton, 1988] Richard S Sutton. Learning to predict by the methods of temporal differences. Machine learning, 3(1):9–44, 1988. [Szer et al., 2005] Daniel Szer, Francois Charpillet, and Shlomo Zilberstein. Maa*: A heuristic search algorithm
1812.00651	2	1812	2019-04-10T11:45:37	Towards Agent-based Models of Rumours in Organizations: A Social Practice Theory Approach	[ "cs.MA" ]	Rumour is a collective emergent phenomenon with a potential for provoking a crisis. Modelling approaches have been deployed since five decades ago; however, the focus was mostly on epidemic behaviour of the rumours which does not take into account the differences of the agents. We use social practice theory to model agent decision making in organizational rumourmongering. Such an approach provides us with an opportunity to model rumourmongering agents with a layer of cognitive realism and study the impacts of various intervention strategies for prevention and control of rumours in organizations.	cs.MA	cs	Towards Agent-based Models of Rumours in Organizations: A Social Practice Theory Approach∗ Amir Ebrahimi Fard, Rijk Mercuur, Virginia Dignum, Catholijn M. Jonker, and Bartel van de Walle Delft University of Technology [email protected], [email protected], [email protected], [email protected] and [email protected] Abstract. Rumour is a collective emergent phenomenon with a poten- tial for provoking a crisis. Modelling approaches have been deployed since five decades ago; however, the focus was mostly on epidemic behaviour of the rumours which does not take into account the differences of the agents. We use social practice theory to model agent decision making in organizational rumourmongering. Such an approach provides us with an opportunity to model rumourmongering agents with a layer of cogni- tive realism and study the impacts of various intervention strategies for prevention and control of rumours in organizations. Keywords: Rumour · Organization · Social Practice Theory · Agent- based Model 1 Introduction The phenomenon of rumourmongering has malicious impacts on societies. Ru- mours make people nervous, create stress, shake financial markets and disrupt aid operations [27]. In organizations, rumours lead to unpleasant consequences such as, breaking the workplace harmony, reduction of profit, drain of produc- tivity and damaging the reputation of a company [7, 20]. Recent work on the McDonald's wormburger rumour and the P&G Satan rumour confirm the neg- ative impact of rumours on the productivity of firms [7]. For 120 years, scholars from a wide range of disciplines are trying to un- derstand different dimensions of this phenomenon. Research in rumour stud- ies can be classified according to the approach followed: a case-based approach and a model-based approach. In the case-based approach, results are based on case studies, not on models, making it hard to generalize their conclusions. The model-based approach tries to explain the phenomenon of rumours by model- based based simulations. The model-based approaches, so far, focus only on the ∗This paper has been peer-reviewed and accepted for the Social Simulation Con- ference 2018 in Stockholm. The final authenticated version will be available online at Springer LNCS. The DOI will be provided when available. 2 A. Fard et al. dynamic of the spread, while rumour is a collective phenomenon and the acts of individuals can influence the whole system. Rumours in organizations have been mainly approached with case-based studied and dynamic spreading model. To our knowledge there are no studies where the cognition of the individual is taken into account. In our agent-based approach, we study the dynamics of the spread of rumours in organizations as an emergent (collective) behavior resulting from the behavior of individual agents using social practice theory. We use the proposed model to study the impact of change in organizational layout on control of organizational rumour. The concept of social practices stems from sociology, and aims to depict our 'doings and sayings' [24, p. 86], such as dining, commuting and rumourmonger- ing. This paper uses the semantics of the social practice agent (SoPrA) model [19] to gain insights in rumourmongering in organizations.1 SoPrA provides an unique tool to combine habitual behavior, social intelligence and interconnected practices in one model. This makes SoPrA especially well-suited for studying the spread of rumours in organizations as this practice is largely habitual, so- cial [10] and interconnected with practices as working and moving around. To build the model with SoPrA, owing to lack of available empirical dataset, we give a proof-of-concept on how to collect data by doing eight semi-structured interviews. This paper is organized as follows. The next section provides an overview of the research on rumours with an emphasis on studies of organizational rumour. Section 3 describes the context for our experiment, and the methodology of data collection and data preprocessing. The model is introduced in Section 4. One possible experiment is described in Section 5 and Section 6 presents our conclusions, discussion and ideas for future work. 2 Background & Related Work Rumours are unverified propositions or allegations which are not accompanied by corroborative evidence [7]. Rumours take different forms such as exaggerations, fabrications, explanations [23], wishes and fears [14]. Rumours have a lifecycle and change over the time. Allport and Postman in their seminal work psychology of rumour concluded that, as a rumour travels, it grows shorter, more concise, more easily grasped and told. [12]. Buckner considers rumour a collective be- haviour which is becoming more or less accurate while being passed on as they are subjected to the individuals' interpretations which depends on the structure of the situation in which the rumour originates and spreads subsequently [4]. Rumours are conceived to be unpleasant phenomena that should be curtailed. 1Mercuur et al. [19] provides a static model of SoPrA based on literature and argued modelling choices. This paper applies this model to the domain and extends it by including competences and affordances and modelling a dynamic component based on [18]. Note that Mercuur et al. [19] is still under review and only available as pre-print at the moment of writing. Rumourmongering in Organizations 3 Therefore, a number of strategies have been proposed to prevent and control them [4, 22, 15]. One of the rumour contexts that has received attention from researchers for almost four decades is organizations. Like rumour in general context which is explained in above paragraph, rumour in organizations has different types and follow its own life-cycle [7, 8, 1, 13, 2, 3]. Also, to quell credible and non-credible organizational rumours, a number of different techniques and strategies have been suggested [7, 13]. The research approach also follow the same pattern, with a slight difference which to best of our knowledge is qualitative without adopting any modelling approach. The related literature reported above are based on case studies or experi- ments in the wild. This pertains to the types of rumour, dynamics of rumour and strategies to control rumours, either in general or in organizational contexts. These case studies and experiments are to inform the construction of theories and models underlying the phenomenon of rumourmongering. Theories and models, in turn, should be tested in case studies and simulations. Model-based approaches do just that. However, the current state-of-the-art in model-based simulations of rumourmongering focus only on the dynamics of the rumourmongering, compa- rable to the epidemic modelling and spread of viruses [6, 31, 21, 28, 26]. These models do not consider the complexities of the agents that participate in ru- mourmongering. The research area of agent-based social simulations (ABSS) specializes on simulating the social phenomena as phenomena that emerge from the behaviour of individual agents. ABSS is a powerful tool for empirical research. It offers a natural environment for the study of connectionist phenomena in social science. This approach permits one to study how individual behaviour give rise to macro- scopic phenomenon [9]. Such an approach is an ideal way to study the macro effects of various social practices, because it can capture routines which are practiced by individuals on a regular basis in micro level and see their collective influence in a macro level. 3 Domain This research investigates the daily routine of rumourmongering in a faculty building on the campus of a Dutch University. In this faculty, students, re- searchers and staff work in offices with capacity of one to ten people. Aside from the actual work going on in the building, filling a bottle with water, getting cof- fee from the coffee machine, having lunch at the canteen and going to the toilet are among the most obvious practices that every employee in this faculty does on a daily basis. Nevertheless, there are other daily routines in the organization which are not that obvious. One of these latent routines is rumourmongering. Rumours or unverified information are transferred between students, researchers and staffs on a daily basis, during lunch, while queuing for coffee, when seeing each other in the hallways, and when meeting in classrooms and offices. All these situations 4 A. Fard et al. are potential contexts for casual talks and information communication without solid evidence. For data collection we conducted semi-structured explorative interviews with people from the above-mentioned faculty. Semi-structured interviews allows us to ask questions that are specifically aimed at acquiring the content needed for the SoPrA model, while still giving the freedom to ask follow up questions on unclear answers. The data collection can be improved in future works by increasing the number of interviewees and diversifying them (Not only asking from students). For demographic information, the reader is referred to Table 3. We prepared following question set to ask from each interviewee based on the meta-model which will be explained in the next section: 1. What are the essential competencies for rumourmongering? 2. What are the associated values with rumourmongering? 3. What kind of physical setting is associated with rumourmongering? Table 1. Interviewees demography. Number of Interviews 8 Number of Different Countries 6 Lowest Educational Level Mean age Female % MSc 28 50 Given the thin line between personality traits and competences, we used the Big Five model [11] to differentiate between personality traits and competences. For Question 2, we asked the interviewees to choose the relevant values from Schwartz's Basic Human Values model [25]. We asked the same set of questions about fact-based talk. We processed the collected data in two ways before using it in the model. Firstly, we clustered answers that point to the same concept. For example, in Question 3, interviewees gave answers such as cafeteria, coffee shop and cafe to point to a place where people can get together and drink coffee. In the coffee example, we clustered answers under the term of "coffee place". Secondly, we classified the answers to Question 2. As mentioned, for that question, we asked interviewees to pick associated values from Schwartz's Basic Human Values model. We used the third abstraction level of the model which is more fine-grained and compared to other levels, and gave the interviewees a better idea of what they point to. However, a model based on level three, would not allow us to compare the agents effectively. Therefore, we decided to wrap the answers and classify them based on second abstraction level. Using a clas- sification based on the first abstraction level would have been too homogeneous in the sense that the agents would behave too similar, which would loose the effectiveness of the simulation. Rumourmongering in Organizations 5 4 Model The model has two main parts: (i) static part and (ii) dynamic part. In the static part, the components of the model and their properties are described, and in the dynamic part we explain the interaction of those components. 4.1 Static Part This section describes the SoPrA meta-model which is used as the groundwork for our agent-based model, how we use empirical data to initiate the model, the model choices we make and how we tailor the model to the context of organiza- tion. The SoPrA meta-model was introduced by Mercuur et al. [19] and describes how the macro concept of social practices can be connected to micro level agent concepts. Figure 1 shows SoPrA in a UML-diagram. The main objects in a SoPrA model are activities (e.g., fact talk, rumourmongering), agents (e.g., PhD stu- dents, supervisors), competences (e.g., networking, listening), context elements (e.g., office, cafetaria) and values. Values here refer to human values as found by the earlier stated Schwartz model, such as, power or conformity. The social practice is an interconnection of (1) activities and (2) related associations as de- picted by the grey box in Figure 1. For example, the practice of talking consists of two possible activities fact talk or rumourmongering. The social practice con- nects these different activities with the Implementation association. If activity A implements activity B this means that A is a way of or a part of doing B. The Implementation association is the first of several associations that are related to an activity (see Table 2). Most associations are fairly self-explanatory, however the Trigger and Strategy association are a bit more complex. Follow- ing Wood and Neal [29], triggers are the basis for habitual behaviour. If an agent is near a context element that has a trigger association with an activity, then it will do that activity automatically (without for example considering its values). Following Crawford and Ostrom [5], strategies are related to norms and signify that something is the normal way to do something.. If an agent believes that activity A is a strategy for activity B, then it believes that other agents usually implement activity B by doing activity A. The SoPrA meta-model does not only relate the activities to other classes, but the agent itself also has two types of associations: HasCompetence and ValueAdherence which plays a role in choosing the activities it will do:. The HasCompetence association links possible skills to the agent who masters those. The ValueAdherence association captures if an agent finds that value important. The model can be initiated using empirical data. Note that in this study we focussed on a small set of explorative interviews. We show with this initial data a proof-of-concept of how the model can be initiated. To properly ground the model a larger and more rigorous empirical study is necessary. The activity class has three instances: talking, rumourmongering and fact talk. The number of instances of agent can vary in the different experiments (see Section 5). The instances of the context element, competence and values 6 A. Fard et al. Fig. 1. The social practice meta-model captured in the Unified Modelling Language, including classes (yellow boxes), associations (lines), association classes (transparent boxes), navigability (arrow-ends) and multiplicity (numbers). class are based on the gathered data and can be found in Table 3 and 4.2 The complete static model consists both of object instances and associations between these instances. An example focusing on one agent (i.e., Bob) and one activity (i.e., rumourmongering) is shown in Figure 2. Bob beliefs that the activity of rumourmongering is related to the value of privacy, curiosity and social power. He thinks it requires the competence of networking and noticing juicy details and thinks the activity is triggered (to some extent) by the hallway, restaurant and another agent named Alice. Furthermore, he himself has the competence of networking and adheres strongest to the value of ambition and weakest to the value of pleasure. 2The context-element 'Friend' and 'Colleague' are special cases; these are rather attributes of context-elements (i.e., agents) than context-elements themselves. In our model these are to some extent implicitly captured, because the agents who one sees most often (i.e., friends, colleagues) are mostly likely to be habitually associated with an action. Table 2. The associations attached to the activity and their specification. Rumourmongering in Organizations 7 Specification which activities are a way of or a part of doing the activity which context elements are needed to do the activity Association Implementation Affordance RequiredCompetence which competences are needed to do the activity Knowledge Belief RelatedValue Trigger Strategy which activities an agent knows about which personal beliefs an agent has about the activity which values are promoted or demoted by the activity which context elements habitually start the activity which activities usually implement the activity Table 3. The elements associated with the rumourmongering activity. Rumourmongering Context Elements Meaning Competence Self-Direction Sneaky Skills Network Skills Talking Skills Coffee place Hallway Friend Power Hedonism Restaurant Office Phone Computer Achievement Observing Skills Benevolence Table 4. The elements associated with the fact talk activity. Fact talk Context Elements Meaning Competence Colleague Academic Staff Office Conference Universalism Being knowledgeable Self-Direction Benevolence Critical Thinking Skills Achievement Communication Skills Listening Skills Meeting room Tradition Classroom Restaurant Phone Computer Pen Coffee 8 A. Fard et al. The agents differ in which activity they associate with which element. In other words, the SoPrA meta-model does not initiate one social practice that all agents share, but one social practice per agent. The chance that an agent relates an activity to a competence is based on the empirical data we gathered in the interviews. For example, if 50% of the interviewees linked critical thinking skills to fact talk the chance an agent makes this association depends on a binomial distribution with p = 0.5. For relatedValue association and HabitualTrigger association all agents make the associations as mentioned in Table 3 and 4. However, the weights differ per agent. The weights for the relatedValue asso- ciation are picked from a normal distribution between 0 and 1. Given the lack of empirical data on the relation between activities and human values, we fol- low the related finding of the World Value Survey that people adhere to values with roughly a normal distribution [30]. The weights for HabitualTrigger are picked on a logarithmic distribution based on the empirical work of [16]. One interesting modelling choice we made was to drop the Affordance assocations in the conceptual model. The SoPrA meta-model conceptualizes two associations with context elements. The HabitualTrigger association representing that some context element can automatically lead to a reactive action and the Affordance association representing that some context elements are a pre-condition to enact a certain behaviour. None of our interviewees mentioned a possible context ele- ment that affords rumourmongering fact talk. As such this association seemed irrelevant for our model. The associations related to the agents themselves are based on random distri- butions. Each competence has a 50% chance to be related to an agent. Each value is associated to each agent, but the weights differ. The weights for the hasValue association strength is based on a correlated normal distribution. Schwartz [25] shows that the strength to which people adhere to values is correlated. For exam- ple, people who positively value universalism usually negatively value achieve- ment. We use the correlations found by Schwartz [25] to simulate intercorrelated normal distribution from which we pick the weights. In future work, we aim to extent our interviews to also gather data that can inform these weights. For our modelling context, we need to extend the SoPrA model with a spatial component. We do this by adding two attributes to the ContextElement class called x-coordinate and y-coordinate. These coordinates can be used by the agent to sense which objects are near. Note that every agent is also a context element as indicated with the 'generalization' association in the UML-diagram. 4.2 Dynamic Part This section describes the dynamic part of the model which on each tick com- prises: 1. An agent decides on its location using the moving submodel and updates its coordinate attributes. 2. An agent decides if it will engage in fact talk or rumourmongering based on the choose-activity submodel. Rumourmongering in Organizations 9 The moving submodel has four components that agents can transfer between. As it is shown in Figure 3 the initial state is offices and from that state agents can leave their offices and pass the hallway to either have lunch at the restaurant or grab a cup of coffee at the coffee place. During the interviews, we discovered most of the people do those daily routines around the same period of time and only a few people do not follow this pattern and leave their offices out of usual time periods, so we concluded the transition of agents between different locations is a random phenomenon which follows a normal probability distribution. Fig. 3. The moving model for agents The choose-activity submodel is based on Mercuur et al. [17] and has three stages. The submodel is depicted in Figure 4. The agent starts by considering both rumourmongering and fact talk. At each stage the agent makes a decision on one cognitive aspect. If this aspect is not conclusive it will prolong the deci- sion to the next stage. In the first stage, the agent compares its own competences to the competences that it beliefs to be required for the activity. In our exam- ple model depicted in Figure 2, Bob would decide it cannot do the activity of rumourmongering, because it requires a competence he does not have: noticing juicy details. As such, Bob will engage in fact talk. (Note that if Bob does not have the skill to do either activity, then the decision is also prolonged to the next stage.) In the second stage, an agent tries to make a decision based on its habits. It will survey its context and decide which context elements are near, i.e., resources, places or other agents. If it has a habitual trigger association with a particular strong strength between one of those context elements and either rumourmongering or fact talk it will automatically do that action. In the last stage, the agent will consider how strongly it relates certain values to both ac- tivities and how strongly it adheres itself to these values. Consequently, it makes a comparison between the two activities and decides which best suits its values. 10 A. Fard et al. For the complete implementation of the habitual model and value model we refer to [18]. Fig. 4. The choose-activity submodel and the three stages the agent uses to decide on its activity: competences, habits and values. 5 Experiment The proposed rumour model with elements associated with physical settings, individuals values and competencies enables us to investigate impacts of a vari- ation of settings and interventions on the spread of rumours in organizations. One of the open questions in organizational rumour literature is the effec- tiveness of different prevention and control strategies. In our approach we only need to extend the model with the specific elements and characteristics of the case that we would like to study. In this paper we study the effect of organiza- tional layout on rumour dynamics. In our case, we take the size of offices and number of coffee places as the proxies for organizational layout and juxtapose two organizational layouts cases (Figure 4) to understand the impact of layout on rumourmongering dynamic. To setup the model, we determine the number of agents, then initialize the context and agents. In the organization that we studied each section has on average 50 people, therefore, we pick 50 as the number of the agents. For context initialization, we design the layouts and assign agents to different locations, then we initialize agents with probability distributions for routines such as grab a cup of coffee or having lunch. After the model setup, it can be executed. Rumourmongering in Organizations 11 Fig. 5. (a) In this case, we study the impact of office size on dynamics of rumourmon- gering (b) In this case we study the impact of number of coffee places on the dynamics of rumourmongering 6 Discussion & Future Research Modelling rumourmongering has been studied since 1964. So far, the modelling did not consider the complexities of individual agents, and mostly focused on the spreading behaviour of the phenomenon. In the model proposed in this pa- per, agents have a cognitive layer that deploys social practice theory and views rumour as a routine with associated competencies, values and a physical setting. In this research, we narrowed our study to the context of organization and after introducing the generic model, we tailored our model to the context of or- ganization via empirical data collected though interviews conducted in a Dutch University. Based on explorative interviews we established that social practice theory are likely to be applicable as people shared a view on rumour, and their habits regarding rumour and rumours seem to be intertwined with other activi- ties. 12 A. Fard et al. Our model can be used to study a wide range of topics in organizational rumour studies, in particular for testing the effectiveness of interventions for prevention and control of rumours in organizations. Future work is to extend the questionnaire by asking about associations, conduct more and more rigorous interviews, implement the model and run the proposed experiments that explore different organization layouts. Furthermore, we aim to validate our model by looking at how rumours travel from person to person in the organization during a pre-selected time period. Contributions & Acknowledgements Fard & Mercuur wrote the first draft. Fard provided the domain knowledge and collected most data, whereas Mercuur provided the meta-model and methodolog- ical knowledge. Dignum, Jonker and van der Walle. supervised the process and contributed to the draft by providing comments, feedback and rewriting. This research was supported by the Engineering Social Technologies for a Responsible Digital Future project at TU Delft and ETH Zurich. Rumourmongering in Organizations 13 Fig. 2. An instance of the SoPrA meta-model for the activity of rumourmongering and one agent. For illustration purposes the assocations related to the activity 'talking and the agent 'alice' are omitted. Bibliography [1] Bordia, P., Hobman, E., Jones, E., Gallois, C., Callan, V.J.: Uncer- tainty during organizational change: Types, consequences, and management strategies. Journal of Business and Psychology 18(4), 507 -- 532 (2004) [2] Bordia, P., Jones, E., Gallois, C., Callan, V.J., Difonzo, N.: Management are aliens!: Rumors and stress during organizational change. Group and Organization Management 31(5), 601 -- 621 (10 2006) [3] Bordia, P., Kiazad, K., Restubog, S.L.D., DiFonzo, N., Stenson, N., Tang, R.L.: Rumor as Revenge in the Workplace. Group and Organization Man- agement 39(4), 363 -- 388 (8 2014) [4] Buckner, H.T.: A Theory of Rumor Transmission. Public Opinion Quarterly 29(1), 54 -- 70 (1965) [5] Crawford, S.E.S., Ostrom, E.: A Grammar of Institutions. Political Science 89(3), 582 -- 600 (2007), http://www.jstor.org/stable/2082975 [6] Daley, D.J., Kendall, D.G.: Epidemics and Rumours. Nature 204(4963), 1118 (1964) [7] DiFonzo, N., Bordia, P., Rosnow, R.L.: Reining in rumours. Organizational Dynamics, 2, 4762 (1994) [8] DiFonzo, N., Bordia, P.: A tale of two corporations: Managing uncertainty during organizational change. Human Resource Management 37(3-4), 295 -- 303 (1998) [9] Epstein, J.M.: Agent-based Computational Models and Generative Social Science. Generative Social Science: Studies in Agent-Based Computational Modeling 4(5), 4 -- 46 (1999) [10] Gersick, C.J., Hackman, J.R.: Habitual routines in task-performing groups. Organizational Behavior and Human Decision Processes 47(1), 65 -- 97 (1990) [11] Goldberg, L.R.: The Structure of Phenotypic Personality Traits. American Psychologist 48(1), 26 -- 34 (1993) [12] Hart, B.: The psychology of rumor. Psychiatry: Interpersonal and Biological Processes 28, 1 -- 26 (1916) [13] Kimmel, A.J.: Rumors and rumor control : a manager's guide to under- standing and combatting rumors. Lawrence Erlbaum, Mahwah, NJ : (2004) [14] Knapp, R.H.: A Pyschology of Rumour. Public Opinion Quarterly 8(1), 22 -- 37 (1944) [15] Knopf, T.A.: Beating the Rumors: An Evaluation of Rumor Control Cen- ters. Policy Analysis 1(4), 599 -- 612 (1975) [16] Lally, P., Van Jaarsveld, C.H.M., Potts, H.W.W., Wardle, J.: How are habits formed: Modelling habit formation in the real world. European Journal of Social Psychology 40(6), 998 -- 1009 (2010) [17] Mercuur, R., Dignum, F., Kashima, Y.: Changing Habits Using Contextual- ized Decision Making. In: Advances in Intelligent Systems and Computing. vol. 528 (2017) Rumourmongering in Organizations 15 [18] Mercuur, R.: Interventions on Contextualized Decision Making : an Agent- Based Simulation Study. Ph.D. thesis, Utrecht University (2015), https: //dspace.library.uu.nl/handle/1874/323482 [19] Mercuur, R., Dignum, V., Jonker, C.M.: Modelling Agents Endowed with Social Practices: Static Aspects. (Unpublished) pp. 1 -- 22 (2018) [20] Michelson, G., Mouly, S.: Rumour and gossip in organisations: a conceptual study. Management Decision 38(5), 339 -- 346 (6 2000) [21] Nekovee, M., Moreno, Y., Bianconi, G., Marsili, M.: Theory of rumour spreading in complex social networks. Physica A: Statistical Mechanics and its Applications 374(1), 457 -- 470 (2007) [22] Oh, O., Agrawal, M., Rao, H.R.: Community Intelligence and Social Media Services: A Rumor Theoretic Analysis of Tweets During Social Crises. MIS Quarterly 37(2), 407 -- 426 (2013) [23] Prasad, J.: The Psychology of Rumour: A Study Relating to the Great Indian Earthquake of 1934. British Journal of Psychology. General Section 26(1), 1 -- 15 (1935) [24] Schatzki, T.R.: Social Practices. A Wittgensteinian approach to human ac- tivity and the social. Cambridge University Press (1996) [25] Schwartz, S.H.: An Overview of the Schwartz Theory of Basic Values. Online readings in Psychology and Culture 2, 1 -- 20 (2012) [26] Turenne, N.: The rumour spectrum. PLoS ONE 13(1), e0189080 (1 2018) [27] Vosoughi, S., Roy, D., Aral, S.: The spread of true and false news online. Science 359(6380), 1146 -- 1151 (3 2018) [28] Wang, C., Tan, Z.X., Ye, Y., Wang, L., Cheong, K.H., Xie, N.g.: A rumor spreading model based on information entropy. Scientific Reports 7(1), 9615 (2017) [29] Wood, W., Neal, D.T.: A new look at habits and the habit-goal interface. Psychological review 114(4), 843 -- 863 (2007) [30] World Values Survey Association: WORLD VALUES SURVEY (0), http: //www.worldvaluessurvey.org/wvs.jsp [31] Zanette, D.H.: Dynamics of rumor propagation on small-world networks. Physical Review E - Statistical Physics, Plasmas, Fluids, and Related In- terdisciplinary Topics 65(4), 9 (3 2002)
1812.03557	1	1812	2018-12-09T20:59:01	Distributed Task Management in Cyber-Physical Systems: How to Cooperate under Uncertainty?	[ "cs.MA", "cs.GT" ]	We consider the problem of task allocation in a network of cyber-physical systems (CPSs). The network can have different states, and the tasks are of different types. The task arrival is stochastic and state-dependent. Every CPS is capable of performing each type of task with some specific state-dependent efficiency. The CPSs have to agree on task allocation prior to knowing about the realized network's state and/or the arrived tasks. We model the problem as a multi-state stochastic cooperative game with state uncertainty. We then use the concept of deterministic equivalence and sequential core to solve the problem. We establish the non-emptiness of the strong sequential core in our designed task allocation game and investigate its characteristics including uniqueness and optimality. Moreover, we prove that in the task allocation game, the strong sequential core is equivalent to Walrasian equilibrium under state uncertainty; consequently, it can be implemented by using the Walras' tatonnement process.	cs.MA	cs	Distributed Task Management in Cyber-Physical Systems: How to Cooperate under Uncertainty? Setareh Maghsudi and Mihaela van der Schaar 1 8 1 0 2 c e D 9 ] A M . s c [ 1 v 7 5 5 3 0 . 2 1 8 1 : v i X r a Abstract We consider the problem of task allocation in a network of cyber-physical systems (CPSs). The network can have different states, and the tasks are of different types. The task arrival is stochastic and state-dependent. Every CPS is capable of performing each type of task with some specific state-dependent efficiency. The CPSs have to agree on task allocation prior to knowing about the realized network's state and/or the arrived tasks. We model the problem as a multi-state stochastic cooperative game with state uncertainty. We then use the concept of deterministic equivalence and sequential core to solve the problem. We establish the non-emptiness of the strong sequential core in our designed task allocation game and investigate its characteristics including uniqueness and optimality. Moreover, we prove that in the task allocation game, the strong sequential core is equivalent to Walrasian equilibrium under state uncertainty; consequently, it can be implemented by using the Walras' tatonnement process. Keywords: Cyber-physical systems, distributed task allocation, stochastic cooperative games, uncertainty. I. INTRODUCTION With the emergence of complex systems designed to perform a wide variety of tasks, in recent years there has been a significant growth in research efforts on developing efficient task allocation schemes, which find application in a large body of real-world scenarios. This includes domains like web services, smart grid, internet of things, cyber-physical systems, and cloud computing, among many others. In the following section, we briefly review the state-of-the-art research. A. State-of-the-Art Currently, a body of research works focuses on heuristic task management. For example, community- aware task allocation for networked multi-agent systems is considered in [1] and [2], where each agent can negotiate only with its intra-community member agents. Heuristic distributed algorithms are developed to solve the problem. The aforementioned model stands in contrast with global-aware task allocation, where every agent communicates with all other agents in the network. In [3], a heuristic distributed task allocation method is designed for multi-vehicle multi-task problems, where each machine is able to perform multiple tasks, against the single-vehicle models. The authors also investigate the application of the developed solution in search and rescue scenarios. The authors of [4] investigate task allocation and load-balancing in complex systems. Here the tolerable load of every network entity depends not only on its own resources, but also on the contextual resources that can be acquired via negotiations. In [5], the authors study the task management problem in a grid. They propose two allocation schemes: (i) Task-based algorithms, that greedily allocate tasks to resources, and (ii) Workflow-based algorithms, that search for an efficient allocation for the entire workflow. Moreover, Markovian methods are also applied to model and solve the task allocation problem. As an example, in [6], task allocation for humanmachine collaborative systems is studied. By modeling the human fatigue as a continuous-time Markov decision process, they show that the optimal task assignment can be solved by linear programming. Article [7] investigates reliability-based task proportioning and resource allocation. For solving the formulated problem, the authors suggest an algorithm based on graph theory, Bayesian approach, and the evolutionary optimization approach. Setareh Maghsudi is with the Department of Electrical Engineering and Computer Science, Technical University of Berlin, 10623 Berlin, Germany (e-mail: [email protected]). Mihaela van der Schaar is with the Department of Engineering Science, University of Oxford, Oxford OX1 2JD, United Kingdam. A part of this paper appeared in IEEE Global Communications Conference 2018. 2 Similar to many other resource management and cost-sharing problems, models adapted from economic theory are widely utilized to address the task allocation and job scheduling problems. In [8], the authors investigate repeated task allocation based on prospect theory. They study the effect of the history of task allocation on future agents participation, and the influence of agents' participation on long-term system performance. Location-dependent task allocation for crowd-sourcing is considered in [9]. Therein, the authors formulate task allocation as an orienteering problem. In addition, auction theory (both single-item and bundle) and mechanism design are among popular theoretical tools to address the task allocation problem. For instance, self-adaptive auction is the basis of task-bundle allocation in [10]. Reference [11] uses a mechanism design approach to address a similar problem. Similar to the auction theory, game theory has been utilized to design distributed task allocation methods. Reference [12] discusses a series of game- theoretical models to address resource management problems in IoT networks, including task allocation. Pilloni et al. [13] formulate the distributed task assignment problem as a non-cooperative game, where neighboring nodes engage in negotiations to maximize their own utility scores, resulting in some task allocation. Reference [14] addresses the problem by using a model based on Stackelberg convention game model. Distributed allocation of complex tasks in social networks is also considered in [15]. Two methods, for cooperative and non-cooperative agents, are developed. Task allocation using coalition formation has been investigated by some research works so far. Manisterski et al. [16] propose a centralized task allocation scheme based on cooperation among coalitions of agents, which are built by utilizing minimum- weighted perfect matching. Similarly, in [17], tasks are executed by coalitions of agents. Every agent decides to join a coalition or not through a decentralized auction process. Distributed algorithms for task allocation via coalition formation are studied also in [18]. In this work, every task is assigned to one coalition. In addition to the conventional model in which each agent can belong to one coalition only, the authors also study the case in which an agent might belong to multiple coalitions (overlapping coalition formation). Furthermore, [19] formulates the task allocation problem as a repeated game. The authors develop a distributed mechanism, which uses a mediator to allocate tasks to agents based on a gradient ascent learning algorithm. The goal is to minimize the average turn-around time. In [20], the authors cast the task allocation problem as a cooperative game and apply the Nash bargaining solution to solve the formulated problem. The topic of [21] is resource pricing for wireless grid computing. This paper, models a resource pricing strategy using a non-cooperative bargaining game for resource allocation considering dynamics in the grid market. Reference [22] investigates the load balancing problem in homogenous mini- grids. The agent-based load balancing is regarded as agent distribution, and two quantities are studied: the number and the size of teams, where agents (tasks) queue. Based on a macroscopic modeling, the load balancing mechanism is characterized using differential equations. By solving the equation, the authors show that the load balancing always converges to a steady state. B. Our Contribution In the emerging complex networks, autonomous entities are often required to perform tasks of different types. The tasks to be performed vary over a wide range, including transmission, sensing, measurement, signal processing, computation, and the like. While the type and the load of the jobs arrived at every CPS is in general random, in an inhomogeneous network each CPS might be inclined towards some specific types of tasks, as a consequence of its own characteristics. In other words, non-identical CPSs do not have the same task preferences, or might not be able to perform some randomly-arrived jobs in full due to a lack of resources. Fortunately, in the presence of ubiquitous connectivity, the jobs can be performed cooperatively despite the initial random arrival. In order to become promoted, such cooperation shall be beneficial not only at the individual systems' level, but also from the platform's perspective (i.e., with respect to the social welfare). Coordinating such cooperation is however challenging, since often an agreement shall be achieved prior to job arrivals. Moreover, the characteristics of each CPS might depend on the global network's state and vary over time. In general, the state is also unknown at the time of allocation. In this paper, we investigate the distributed task management problem in a network of cyber-physical 3 systems, where the stochastic task arrival, as well as the systems' characteristics, are contingent on a random state that is a priori unknown. We model the problem as a stochastic cooperative game with state uncertainty. We first eliminate the randomness in the task arrival by using a notion of expectation, referred to as the certainty equivalence. In the resulted deterministic cooperative game, we establish the existence of a special type of core, which is self-enforcing and resistant to the uncertainty in the state. This type of core is called the strong sequential core, and remains stable at every possible state. We investigate its characteristics, including the non-emptiness and optimality. Moreover, we prove that in the developed task allocation game, the strong sequential core is equivalent to the Walrasian equilibrium under uncertainty; i.e., a steady state which is achieved even if the state is a priori unknown. Consequently, it can be implemented by using the Walras' tatonnoment process, also known as the Walrasian auction. To the best of our knowledge, the developed analytical framework appears in the literature for the first time. Our work extends the state-of-the-art in the following aspects: • Our system model is general and consistent with real-world scenarios, as explained in the following: (i) For every CPS (autonomous agent), the task arrival is random. Tasks are inhomogeneous (transmis- sion, computation, measurement, etc.) and impose different demands such as requirement for resource usage (power, memory, etc.); (ii) The CPSs are distinct in types and capabilities (storage capacity, CPU cycle, available radio resources, etc.); (iii) Each CPS has its own preferences in selecting jobs, determined by utility scores; (iv) Tasks are dividable and can be performed by multiple CPSs each; (v) Network's state is a random variable and a priori unknown. In conclusion, the model accommodates most features of complex systems, also those involving human agents. Table I compares the features of our system model with state-of-the-art literature. • The allocation method can be implemented in a distributed manner, which is important in particular in conjunction with the generality of our system model, in particular, the involved uncertainty. As it is shown by some previous works, a distributed implementation imposes less complexity and overhead compared to fully-centralized schemes. • The allocation is efficient at both individual and network levels, despite limited information availability and the presence of randomness/uncertainty. In addition, the solution is stable, in the sense that no individual benefits from not accepting the proposal to join the grand coalition initially or by a unilateral deviation from it after the allocation. Thus all CPSs cooperate in sharing the cost and executing the jobs arrived at the network. Therefore, the approach is applicable also when CPSs belong to independent stakeholders. • The proposed method results in self-enforcing allocation. This means that no CPS benefits by deviating from the agreements, even after the resolution of the uncertainty. Therefore, they are binding also in the absence of a central authority that obligates the CPSs to follow the agreements. • The developed analytical framework can be used in conjunction with a large class of utility functions that satisfy some fairly common conditions. Thus the model and solution can be applied to a wide range of resource allocation problems beyond task management. C. Organization The rest of the paper is organized as follows. Section II presents the system model and basic assumptions. In Section III, we describe an exemplary problem to clarify the application of the considered system model. In Section IV, we formulate the task allocation problem in the presence of random task arrival and state uncertainty. Moreover, we provide a general overview of the proposed solution scheme. We introduce multi-state stochastic cooperative games in Section V. In addition, we describe the concept of certainty equivalence. We model the task allocation problem as a two-stage stochastic game with a priori unknown states and derive its deterministic equivalent. The state uncertainty is addressed in Section VII, where we characterize the strong sequential core of the formulated task allocation game. In Section VIII, we show that the strong sequential core can be implemented using the Walras' tatonnement process. Section IX includes numerical results. Section X summarizes the paper and adds some concluding remarks. COMPARISON OF STATE-OF-THE-ART SYSTEM MODELS TABLE I Systems' Types Task's Type Random Network's State Random Task Arrival Distributed 4 ❵ ❵ ❵ ❵ Reference Feature ❵ ❵ ❵ ❵ [1] [2] [3] [8] [9] [10] [16] [19] [20] [21] Our Work ❵ ❵❵ Arbitrary Arbitrary Arbitrary Arbitrary Arbitrary Arbitrary Arbitrary Arbitrary Arbitrary Identical Arbitrary Arbitrary Arbitrary Arbitrary Identical Identical Identical Arbitrary Arbitrary Arbitrary Identical Arbitrary ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✓ II. SYSTEM MODEL ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✓ ✓ ✗ ✓ ✗ ✓ ✗ ✓ ✗ ✗ ✓ Throughout the paper, by cyber-physical system (CPS), we refer to an autonomous entity consisting of physical and computational elements, which is capable of performing a variety of tasks, possibly at different efficiency levels. This includes measurement, computation, signal processing, and transmission. Consider a network of N CPSs, gathered in a set N . The network can be in one of the S different states, modeled as the outcomes of some random variable with some arbitrary distribution. The state is unknown a priori.1 The set of network's states, S, is finite, mutually exclusive, exhaustive, and known to all CPSs. The CPSs engage in executing some divisible tasks of M different types, collected in a set M. For S states of the network and M tasks, there exist MS state contingent tasks. In simple words, state contingency means that a task m in state s ∈ S is regarded as another task, say m′, when the state changes to s′ ∈ S. It is worth mentioning that the analysis remains valid also if the number of task types is state-dependent; That is, if at every state s ∈ S, the CPSs engage in performing M (s) different tasks. n1 , ..., ρ(s) In such case there exists Ps∈S M (s) state contingent tasks. n =(cid:16)ρ(s) At every state s ∈ S, each CPS n ∈ N is characterized by a performance index (or type) vector (s) nm ∈ R>0 − {∞} of the type vector represents the ability of CPS ρ n ∈ N to perform a task of type m ∈ M at state s ∈ S, in the sense that larger performance index implies higher efficiency. Naturally, ρ(s) nm depends on a variety of factors such as the quality of the available transmission channel, computational capacity, measurement precision, and the like. Note that every CPS knows its own type at every possible state, but does not know which state will be realized in future. In the following we provide an intuitive example to clarify the model. nM(cid:17). Each element ρ(s) At state s ∈ S, every CPS n ∈ N randomly receives q(s) nm unit(s) of task m ∈ M to perform. We have q(s) n =(cid:16)q(s) n1 , ..., q(s) nM(cid:17), and for all m ∈ M, Q(s) m = Xn∈N q(s) nm. (1) m is the total load of task m ∈ M at state s ∈ S in the network. At every state s ∈ S, the That is, Q(s) systems redistribute Q(s) Assumption 1. We assume that each CPS has a positive initial load of each task; that is, q(s) all n ∈ N , m ∈ M, and s ∈ S. m , m ∈ M, among themselves to perform. nm > 0 for 1The model can be explained by considering the network as a player whose strategy set is the set of states. At every round, the player selects one of the available actions randomly, rather than strategically, based on the payoff. In the language of game theory, such player is referred to as the nature. For each CPS n ∈ N and at every state s ∈ S, the initial arrived load of type m ∈ M is a random variable that follows an exponential distribution with parameter λ(s) nm. The CPSs are connected to each other (for example via internet), and the cost of exchanging data between the systems is negligible. Consequently, CPSs are able to cooperate in performing all tasks by redistributing Q(s) m , m ∈ M. The system model is depicted in Fig. 1. 5 Fig. 1. Network of cyber-physical systems at some state s ∈ S. At every state s, each CPS n ∈ N is characterized by a type (efficiency) nM(cid:17). The total load of every task in the network vector ρ follows from (1). The tasks are then redistributed as described throughout the paper and summarized in Fig. 3. nM(cid:17), and receives the initial task vector q n = (cid:16)ρ n = (cid:16)q (s) n1 , ..., ρ (s) n1 , ..., q (s) (s) (s) (s) Definition 1 (Allocation). An allocation of state contingent task ms corresponds to an exhaustive (i.e., without any residue) division of Q(s) where x(s) only if state s ∈ S occurs. nm ∈ R≥0 describes the share of task m ∈ M that should be performed by CPS n ∈ N , if and m between N CPSs. We denote any allocation by x(s) n =(cid:16)x(s) nM(cid:17), n1 , ..., x(s) Note that we investigate one round of task allocation; As a result, throughout the paper, we do not use any notion of time. Naturally, the allocation profile can be used in multiple rounds of task arrival. Upon performing a share x(s) nm of any task m ∈ M at state s ∈ S, a CPS n ∈ N receives a utility denoted by u(s) nm is random as well. We model the network by a multi-agent system, where each CPS is a risk-averse player. The assumption of risk aversion, modeled by a concave utility function, is widely used in multi-agent systems [23], [24]. We capture the risk-aversion by an exponential utility function as2 nm(cid:17). Note that as the task arrival is random, x(s) nm(cid:16)x(s) u(s) nm(cid:0)x(s) nm(cid:1) = ρ(s) − 1 (s) nm ρ nm(cid:18)1 − e x (s) nm(cid:19) . (2) Fig. 2 shows the relation between the utility and the efficiency for a given amount of task. Intuitively, this choice of utility function guarantees that a CPS with higher efficiency makes larger utility by performing a specific share of a given task, compared to a CPS with lower efficiency. Thus, in a Pareto-optimal task allocation, CPSs with better performance index tend to receive more tasks to perform, which improves the efficiency of the platform. Note that the utility involves the cost of performing task implicitly through types or performance indexes. In other words, high performing cost adversely affects the performance index, which reduces the utility. Exponential utility function is also used by some previous works to model the satisfaction level of the agents in multi-agent systems. Examples include [25], [26], [27], [28]. The total utility follows as (3) where x(s) n1 , ..., x(s) n =(cid:16)x(s) u(s) u(s) n (cid:0)x(s) n (cid:1) = Xm∈M nm(cid:0)x(s) nm(cid:1) , nM(cid:17). By (3), the utility is task-separable. 2Note that any cost-sharing game can be modeled similarly, by considering costs as negative utility. Note that other functions can be used to model the utility as long as some regularity conditions are satisfied. 6 0.7 0.6 0.5 0.4 0.3 0.2 0.1 y t i l i t U 0 0 0.1 0.2 0.3 0.5 0.4 0.6 Efficiency Index ( ) 0.7 0.8 0.9 1 Fig. 2. Relation between the utility (u(·)) and one-dimensional efficiency index (type, ρ), for a fixed amount of task x. It can be seen that larger efficiency index results in larger utility. This implies that the CPSs with the utility function (2) attempt to receive a higher share of the tasks in which they are more efficient, thereby contributing to the overall network efficiency. Proposition 1. For every n ∈ N , the utility function defined by (3) satisfies the following: (a) At every state s ∈ S, u(s) n is continuous, concave and monotonically increasing; n is continuous and state-separable. (b) The total utility function un =Ps∈S u(s) Proof: See Appendix XI-D. III. AN EXEMPLARY APPLICATION: ENERGY HARVESTING SMALL CELL NETWORK WITH EDGE COMPUTING Next generation networks are foreseen to be cognitive, meaning that the involved entities are expected to come to a consensus in distributed decision-making processes, where the outcomes affect all of them jointly. In such networks, many applications include some variables that affect (or even exclusively determine) the demands for the tasks to be done by the network and/or the performance efficiency of the involved entities. Although such variables play a major role in optimal decision-making, sometimes they are unknown at the time of negotiations. In such circumstances, we propose to consider every possible scenario with respect to the unknown variables as a state. Therefore, the state is a random variable whose realization is determined simply at random, by a (virtual) player referred to as the nature. Moreover, there are many circumstances in which the utility of each network entity is not deterministic, but rather a random variable. The randomness in the utility function arises as a result of randomness in numerous characteristics of the network. In the following, we provide a high-level example of the proposed system model, in order to demonstrate its general applicability. In Section IX-A, we provide numerical analysis for the described scenario. Consider a heterogeneous wireless small cell network with a set N of small base stations (SBS). Driven by the ever-increasing demand for delay-sensitive computation and cost-effective control, each SBS is foreseen to be an intelligent unit, able to perform low-complexity computation, transmission and decision-making at the network's edge. As mentioned in Section II, here we define a CPS as an engineered entity in which computation, communication and control units are integrated. From this perspective, in this application scenario, we consider each SBS as a CPS. Since the SBSs are able to communicate with each other, we observe a set of them as a network of CPSs. Primarily, each SBS acts as an amplify-and-forward relay to improve the transmission performance. As it is conventional in small cell networks, each SBS has access to a frequency channel that can be used only if it is idle. The status of the frequency channel (busy, idle) is determined by the activity of the primary users, therefore it is a random variable. For every SBS n ∈ N , the availability of transmission channel is modeled by a Bernoulli random variable with parameter (success probability) αn ∈ (0, 1]. Beside transmission, the SBSs act as edge servers to perform some low-load computational tasks, if enough energy is available. Thus, in this exemplary application, two tasks are divided among SBSs, namely transmission over idle channels and computation at the edge. As such, the relevant SBS's characteristics include the availability of idle transmission channels as well as the amount of energy resource. 7 The SBSs are deployed randomly. The random deployment increases the coverage, but it eliminates the access to the power grid. As a result, every SBS uses the power from a built-in battery to perform transmissions, which is its main duty. Moreover, it uses energy harvesting units to possibly obtain some additional energy for computational tasks. Since energy harvesting depends on the weather that is random in nature, the amount of available energy for computation is random. We consider a set S of possible states for the weather. At every state, the energy arrival at every SBS approximately follows a normal distribution with mean µ(s) n > 0 [29]. Moreover, some user devices have built-in (micro) solar cells and are capable of ambient energy harvesting in addition to accessing the battery as a source of energy. The transmission requests are always forwarded to the SBSs, irrespective of the weather. We denote the (initial) transmission requests to every SBS n ∈ N by q(s) n1 , and model it by a Poisson process with rate λn1 > 0 for all s ∈ S. In contrast, the intensity of the offloading demands depends on the weather. In the sunny weather, the devices might use their own harvested energy for computation, thereby reducing the demand for computation offloading; However, if the weather is windy, the devices would prefer to offload the computation to the SBSs. We denote the (initial) computation offloading demands by q(s) n2 , and model it by a Poisson process with rate λ(s) n2 > 0, s ∈ S. Naturally, every SBS or server is reimbursed on the basis of the quality of service (QoS) that it provides to the users. As a result, we model the utility of each SBS with (2), while assuming that ρnm is directly proportional to the resource availability, based on which the QoS is determined. Intuitively, the SBSs with a larger αn can be more successful in addressing the transmission requests, since they are more likely to have access to an idle channel. Therefore, we model the utility of SBS n ∈ N with the utility function given in (2), assuming that ρ(s) n1 = αn for all s ∈ S. Furthermore, it is natural that the utility of computation in every state depends on the energy-availability in that state. Consequently, we model the utility of SBS n ∈ N with the utility function given in (2), assuming that ρ(s) n , for every s ∈ S. n2 = µ(s) IV. THE TASK ALLOCATION PROBLEM Under state uncertainty, the distributed task management can be described in two consecutive stages: (i) At Stage 0, the state of the network is not known. However, all agreements among CPSs with regards to task distribution are made at this stage; (ii) At Stage 1, the state is revealed and the agreements related to the realized state becomes valid. All other agreements are void. The sequence of events is summarized in Algorithm 1. This model corresponds to the two-stage economy model [30]. Algorithm 1 Sequence of Events under State Uncertainty 1: At Stage 0, the network's state is unknown. Commitments are however made in this stage, i.e., under state uncertainty. This means that, at Stage 0, CPSs make commitments (to form coalitions, perform a share of a task, etc.) for every possible state. 2: At Stage 1, the state s ∈ S is revealed. The agreements for the realized state s are executed, and all others become void. n , ..., a(S) Let an =(cid:16)a(1) n (cid:17) indicate a probability distribution over the set of states S, where a(s) is the likelihood of occurrence of state s ∈ S, as predicted by CPS n ∈ N . Naturally, Ps∈S a(s) expected utility then yields (4) n , s ∈ S, n = 1. The vn(cid:16)(cid:2)x(s) n (cid:3)s∈S(cid:17) =Xs∈S a(s) n u(s) n (cid:0)x(s) n (cid:1) , where hx(s) n is∈S is the collection of x(s) n for all s ∈ S. Note that an is not necessarily known by CPSs. In case an is unknown, some probability distribution, for instance the uniform distribution, can be used. Moreover, the CPSs do not need to agree on a specific distribution a.3 In (4), two random variables can be observed: tasks' loads and network's state. 3The effect of private prior on equilibrium is discussed in [31]. 8 In order to formalize the cooperative task allocation, we model the CPS network with an stochastic exchange economy, also under state uncertainty. The model consists in a set of consumers n ∈ N (representing CPSs), a set of divisible commodities m ∈ M (representing tasks), and a set of states S. The exchange economy model captures the idea of exchanging goods, without production, where the allocation of a given amount of each commodity implies its final consumption, associated with some utility score. Since utilities and initial endowments are the main blocks of an exchange economy, we denote the exchange economy model of task allocation as Ω : nhu(s) n is∈S ,hq(s) n is∈Son∈N of exchange economy, all variables are deterministic and all information is provided to all agents a priori; in contrast, in our model, the nature (network) can have different states, modeled as the outcomes of some random variable, unknown a priori. Moreover, parameters such as initial endowments are random variables and their statistical characteristics very over states. . In a conventional model In order to describe the redistribution (exchange) mechanism, we use virtual prices to quantify the value m has to be paid for every of each task based on its popularity over CPSs. In this virtual model, a price p(s) unit of a state contingent task ms, m ∈ M and s ∈ S. We denote the price vector by p = (cid:2)p(s)(cid:3)s∈S, where p(s) = (cid:16)p(s) >0. Given prices, for every n ∈ N , the initial endowments maps to a M(cid:17) ∈ RM 1 , ..., p(s) budget set as (5) B(s) n (cid:0)p(s)(cid:1) =(cid:8)x(s) n : p(s) · x(s) n ≤ p(s) · q(s) n (cid:9) . The CPSs are expected-utility maximizer; that is, at stage 0, every CPS n ∈ N would like to agree on be the set of so as to maximize its expected utility. Let X (s) n its state-dependent load of tasks hx(s) n is∈S all possible demands for CPS n ∈ N . Each CPS n ∈ N solves the following optimization problem: ∈X (s) maximize n is∈S hx (s) s.t. p(s) · x(s) n vn(cid:16)(cid:2)x(s) n (cid:3)s∈S(cid:17) n ≤ p(s) · q(s) n , s ∈ S. (6) Let X be the set of all possible task allocations. From the platform's perspective, the objective is to maximize the aggregate utility performance (social welfare), i.e., maximize n is∈S,n∈N (s) hx s.t. Xn∈N ∈X Xn∈N n ≤ Xn∈N x(s) vn(cid:16)(cid:2)x(s) n (cid:3)s∈S(cid:17) q(s) n , s ∈ S. (7) In the rest of this paper, we model and solve the task allocation problem using stochastic cooperative games under state uncertainty. The procedure is summarized in Fig. 3. Briefly, we first replace the stochastic state game with its deterministic equivalent. Afterward, we use the expected utility and Walrasian auction model to implement the strong sequential core. Details are discussed in the upcoming sections. V. STOCHASTIC COOPERATIVE GAMES IN TWO-STAGE ECONOMIES As mentioned before, in our model, CPSs are connected in a network. Consequently, they are able to share data and thereby to cooperate in performing tasks. Hence we solve the task management problem using a stochastic cooperative game model under state uncertainty, where players form coalitions, engage in performing joint actions, and each receive some utility, all given no prior information about the future state and/or arrived jobs. Let C indicate the set of all possible coalitions among N CPSs with its cardinality being 2N . The state game, i.e., the game in any state s ∈ S, is defined as a tuple G(s) :(cid:16)N ,hw(s) c ic∈C ,hu(s) n in∈N • N is the set of N players, here CPSs; , [>n]n∈N(cid:17), where 9 Fig. 3. A diagram of the proposed solution for the task allocation problem. Briefly, we first replace the stochastic state game with its deterministic equivalent. Afterward, we use the expected utility and Walrasian auction model to implement the strong sequential core. • w(s) c is the characteristics function 2N → RM ≥0, which assigns a value to each of the 2N coalitions c ∈ C at state s ∈ S. This implies that the value of each coalition is state-dependent; nm is the utility function of player n ∈ N by performing task m ∈ M at state s ∈ S; • u(s) • >n is the preference relations of player (CPS) n ∈ N over the set R≥0 of rewards. We define preference relation >n in terms of Von Neumann-Morgenstern preference based on the utility function unm(x) : R>0 → R, where for x, y ∈ R>0, x >n y iff E [unm(x)] > E [unm(y)] , (8) where expectation is taken with respect to the random reward; That is, a stochastic reward is preferred to the other if it yields a larger utility in expectation. This definition of preference corresponds to rationality. In our problem, the initial wealth of each singleton coalition consisting of CPS n ∈ N yields q(s) n , n . After the allocation, i.e., after is stochastic so that the initial wealth of every coalition c ∈ C is given as Pn∈c q(s) c = Pn∈c x(s) Let coalition c include Nc CPSs, collected in a set Nc. At state s, each member n ∈ Nc receives a task n , the value of coalition c ∈ C follows as w(s) determining x(s) with finite first moment. n . Therefore w(s) c vector [32] where x(s) n,c = r(s) n,cw(s) c , r(s) n,c = 1, Xn∈Nc with r(s) n,c ≥ 0. The pay off of member n ∈ Nc then yields n,c + u(s) u(s) where, for utility function (2), at any Pareto-optimal allocation it holds [32], [33] n,cw(s) n,c(cid:0)r(s) c (cid:1) , n,c(cid:0)x(s) n,c(cid:1) = f (s) Xn∈Nc f (s) n,c = 0. (9) (10) (11) (12) n,c is the deterministic part of the payoff. It is in general used to give some coalition members n = 0 for all n ∈ N and s ∈ S. . The term f (s) higher priority, or in other words, larger rewards. In this paper we assume f (s) The term r(s) is stochastic. For coalition c ∈ C, we show the allocation profile as x(s) At state s ∈ S, the set of all possible allocations for coalition c ∈ C is denoted by X (s) s ∈ S, each CPS n ∈ N evaluates its payoff with some utility function, as discussed in Section II. c =hx(s) n,cin∈Nc . At every state n,cw(s) c c A. Certainty (Deterministic) Equivalence 10 As we mentioned in Section II, the optimization problem (4) includes two random variables: the tasks' loads and the network's states. In order to approach the problem, we first eliminate the randomness in x(s) (tasks' loads), n ∈ N . To this end, at every state, the stochastic cooperative game is transformed n to a deterministic cooperative game using the concept of certainty (deterministic) equivalence [32]. In order to transform a stochastic game to a deterministic one, for each agent n ∈ N , we need to specify the deterministic share of the coalition value, in a way that every player is indifferent between receiving the stochastic reward x(s) following definition characterizes the deterministic equivalent of a stochastic game G(s) to guarantee such indifference as well as Pareto-optimality.4 n,c and its deterministic equivalent shown by d(s) n,c(cid:17) = (cid:16)d(s) nM(cid:17). The n (cid:16)x(s) n1 , ..., d(s) Definition 2 (Deterministic Equivalence [32]). Consider the stochastic game G(s) defined by the tuple (cid:16)N ,hw(s) given by G(s) c ic∈C ,hu(s) n in∈N D :(cid:16)N ,hw(s) [>n]n∈N(cid:17). Then the associated cooperative game with deterministic payoff is c,Dic∈C ,hu(s) n,Din∈N w(s) c,D = maximum (s) n,c∈X (s) x , [>n]n∈N(cid:17), where c Xn∈Nc dn(cid:0)x(s) n,c(cid:1) , nm(cid:0)d(s) nm(cid:1) . n,D = Xm∈M u(s) u(s) (13) (14) and n,c ∈ RM >0 : x(s) Moreover, for every n ∈ N , d(s) (a) ∀ x(s) n,c ≈n d(s) (b) ∀ x(s) (c) ∀ k ∈ RM with k being deterministic: d(s) (d) ∀ x(s) (e) ∀ x(s) n (cid:16)x(s) n,c(cid:17)(cid:17) = 0; n,c >n y ⇐⇒ d(s) n (cid:16)x(s) n (cid:16)x(s) n,c, y ∈ RM >0 : x(s) n,c ∈ RM n,c ∈ RM n,c − d(s) >0 : d(s) >0 and deterministic k, k′ ∈ RM , with k < k′ : d(s) n (cid:16)x(s) n,c(cid:17) in (13) satisfies the following conditions: n (cid:16)x(s) n,c(cid:17); n (y); n,c(cid:17) ≥ d(s) n (k) = k; f (s) n,c = 0; Thus, in (13), the maximum is actually taken over the set n (cid:16)x(s) n,c + k(cid:17) < d(s) n (cid:16)x(s) n,c + k′(cid:17). As mentioned before, Pn∈Nc {r ∈ [0, 1]M Pn∈Nc rn = 1}. Proposition 2. Let the utility function of a strictly risk-averse player n ∈ N be given by (2). Then the certainty equivalent of the random reward x(s) nm,c is given by d(s) nm(cid:0)x(s) nm,c(cid:1) = u−1 = −ρ(s) nm,c(cid:0)x(s) −1 (s) nm nm,c(cid:0)E(cid:2)u(s) nm ln(cid:18)E(cid:20)e (cid:16)x nm,c(cid:1)(cid:3)(cid:1) nm,c(cid:17)(cid:21)(cid:19) . (s) ρ (15) Proof: See Appendix XI-E. In the rest of the paper, we will analyze the stochastic two-stage game based on the notion of deterministic equivalence. VI. ANALYSIS OF THE TWO STAGE GAME In order to analyze the multi-state two-stage game, we start from its basis, which is the stochastic game at every state. For a single-state cooperative game with deterministic rewards, a well-known solution concept is the core [34]. We state the formal definition of the core below. 4Note that the deterministic equivalent is developed independently for every task m ∈ M at every state s ∈ S. 11 The core, denoted by C, is a set of feasible payoff allocations for the grand coalition [¯xn]n∈N satisfying the following conditions: Definition 3 (Core). Consider a (deterministic) cooperative game G : (cid:0)N , [wc]c∈C , [un]n∈N , [≥n]n∈N(cid:1). (a) Efficiency: Pn∈N ¯xn,c = wc; (b) Coalitional rationality: There is no coalition c ⊂ N and allocation [yn]n∈N which is preferred by every n ∈ c over the grand coalition and allocation [¯xn]n∈N ; formally, there is no [yn]n∈N so that un(¯xn) ≤ un(yn) ∀ n ∈ c and ∃ n ∈ c such that un(¯xn) < un(yn). In words, in a core solution of a deterministic cooperative game, all the wealth is allocated in a way that there is no coalition of players in which all members benefit by deviating from the grand coalition. As discussed before, in our setting, the state games are stochastic. In order to find the core of the stochastic state game, we use the following theorem. The theorem states the relation between the core of a stochastic game and that of its deterministic equivalent. Theorem 1 ([32]). The core of stochastic game G(s) is identical to the core of its deterministic equivalent G(s) D . By Theorem 1, the problem of finding the core of the stochastic game at every state boils down to finding the core of its certainty equivalent. Consequently, when analyzing the two-stage game, we can replace the stochastic state game with its deterministic equivalent. As mentioned before, at every state s ∈ S, the random initial value of each singleton coalition including q(s) n , CPS n ∈ N equals q(s) so that the random value of the grand coalition yields Q(s). Thus problem (6) can be transformed as n . Moreover, the value of every coalition c with Nc members is given asPn∈Nc where and vn,D(cid:16)(cid:2)r(s) n Q(s)(cid:3)s∈S(cid:17) =Xs∈S a(s)u(s) n (cid:16)d(s) n Q(s)(cid:3)s∈S(cid:17) , Moreover, problem (7) is converted to After reformatting the game by using the concept of certainty equivalence, we face a multi-state deterministic cooperative game under state uncertainty. For such game, the conventional core concept (Definition 3) does not suffice, and new solutions are developed in the literature. This includes two-stage core, strong sequential core, and weak sequential core. For a comparative study see [35] and [36]. In this paper we focus on the strong sequential core (SSC), defined below. Note that, since we eliminated the randomness by using deterministic equivalent, in the rest of the paper, x(s) is a deterministic variable n which simply maps to r(s) n , as x(s) n = r(s) n Q(s). maximize n is∈S (s) ∈[0,1]M hr vn,D(cid:16)(cid:2)r(s) n Q(s)(cid:3)s∈S(cid:17) s.t. p(s) · r(s) n ≤ p(s) · , s ∈ S, q(s) n,D Q(s) D (s) n (cid:2)r(s) n (cid:17)(cid:21)(cid:19) . n Q(s)(cid:3)s∈S(cid:17) q(s) n,D = −ρ (s) n ln(cid:18)E(cid:20)e −1 (s) n ρ (cid:16)q maximize hr (s) n is∈S,n∈N ∈[0,1]NM Xn∈N vn,D(cid:16)(cid:2)r(s) r(s) n = 1, s ∈ S. s.t. Xn∈N (16) (17) (18) (19) Definition 4 (Strong Sequential Core [35]). The strong sequential core of the deterministic game G is for the grand coalition (N ), for which the following 12 n,Nis∈S,n∈N holds: the set of all feasible allocations ¯x =h¯x(s) (a) ¯x(s) ∈ C(cid:0)G(s)(cid:1) for all s ∈ S; n,cis∈S(cid:17) > vn,c(cid:16)h¯x(s) vn,c(cid:16)hx(s) n,Nis∈S(cid:17) for all n ∈ c. (b) There is no coalition c ⊆ N and allocation xc such that w(s) for all s ∈ S, and c = Pn∈c q(s) c In words, for an allocation to belong to the strong sequential core, the following must hold: (a) The allocation should belong to the core of the state game in every state s ∈ S; That is, it must be stable against all the deviations ex-post (after revealing the uncertainty). (b) No coalition of agents (including the grand coalition) shall be able to improve upon ¯x ex-ante (before revealing the uncertainty). A coalition can improve upon an allocation ¯x with allocation x, if x is feasible and gives each CPS of the deviating coalition a higher expected utility.5 Before revealing the uncertainty in the state of the network, every agent joins/leaves coalitions with the goal of maximizing its expected utility. Nonetheless, a coalition which has no incentive to block a specific allocation ex-ante, may object against it ex-post, i.e., after the uncertainty is resolved. Mostly, it is assumed that the agents remain committed to the agreements also after the resolution of uncertainty, but clearly such assumption is restrictive and in general difficult to implement. The strong sequential core, however also guarantees ex-post commitment, which makes it desirable as a solution concept. In the next section we describe a method to implement the strong sequential core, and we establish some of its characteristics, including the non-emptiness, in our designed task allocation game. VII. WALRASIAN EQUILIBRIUM UNDER UNCERTAINTY AND STRONG SEQUENTIAL CORE Till now, we showed that some stochastic cooperative games can be transformed into deterministic games using the notion of certainty equivalence (Definition 2). If the game is played under uncertainty, i.e., if the nature randomly takes one of some a priori unknown states, the deterministic equivalent should be calculated separately for every possible state. The resulted game is then a multi-state deterministic cooperative game, but with state uncertainty. As discussed before, a general solution concept for such games is the strong sequential core (Definition 4). However, similar to the conventional core concept developed for games without uncertainty, it is necessary to characterize such equilibrium, for instance to establish its non-emptiness in the game under investigation. Upon existence, it should be discussed that how such equilibrium can be implemented. For deterministic cooperative games without uncertainty, it is known that Walrasian equilibrium lies in the core of the corresponding cooperative game [37]. In other words, with implementing a Walrasian equilibrium, one achieves a core solution. In what follows, we first provide the definition of Walrasian equilibrium under state uncertainty, and then describe its characteristics. Afterward, we establish that under few reasonable assumptions, Walrasian equilibrium under uncertainty belongs to the strong sequential core of the corresponding deterministic cooperative game with state uncertainty. Moreover, since we have used certainty equivalence, it is also a core solution for the initial stochastic cooperative game under uncertainty. We also describe how to implement this solution. A. Exchange Economy and Walrasian Equilibrium under Uncertainty In an exchange economy under uncertainty, the equilibrium notion is the Walrasian equilibrium under uncertainty, also called Arrow-Debreu equilibrium, defined in the following. Definition 5 (Walrasian Equilibrium under Uncertainty). A set of allocation matrices ¯x(s), together with a price vector p(s), s ∈ S, are Walrasian equilibrium under uncertainty if 5The definitions of feasible allocation and deviation are included in the appendix (Definition 6 and Definition 7, respectively), in order to maintain the consistency and readability. 13 (20) q(s) n . Xn∈N ¯x(s) n = Xn∈N (a) ∀ n ∈ N , h¯x(s) n is∈S (b) Market clears, that is, for all s ∈ S maximizes vn(·) on Bn; In the following, we investigate the existence and characteristics of Walrasian equilibrium under uncer- tainty, in our designed task allocation economy. Proposition 3. In the deterministic equivalent of our designed multi-state stochastic task exchange econ- omy, Walrasian equilibrium under state uncertainty exists. Moreover, it is unique, Pareto-efficient, and thus social-optimal. Proof: See Appendix XI-F. B. Relation to The Strong Sequential Core It is obvious that in an exchange economy with no uncertainty, the definition of Walrasian equilibrium collapses to that of Walrasian equilibrium under uncertainty when only one state exists. The following theorem declares that in such economy, the Walrasian equilibrium lies inside the core. Theorem 2 ([38]). Let allocation ¯x together with price vector p be a Walrasian equilibrium for an ex- change economy Ω : {un, qn}n∈N (with no state uncertainty), where ¯x = [¯xn]n∈N and ¯xn = (¯xn1, ..., ¯xnm). If each ¯xnm, n ∈ N and m ∈ M, is locally non-satiated (LNS), then ¯x lies in the core of Ω. The definition of locally non-satiated can be found in Appendix XI-B, Definition 11. Informally, non- satiation means that greater quantities provide higher levels of satisfaction to individuals. In should be mentioned that non-satiation is implied by the monotonicity, whereas the reverse does not hold. The following theorem states the conditions under which any Walrasian equilibrium with state uncertainty lies in the strong sequential core. Theorem 3. In our designed multi-state stochastic task exchange economy, the Walrasian equilibrium (¯x, p) lies in strong sequential core. Proof: See Appendix XI-G. VIII. IMPLEMENTING THE STRONG SEQUENTIAL CORE: WALRAS' TATONNEMENT PROCESS In the previous section, we showed that Walrasian equilibrium under uncertainty lies in strong sequential core. Thus, in order to implement a core solution, it suffices to implement a Walrasian equilibrium. In a competitive market, every self-interested agent selects its demand so as to maximize its own utility score. Such selfish behavior yields a conflict which degrades the network's performance. Consequently, a mechanism should be used to guide the agents to a stable and efficient operating point. One such mechanism is the Walras' tatonnement process, also called the Walrasian auction. The process requires a coordinator (auctioneer), which, at each round, announces the prices, starting at some random initial point. Afterward, agents disclose their demands at the given prices, so that the auctioneer adjusts prices to claimed demands. The process continues until the market clears; that is, when a set of prices yields a demand equal to supply. At this point, prices and demands are final, and the auction process terminates, i.e., trade occurs [39]. Let z (p) be the excess demand given price vector p, given by The price adjustment rule is given by [40] z(s) m (p) = Xn∈N x(s) nm(p, p.qn) −Xn∈N q(s) nm. p(s)(t + 1) = p(s)(t) + αz(cid:0)p(s)(t)(cid:1) , (21) (22) for a sufficiently small α > 0. Note that the variable t is a local variable that counts the number of iterations of the auction process; i.e., the number of negotiation rounds until the systems achieve some agreement on the task allocation. Clearly, the only stationary points of this process are prices p at which z(p) = 0, i.e., equilibrium prices [40]. The procedure is summarized in Algorithm 2. 14 Algorithm 2 Walrasian Auction 1: Select the price adjustment factor α → 0. 2: Initialize the price of each commodity, p 3: repeat 4: • (s) m → 0, m ∈ M and s ∈ S. The auctioneer announces the prices. Each CPS declares its demand. The auctioneer observes excess demands, and adjusts the prices using (22). • • 5: until Market clears. 6: Trades occur (Agreements are final). Proposition 4. Let h¯x(s) Walras' tatonnement process with price adjustment rule (22) converges to h¯x(s) n , p(s)is∈S Proof: See Section XI-H. be a Walrasian equilibrium under uncertainty. Then, in our setting, the n , p(s)is∈S as t → ∞. The process described in Algorithm 2 requires a coordinator; however, in the absence of such a coordinator, the auction process can be implemented in a fully distributed manner. In doing so, the initial price and the price adjustment factor is known to all systems. Then every system (i) announces its demands, (ii) updates the excess demand based on the announced demands of all other bidders as well as its own, and (iii) updates the prices based on the excess demand. The process continues until convergence. Note that in engineering applications, for example in our system model when CPSs communicate over the wireless medium, the auction process does not cause heavy overhead, due to the following reason: Since the demand is not considered as private information, it can be simply announced through a control channel. Due to the broadcast nature of wireless medium, the signals are heard by all systems; in other words, no pairwise communication is required. After the convergence of the process, the channel becomes free. Usually, the convergence is fast, implying that the channel is not occupied for a long time. As conventional, we assume the existence of a noiseless control channel, implying that the signals arrive error-free at CPSs and/or the auctioneer. Such an assumption is justified by the existence of coding schemes that reduce the transmission error probability to almost zero [41]; therefore, it is widely used in game theory and networks literature (see [23] and [42], among many others). While the analysis of the problem with noisy communication channel is beyond the scope of our work, there are some research works that study such setting. For example, in [43], the authors assume that the channel from bidders to the auctioneer is noisy, whereas the reverse channel is noise-free. In this setting, they investigate a spectrum auction problem. A. Complexity and Convergence Speed For divisible goods, the convergence of the tatonnement process given by (22) is guaranteed only asymptotically, similar to many other tatonnement procedures. Thus, most often the auction is implemented to stop when the excess demand is only almost-zero, resulting in an approximate equilibrium. Moreover, the convergence speed of the Walrasian auction depends mainly on the price adjustment factor α as well as agents' utility functions, and thus cannot be determined rigorously. It is however important to note that α being too small slows down the convergence dramatically, whereas selecting α too large prevents the process from convergence. On the agents' side, the problem of calculating demands given the prices (budget) is a non-linear contentious knapsack problem with the number of inputs being the number of tasks M at every state. 15 For such problem polynomial-time algorithm exists [44]. On the auctioneer side, calculating the excess demand and adjusting the price is linear in the number of CPSs N . IX. NUMERICAL RESULTS The numerical evaluation consists of two parts. In the first part, we implement the application scenario discussed in Section III. We provide some numerical results to describe the applicability of the model and solution in an intuitive manner. In the second part, we consider a general example to demonstrate the theoretical results also by numerical analysis. The evolution of price and demand, the convergence rate, the incentive to form the grand coalition and other theoretical issues are discussed as well. A. A Toy Example Recall the application model we described in Section III. There exist two SBSs in the network, i.e., N = 2. Moreover, let α1 = 0.9 and α2 = 0.7. In addition, there are two possible states with respect to the weather, namely S = {Sunny, W indy}; In other words, the weather can be either sunny or windy. For the simplicity of notation, we label the sunny and the windy weather as State 1 and State 2, respectively. SBS 1 is provided with a number of solar cells as energy harvesting units. SBS 2 has a lower number of solar cells compared to SBS 1; instead, it has access to few wind-turbines. While a sunny weather is ideal to utilize solar cells for energy harvesting, it is not productive for wind-turbines. Consequently, we select ρ(1) 12 = 0.9, ρ(2) nm = 1 for n, m, s ∈ {1, 2}. The probability distributions assigned by the SBSs to the weather are selected randomly as a1 = (0.20, 0.80) and a2 = (0.40, 0.60). 22 = 0.6. For simplicity, we let λ(s) 12 = 0.1, ρ(1) 22 = 0.4, ρ(2) In Fig. 4, we depict the task allocation resulted by the proposed method. It can be seen that in both states, a large fraction of transmission is allocated to SBS 1. Intuitively, such allocation is justified since SBS 1 is more likely to have a free channel to perform the relaying compared to SBS 2. Moreover, in State 1, i.e., if the weather is sunny, SBS 1 receives the larger part of the computation task as it has a larger number of solar cells, thereby more energy. In State 2, however, SBS 1 can afford only a low energy resource to perform the computation, due to the lack of sun; as a result, a part of the task is transferred to SBS 2, which can harvest the energy also from the wind. Therefore, in State 2, SBS 2 performs a large part of the computation task. Thus, the allocation of the computation tasks is along with the interests of SBSs, although it is performed before the weather is known. Hence, the SBSs are unlikely to deviate from it when the state is observed at a later point. It should be noted that, unlike the allocation profile of the computation task, the allocation profile of the transmission task is affected by the weather condition only slightly. This is due to the fact that, as described in Section III, the required transmission power is supplied by the battery. The slight changes assure appropriate load-balancing and fairness in utility. 0.8 0.6 0.4 0.2 0 0.8 0.6 0.4 0.2 0 Task 1, State 1 SBS 1 SBS 2 Task 1, State 2 SBS 1 SBS 2 0.8 0.6 0.4 0.2 0 1 0.5 0 Task 2, State 1 SBS 1 SBS 2 Task 2, State 2 SBS 1 SBS 2 Fig. 4. Allocation of the computation and transmission tasks in State 1 and State 2. 16 B. A General Example We consider a CPS network consisting of four CPSs (N = 4), three task types (M = 3), and three states (S = 3). The CPSs' performance indexes for carrying out tasks in different states are summarized in Table II. The values are generated simply at random. For every CPS n ∈ N , the initial arrival of any task m ∈ M at every state s ∈ S is a random variable following an exponential distribution with parameter (rate) λ(s) nm = n. The probability distributions assigned by CPSs to states are selected randomly from the 3-dimensional probability space, and follow as a1 = (0.10, 0.30, 0.60), a2 = (0.20, 0.50, 0.30), a3 = (0.34, 0.33, 0.33), and a4 = (0.90, 0.05, 0.05). The price adjustment factor is selected as α = 0.01. PERFORMANCE INDEX OF EVERY CPS n ∈ N FOR EVERY TASK m ∈ M AT EVERY STATE s ∈ S TABLE II (a) State 1 (b) State 2 Task CPS 1 CPS 2 CPS 3 CPS 4 0.42 0.78 0.71 0.72 0.22 0.65 0.80 0.21 0.26 0.34 0.68 0.19 1 2 3 Task CPS 1 CPS 2 CPS 3 CPS 4 0.90 0.61 0.62 0.74 0.50 0.48 0.70 0.19 0.20 0.90 0.89 0.33 1 2 3 (c) State 3 Task CPS 1 CPS 2 CPS 3 CPS 4 0.98 0.71 0.50 0.21 0.24 0.89 0.19 0.49 0.21 0.86 0.81 0.88 1 2 3 First we investigate the effect of using deterministic equivalence in the analysis. We initially consider stochastic task arrival, where CPSs apply the proposed approach to agree on task allocation. In Fig. 5, we show the ex-ante and ex-post utility of each CPS n ∈ N , normalized by the aggregate utility, i.e., . In essence, this is the utility achieved by the task allocation resulted from using the concept of un Pn∈N un deterministic equivalence (i.e., dictated byhr(s) n in∈N ,s∈S ). In another experiment, we assume that the tasks arrived at every CPS are deterministic and known. We perform only the conventional auction process, and show the utility in Fig. 5, as described before. From this figure, it can be concluded that by using the deterministic equivalence, the effect of stochastic task arrival is almost eliminated; that is, the users are indifferent between receiving the deterministic equivalent payoff or the stochastic payoff, as expected. Moreover, in Fig. 6, we illustrate the relation between the performance index and task allocation. As expected, a CPS with high performance index for some task is more likely to receive larger share of that task in the final allocation, compared to a CPS with a low performance index for that task. This results in higher utility for CPSs, and also improves the overall performance of the platform. y t i l i t U y t i l i t U y t i l i t U y t i l i t U 0.5 0 0.4 0.2 0 0.4 0.2 0 0.4 0.2 0 CPS 1 Deterministic Equivalence Deterministic ex-post, S 1 ex-post, S 2 ex-post, S 3 ex-ante CPS 2 ex-post, S 1 ex-post, S 2 ex-post, S 3 ex-ante CPS 3 ex-post, S 1 ex-post, S 2 ex-post, S 3 ex-ante CPS 4 ex-post, S 1 ex-post, S 2 ex-post, S 3 ex-ante Fig. 5. The utility resulted from deterministic equivalence allocation for the stochastic problem, compared to the deterministic game. 17 1 0.5 0 1 0.5 0 1 0.5 0 State 1 Normalized Perfromance Index Task Share T1C1 T1C2 T1C3 T1C4 T2C1 T2C2 T2C3 T2C4 T3C1 T3C2 T3C3 T3C4 Task-CPS State 2 T1C1 T1C2 T1C3 T1C4 T2C1 T2C2 T2C3 T2C4 T3C1 T3C2 T3C3 T3C4 Task-CPS State 3 T1C1 T1C2 T1C3 T1C4 T2C1 T2C2 T2C3 T2C4 T3C1 T3C2 T3C3 T3C4 Task-CPS Fig. 6. The relation between efficiency (relative performance index) and allocated task. agreement process ends when the market clears for all state contingent tasks, i.e., Pn∈N r(s) Fig. 7 shows the relation between the excess demand and the price of two exemplary state contingent tasks. It can be seen that, as expected, the excess demand reduces as the price increases. The entire n = 1 for all s ∈ S. The number of iterations required for the market clearing for every state contingent task is shown in Fig. 8. Note that as described in Section VIII-A, we implement the tatonnement process to stop when the excess demand is very small (here 0.01). From the figure, the entire allocation process converges in 60 iterations. Note that as we discussed in Section VIII-A, the convergence speed cannot be determined rigorously, and may vary based on many factors such as the price adjustment coefficient, random initial task loads, performance indexes, etc. 0.0103 e c i r P 0.0102 0.0101 0.01 0 0.05 0 -0.05 d n a m e D s s e c x E -0.1 0 Task 2, State 1 2 4 6 8 Iteration 2 4 6 8 Iteration 0.015 0.0145 0.014 e c i r P 0.0135 0 0.4 0.2 0 d n a m e D s s e c x E -0.2 0 Task 1, State 2 5 Iteration 10 5 Iteration 10 Fig. 7. The changes in price and excess demand for two exemplary state-contingent tasks. s n o i t a r e t I f o r e b m u N 60 50 40 30 20 10 0 State 1 State 2 State 3 Fig. 8. Required number of iterations to converge (market clearing). Next, we show that the task allocation belongs to the strong sequential core; that is, joining the grand coalition is beneficial for every CPS n ∈ N ex-ante, and the ex-ante allocation lies in the core of the ex-post games. Since the efficiency prerequisite is satisfied by the market clearing condition (every task is allocated in full), we only need to check the coalitional deviations prerequisites. This means that there should be no coalition in which at least one member is better off leaving the grand coalition while all other members remain indifferent. 18 To perform the experiment, we first simulate some task arrival according to the described exponential distribution. In Figs. 9(a)-9(d), we illustrate the maximum ex-ante and ex-post achievable reward of every CPS at every possible coalition, including the grand coalition. For each CPS, all values are normalized n in∈N ,s∈S is agreed upon before the task arrival. The values which do not appear in the diagrams by its utility resulted from our proposed task allocation scheme, i.e., as dictated by hr(s) that hr(s) are less than 95% of the reward of grand coalition, and therefore cannot be observed. Fig. 9(a) shows the normalized expected utility of every CPS ex-ante. It can be observed that for the grand coalition, the ratio is always equal to one. This means that our allocation scheme achieves the best performance, despite being performed prior to the actual job arrival. Moreover, there is no coalitional deviation (including singleton coalitions) in which at least one member benefits, while other members remain indifferent. Thus the allocation is stable. In Figs. 9(b), 9(c), and 9(d), we assume that state 1, state 2, and state 3 are realized, respectively. The normalized utility for the ex-post games are then depicted. The figures show that the grand coalition is also ex-post stable, i.e., no coalition of CPSs benefits from a deviation even after the realized state is revealed. Together with the efficiency condition, we can conclude that the allocation n in∈N ,s∈S . Note cannot be improved upon by any other allocation ex-ante and also ex-post. dictated by hr(s) n in∈N ,s∈S Therefore it lies in the strong sequential core. Fig. 9. Ex-ante and ex-post coalitional rationality. y i l i t U y i l i t U y i l i t U y i l i t U 1.05 1 0.95 1.05 1 0.95 1.05 1 0.95 1.05 1 0.95 y i l i t U y i l i t U y i l i t U y i l i t U 1.05 1 0.95 1.05 1 0.95 1.05 1 0.95 1.05 1 0.95 CPS 1 C1 C1C2 C1C3 C1C4 C1C2C3 C1C2C4 C1C3C4 Grand CPS 2 C2 C2C1 C2C3 C2C4 C2C1C3 C2C1C4 C2C3C4 Grand CPS 3 C3 C3C1 C3C2 C3C4 C3C2C1 C3C1C4 C3C2C4 Grand CPS 4 C4 C4C1 C4C2 C4C3 C4C2C1 C4C1C3 C4C2C3 Grand (a) Ex-Ante CPS 1 C1 C1C2 C1C3 C1C4 C1C2C3 C1C2C4 C1C3C4 Grand CPS 2 C2 C2C1 C2C3 C2C4 C2C1C3 C2C1C4 C2C3C4 Grand CPS 3 C3 C3C1 C3C2 C3C4 C3C2C1 C3C1C4 C3C2C4 Grand CPS 4 C4 C4C1 C4C2 C4C3 C4C2C1 C4C1C3 C4C2C3 Grand y i l i t U y i l i t U y i l i t U y i l i t U 1.05 1 0.95 1.05 1 0.95 1.05 1 0.95 1.05 1 0.95 y i l i t U y i l i t U y i l i t U y i l i t U 1.05 1 0.95 1.05 1 0.95 1.05 1 0.95 1.05 1 0.95 CPS 1 C1 C1C2 C1C3 C1C4 C1C2C3 C1C2C4 C1C3C4 Grand CPS 2 C2 C2C1 C2C3 C2C4 C2C1C3 C2C1C4 C2C3C4 Grand CPS 3 C3 C3C1 C3C2 C3C4 C3C2C1 C3C1C4 C3C2C4 Grand CPS 4 C4 C4C1 C4C2 C4C3 C4C2C1 C4C1C3 C4C2C3 Grand (b) Ex-post, State 1 CPS 1 C1 C1C2 C1C3 C1C4 C1C2C3 C1C2C4 C1C3C4 Grand CPS 2 C2 C2C1 C2C3 C2C4 C2C1C3 C2C1C4 C2C3C4 Grand CPS 3 C3 C3C1 C3C2 C3C4 C3C2C1 C3C1C4 C3C2C4 Grand CPS 4 C4 C4C1 C4C2 C4C3 C4C2C1 C4C1C3 C4C2C3 Grand (c) Ex-post, State 2 (d) Ex-post, State 3 Finally, we evaluate the performance of our approach in terms of the expected social welfare (expected aggregate utility), by comparing it with the following allocation methods: • Weighted Matching: We construct a bipartite graph consisting of CPSs on the one side and tasks on the other side. The performance indexes (types) of CPSs are used as the edges' weights. By using a bipartite maximum matching algorithm [29], each task is allocated to only one CPS, in a way that the sum of the weights of assignment edges is maximized. • Random Allocation: The tasks are divided between the CPSs randomly; • Equal Allocation: Each task is equally divided between the CPSs. The results are depicted Fig. 10, which shows the superior performance of our proposed method. Naturally, the achieved gain is not fixed and might vary depending on many parameters such as CPSs' types. 19 i ) d e z a m r o N l ( y t i l i t U e t g a e r g g A d e t c e p x E Fig. 10. Performance Comparison. 1.05 1 0.95 0.9 0.85 0.8 0.75 0.7 0.65 0.6 Proposed Equal-Disvision Random Matching X. CONCLUSION We considered multi-state stochastic cooperative games, where (i) the reward of every coalition is a state-dependent random variable and (ii) the nature's state is random. The players, characterized by state-dependent utility functions, have to divide the stochastic reward, before the state is realized. We applied the strong sequential core as a desirable solution concept. We then solved the problem by using the concept of deterministic equivalence and expected utility. For an exemplary application, namely task allocation in cyber-physical systems, we established the non-emptiness of core and characterized it with respect to uniqueness and optimality. We proved that in our setting, the Walrasian equilibrium under state uncertainty lies in the strong sequential core, and thus the core can be implemented by using the Walras' auction. Numerical results confirm the theoretical analysis and showed the applicability of the proposed model and solution. XI. APPENDIX Before proceeding to proofs, we state some auxiliary definitions and results, which were left out from the main text in order to maintain the readability and consistency. A. Technical Preliminaries for Cooperative Games Definition 6 (Feasible Allocation). Let some allocation ¯x be given. The allocation xc =hx(s) feasible for a coalition c ∈ C at state s ∈ S if x is the value of the coalition at state s ∈ S. In words: (i) The members of the coalition take allocations outside s ∈ S as given; (ii) In every state s ∈ S, the members of a coalition can redistribute at most their wealth. For an allocation to be feasible at stage 0 (before knowing the state), the second condition must hold for every state s ∈ S. and Pn∈c x(s) n,cis∈S,n∈N −(s) −(s) c = ¯x c n,c ≤ w(s) c , where w(s) c is Definition 7 (Deviation). Let some allocation ¯x be given. A coalition c ∈ C can deviate from ¯x at state s ∈ S if there exists a feasible allocation x(s) for c at state s such that un,c (xn,c) > un,c (¯xn,c) for all n ∈ C. Such allocations are called deviations, which are not necessarily self-enforcing. c B. Technical Preliminaries for Exchange Economy Definition 8 (Feasible Exchange). In an exchange economy Ω, an allocation ¯x is a feasible exchange if Pn∈N ¯xn ≤Pn∈N qn. Definition 9 (Pareto-Efficient Exchange). In an exchange economy Ω, a feasible exchange ¯x is Pareto- efficient if there is no other feasible exchange x such that un(xn) ≥ un(¯xn) for all n ∈ N with strict inequality for some n ∈ N . Definition 10 (Social-Optimal Exchange). A feasible exchange profile ¯x(s), s ∈ S, is social-optimal if it maximizes the aggregate agents' utility (social welfare). Formally, any feasible solution to the following optimization problem is a social-optimal exchange. 20 maximize x(s)∈X (s) Xn∈N un(cid:16)(cid:2)x(s)(cid:3)s∈S(cid:17) . (23) Definition 11 (Local Non-Satiated). If X is a consumption set, then for any x ∈ X and every ǫ > 0, there exists a y ∈ X such that ky − xk ≤ ǫ and y is preferred to x. Definition 12 (Gross Substitutes [40]). A demand function xn satisfies the gross substitutes property if, for any two price vectors p and p′ such that p′ l ≥ pl for all l 6= k, then xl(p′) > xl(p) for all l 6= k. k ≥ pk and p′ Lemma 1 ([37]). Additive concave and separable additive utility functions satisfy the gross substitutes property. Lemma 2 ([40]). If each individual has a utility function satisfying the gross substitutes property, then both the individual and aggregate excess demand functions satisfy the gross substitutes property as well. C. Technical Preliminaries for Walrasian Equilibrium In this section we describe some results regarding the existence and optimality of Walrasian (Arrow- Debreu) equilibrium. Note that the following theorems were originally established for competitive markets without uncertainty (single-state), but were later shown in [30] to also hold for competitive market with uncertainty. Theorem 4 (Existence of Walrasian Equilibrium [31]). Walrasian equilibriumhx(s) exchange economy with divisible goods if (i) for every agent, the utility function is continuous, increasing, concave; and (ii) q(s) n , p(s)is∈S n > 0 for all s ∈ S and n ∈ N . exists for an Theorem 5 (Efficiency of Walrasian Equilibrium [37]). Let hx(s) n , p(s)is∈S Then it is Pareto-efficient for an exchange economy with divisible goods if the utility function is increasing. be a Walrasian equilibrium. Theorem 6 (Uniqueness of Walrasian Equilibrium [40]). For an exchange economy Ω with divisible goods, if the aggregate excess demand function z(·) satisfies the gross substitutes property, then the economy has at most one Walrasian equilibrium, i.e., z (p) = 0 has at most one (normalized) solution. Theorem 7 (Convergence of Walras' Tatonnement [40]). Consider an exchange economy Ω and suppose the aggregate excess demand function z(·) satisfies the gross substitutes property. Then the tatonnement that (cid:2)p(s)(cid:3)s∈S is a Walrasian equilibrium price vector. Suppose that the commodities are divisible and process with price adjustment rule (22) converges to the relative prices of (cid:2)p(s)(cid:3)s∈S as t → ∞ for any initial condition p(t = 0). D. Proof of Proposition 1 Recall that by (2), we have u(s) nm(cid:16)x(s) nm(cid:17) = ρ(s) nm(cid:18)1 − e − 1 (s) nm ρ x (s) nm(cid:19). Therefore, • The continuity of the function (3) follows by the continuity of the exponential function e for all m ∈ M. The continuity of the exponential function is a standard result proved by a straightforward application of the definition of continuity. Since the result exists in standard mathematics text books, we do not state it here to save the space. − 1 (s) nm ρ (s) nm x • The first derivation of (3) yields u′(s) implying that the function is monotone increasing. • The second derivative is given by u′′(s) that the function is concave. nm(cid:16)x(s) nm(cid:16)x(s) nm(cid:17) = Pm∈M e nm(cid:17) =Pm∈M −1 (s) ρ nm The second part of the proposition simply follows by the definition of un (xn). 21 (s) nm x , which is positive for x(s) nm > 0, − 1 (s) nm ρ (s) nm x − 1 (s) e nm ρ , which is negative, meaning E. Proof of Proposition 2 We follow the line suggested in [32]. For simplicity of notation, define α1 = ρ(s) nm, and . Recall that the utility function is given by u(x) = α1 + α2ebx. Moreover, for x, y ∈ R>0, x ≥ y nm, α2 = −ρ(s) b = −1 if E [x] ≥ E [y]. It is obvious that (s) nm ρ We define [32] Clearly, By (25) and (26), we have Consequently, E [u(x)] = α1 + α2E(cid:2)ebx(cid:3) . d(x) = u−1 (E [u(x)]) . u−1(τ ) = α2 1 b α2 (cid:19) . ln(cid:18) τ − α1 (cid:19) ln(cid:18) E [u(x)] − α1 ln α1 + α2E(cid:2)ebx(cid:3) − α1 ln(cid:0)E(cid:2)ebx(cid:3)(cid:1) . α2 d(x) = = = 1 b 1 b 1 b u (d(x)) = α1 + α2ebd(x) = α1 + α2E(cid:2)ebx(cid:3) (24) (25) (26) (27) (28) ! where the second inequality simply follows by using (27). In order to prove that the d(x) given by (27) is the certainty equivalent of x, in what follows we show that all requirements stated in Definition 2 are satisfied. (a) By (24) and (28), we see that u (d(x)) = E [u(x)]. Thus, the player is indifferent between receiving the stochastic reward x or the deterministic reward d(x), so that x ≈ d(x). Thus the first condition holds. (b) The Von Neumann-Morgenstern preference relation is stated in (8). Consequently, Thus x >n y if α1 + α2E(cid:2)ebx(cid:3) > α1 + α2E(cid:2)eby(cid:3) . ln(cid:0)E(cid:2)ebx(cid:3)(cid:1) ≥ ln(cid:0)E(cid:2)eby(cid:3)(cid:1) , which, by (27), implies d(x) ≥ d(y). The reverse follows similarly, hence condition two holds. (c) The third condition holds since by using (27), for deterministic k we have d(k) = 1 b ln(cid:0)E(cid:2)ebk(cid:3)(cid:1) = k. (29) (30) (31) (d) By using (27) we have d (x − d(x)) = = 1 b 1 b ln(cid:0)E(cid:2)eb(x−d(x))(cid:3)(cid:1) ln(cid:0)e−bd(x)(cid:1) + 1 b = −d(x) + d(x) = 0. ln(cid:0)E(cid:2)ebx(cid:3)(cid:1) Therefore condition four holds. (e) Consider deterministic k, k′ ∈ RM >0, with k < k′. Then which by (27) implies d(x + k) ≤ d(x + k′). Thus condition five holds. ln(cid:0)E(cid:2)ebkebx(cid:3)(cid:1) ≤ ln(cid:16)Ehebk′ ebxi(cid:17) , 22 (32) (33) F. Proof of Proposition 3 In our designed multi-state stochastic task exchange economy, for each CPS n ∈ N and in every state s ∈ S, the utility of performing some task m ∈ M is given by (2). Moreover, for CPS n ∈ N at state s ∈ S, the initial arrived load of type m ∈ M follows an exponential distribution with parameter λ(s) nm. Also, the total utility at state s, i.e., u(s) n , follows by (3). To establish the proposition, first we need to calculate the deterministic equivalent. By (15), the deterministic (equivalent) utility function of CPS n ∈ N is given as d(s) nm = −ρ(s) Thus we first need to solve for E(cid:20)e E(cid:20)e ρ −1 (s) nm ρ (cid:16)r (s) nmQ (cid:16)r (s) nmQ −1 (s) nm ρ (s) (s) (cid:16)r (s) nmQ −1 (s) nm m (cid:17)(cid:21)(cid:19) nm ln(cid:18)E(cid:20)e m (cid:17)(cid:21). Due to (1), we have m (cid:17)(cid:21) = E(cid:20)e jm(cid:17)(cid:21) jm(cid:21), = Yj∈N E(cid:20)e nm Pj∈N q −1 (s) nm −1 (s) nm (s) nmq (cid:16)r (s) (s) (s) (s) r ρ ρ (34) (35) where the equality follows due to the independence of q(s) an exponential distribution with parameter λ(s) jm. Then jm, j ∈ N . Let q(s) jm be distributed according to (s) nmq r −1 (s) nm ρ E(cid:20)e (s) jm(cid:21) =Z ∞ 0 (s) nm r −1 (s) nm ρ e λ(s) jme−λ (s) jmq (s) jmdq By (35) and (36), (34) can be written as = λ(s) λ(s) jm + 1 jm (s) nm ρ . r(s) nm (36) (37) r(s) nm r(s) nm     . d(s) nm = ρ(s) nm ln Yj∈N ln nmXj∈N  λ(s) jm + 1 λ(s) ρ jm (s) nm λ(s) jm + 1 λ(s) ρ jm (s) nm = ρ(s) 23 Since the utility function defined by (2) is continuous and monotonically increasing, maximizing u(s) is equivalent to maximizing d(s) nm. Moreover, the function defined by (37) is continuous, monotonically increasing and concave. Thus, by (14), the total utility is the sum of continuous, monotonically increasing and concave functions, hence demonstrating the characteristics required by Theorem 4. Moreover, by Assumption 1, the requirement on q(s) D is satisfied as well. Hence, by Theorem 4, Walrasian equilibrium exists, and is also Pareto-efficient due to Theorem 5. nm(cid:17) nm,D(cid:16)d(s) Since the utility function given by (14) is additive separable, it satisfies the gross substitutes property, according to Lemma 1. Then, due to Lemma 2, the excess demand function also satisfies the gross substitutes property. The result therefore follows from Theorem 6. Moreover, by Definition 9 and Definition 10, every social-optimal allocation is also Pareto-efficient. Consequently, since there exists only one Pareto- efficient allocation, it is social-optimal as well. G. Proof of Theorem 3 In order to proof this proposition, it suffices to establish that both axioms of Definition 4 hold. By (17), we know that vn,D(cid:16)hx(s) n is∈S(cid:17) = PS s=1 a(s)u(s) n,D(cid:16)x(s) n (cid:17). Thus, Proposition 1 also holds for vn,D; that is, it is state-separable, are concave and monotonically increasing. Consequently, given Assumption 1, it follows from Theorem 4 that for the game GD (N , vn,D), Walrasian equilibrium exists. Moreover, as established in the proof of Proposition 3, the excess demand function (21) satisfies the gross substitutes property. Together with Theorem 6, this results in the uniqueness of Walrasian equilibrium. Finally, from Theorem 5 and the uniqueness of equilibrium, it follows that it is Pareto-efficient and social- optimal. As a result, no coalition (including the grand coalition) is able to improve by means of deviating and blocking the Walrasian equilibrium. In fact, due to Theorem 2, the Walrasian equilibrium lies inside the core of the ex-ante cooperative game GD (before revealing the state uncertainty). Thus the second condition of Definition 4 is satisfied. Now we prove that the first condition is satisfied as well. That is, the Walrasian equilibrium under D for every state s ∈ S, and thus there is no incentive uncertainty belongs to the core of the state game G(s) for deviation ex-post. Consider the ex-ante game GD (N , vn,D) and let ¯x := ¯xN be the Walrasian equilibrium under un- certainty. Following the discussion above, we know that ¯xN ∈ C (GD). Since ¯xN is a competitive equilibrium, ¯x(s) n,N , i.e., the reward allocated to agent n ∈ N at state s ∈ S, maximizes vn on the budget set Bn(p) = {xn : p · ¯xn ≤ p · qn}. By (17), we know that vn,D(cid:16)h¯x(s) Moreover, vn is state-separable and u(s) 1. Thus, ¯x(s) n ≤ p(s) · q(s) The rest of the proof is similar to proving that the competitive equilibrium lies inside the core [38]. N is the unique maximizer of u(s) n,D on the budget set B(s) n : p(s) · ¯x(s) s=1 a(s)u(s) n is∈S(cid:17) = PS n,D, s ∈ S, are concave and monotonically increasing by Proposition n,D(cid:16)¯x(s) n (cid:17). n,Do. D (cid:17). This implies that at state s, there is some coalition c that can N does not belong to the core of the state game after the resolution n (p) =nx(s) of the uncertainty; i.e, ¯x(s) unilaterally improve upon ¯x(s) N /∈ C(cid:16)G(s) Assume that at some state s ∈ S, ¯x(s) N via an allocation yc. Formally, q(s) n,D; y(s) n,c ≤Xn∈c Xn∈c n,D(cid:16)y(s) n,c(cid:17) ≥ u(s) n,D(cid:16)y(s) n,c(cid:17) > u(s) n,D(cid:16)¯x(s) n,N(cid:17) ; n,N(cid:17) . n,D(cid:16)¯x(s) ∀ n ∈ c : u(s) ∃ n ∈ c : u(s) As argued before, x(s) N maximizes u(s) n,D on the budget set. Thus, (40) ensures that ∃ n ∈ c : p(s) · y(s) n,c > p(s) · q(s) n,D (38) (39) (40) (41) n,c < p(s) · q(s) n,D, there would be a neighborhood A of y(s) Since if p(s) · y(s) n,c for which there exists some xn,c ∈ A so that p(s) · x(s) n,D, and by the locally non-satiated (LNS) property, or due to monotonically increasing assumption of utility (see Proposition 1), such a neighborhood contains an x(s) that satisfies u(s) n,N maximizing u(s) on the budget set. Then (39) yields n,N(cid:17), which is inconsistent with ¯x(s) n,c(cid:17) ≥ u(s) n,D(cid:16)y(s) n,D(cid:16)x(s) n,c(cid:17) > u(s) n,D(cid:16)¯x(s) n,c < p(s) · q(s) n,c n,D 24 By summing the inequalities (41) and (42) over coalition c, we have ∀ n ∈ c : p(s) · y(s) n,c > p(s) · q(s) n,D (42) (43) p(s) · q(s) n,D n,c >Xn∈c p(s) · y(s) Xn∈c ⇒ p(s) ·Xn∈c n,c >Pn∈c q(s) y(s) n,c > p(s) ·Xn∈c q(s) n,D n,D for at least one n ∈ c, which contradicts (38). Since p(s) ∈ R≥0, we have Pn∈c y(s) H. Proof of Proposition 4 As described in the proof of Proposition 3, the utility function given by (17) satisfies the gross substitutes property. Then by Lemma 2, the excess demand function satisfies the gross substitutes property as well. Therefore, the result follows by Theorem 7. REFERENCES [1] W. Wang and Y. Jiang, "Community-aware task allocation for social networked multiagent systems," IEEE Transactions on Cybernetics, vol. 44, no. 9, pp. 1529 -- 1543, Sept 2014. [2] M.M. de Weerdt, Y. Zhang, and T. Klos, "Multiagent task allocation in social networks," Autonomous Agents and Multi-Agent Systems, vol. 25, no. 1, pp. 46 -- 86, Jul 2012. [3] W. Zhao, Q. Meng, and P.W.H. Chung, "A heuristic distributed task allocation method for multivehicle multitask problems and its application to search and rescue scenario," IEEE Transactions on Cybernetics, vol. 46, no. 4, pp. 902 -- 915, April 2016. [4] Y. Jiang and J. Jiang, "Contextual resource negotiation-based task allocation and load balancing in complex software systems," IEEE Transactions on Parallel and Distributed Systems, vol. 20, no. 5, pp. 641 -- 653, May 2009. [5] J. Blythe, S. Jain, E. Deelman, Y. Gil, K. Vahi, A. Mandal, and K. Kennedy, "Task scheduling strategies for workflow-based applications in grids," in IEEE International Symposium on Cluster Computing and the Grid, May 2005, vol. 2, pp. 759 -- 767. [6] B. Hu and J. Chen, "Optimal task allocation for human-machine collaborative manufacturing systems," IEEE Robotics and Automation Letters, vol. 2, no. 4, pp. 1933 -- 1940, Oct 2017. [7] Y. Dai and G. Levitin, "Optimal resource allocation for maximizing performance and reliability in tree-structured grid services," IEEE Transactions on Reliability, vol. 56, no. 3, pp. 444 -- 453, Sept 2007. [8] Q.C. Ye and Y. Zhang, "Participation behavior and social welfare in repeated task allocations," in 2016 IEEE International Conference on Agents (ICA), Sept 2016, pp. 94 -- 97. [9] S. He, D. H. Shin, J. Zhang, and J. Chen, "Near-optimal allocation algorithms for location-dependent tasks in crowdsensing," IEEE Transactions on Vehicular Technology, vol. 66, no. 4, pp. 3392 -- 3405, April 2017. [10] H. Zhao and X. Li, "Efficient grid task-bundle allocation using bargaining based self-adaptive auction," in IEEE/ACM International Symposium on Cluster Computing and the Grid, May 2009, pp. 4 -- 11. [11] F. Fu and M. van der Schaar, "Noncollaborative resource management for wireless multimedia applications using mechanism desig," IEEE Transactions on Multimedia, vol. 9, no. 4, pp. 851 -- 868, May 2007. [12] P. Semasinghe, S. Maghsudi, and E. Hossain, "Game theoretic mechanisms for resource management in massive wireless iot systems," IEEE Communications Magazine, vol. 55, no. 2, pp. 121 -- 127, February 2017. [13] V. Pilloni, P. Navaratnam, S. Vural, L. Atzori, and R. Tafazolli, "TAN: a distributed algorithm for dynamic task assignment in WSNs," IEEE Sensors Journal, vol. 15, no. 4, pp. 1266 -- 1279, April 2014. [14] J. Park and M. Van Der Schaar, "Stackelberg contention games in multiuser networks," EURASIP Journal on Advances in Signal Processing, pp. 6:1 -- 6:15, Jan 2009. [15] W. Wang and Y. Jiang, "Multiagent-based allocation of complex tasks in social networks," IEEE Transactions on Emerging Topics in Computing, vol. 3, no. 4, pp. 571 -- 584, Dec 2015. [16] E. Manisterski, D. Esther, K. Sarit, and N.R. Jennings, "Forming efficient agent groups for completing complex tasks," in Proceedings of the Fifth International Joint Conference on Autonomous Agents and Multiagent Systems. 2006, pp. 834 -- 841, ACM. [17] S. Kraus, O. Shehory, and G. Taase, "Coalition formation with uncertain heterogeneous information," in Proceedings of the Second International Joint Conference on Autonomous Agents and Multiagent Systems. 2003, pp. 1 -- 8, ACM. [18] O. Shehory and S. Kraus, "Methods for task allocation via agent coalition formation," Artificial Intelligence, vol. 101, no. 1-2, pp. 165 -- 200, May 1998. 25 [19] S. Abdallah and V. Lesser, "Learning the task allocation game," in Proceedings of the Fifth International Joint Conference on Autonomous Agents and Multiagent Systems. 2006, pp. 850 -- 857, ACM. [20] D. Grosu, A. T. Chronopoulos, and M.-Y. Leung, "Load balancing in distributed systems: an approach using cooperative games," in International Parallel and Distributed Processing Symposium, April 2002, pp. 10 pp -- . [21] M. N. Birje, S. S. Manvi, and S. K. Das, "Resource pricing strategy in wireless grid using non-cooperative bargaining game," in 2012 2nd IEEE International Conference on Parallel, Distributed and Grid Computing, Dec 2012, pp. 61 -- 66. [22] J. Liu, X. Jin, and Y. Wang, "Agent-based load balancing on homogeneous minigrids: macroscopic modeling and characterization," IEEE Transactions on Parallel and Distributed Systems, vol. 16, no. 7, pp. 586 -- 598, July 2005. [23] S. Maghsudi and E. Hossain, "Distributed user association in energy harvesting small cell networks: An exchange economy with uncertainty," IEEE Transactions on Green Communications and Networking, vol. 1, no. 3, pp. 294 -- 108, Sep 2017. [24] S. Ranadheera, S. Maghsudi, and E. Hossain, "Minority games with applications to distributed decision making and control in wireless networks," IEEE Wireless Communications, vol. 24, no. 3, pp. 184 -- 192, Oct 2018. [25] S. Khan, M. I. Ahmad, and F. Hussain, "Exponential utility function based criteria for network selection in heterogeneous wireless networks," Electronics Letters, vol. 54, no. 8, pp. 529 -- 531, 2018. [26] Y. Chen, B. Wang, and K. J. R. Liu, "Multiuser rate allocation games for multimedia communications," IEEE Transactions on Multimedia, vol. 11, no. 6, pp. 1170 -- 1181, Oct 2009. [27] S. He, J. Chen, Y. Sun, D. K. Y. Yau, and N. K. Yip, "On optimal information capture by energy-constrained mobile sensors," IEEE Transactions on Vehicular Technology, vol. 59, no. 5, pp. 2472 -- 2484, Jun 2010. [28] G. B. Shrestha, B. K. Pokharel, T. T. Lie, and S. Fleten, "Management of price uncertainty in short-term generation planning," IET Generation, Transmission Distribution, vol. 2, no. 4, pp. 491 -- 504, July 2008. [29] S. Maghsudi and E. Hossain, "Distributed user association in energy harvesting small cell networks: A probabilistic bandit model," IEEE Transactions on Wireless Communications, vol. 16, no. 3, pp. 1549 -- 1563, March 2017. [30] G. Debreu, Theory of Value: An Axiomatic Analysis of Economic Equilibrium, Yale University Press, 1987. [31] A. Mas-Colell, M.D. Whinston, and J.R. Green, Microeconomic Theory, Oxford University Press, 1985. [32] J. Suijs and P. Borm, "Stochastic cooperative games: Superadditivity, convexity, and certainty equivalents," Games and Economic Behavior, vol. 27, no. 2, pp. 331 -- 345, 1999. [33] R. Wilson, "The theory of syndicates," Econometrica, vol. 36, no. 1, pp. 119 -- 132, Jan 1968. [34] D.B. Gillies, "Solutions to general non-zero-sum games," Annals of Mathematics Studies, vol. 40, pp. 4785, 1959. [35] P.J.J. Herings, A. Predtetchinski, and A. Perea, "The weak sequential core for two-period economies," International Journal of Game Theory, vol. 34, no. 1, pp. 55 -- 65, 2006. [36] H. Habis and P.J.J. Herings, "Transferable utility games with uncertainty," Journal of Economic Theory, vol. 146, no. 5, pp. 2126 -- 2139, 2011. [37] F. Gul and E. Stacchetti, "Walrasian equilibrium with gross substitutes," Journal of Economic Theory, vol. 87, no. 1, pp. 95 -- 124, July 1999. [38] G. Debreu and H. Scarf, "A limit theorem on the core of an economy," International Economic Review, vol. 4, no. 3, pp. 235 -- 246, Sep 1963. [39] H. Uzawa, "Walras' tatonnement in the theory of exchange," The Review of Economic Studies, vol. 27, no. 3, pp. 182 -- 194, June 1960. [40] J. Levin, "General equilibrium," 2006. [41] S. Maghsudi and S. Stanczak, "On channel selection for energy-constrained rateless-coded D2D communications," in European Signal Processing Conference, Aug 2015, pp. 1028 -- 1032. [42] R. Mochaourab, B. Holfeld, and T. Wirth, "Distributed channel assignment in cognitive radio networks: Stable matching and walrasian equilibrium," IEEE Transactions on Wireless Communications, vol. 14, no. 7, pp. 3924 -- 3936, July 2015. [43] D. S. Palguna, D. J. Love, and I. Pollak, "Secondary spectrum auctions for markets with communication constraints," IEEE Transactions on Wireless Communications, vol. 15, no. 1, pp. 116 -- 130, Jan 2016. [44] D.S. Hochbaum, "Complexity and algorithms for nonlinear optimization problems," Annals of Operations Research, vol. 153, no. 1, pp. 257 -- 296, 2007.
1210.0437	1	1210	2012-10-01T15:07:16	Multi-Agent Programming Contest 2012 - The Python-DTU Team	[ "cs.MA" ]	We provide a brief description of the Python-DTU system, including the overall design, the tools and the algorithms that we plan to use in the agent contest.	cs.MA	cs	Multi-Agent Programming Contest 2012 -- The Python-DTU Team Jørgen Villadsen, Andreas Schmidt Jensen, Mikko Berggren Ettienne, Steen Vester, Kenneth Balsiger Andersen, and Andreas Frøsig Department of Informatics and Mathematical Modelling Technical University of Denmark Richard Petersens Plads, Building 321, DK-2800 Kongens Lyngby, Denmark Abstract. We provide a brief description of the Python-DTU system, including the overall design, the tools and the algorithms that we plan to use in the agent contest. Updated 1 October 2012: Appendix with comments on the contest added. Introduction 1. The name of our team is Python-DTU. We participated in the contest in 2009 and 2010 as the Jason-DTU team [2,3], where we used the Jason platform [1], but this year we use just the programming language Python as we did in 2011 [4]. 2. The members of the team are as follows: -- Jørgen Villadsen, PhD -- Andreas Schmidt Jensen, PhD student -- Mikko Berggren Ettienne, MSc student -- Steen Vester, MSc student -- Kenneth Balsiger Andersen, BSc student -- Andreas Frøsig, BSc student We are affiliated with DTU Informatics (short for Department of Informatics and Mathematical Modelling, Technical University of Denmark, and located in the greater Copenhagen area). 3. The main contact is associate professor Jørgen Villadsen, DTU Informatics, email: [email protected] 4. We expect that we will have invested approximately 200 man hours when the tournament starts. 1 System Analysis and Design 1. The competition is built on the Java MASSim platform and the Java EIS- MASSim framework is distributed with the competition files. This framework is based on EIS and abstracts the communication between the server and the agents to simple Java method calls and callbacks. We decided to skip EISMASSim to instead follow a much cleaner Python-only implementation. Even though some work was needed to implement the protocol specific parts which EISMASSim handled, this left us with a more flexible implementa- tion of which we have complete knowledge and control of every part of the implementation. 2. We do not use any existing multi-agent system methodology. 3. We do not plan to distribute the agents on several machines mainly for two reasons. Firstly, we had no need to, as we have enough computation power on a single machine to reason and send the action messages before the deadlines. Secondly, the shared data structure in our implementation would have to be replaced by a message server and a simple protocol. Due to limited time we have to prioritize differently. 4. We do not plan a solution with a centralization of coordination/information on a specific agent. Rather we plan a decentralized solution where agents share percepts through through shared data structures and coordinated ac- tions using distributed algorithms. 5. Our communication strategy is to share all new percepts to keep the agents internal world models identical. Furthermore our agreement based auction algorithm heavily relies on communication and is part of how agents decide on goals. 6. We hope to achieve the following properties when designing an algorithm to assign goals to agents: (a) The total benefit of the assigned goals should be as high as possible. Preferably optimal or close to it. (b) The running time of the algorithm should be fast, since we need to assign goals to agents at every time step in the competition and still have time left for other things such as environment perception, information sharing, etc. (c) The algorithm should be distributed between the agents resembling a true multi-agent system. (d) It should not be necessary for the agents to have the same beliefs about the state of the world in order to agree on an assignment. (e) The algorithm should be robust. If it is possible, our agents should be able to agree on an assignment even if some agents break down or some communication channels are broken. 2 7. Each agent acts on its own behalf based on its local view of the world which is updated through percepts and is thus autonomous and reactive. This is implemented as an agent-control-loop in which the agents decide which actions to execute based on their current view of the world. When a repairer and a disabled agent moves towards each other the repairer decides and announces who should take the last step so they won't miss each other. This proactiveness is implemented by considering the current energy and the paths of the agents. Software Architecture 1. We do not use any multi-agent programming language. We implement the multi-agent system using just the programming language Python. We choose Python as our programming language, as we think it has some advantages over for example Java, mainly in development speed/programmer effectiveness. Some of the reasons being that Python in contrast to Java: -- is dynamically typed -- is concise -- is compact -- supports multiple programming paradigms (object-oriented, imperative, functional) -- is popular for scripting -- does not need to be compiled before execution 2. We use Python 3.0 on Linux and Mac OS X as the development platforms and GEdit, Eclipse and TextMate as code editors/IDEs. 3. As the runtime platform for the competition we plan to use a suitable Linux system with Python 3.0. 4. Our implementation has mainly relied on custom best-first searches and a distributed auction-based agreement algorithm and a custom pathfinding algorithm tweaked for this domain. References 1. Rafael H. Bordini, Jomi Fred Hubner, and Michael Wooldridge. Programming Multi- Agent Systems in AgentSpeak Using Jason. John Wiley & Sons, 2007. 2. Niklas Skamriis Boss, Andreas Schmidt Jensen, and Jørgen Villadsen. Building Multi-Agent Systems Using Jason. Annals of Mathematics and Artificial Intelli- gence, 59: 373-388, Springer 2010. 3. Steen Vester, Niklas Skamriis Boss, Andreas Schmidt Jensen, and Jørgen Villadsen. Improving Multi-Agent Systems Using Jason. Annals of Mathematics and Artificial Intelligence, 61: 297-307, Springer 2011. 3 4. Mikko Berggren Ettienne, Steen Vester, and Jørgen Villadsen. Implementing a Multi-Agent System in Python with an Auction-Based Agreement Approach. In Louise A. Dennis, Olivier Boissier, and Rafael H. Bordini (Eds.): ProMAS 2011, LNCS 7217, 185-196, Springer 2012. Acknowledgements Thanks to Per Friis for IT support. More information about the Python-DTU team is available here: http://www.imm.dtu.dk/~jv/MAS Appendix The aim of the annual agent contest is to stimulate research in the area of multi- agent systems, to identify key problems and to collect suitable benchmarks. The 2012 contest was organized by Tristan Behrens, Jurgen Dix, Michael Koster and Federico Schlesinger at the Clausthal University of Technology, Germany. The scenario and schedule were announced 20 February 2012 and the tournament took place 10-12 September 2012. The 2012 winner was the Jason-UFSC team led by Jomi Fred Hubner, Federal University of Santa Catarina, Brazil. Like in 2011 we came in second. Both teams won all matches against the 5 other teams but we lost 1-2 against the winner. The 5 other teams came from Brazil, China, Germany, Iran and USA. In addition 2 teams from Germany and Ireland did not make it in the qualification phase where the stability of the teams had to be proved. Further details are available here: http://multiagentcontest.org 4
1204.6638	1	1204	2012-04-30T13:58:47	Modelling the emergence of spatial patterns of economic activity	[ "cs.MA", "cs.SI", "physics.soc-ph", "q-fin.GN" ]	Understanding how spatial configurations of economic activity emerge is important when formulating spatial planning and economic policy. A simple model was proposed by Simon, who assumed that firms grow at a rate proportional to their size, and that new divisions of firms with certain probabilities relocate to other firms or to new centres of economic activity. Simon's model produces realistic results in the sense that the sizes of economic centres follow a Zipf distribution, which is also observed in reality. It lacks realism in the sense that mechanisms such as cluster formation, congestion (defined as an overly high density of the same activities) and dependence on the spatial distribution of external parties (clients, labour markets) are ignored. The present paper proposed an extension of the Simon model that includes both centripetal and centrifugal forces. Centripetal forces are included in the sense that firm divisions are more likely to settle in locations that offer a higher accessibility to other firms. Centrifugal forces are represented by an aversion of a too high density of activities in the potential location. The model is implemented as an agent-based simulation model in a simplified spatial setting. By running both the Simon model and the extended model, comparisons are made with respect to their effects on spatial configurations. To this end a series of metrics are used, including the rank-size distribution and indices of the degree of clustering and concentration.	cs.MA	cs	Modelling the emergence of spatial patterns of economic activity Jung-Hun Yang, Dick Ettema, Koen Frenken Urban and Regional Research Centre, Faculty of Geosciences, Utrecht University P.O.Box 80115, Utrecht, 3508 TC, the Netherlands e-mail: [email protected], [email protected], [email protected] 1. Introduction The spatial pattern of economic activities is an important determinant of urban development. Locations of firms influence where workers will live, where consumers will buy products and where other firms are located. The locations of firms also impact on transportation flows, since they are important attractors and producers of both personal and freight traffic. Finally, the spatial pattern of firms obviously has a profound impact on the economic viability and conditions for economic growth in a region. Through the decades, therefore, researchers have developed models that describe and predict how spatial patterns of economic activity emerge. Without intending to exhaustively review all approaches taken, we will here review some modelling approaches that are relevant to our study. In particular, we will review micro- simulation and agent based approaches that take the individual firm as the unit of analysis. A first type of models (UrbanSim, SimFirms, ILUMASS) describes the evolution of spatial economic systems as a stochastic process, in which events such as firm growth, firm relocation, spin offs and take place with a probability that is predominantly a function of firm characteristics. In UrbanSim, economic activity is J.H. Yang, D. Ettema, K. Frenken Page 2 represented in terms of individual jobs, which are taken from an independent economic forecasting model, and are exogenous to the model. The jobs are treated as independent entities (i.e. not organised in firms), which are distributed across grid cells. ILUMASS (Moeckel, 2005) applies a more elaborate economic component. In particular, it uses a synthetic database of firms, which may take decisions regarding relocation, growth and closure. In addition, new firms may emerge at a particular birth rate, which is specific per sector and dependent on general economic growth rates. One of the most elaborate micro-simulation models of firms’ developed to date is SIMFIRMS (Van Wissen, 2000). This model distinguishes the same events as ILUMASS (birth, growth, (re-) location, closure) but uses more sophisticated behavioural rules, accounting for such factors as market stress, spin offs of existing firms, age effects and spatial inertia in the case of relocation. Market stress is related to the concept of carrying capacity, which, analogous to the ecological concept, indicates the maximum number of firms that an urban system can contain. Carrying capacity is operationalised as the difference between market supply and market capacity, which is based on aggregate input-output models. Thus, the measure is the outcome of aggregate conceptualisations, rather than on firms’ perception of demand and supply. In general the micro-simulation approaches are especially insightful to study demographic processes. For instance, they suffice to describe what the distribution across sectors in a region will be given some initial setting and given birth rates, spin-off probabilities etc. An element that is much less developed in these models is the role of spatial proximity. The fact that firms cluster in order to achieve agglomeration advantages is not well represented. Structural changes in spatial economics structures (e.g. the emergence of new economic centres due to changes in industries) are not well represented. A second type of models focuses on the emergence of hierarchies of concentrations (of firms or population) as a result of simple reproduction and migration rules. Simon (1955) shows that by assuming fixed reproduction rates and relocation probabilities, and assuming that larger concentrations attract more migrants than lower concentrations, a hierarchy of concentrations emerges that follows a power law distribution. Remarkably, such power law distributions match existing hierarchies in economic concentration (Frenken et al., 2007) and population concentrations (Pumain, 2006) very well. Although apparently these simple reproduction and migration rules 48th Congress of the European Regional Science Association, 27~31 August 2008, Liverpool, U.K. J.H. Yang, D. Ettema, K. Frenken Page 3 touch upon general principles of spatial organisation, the theoretical underpinning of the models is somewhat cumbersome (Krugman, 1996). In their most basic form, models as suggested by Simon are non-spatial. That is to say, the relative position of a concentration (e.g. a city or a commercial area) to other concentrations does not matter, since locational preferences of migrants only depend on the size of the concentration and not on its surroundings. As a result, a big city on an isolated place would be equally attractive as an equally big city surrounded by other cities. This assumption is problematic since it ignores the impact of proximity. For instance, studies in evolutionary economics (Boschma et al., 2002) suggest that proximity to other firm maters for their productivity and innovative capacity, and that this proximity exceeds the purely local scale. In particular, regions play an important role in processes of economic innovation, where the size of a region differs between types of industries. Thus, although correctly reproducing the rank size distribution of existing economic and population concentrations, the Simon model falls short in describing the emergence of clusters of economic development on a regional level. From the above, we conclude that existing models of spatial economic development have some important limitations. Most importantly, the role of spatial proximity to other firms is not well represented in the models. This proximity includes both the availability of other firms and the distance to these firms. Given this shortcoming, the objective of this paper is to propose a model of spatial economic development that is capable of representing the impact of spatial proximity on emerging spatial patterns. To this end, a theoretical framework is developed in which market potential, agglomeration benefits and congestion affect locational decisions on different spatial scales. The model of location behaviour is embedded in a demographic model of firm growth and spin-off processes. The model is tested in a stylised spatial setting, to illustrate how different parameter settings lead to different spatial configurations. The paper is organised as follows. Section 2 outlines a model that describes the behaviour divisions and the utility of spatial proximity. Section 3 describes the application of the model in a series of simulations. Section 4 analyses the impacts of spatial proximity, weighting effects and relocation probabilities on the emerging patterns of economic activity. 48th Congress of the European Regional Science Association, 27~31 August 2008, Liverpool, U.K. J.H. Yang, D. Ettema, K. Frenken Page 4 2. Model description In line with the models reviewed above, our model describes the spatial behaviour of firms dynamically over a number of time steps. However, as firms may consist of multiple establishments and divisions, that may take individual locational decisions, we take the division as the unit of analysis. We define divisions as coherent working units with a minimum and maximum size, dependent on the type of firm. The behaviours described by the model are growth, spin-off and relocation. With respect to internal growth, we assume that divisions in a certain sector grow uniformly with a fixed amount per year. In reality, growth rates will differ between firms due to factors such as quality of management, position in a network of firms and geographical position. Although we recognise the existence of such heterogeneity, we will not include it in this study. In particular, we assume that division size in year t+1 equals: Division ( InternalGr owth :) Div size =+ t 1 Div size t + 1 (ex.1) where is the size of a division in a year t+1. We further assume that divisions tDiv 1+ size have a maximum size and that growth beyond this maximum results in less effective functioning of divisions, e.g. through increasing overheads. Hence we assume that if the maximum size is reached the division will split, resulting in a new division (spin-off). To reflect developments in product and sector lifecycles a spin-off does not necessarily result in a division of the same type as the parent division. For instance, a spin-off of an industrial division may be a division in services or high-tech. This reflects ongoing shifts in economies from traditional industries to high-tech and from manufacturing to services. In the models tested in this study we will assume the existence of a traditional and an innovative industry, in which all spin-offs (both from traditional and new industries) are innovative industries. The rule for occurrence of spin-offs is: Spinoff Division ( SpinoffDiv type ⎧ t 1 + ⎪ SpinoffDiv type ⎨ t 1 + ⎪ SpinoffDiv type ⎩ t 1 + = ) = = = Div old & size t 1 + Div new & size t 1 + Div new if _ (( Div if (( _0 type = t Div if _0 (( type = t old type (&) = t Div old )) (&) size δ > = t old Div new (&) )) size δ = > new t Div random Div size (&) δ = ∈ t old t (ex.2) )) ( φ 48th Congress of the European Regional Science Association, 27~31 August 2008, Liverpool, U.K. J.H. Yang, D. Ettema, K. Frenken Page 5 where is the type of is the type of division in time step t+1 and SpinoffDiv 1+ tDiv 1+ type type t newδ are constants indicating the maximum size of oldδ and division which is split. divisions of old and new type respectively. The first and second line of expression 2 imply that if the maximum size is reached the division is split as the same type and size and is initialised with size zero. If an old type division reaches its maximum size with probability φ, the spin-off is the new type. Apart from such demographic processes, the model describes firms’ relocation behaviour. Relocation of firms may take place for many reasons, which are usually concerned with internal processes, such as growth or suitability of the building. In such cases, the relocation is likely to take place within the same municipality or region, without structurally changing the spatial structure of the economy. In this study, however, we are particularly interested in the more strategic relocation, in which divisions seek to improve their access to markets and resources by moving to another geographic location. In this respect, we assume that each division has a certain probability to evaluate its current geographical position against alternative positions to test whether relocation results in an improvement of its conditions. Two options are distinguished. First, a division may investigate relocation to an existing city (defined as an existing concentration of firms) or to a new place without a current concentration of firms. We assume that the probabilities of not exploring relocation, investigating relocation to an existing and to a new city are 0.9, 0.09 and 0.01 respectively. It is recognised that different probabilities of relocation may result in different spatial patterns and different development speeds. This relation can be described as: (ex.3) W t ⋅+ κπ ⋅ W t (ex.4) ⋅+ κ + 1( W t κπ) ⋅ ⋅ − W t =+ 1 = = W t W t 1( − W () γ ⋅ t 1( −+ W ( t = λ 1 ⋅ 1( −+ ) κπ ⋅ ⋅ W t ) ⋅+ γ 1( −+ ) κπ ⋅ ⋅ W t ) ⋅+ κπ ⋅ ) κπ ⋅ ⋅ W t ) + λ 2 ⋅ W ( t ) κπ ⋅ ⋅ W t ) + κλ ⋅ ⋅ 3 W t W ( t 1( −+ (ex.5) W t (ex.6) where Wt is the total number of divisions in a year t and κ is the growth rate of the industry. The second term of expression 4, which plays an important role in the persistence and self-reinforcement of clusters, is the number of division that is attached 48th Congress of the European Regional Science Association, 27~31 August 2008, Liverpool, U.K. J.H. Yang, D. Ettema, K. Frenken Page 6 to an existing city by 1−π and the third term is the number of divisions moving to a new city (vacant area) based on the rate π. γ is the rate of moving to another city. The first term of expression 5 is the number of divisions not moving to another city and the second term is the number moving to another city. The third term is the number of 1λ , 2λ and 3λ are the probabilities of not exploring divisions moving to new city. relocation, moving to an existing and to a new city respectively. 2λ can be 1λ and 3λ is also equal to π of expression 5. However, since the focus of our study is on the role of spatial proximity in the emergence of spatial patterns, we use the above values, which proved to work well in other studies. The rule of relocation can be therefore defined as: matched with the 1− γ and γ. ( relocation ) Division ⎧ ⎪ ⎪ ⎨ ⎪ ⎪ ⎩ Div move t 1 + Div move t 1 + Div move t 1 + = ← ← ← Don tMove if Div (_ ' t Div U if (max (_) ∈ t m 1 = U (max n 1 = m if (_) Div t ∈ n (ex.7) random ( )) ∈ λ 1 random )) ( λ 2 )) random ( λ 3 where is the utility of city m and is the utility of vacant area n. mU nU Evaluation of alternative locations and relocation take place as follows. A firm will evaluate all locations to find the location with the highest utility. If the utility is higher than the utility of the current location, the division will move to this new location, otherwise, it will stay in its current place. The central issue when discussing the impact of spatial proximity is how utility is defined. In a non-spatial model, utility of each location would be equal, suggesting a random spatial process. The spatial sensitivity of the model is improved if the locational preference depends on the size (number of divisions) in the destination. In this case utility is defined as: U = i N i (ex.8) where is the utility of area i and iU iN this is the model proposed by Simon, which leads to the well known power law is the number of division in an area i. In essence, 48th Congress of the European Regional Science Association, 27~31 August 2008, Liverpool, U.K. J.H. Yang, D. Ettema, K. Frenken Page 7 distribution of concentrations. As noted before, the Simon model is local in terms of its utility function, since it only accounts for firms in a certain location, and not in the surroundings are taken into accoun t. In this respect, this study aims at proposing and testing utility formulations that not only take into account locational characteristics, but also characteristics of the surroundings, such as the proximity to other firms. Looking at locational characteristics of firms, the literature suggests various factors relating to proximity of firms that clearly exceed the purely local level. A first factor concerns market potential. Firms make profits from selling products of services to other f irms or to individuals. The shorter the travel distance to these c lients, the lower the costs and the higher the profit. In addition, the more clients can be reached within acceptable travel distance from a location, the larger the market potential and the more attractive the location is to settle. In this respect, the sensitivity to distance is the factor determining the spatial configuration. For common goods, such as groceries, willingness to travel is low. For more specialised goods/services, the willingness to travel and the market area will be larger. Such differences in distance decay will have a large impact on the emerging spatial patterns of economic activity. To operationalise this factor we assume that firms from different sectors buy each others products and that firms also serve as a proxy for the number of consumers that are wiling to buy goods or services. Hence, market potential (MP) can be defined as: MP i 1α−∑= ijd eN j j 1= ( ex.9) where MPi is the market potential in area i and Nj is po pulation (number of divisions) in ity j c . 1α is a parameter for controlling the distance decay and ijd is distance between area i and city j. A second factor related to spatial proximity is agglomeration advantages. Many studies suggest that firms benefit from prox im ity to similar firms. One reason is that they may profit from shared facilities and suppliers. In addition, some firms may be better able to attract clients and employees jointly than individually. Another important issue is that firms form networks in which knowledge is exchanged, projects are carried out and market information is exchanged, in order to achieve 48th Congress of the European Regional Science Association, 27~31 August 2008, Liverpool, U.K. J.H. Yang, D. Ettema, K. Frenken Page 8 competitive advantages. Such agglomeration advantages suggest that firms will prefer to locate near other firms from the same sector. In equation, agglomeration effects are expressed as: type AP i = eN type j d α (ex.10) − ij 2 ∑ j 1 = where N type j type is the agglomeration potential of division of specific type in area i and iAP is a population of divisions of a specific type in city j. 2α is a parameter for is distance between area i and city j. It is controlling the distance effect and ijd eration advantages for different sectors ma recognised that agglom y differ in im portance, e.g. due to the relative importance of knowledge and innovation in a sector. Also the scale of agglomeration advantages may differ, due to the type of interaction. E.g. having similar consumers asks for immediate physical proximity, whereas exchange of knowledge via personal meetings allows a longer travel time. Finally, having noted the advantages of being close to other firms and clients, we note that there will also be disadvantages. Increasing densi ty leads to congestion of infrastr ucture and facilities, but also to higher prices and increasing competition for employees and other resources. Note that congestion is not sector specific, in the sense that firms suffer from congestion caused by all other firms. In equations: CP i = ∑ − 3d α eN ij j 1j = (ex.11) p arameter for controlling distance effect and where CPi is the congestion effect in area i and Nj is population in city j. 3α is distance between area i and city j. ijd Again, we note that sensitivity or congestion may differ between firm types, due to their need for space and infrastructure and the required qualifications of their employees. However, also the advantage of agglomeration will be weighted off against the disadvantage of congestion. As a result, the Simon utility of expression 8 can be transformed as: is a 48th Congress of the European Regional Science Association, 27~31 August 2008, Liverpool, U.K. J.H. Yang, D. Ettema, K. Frenken Page 9 U i = = MP β i 1 ∑ j 1 = β 1 + β 2 AP i type eN j d α − 1 ij + β 2 eN type j CP β + i 3 ∑ j 1 = (ex.12) ∑ j =1 (ex.13) eN j β 3 α − 2 + d α − 3 d ij ij where β is a coordinating parameter and MPi is the market potential in an area i . iAP the agglomeration potential of type t and CPi is the congestion potential in the area i. is The relocation probability can then be defined as: type P i = ∑ U ) exp( i U exp( j 1 = = ) j ∑ MP exp( + β i 1 MP exp( β j 1 β 2 + j 1 = AP type i AP β i 2 + β 3 type + CP ) i CP β j 3 ) (ex.14) where Pi is the probability of area i for relocation. This function is applied both for igration to existing cities and for migration to new cities. m To summarise, our model assumes that apart from locational and building uided by various variables specific characteristics, firms’ locational preferences are g that ex press the proximity to other firms, of similar and other sectors. In particular, we assume that market potential, agglomeration advantages and congestion effects, as defined in the above influence more strategic decisions about where firms are located. At the same time, we assume that firms differ with respect to the importance of these effects and the spatial scale at which they play. We hypothesize that the preferences of firms with respect to proximity will determine the spatial configuration of economic activities. For instance, agglomeration advantages on a small scale will lead to multiple centres of economic activity, whereas agglomeration advantages on a larger scale may lead to a single centre. In the remainder of this paper we will test to what extent differences in the spatial scale will lead to different spatial configurations. 3 The objective of this study is to test to what extent differences in firms’ preferences w ith respect to spatial proximity lead to different spatial patterns of economic activities. . Study design 48th Congress of the European Regional Science Association, 27~31 August 2008, Liverpool, U.K. J.H. Yang, D. Ettema, K. Frenken Page 10 In addition, we want to find out to what extent the impact of firms’ preferences is affected by factors such as growth rates per sector and the flexibility of relocation. Although we recognise that many factors other than discussed before (such as the availability of facilities, path dependency etc.) impact on firms’ location choice, we will use a stylised setting in which we will test some fundamental relationships between individual preferences of firms on the one hand, and aggregate spatial patterns on the other hand. In particular, we assume that firms operate in a landscape that is homogeneous in terms of travel speeds and quality of locations, and only varies in terms of the presence of other firms. The landscape consists of a square of 50x50 cells. Initially, at t=0, the landscape is filled with 2500 divisions, which in each time step will grow and with some probability relocate. The likelihood and effectuation of these events is determined by the equations described in the above. To test the impact of different preferences of spatial proximity, the model will be run with different parameters during 210 Time steps, after which the resulting pattern is analysed. This analysis will include three elements. First, the resulting patterns will be interpreted visually in terms of the number and size of emerging clusters of economic activity. Second, the distribution of rank sizes will be plotted, to see whether the resulting patterns follow the power law distribution typical for urban and economic distributions (Simon, 1955; Pumain, 2006). Third, the degree of clustering is expressed using the formula: K extension d )( = ∑ = i 1 dsCSnoN , ( [ ⋅ ∈ i i N )] (ex.15) dL )( = K extension d )( π − d (ex.16) extension K is the cluster density and N is the total number of divisions. C(si, d) is a where ircle with distance d from si. This index has a higher value if more divisions are closer c to one another. The index is calculated with the distance 10 for the purpose of our study. Starting point of the analyses is a base specification of the model, with parameters specified as in Table 1. Relative to this base model, the following analyses are carried out. First, starting from a Simon-type model without locational preferences 48th Congress of the European Regional Science Association, 27~31 August 2008, Liverpool, U.K. J.H. Yang, D. Ettema, K. Frenken Page 11 (beta are zero), various spatial factors are added stepwise, to see how this changes the resulting pattern. Second, the impact of different spatial factors will be varied by changing the beta parameter, in order to find out how this relative impact affects the spatial pattern of economic activity. Finally, the base model will be run with varying values for the parameters lambda, to see how that influences the resulting spatial pattern. 4. s shown in figure 1, the initial state of simulation is that the division of old type is A equally distributed acro ss all regions which consists of 2500 cells (50 by 50). A division the old industry may grow in each time step (ex.1) leading to a spin-off (ex.2) or in Simulation results Table 1. Parameter for Model 1~7 Parameter Type MP AP CP effe MP ct effe AP ct CP effect ximum Ma Growth Migration Time Step 1α 2α 3α 1β 2β 3β oldδ newδ 2λ 3λ - Old New Old New Old New Old New Old New Old New Old New y ot her cit new city - Empty Only MP MP+ AP Larger MP Larger AP Model Model Model Model Model Model Model 7 4 1 2 3 5 6 Larger MP+ MP+ AP+ AP CP 0.5 0.5 0.2 0.4 0.5 0.5 0.2 0.4 0.5 0.5 0.4 0.4 1 1 1 1 0.5 0.5 0.5 0.5 − − 1 1 − 1 − 1 50 50 10 10 19 19 0.3 0.3 210 210 0.5 0.4 0.5 0.4 0.5 0.4 0 0 0 0 0 0 50 10 19 0.3 210 0.5 0.4 0.5 0.2 0.5 0.4 1 1 0.5 0.5 − 1 − 1 50 10 19 0.3 210 0.5 0.2 0.5 0.4 0.5 0.4 1 1 0.5 0.5 − 1 − 1 50 10 19 0.3 210 0.5 0.4 0.5 0.4 0.5 0.4 1 1 0 0 0 0 50 10 19 0.3 210 0.5 0.4 0.5 0.4 0.5 0.4 1 1 0.5 0.5 0 0 50 10 19 0.3 210 48th Congress of the European Regional Science Association, 27~31 August 2008, Liverpool, U.K. J.H. Yang, D. Ettema, K. Frenken Page 12 relocation (ex.7). he expressions obability fo also igrat an T utility r m d pr ion follows the 13 a nd 14 re spectively. The figure 1 shows that the spatial pattern of old industry changed from an evenly distributed pattern red spatial pattern. Concerning into a cluste the emergence of a new industry, we assume a growth rate five times higher than that of the old industry. The different parameters for the old and new industry clearly affect their pattern of evolution. The new industry emerges both in a new agglomeration and in the existing agglomeration. This can be understood from the fact that the new industry profits from agglomeration economies of co-location (which explains the emergence of the new agglomeration) as well as from proximity to demand (explaining the growth of the new industry in the existing agglomeration). The resulting spatial pattern after the new industry has emerged, has become more "Zipf-like" in the sense (density: green < red < blue) Fig. 1. Time series for spatial pattern of old and new type division 3λ = 0.2 and Model 4) 2λ = 7, (based on Fig. 2. Time series for cluster indicator L of Model 1~7 Fig. 3. Time series for population of division 48th Congress of the European Regional Science Association, 27~31 August 2008, Liverpool, U.K. J.H. Yang, D. Ettema, K. Frenken Page 13 that we can witness cities of different sizes with the frequency of particular size decreasing with increasing size. Our model thus underlines the need to understand the spatial structure of an economy as a historical process of structural change leading to a progressive diversification of the economy. Figure 2 shows the cluster degree of each model run. Models 2 to 4 of red colour have a higher value than models 5 to 7 of green colour, though differences are small. The model 1 has the lowest value because any spatial interaction structures are absent. In addition, the total population of division represents the exponential growth. 4.1. The impact of spatial proximity The first model (model 1) that is tested is only based on growth and spin-off processes, lacking spatial preferences. This model is rather similar to the Simon model, except for the fact that Simon’s model assumes that cells with more di visions are more likely to attract newcomers, whereas in our model all cells have equal probability. As seen in figure 4, this model results in a pattern without centres, with divisions scattered out over space and filling all cells. The second model (model 2) includes the proximity to both old and new industry firms, representing market potential. The second picture of figure 4 suggests that adding market power results in a more clustered configuration, with one large centre. In time step 210, two smaller subcentres have emerged, which may in time develop to new centres. The rank size distribution clearly shows th e power law distribution in figure 5. (model 3) includes market potential and agglomeration effects, The thir d model here agglomeration effects only relate to proximity to similar firm types. The w simulation suggests that this model also leads to clustering of divisions, with one large centre. However, agglomeration may, since it is only focussed on similar firm types, more easily result in local clusters, such as the cluster in the lower left corner. As a consequence, the cluster indicator L is lower than in the case where only market potential plays a role. Again, the rank size distribution represents the zipf distribution more than a model 2. Finally, (model 4) adding the impact of congestion to the model results in a pattern with one centre like the fourth pictu re of figure 4. Rem arkably, this pattern is 48th Congress of the European Regional Science Association, 27~31 August 2008, Liverpool, U.K. J.H. Yang, D. Ettema, K. Frenken Page 14 less fragmented than the model with market potential and agglomeration effects, although the congestion is supposed to lead to more dispersed locations. Overall, we observe that centripetal forces such as market potential and agglomeration lead to a higher degree of clustering of economic activity in our model. The effect of congestion, however, is limited. L = 107.14 Model 1 L = 174 .82 Model 2 L = 167.25 Model 3 L = 174.82 Model 4 (density: green < red < blue) L = 167.44 L = 182.64 L = 166.37 Model 5 Model 6 Model 7 Fig. 4. Spatial Pattern: The impact of spatial proximity and weighting effects Model 1 Model 2 Model 3 Model 4 Fig. 5. Rank size distribution 48th Congress of the European Regional Science Association, 27~31 August 2008, Liverpool, U.K. J.H. Yang, D. Ettema, K. Frenken Page 15 4.2. Weighting effects In this section various models in which the relative impact of market potential and agglomeration are varied, are discussed. As seen in figure 4, Model 5 has a lower 1α for the market potential of new industries, implying that products are delivered in a larger area (e.g. because transport costs are lower). This model results in a spatial pattern with ne large centre, which is visually hard to distinguish from the base model. The cluster o index, however, suggests that the degree of clustering is lower as compared to the base model. This would be logical, given that the market area is larger, reducing the need to be in the immediate proximity of clients. 2α Model es he s 6 increas t patial sca le of ag on effects through a lower glomerati for the agglomeration potential of new industries. This results in a pattern with one larger and one smaller centre. Apparently, the larger spatial reach increases the attractiveness of less densely ‘populated’ areas, increasing also the probability of subcentres emerging. Yet, the cluster index has a high value, suggesting that within and around the cl This is also the result of the r firms essibili is high. y to oth e t sters ac c u rela tive closeness of the clusters. Model 7, in which both market potential and agglomeration potential have a larger spatial reach, clustering becomes less, as expected. The larger spatial reach facilitates the emergence of two clusters that are more equal in size than in the other models. To conclude, the spatial reach of the three identified effects has an impact on the emerging pattern. In particular, it seems more likely that multiple clusters emerge, since market and agglomeration advantages are available in a larger area. The impact this has on the cluster index varies, depending on the locations where clusters emerge and also depending on the scale of the cluster index. 4.3. The impact of relocation probabilities To test the impact of relocation probabilities, the variables 3λ were varied to 2λ and 2λ ) and a represent a higher relocation probability in general (mostly determined by higher probability of moving to a new city ( 3λ ). For each ion, the emerging spatial pattern as well as the cluster index are displayed (fig combinat ure 7). Visual inspection of 48th Congress of the European Regional Science Association, 27~31 August 2008, Liverpool, U.K. J.H. Yang, D. Ettema, K. Frenken Page 16 has a significant impact on the the emerging patterns suggests that especially 3λ outcomes. In particular, a higher probability of moving to a new city results in a setting with one large centre without subcentres, whereas a lower value of 3λ leads to ore m subcentres. Apparently, constraining the opportunity to settle down in a new city Table 2. The average of indicator L by 4 times simulation based on Model 4 λ 3 0.2 % 0.3 % 2λ 7 % average 11 % average 15 % average 19 % average 0.4 % 0.5 % 148.44 148.91 151.93 153.76 150.76 146.47 162.06 157.11 149.83 153.87 121.67 136.33 164.42 153.07 143.87 164.92 153.78 144.99 128.67 148.09 149.90 152 .33 162.70 165.46 157.60 155.31 169.77 156.50 167.05 162.15 174.38 190.49 187.67 138.46 172.75 160.58 191.16 156.58 170.60 169.73 188.54 176.21 202.97 209.16 194.22 200.71 174.95 212.44 190.31 194.60 201.53 209.07 181.41 202.36 198.59 214.86 186.33 206.44 174.24 195.47 228.66 245.22 219.69 217.42 227.75 210.01 201.90 218.96 193.74 206.15 226.93 163.68 225.91 205.03 205.39 255.57 205.20 171.43 208.32 210.13 Fig. 6. Indicator L by 2λ and 3λ based on Model 4 (ref. Table 2) 48th Congress of the European Regional Science Association, 27~31 August 2008, Liverpool, U.K. J.H. Yang, D. Ettema, K. Frenken Page 17 increases the probability that subcentres emerge in existing cities. In most cases, increasing the probability of moving to a new city ( 3λ ) results in a higher degree of clustering (figure 6 and table 2), as a result of the more centralised configuration. It is noted however, that this effect is less obvious in case of a lower probability of moving to an existing city ( 2λ ). 5. Conclusion and discussion ifferences in preferences strated in In this paper we have sed setting, a styli demon how d to diffe rent spatial patterns of eco mity lead with r to spatia espect l proxi nomic activity. 0. 2 % 0.3 % 0.4 % 0.5 % ( density : green < red < blue) 153.76 L = L = 146.47 L = 164.42 L = 1 52.33 L = 2 02.97 L = 22 8.66 L = 156.50 L = 200.71 L = 201.90 L = 174.38 L = 201.53 L = 205.03 λ 3 2λ 7 % 11 % 15 % 19 % L = 153.78 L = 160.58 Fig. 7. Spatial Pattern: The impact of relocation probabilities (based on Model 4) L = 208.32 L = 206.44 48th Congress of the European Regional Science Association, 27~31 August 2008, Liverpool, U.K. J.H. Yang, D. Ettema, K. Frenken Page 18 In this respect, a preference to achieve a high market potential and to profit from agglomeration advantages results in more centralise d settings. However, the spatial scale of the market and agglomeration effects matters. In particular, if agglomeration advantages stretch out over a longer distance, more subcentres emerge. Somewhat surprisingly, conge ion st seems to have a minor impact on the emerging patterns. lthough the simulation outcomes are intuitively plausible, they also articulate the need A r validation of the behavioural decision rules. If outcomes are determined by the fo th, although not tested in this paper) and spatial reach of presence (and potentially streng arket potential, agglomeration and congestion effects, it is important to investigate m how firms of different types valuate these factors in their location choice behaviour. In particular, it is important how the valuation of these factors varies with firm characteristics such as type of activities, ze, history and the position in economic si networks. Such information would be necessary to apply the above approach in a more realistic setting as a policy support tool. A second conclusion that can be drawn from the simulations is that relocation probability to existing and new cities impacts on the emerging patterns. This finding is highly policy relevant, since it suggests that the availability of locations where firms/divisions can move has a significant impact on spatial patterns of economic activity. If this is confirmed by validation studies, it would suggest that spatial planning is a to ol that can direc tly impact on the economic structure of regions and will influence firms’ performance and thereby regional economic develop ment. Although this study provides first insights into the emergence of spatial patterns of economic activ h more research is needed to develop this bviou ity, it is o s that muc approach into a tool that can be readily used for policy analysis. This research should address the following issues. First, the behavioural rules applied in this test of concept need to be verified and refined. In particular, multivariate analyses are needed that relate firm characteristic s to the d gr e ee and spatial reach of proximity preferences. This will require dedicated data to be collected from individual firms. Second, it should be recognised that firms do not operate in isolation, but interact with households and individuals (as clients and employees), institutions (such as government agencies, universities, schools etc.) and react to the physical environment (landscape, quality of residential environment, pollution and noise). A proper model for policy evaluation 48th Congress of the European Regional Science Association, 27~31 August 2008, Liverpool, U.K. J.H. Yang, D. Ettema, K. Frenken Page 19 should include a representation of how proximity concerns are traded off against these other factors. Reference Boschma, R.A., K. Frenken and J.G. Lambooy (2002) Evolutionaire Economie. Een inleiding. Bussum: Coutinho. Frenken, K., and R.A. Boschma. (2007) A Theoretical Framework for Evolutionary Economic Geography: Industrial Dynamics and Urban Growth as a Branching Process. Journal of Economic Geography 7: 635-649. Krugman, P.R. (1996) The Self-Organizing Economy. Oxford: Blackwell Publishers. Moeckel, R. (2005) Microsimulation of firm location decisions, Proceedings of 9th International Conference on Computers in Urban Planning and Urban Management, June 29-July 1, University College London, U.K. Pumain, D. et al. (2006) Alternative Explanation of Hierarchical Differentiation in Urban System. In: Hierarchy in Natural and Social Sciences. D. Pumain, ed., Springer. Simon, H. (1955) On a class of skew distribution functions. Biometrika 44: 425-440. van Wissen, L. (2000) A micro-simulation model of firms: Applications of concepts of the demography of the firm. Papers in Regional Science 79: 111-134. 48th Congress of the European Regional Science Association, 27~31 August 2008, Liverpool, U.K.
1809.04240	5	1809	2019-05-29T05:21:39	Towards Efficient Detection and Optimal Response against Sophisticated Opponents	[ "cs.MA" ]	Multiagent algorithms often aim to accurately predict the behaviors of other agents and find a best response accordingly. Previous works usually assume an opponent uses a stationary strategy or randomly switches among several stationary ones. However, an opponent may exhibit more sophisticated behaviors by adopting more advanced reasoning strategies, e.g., using a Bayesian reasoning strategy. This paper proposes a novel approach called Bayes-ToMoP which can efficiently detect the strategy of opponents using either stationary or higher-level reasoning strategies. Bayes-ToMoP also supports the detection of previously unseen policies and learning a best-response policy accordingly. We provide a theoretical guarantee of the optimality on detecting the opponent's strategies. We also propose a deep version of Bayes-ToMoP by extending Bayes-ToMoP with DRL techniques. Experimental results show both Bayes-ToMoP and deep Bayes-ToMoP outperform the state-of-the-art approaches when faced with different types of opponents in two-agent competitive games.	cs.MA	cs	Towards Efficient Detection and Optimal Response against Sophisticated Opponents Tianpei Yang1 , Jianye Hao1∗ , Zhaopeng Meng1 , Chongjie Zhang2 , Yan Zheng1 , Ze Zheng3 1College of Intelligence and Computing, Tianjin University 2MMW, Tsinghua University, China {tpyang, jianye.hao, mengzp, yanzheng}@tju.edu.cn, [email protected], 3Beifang Investigation, Design & Research CO.LTD [email protected] Abstract Multiagent algorithms often aim to accurately pre- dict the behaviors of other agents and find a best response accordingly. Previous works usually as- sume an opponent uses a stationary strategy or randomly switches among several stationary ones. However, an opponent may exhibit more sophis- ticated behaviors by adopting more advanced rea- soning strategies, e.g., using a Bayesian reasoning strategy. This paper proposes a novel approach called Bayes-ToMoP which can efficiently detect the strategy of opponents using either stationary or higher-level reasoning strategies. Bayes-ToMoP also supports the detection of previously unseen policies and learning a best-response policy accord- ingly. We provide a theoretical guarantee of the optimality on detecting the opponent's strategies. We also propose a deep version of Bayes-ToMoP by extending Bayes-ToMoP with DRL techniques. Experimental results show both Bayes-ToMoP and deep Bayes-ToMoP outperform the state-of-the-art approaches when faced with different types of op- ponents in two-agent competitive games. 1 Introduction In multiagent systems, the ideal behavior of an agent is contingent on the behaviors of coexisting agents. How- ever, agents may exhibit different behaviors adaptively de- pending on the contexts they encounter. Hence, it is crit- ical for an agent to quickly predict or recognize the be- haviors of other agents, and make a best response accord- ingly [Powers and Shoham, 2005; Fern´andez et al., 2010; Albrecht and Stone, 2018; Hernandez-Leal et al., 2017a; Yang et al., 2019]. One efficient way of recognizing the strategies of other agents is to leverage the idea of Bayesian Policy Reuse (BPR) [2016], which was originally proposed to determine the best policy when faced with different tasks. Hernandez-Leal et al. [2016] proposed BPR+ by extending BPR to multiagent learning settings to detect the dynamic changes of an oppo- nent's strategies. BPR+ also extends BPR with the ability ∗Corresponding Author to learn new policies online against an opponent using previ- ously unseen policies. However, BPR+ is designed for single- state repeated games only. Later, Bayes-Pepper [2017] is proposed for stochastic games by combing BPR and Pepper [Crandall, 2012]. However, Bayes-Pepper cannot handle an opponent which uses a previously unknown policy. There are also some deep multiagent RL algorithms investigating how to learn an optimal policy by explicitly taking other agents' [2016] proposed DRON behaviors into account. He et al. which incorporates the opponent's observation into deep Q- network (DQN) and uses a mixture-of-experts architecture to handle different types of opponents. However, it cannot guarantee the optimality against each particular type of oppo- nents. Recently, Zheng et al. [2018] proposed Deep BPR+ by extending BPR+ with DRL techniques to achieve more accu- rate detection and better response against different opponents. However, all these approaches assume that an opponent ran- domly switches its policies among a class of stationary ones. In practice, an opponent can exhibit more sophisticated be- haviors by adopting more advanced reasoning strategies. In such situations, higher-level reasonings and more advanced techniques are required for an agent to beat such kinds of so- phisticated opponents. The above problems can be partially addressed by in- troducing the concept of Theory of Mind (ToM) [Baker et al., 2011; de Weerd et al., 2013]. ToM is a kind of re- cursive reasoning technique [Hernandez-Leal et al., 2017a; Albrecht and Stone, 2018] describing a cognitive mechanism of explicitly attributing unobservable mental contents such as beliefs, desires, and intentions to other players. Previous methods often use nested beliefs and "simulate" the reason- ing processes of other agents to predict their actions [Gmy- trasiewicz and Doshi, 2005; Wunder et al., 2012]. However, these approaches show no adaptation to non-stationary oppo- nents [Hernandez-Leal et al., 2017a]. Later, De Weerd et al. [2013] proposed a ToM model which enables an agent to pre- dict the opponent's actions by building an abstract model of its opponent using recursive nested beliefs. Additionally, they adopt a confidence value to help an agent to adapt to different opponents. However, the main drawbacks of this model are: 1) it works only if an agent holds exactly one more layer of belief than its opponent; 2) it is designed for predicting the opponent's primitive actions instead of high-level strategies, resulting in slow adaptation to non-stationary opponents; 3) it shows poor performance against an opponent using previ- ously unseen strategies. To address the above challenges, we propose a novel al- gorithm, named Bayesian Theory of Mind on Policy (Bayes- ToMoP), which leverages the predictive power of BPR and recursive reasoning ability of ToM, to compete against such sophisticated opponents. In contrast to BPR which is capable of detecting non-stationary opponents only, Bayes-ToMoP in- corporates ToM into BPR to quickly and accurately detect not only non-stationary, and more sophisticated opponents and compute a best response accordingly. Theoretical guar- antees are provided for the optimal detection of the oppo- nent's strategies. Besides, Bayes-ToMoP also supports de- tecting whether an opponent is using a previously unseen pol- icy and learning an optimal response against it. Furthermore, Bayes-ToMoP can be straightforwardly extended to DRL en- vironment with a neural network as the value function ap- proximator, termed as deep Bayes-ToMoP. Experimental re- sults show that both Bayes-ToMoP and deep Bayes-ToMoP outperform the state-of-the-art approaches when faced with different types of opponents in two-agent competitive games. model of the opponent using recursive nested beliefs. ToM model is described in the context of a two-player competitive game where an agent and its opponent differ in their abilities to make use of ToM. The notion of ToMk indicates an agent that has the ability to use ToM up to the k-th order, and we briefly introduce the first two orders of ToM models A zero-order ToM (ToM0) agent holds its zero-order belief in the form of a probability distribution on the action set of its opponent. The ToM0 agent then chooses the action that max- imizes its expected payoff. A first-order ToM agent (ToM1) keeps both zero-order belief β(0) and first-order belief β(1). The first-order belief β(1) is a probability distribution that de- scribes what the ToM1 agent believes its opponent believes about itself. The ToM1 agent first predicts its opponent's ac- tion under its first-order belief. Then, the ToM1 agent inte- grates its first-order prediction with the zero-order belief and uses this integrated belief in the final decision. The degree to which the prediction influences the agent's actions is de- termined by its first-order confidence 0 ≤ c1 ≤ 1, which is increased if the prediction is right while decreased otherwise. 2 Background Bayesian Policy Reuse BPR was originally proposed in [Rosman et al., 2016] as a framework for an agent to quickly determine the best policy to execute when faced with an un- known task. Given a set of previously-solved tasks T and an unknown task τ∗, the agent is required to select the best pol- icy π∗ from the policy library Π within as small numbers of trials as possible. BPR uses the concept of belief β, which is a probability distribution over the set of tasks T , to measure the degree to which τ∗ matches the known tasks based on the signal σ. A signal σ can be any information that is correlated with the performance of a policy (e.g., immediate rewards, episodic returns). BPR involves performance models of poli- cies on previously-solved tasks, which describes the distribu- tion of returns from each policy π on previously-solved tasks. A performance model P (Uτ, π) is a probability distribution over the utility of a policy π on a task τ. A number of BPR variants with exploration heuristics are proposed to select the best policy, e.g., probability of improvement (BPR-PI) heuristic and expected improvement (BPR-EI) heuristic. BPR-PI heuristic utilizes the proba- (cid:80) bility with which a policy can achieve a hypothesized in- crease in performance (U +) over the current best estimate τ∈T β(τ )E[Uτ, π]. Formally, it chooses (cid:80) ¯U = maxπ∈Π the policy that most likely to achieve the utility U +: π∗ = τ∈T β(τ )P (U +τ, π) . However, it is not arg maxπ∈Π straightforward to determine the appropriate value of U +, thus another way of avoiding this issue is BPR-EI heuris- (cid:80) tic, which selects the policy most likely to achieve any possible utilities of improvement ¯U < U + < U max: τ∈T β(τ )P (U +τ, π)dU +. π∗ = arg maxπ∈Π BPR [2016] showed BPR-EI heuristic performs best among all BPR variants. Therefore, we choose BPR-EI heuristic for playing against different opponents. (cid:82) U max Theory of Mind ToM model [de Weerd et al., 2013] is used to predict an opponent's action by building an abstract ¯U 3 Bayes-ToMoP 3.1 Motivation Previous works [Hernandez-Leal et al., 2016; Hernandez- Leal and Kaisers, 2017; Hernandez-Leal et al., 2017b; Zheng et al., 2018; He and Boyd-Graber, 2016] assume that an oppo- nent randomly switches its policies among a number of sta- tionary policies. However, a more sophisticated agent may change its policy in a more principled way. For instance, it first predicts the policy of its opponent and then best responds towards the estimated policy accordingly. If the opponent's policy is estimated by simply counting the action frequencies, it is then reduced to the well-known fictitious play [Shoham and Leyton-Brown, 2009]. However, in general, an oppo- nent's action information may not be observable during inter- actions. One way of addressing this problem is using BPR, which uses Bayes' rule to predict the policy of the opponent according to the received signals (e.g., rewards), and can be regarded as the generalization of fictitious play. Therefore, a question naturally arises: how an agent can effectively play against both simple opponents with station- ary strategies and more advanced ones (e.g., using BPR)? To address this question, we propose a new algorithm called Bayes-ToMoP, which leverages the predictive power of BPR and recursive reasoning ability of ToM to predict the strate- gies of such opponents and behave optimally. We also extend Bayes-ToMoP to DRL scenarios with a neural network as the value function approximator, termed as deep Bayes-ToMoP. In the following descriptions, we do not distinguish whether a policy is represented in a tabular form or a neural network unless necessary. We assume the opponent owns a class of candidate sta- tionary strategies to select from periodically. Bayes-ToMoP needs to observe the reward of its opponent which is not an assumption since in competitive settings, e.g., zero-sum games, an agent's opponent's reward is always the opposite of its own. Bayes-ToMoP does not need to observe the ac- tions of its opponent except for learning a policy against an unknown opponent strategy. We use the notation of Bayes- ToMoPk to denote an agent with the ability of using Bayes- ToMoP up to the k-th order. Intuitively, Bayes-ToMoPi with a higher-order theory of mind could take advantage of any Bayes-ToMoPj with a lower-order one (i > j). De Weerd [2013] showed that the reasoning levels deeper than et al. 2 do not provide additional benefits, so we focus on Bayes- ToMoP0 and Bayes-ToMoP1. Bayer-ToMoPk (k > 1) can be naturally constructed by incorporating a higher-order ToM idea into our framework. (cid:80) 3.2 Bayes-ToMoP0 Algorithm We start with the simplest case: Bayes-ToMoP0, which ex- tends ToM0 by incorporating Bayesian reasoning techniques to predict the strategy of an opponent. Bayes-ToMoP0 holds a zero-order belief β(0) about its opponent's strategies {jj ∈ J }, each of which β(0)(j) is a probability that its oppo- nent may adopt each strategy j: β(0)(j) ≥ 0,∀j ∈ J . j∈J β(0)(j) = 1. Given a utility U, a performance model Pself (Uj, π) describes the probability of an agent using a policy π ∈ Π against an opponent's strategy j. For Bayes-ToMoP0 agent equipped with a policy library Π against its opponent's with a strategy library J , it first initial- izes its performance models Pself (UJ , Π) and zero-order belief β(0). Then, in each episode, given the current belief β(0), Bayes-ToMoP0 agent evaluates the expected improve- ment utility defined following BPR-EI heuristic for all poli- cies and then selects the optimal one. Next, Bayes-ToMoP0 agent updates its zero-order belief using Bayes' rule [Rosman et al., 2016]. At last, Bayes-ToMoP0 detects whether its op- ponent is using a previously unseen policy. If yes, it learns a new policy against its opponent. The new strategy detection and learning algorithm will be described in Section 3.4. Finally, note that without the new strategy detection and learning phase, Bayes-ToMoP0 agent is essentially equiva- lent to BPR and Bayes-Pepper since they both first predict the opponent's strategy (or taks type) and then select the op- timal policy following BPR heuristic, each strategy of the op- ponent here can be regarded as a task in the original BPR. Besides, the full Bayes-ToMoP0 is essentially equivalent to BPR+ since both can handle previously unseen strategies. factor to balance the influence between its first-order predic- tion and zero-order belief. Then, an integration function I is introduced to compute the final prediction results which is de- fined as the linear combination of the first-order prediction j and zero-order belief β(0) weighted by the confidence degree c1 following Equation 1 [de Weerd et al., 2013] (Line 3). (cid:26)(1 − c1)β(0)(j) + c1 (1 − c1)β(0)(j) if j = j otherwise I(β(0), j, c1)(j) = (1) Next, Bayes-ToMoP1 agent computes the optimal policy based on the integrated belief (Line 4). At last, Bayes- ToMoP1 updates its first-order belief and zero-order belief using Bayes' rule [Rosman et al., 2016] (Lines 6-11). Algorithm 1 Bayes-ToMoP1 Algorithm Initialize: Policy library Π and J , performance models Pself (UJ , Π) and Poppo(UΠ,J ), zero-order belief β(0), first-order belief β(1) 1: for each episode do 2: Compute the first-order opponent policy prediction j: arg maxj∈J(cid:82) U max (cid:82) U max (cid:80) (cid:80) π∈Π β(1)(π)Poppo(U +π, j)dU + Integrate j with β(0): I(β(0), j, c1) (see Equation (1)) Select the optimal policy π∗: arg max π∈Π Play and receive the episodic return (cid:104)rself , roppo(cid:105) for each own policy π ∈ Π do Update first-order belief β(1): β(1)(π) = I(β(0), j, c1)(j)Pself (U +j, π)dU + Poppo(roppoπ,j)β(1)(π) (cid:48) ∈Π Poppo(roppoπ ,j)β(1)(π (cid:80) j(cid:48)∈J ¯U (cid:48) ) ¯U π (cid:48) end for for each opponent strategy j ∈ J do Update zero-order belief β(0): Pself (rselfj,π)β(0)(j) β(0)(j) = (cid:48) ∈J Pself (rselfj (cid:80) (cid:48) j ,π)β(0)(j (cid:48) ) end for Update c1 following Equation (2) Detect new opponent strategy 11: 12: 13: 14: end for 3: 4: 5: 6: 7: 8: 9: 10: 3.3 Bayes-ToMoP1 Algorithm Next, we move to Bayes-ToMoP1 algorithm. Apart from its zero-order belief, Bayes-ToMoP1 also maintains a first- order belief, which is a probability distribution that describes the probability that an agent believes his opponent believes it will choose a policy π ∈ Π. The overall strategy of Bayes-ToMoP1 is shown in Algorithm 1. Given the policy library Π and J , performance models Pself (UJ , Π) and Poppo(UΠ,J ), zero-order belief β(0) and first-order belief β(1), Bayes-ToMoP1 agent first predicts the policy j of its opponent assuming the opponent maximizes its own utility based on BPR-EI heuristic under its first-order belief (Line 2). However, this prediction may conflict with its zero-order belief. To address this conflict, Bayes-ToMoP1 holds a first- order confidence c1(0 ≤ c1 ≤ 1) serving as the weighting The next issue is how to update the first-order confidence degree c1. The value of c1 can be understood as the explo- ration rate of using first-order belief to predict the opponent's strategies. In previous ToM model [de Weerd et al., 2013], the value of c1 is increased if the prediction is right while de- creased otherwise based on the assumption that an agent can observe the actions of its opponent. However, in our settings, the prediction works on a higher level of behaviors (the poli- cies), which information usually is not available (agents are not willing to reveal their policies to others to avoid being exploited in competitive environments). Therefore, we pro- pose to use game outcomes as the signal to indicate whether our previous predictions are correct and adjust the first-order confidence degree accordingly. Specifically, in a competitive environment, we can distinguish game outcomes into three cases: win (rself > roppo), lose or draw. Thus, the value of c1 is increased when the agent wins and decreased otherwise by an adjustment rate of λ. Following this heuristic, Bayes- ToMoP1 can easily take advantage of Bayes-ToMoP0 since it can well predict the policy of Bayes-ToMoP0 in advance. However, this heuristic does not work with less sophisticated opponents, e.g., an opponent simply switching among several stationary policies without the ability of using ToM. This is due to the fact that the curve of c1 becomes oscillating when it is faced with an agent who is unable to make use of ToM, thus fails to predict the opponent's behaviors accurately. i−l rself (cid:80)i To this end, we propose an adaptive and generalized mech- anism to adjust the value of c1. We first introduce the con- cept of win rate υi = during a fixed length l of episodes. Since Bayes-ToMoP1 agent assumes its opponent is Bayes-ToMoP0 at first, the value of l controls the number of episodes before considering its opponent may switch to a less sophisticated type. If the average performance till the current episode is better than the previous episode's (υi ≥ υi−1), we increase the weight of using first-order prediction, i.e., in- creasing the value of c1 with an adjustment rate λ; if υi is less than υi−1 but still higher than a threshold δ, it indicates the performance of the first-order prediction diminishes. Then Bayes-ToMoP1 decreases the value of c1 quickly with a de- lg(υi−δ); if υi ≤ δ, the rate of exploring first- creasing factor order belief is set to 0 and only zero-order belief is used for prediction. Formally we have: lgυi l ((1 − λ)c1 + λ)F(υi) ( lg(υi−δ) c1)F(υi) λF(υi) lgυi c1 = if υi ≥ υi−1 if δ < υi < υi−1 if υi ≤ δ (2) where δ is the threshold of the win rate υi, which reflects the lower bound of the difference between its prediction and its opponent's actual behaviors. F(υi) is an indicator function to control the direction of adjusting the value of c1. Intuitively, Bayes-ToMoP1 detects the switching of its opponent's strate- gies at each episode i and reverses the value of F(υi) when- ever its win rate υi is no larger than δ (Equation 3). if (υi ≤ δ & F(υi) = 0) if (υi ≤ δ & F(υi) = 1) (cid:26)1 F(υi) := (3) 0 Finally, Bayes-ToMoP1 learns a new optimal policy if it de- tects a new opponent strategy (detailed in next section). 3.4 New Opponent Detection and Learning The new opponent detection and learning component is the same for all Bayes-ToMoPk agents (k ≥ 0). Bayes-ToMoPk first detects whether its opponent is using a new kind of strate- gies. This is achieved by recording a fixed length of game out- comes and checking whether its opponent is using one of the known strategies at each episode. In details, Bayes-ToMoPk keeps a length h of memory recording the game outcomes at over the most episode i, and uses the win rate θi = recent h episodes as the signal indicating the average perfor- mance over all policies till the current episode i. If the win rate θi is lower than a given threshold δ (θi < δ), it indicates that all existing policies show poor performance against the (cid:80)i i−h rself h current opponent strategy, in this way Bayes-ToMoPk agent infers that the opponent is using a previously unseen policy outside the current policy library. Since we can easily obtain the average win rate θ(πj) of each policy π against each known opponent strategy j, the lowest win rate among the best-response policies (minπ∈Π maxj∈J θ(πj)) can be seen as an upper bound of the value of δ. The value of h controls the number of episodes be- fore considering whether the opponent is using a previously unseen strategy. Note that the accuracy of detection is in- creased with the increase of the memory length h, however, a larger value of h would necessarily increase the detection delay. The optimal memory length is determined empirically through extensive simulations (see supplementary materials). After detecting the opponent is using a new strategy, the agent begins to learn the best-response policy against it. Fol- lowing previous work [Hernandez-Leal et al., 2016], we adopt the same assumption that the opponent will not change its strategy during the learning phase (a number of rounds). Otherwise, the learning process may not converge. For tabu- lar Bayes-ToMoP, we adopt the traditional model-based RL: R-max [Brafman and Tennenholtz, 2002] to compute the op- timal policy. Specifically, once a new strategy is detected, R-max estimates the state transition function T and reward function R with T and R. It also keeps a transition count ) and total reward t(s, a) for all state-action pairs. c(s, a, s Each transition (cid:104)s, a, s , r(cid:105) results in an update for the tran- ) ← c(s, a, s sition count: c(s, a, s ) + 1 and total re- ward: t(s, a) ← t(s, a) + r. The estimates T and R is ) and t(s, a) with a given parame- updated using c(s, a, s ter n: T (s, a, s )/n, R(s, a) = t(s, a)/n if ) ≥ n. R-max computes Q(s, a) = R(s, a) + ) for all state-action pairs and selects the action that maximizes Q(s, a) according to -greedy mechanism. (cid:80) γ(cid:80) (cid:48) T (s, a, s s ) = c(s, a, s (cid:48) c(s, a, s s (cid:48) (cid:48) Q(s ) maxa (cid:48) , a (cid:48) (cid:48) (cid:48) (cid:48) (cid:48) (cid:48) (cid:48) (cid:48) (cid:48) (cid:48) For deep Bayes-ToMoP, we apply DQN [Mnih et al., 2015] to do off-policy learning using the obtained interaction expe- rience. DQN is a deep Q-learning method with experience re- play, consisting of a neural network approximating Q(s, a; θ) and a target network approximating Q(s, a; θ−). DQN draws , r) ∼ U (D) samples (or minibatches) of experience (s, a, s uniformly from a replay memory D, and updates using the following loss function: L(θ) = E(s,a,s (cid:48) ,r),r∼U (D)[(r + γ maxa(cid:48) Q(s(cid:48), a(cid:48); θ−) − Q(s, a; θ))2]. DQN is not a re- quirement, actually, our learning framework is general in which other DRL approaches can be applied as well. How- ever, most DRL algorithms suffer from the sample efficiency problems under some specific settings. This can be ad- dressed by incorporating sample efficient DRL methods in the future. To generate new performance models, we use a neural network to estimate the policy of the opponent based on the observed state-action history of the opponent using supervised learning techniques [Zheng et al., 2018; Foerster et al., 2018]. After the above learning phase, new performance models are generated using rewards obtained from a number of simu- lations of the agent's policy against the opponent's estimated strategy. These values are modeled as a Gaussian distribution to obtain the performance models. Finally, it adds the new policy πnew and the estimated opponent policy to its policy library Π and its opponent's policy library J respectively. 3.5 Theoretical Analysis In this section, we provide a theoretical analysis that Bayes- ToMoP can accurately detect the opponent's strategy from a known policy library and derives an optimal response policy accordingly. Theorem 1 (Optimality on Strategy Detection) If the oppo- nent plays a strategy from the known policy library, Bayes- ToMoP can detect the strategy w.p.1 and selects an optimal response policy accordingly. The proof is given in supplementary materials. 4 Simulations In this section, we present experimental results of Bayes- ToMoP compared with state-of-the-art tabular approaches (BPR+ [Hernandez-Leal et al., 2016] and Bayes-Pepper [Hernandez-Leal and Kaisers, 2017]). For deep Bayes- ToMoP, we compare with the following four baseline strate- gies: 1) BPR+, 2) Bayes-Pepper (BPR+ and Bayes-Pepper use a neural network as the value function approximator), 3) DRON [He and Boyd-Graber, 2016] and 4) deep BPR+ [Zheng et al., 2018]. We first evaluate the performance of Bayes-ToMoP by comparing it with state-of-the-art ap- proaches in both tabular and deep settings. We also compare the performance of Bayes-ToMoP and deep Bayes-ToMoP with previous works against an opponent using previously unseen strategies. The network structure and details of pa- rameter settings are in supplementary materials. 4.1 Game Settings We evaluate the performance of Bayes-ToMoP on the follow- ing testbeds: soccer [Littman, 1994; He and Boyd-Graber, 2016] and thieves and hunters [Goodrich et al., 2003; Cran- dall, 2012]. Soccer (Figure 1) and Thieves and hunters (Fig- ure 2) are two stochastic games both on a 7 × 7 grid. Two players, A and B, start at one of starting points in the left and right respectively and can choose one of the following 5 ac- tions: go left, right, up, down and stay. Any action that goes to grey-slash grids or beyond the border is invalid. 1) In soc- cer, the ball (circle) is randomly assigned to one player ini- tially. The possession of the ball switches when two players move to the same grid. A player scores one point if it takes the ball to its opponent's goals; 2) in thieves and hunters, player A scores one point if two players move to one goal simulta- neously (A catches B), otherwise, it loses one point if player B moves to one goal without being caught. If neither player gets a score within 50 steps, the game ends with a tie. We consider two versions of the above games in both tab- ular and deep representations. For the tabular version, the state space includes few numbers of discrete states, in which Q-values can be represented in a tabular form. For the deep version, each state consists of different dimensions of infor- mation: for example, states in soccer includes coordinates of Figure 1: The soccer game. Figure 2: Thieves and hunters two agents and the ball possession. In this case, we evaluate the performance of deep Bayes-ToMoP. We manually design six kinds of policies for the opponent in soccer (differenti- ated by the directions of approaching the goal) and twenty- four kinds of policies for the opponent in thieves and hunters (differentiated by the orders of achieving the goal). A policy library of best-response policies are trained using Q-learning and DQN for Bayes-ToMoP and deep Bayes-ToMoP. Table 1: Average win rates with std.dev.(±) in soccer. Opponents / Methods BPRs DRON Deep BPR+ Bayes-ToMoP1 Ons OT oM oP0 OT oM oP0-s 49.78±1.71% 99.37±0.72% 66.31±0.57% 74.75±0.19% 76.54±0.16% 75.35±0.18% 71.57±1.26% 99.49±0.51% 78.88±0.76% 99.82±0.18% 98.21±0.37% 98.48±0.54% 4.2 Performance against Different Opponents Three kinds of opponents are considered: 1) a Bayes-ToMoP0 opponent (OT oM oP0); 2) an opponent that randomly switches its policy among stationary strategies and lasts for an un- known number of episodes (Ons) and 3) an opponent switch- ing its strategy between stationary strategies and Bayes- ToMoP0 (OT oM oP0-s). We assume an opponent only selects a policy from the known policy library. Thus Bayes-Pepper is functionally equivalent to BPR+ in our setting and we use BPRs to denote both strategies. Due to the space limitation, we only give experiments on the tabular form of Thieves and hunters and a deep version of soccer. (a) Thieves and hunters (b) Soccer (deep version) Figure 3: Against an opponent OT oM oP0 on different games. Figure 3 (a) shows that only Bayes-ToMoP1 can quickly detect the opponent's strategies and achieve the highest aver- age rewards. In contrast, BPRs fails against OT oM oP0. Simi- lar comparison results can be found for deep Bayes-ToMoP1 (Figure 3 (b)). This is because Bayes-ToMoP1 explicitly con- siders two-orders of belief to do recursive reasoning first and 0100200300400500Episodes-4-2024Average rewardsBayes-ToMoP1BPRs0100200300400500Episodes-2-101Average rewardsDeep Bayes-ToMoP1BPRsDRONDeep BPR+(a) Thieves and hunters (b) Soccer (deep version) (a) Thieves and hunters (b) Soccer (deep version) Figure 4: Against an opponent Ons on different games. Figure 5: Against an opponent OT oM oP0−s on different games. then derives an optimal policy against its opponent. How- ever, BPRs is essentially equivalent to Bayes-ToMoP0, i.e., it is like Bayes-ToMoP0 is under self-play. Therefore, nei- ther BPRs nor Bayes-ToMoP0 could take advantage of each other and the winning percentages are expected to be ap- proximately 50%. Average win rates shown in Table 1 also confirm our hypothesis. The results for Bayes-ToMoP1 un- der self-play can be found in supplementary materials. Deep BPR+ incorporates previous opponent's behaviors into BPR, however, their model only considers which kind of station- ary strategy the opponent is using, thus is not enough to de- tect the policy of opponent OT oM oP0. Figure 3 (b) shows DRON performs better than Deep BPR+ and BPRs since it explicitly considers the opponent's strategies. However, it fails to achieve the highest average rewards because the dy- namic adjustment of DRON cannot guarantee that the opti- mality against a particular type of opponents. Next, we present the results of playing against Ons that switches its policy among stationary strategies and lasts for 200 episodes. Figure 4 (a) shows the comparison of Bayes- ToMoP1 with BPRs, where both methods can quickly and ac- curately detect which stationary strategy the opponent is us- ing and derive the optimal policy against it. We observe that Bayes-ToMoP1 requires longer time to detect than BPRs and similar comparison is found in Figure 4 (b) that deep Bayes- ToMoP1 requires longer time to detect than BPRs and deep BPR+. This happens because 1) Bayes-ToMoP1 needs longer time to determine that the opponent is not using a BPR-based strategy; 2) some stochastic factors such as the random ini- tialization of belief. This phenomenon is consistent with the slightly lower win rate of Bayes-ToMoP1 than BPRs and deep BPR+ (Table 1). The slight performance decrease against Ons is worthwhile since Bayes-ToMoP1 performs much bet- ter than deep BPR+ against other two types of advanced op- ponents (OT oM oP0 and OT oM oP0−s) as shown in Table 1. However, DRON only achieves the average rewards of 0.7 since it uses end-to-end trained response subnetworks, which cannot guarantee that each response is good enough against a particular type of opponent. Finally, we consider the case of playing against OT oM oP0- s to show the robustness of Bayes-ToMoP. Figure 5 (a) shows that only Bayes-ToMoP1 can quickly and accurately detect the strategies of both opponent OT oM oP0 and non-stationary opponent. In contrast, BPRs fails when its opponent's strat- egy switches to Bayes-ToMoP0. A similar comparison can be found in soccer shown in Figure 5 (b) in which both BPRs and deep BPR+ fail to detect and respond to OT oM oP0-s oppo- (a) Thieves and hunters (b) Soccer (deep version) Figure 6: Against an opponent using a new policy on two games. nent. Figure 5 (b) also shows DRON performs poorly against both kinds of opponents due to the similar reason described above. Average win rates are summarized in Table 1 which are consistent with the results in Figure 5 (b). 4.3 New Opponent Detection and Learning In this section, we evaluate Bayes-ToMoP1 against an oppo- nent who may use previously unknown strategies. We manu- ally design a new opponent strategy in soccer (the new policy is different from the policies in the library at starting from a different grid and approaching the goal following a different direction) and thieves and hunters (different from the poli- cies in the library at approaching the four destinations in a different order). The opponent stars with one of the known strategies and switches to the new one at the 200th episode. Figure 6 shows the average rewards of different approaches on two games respectively. Figure 6 (a) shows that when the opponent switches to an unknown strategy, both Bayes-ToMoP1 and BPR+ can quickly detect this change, and finally learn an optimal policy. However, Bayes-Pepper fails. This is because Bayes-Pepper predicts the strategy of its opponent from a known policy li- brary, which makes it fail to well respond to a previously un- seen strategy. The learning curve of Bayes-ToMoP1 is closer to BPR+ since both methods learn from scratch. Similar re- sults can be found for deep Bayes-ToMoP1 and BPR+ in Fig- ure 6 (b). DRON fails against the unknown opponent strategy due to the fact that the number of expert networks is fixed and thus unable to handle such a case. Deep BPR+ performs better than the other three methods since it uses policy distil- lation to transfer knowledge from similar previous policies to accelerate online learning, which can be readily applied to our Bayes-ToMoP to accelerate online learning as future work. 0200400600Episodes-4-2024Average rewardsBayes-ToMoP1BPRs0200400600Episodes-101Average rewardsDeep Bayes-ToMoP1BPRsDeep BPR+DRON0200400600Episodes-8-6-4-2024Average rewardsBayes-ToMoP1BPRs0200400600Episodes-2-101Average rewardsDeep Bayes-ToMoP1BPRsDeep BPR+DRON0200400600800Episodes-101Average rewardsBPR+Bayes-ToMoP1Bayes-Pepper0200400600800Episodes-101Average rewardsDeep BPR+Deep Bayes-ToMoP1DRON5 Conclusion and Future Work This paper presents a novel algorithm called Bayes-ToMoP to handle not only switching, non-stationary opponents and also more sophisticated ones (e.g., BPR-based). Bayes-ToMoP also enables an agent to learn a new optimal policy when en- countering a previously unseen strategy. Theoretical guaran- tees are provided for the optimal detection of the opponent's strategies. Extensive simulations show Bayes-ToMoP outper- forms the state-of-the-art approaches both in tabular and deep learning environments. Bayes-ToMoP can be seen as a generalized framework to reason and detect the policy change of an opponent, in which any recent deep RL algorithms can be incorporated to address large-scale state and action space problems. Currently we only use DQN to handle large-scale state space, while policy- based DRL (e.g., DDPG) can be used for problems with con- tinuous actions. On the other hand, Bayes-ToMoP can only handle two-player games, and it is worthwhile investigating how to apply theory of mind in multiple agents' scenarios to our Bayes-ToMoP framework as future work. Furthermore, how to accelerate the online new policy learning phase and higher orders of Bayes-ToMoP are worth investigating to han- dle more sophisticated opponents and apply to large scale, real scenarios. Acknowledgments The work is supported by the National Natural Science Foun- dation of China (Grant Nos.: 61702362, U1836214), Special Program of Artificial Intelligence, Special Program of Arti- ficial Intelligence and Special Program of Artificial Intelli- gence of Tianjin Municipal Science and Technology Com- mission (No.: 569 17ZXRGGX00150) and Huawei Noah's Ark Lab (Grant No.: YBN2018055043). References [Albrecht and Stone, 2018] Stefano V. Albrecht and Peter Stone. Autonomous agents modelling other agents: A comprehensive survey and open problems. Artificial In- telligence, 258:66 -- 95, 2018. [Baker et al., 2011] Chris L. Baker, Rebecca Saxe, and Joshua B. Tenenbaum. Bayesian theory of mind: Mod- eling joint belief-desire attribution. In Proceedings of An- nual Meeting of the Cognitive Science Society, 2011. [Brafman and Tennenholtz, 2002] Ronen I. Brafman and Moshe Tennenholtz. R-MAX - A general polynomial time algorithm for near-optimal reinforcement learning. Jour- nal of Machine Learning Research, 3:213 -- 231, 2002. [Crandall, 2012] Jacob W. Crandall. Just add pepper: ex- tending learning algorithms for repeated matrix games to repeated markov games. In Proceedings of International Conference on Autonomous Agents and Multiagent Sys- tems, pages 399 -- 406, 2012. [de Weerd et al., 2013] Harmen de Weerd, Rineke Ver- brugge, and Bart Verheij. How much does it help to know what she knows you know? an agent-based simulation study. Artificial Intelligence, 199:67 -- 92, 2013. [Fern´andez et al., 2010] Fernando Fern´andez, Javier Garc´ıa, and Manuela M. Veloso. Probabilistic policy reuse for inter-task transfer learning. Robotics and Autonomous Sys- tems, 58(7):866 -- 871, 2010. [Foerster et al., 2018] Jakob Foerster, Richard Y Chen, Maruan Al-Shedivat, Shimon Whiteson, Pieter Abbeel, and Igor Mordatch. Learning with opponent-learning awareness. In Proceedings of International Conference on Autonomous Agents and MultiAgent Systems, pages 122 -- 130, 2018. [Gmytrasiewicz and Doshi, 2005] Piotr J. Gmytrasiewicz and Prashant Doshi. A framework for sequential planning in multi-agent settings. Journal of Artificial Intelligence Research, 24:49 -- 79, 2005. [Goodrich et al., 2003] M. A. Goodrich, J. W. Crandall, and Issues and J. R. Stimpson. Neglect tolerant teaming: In AAAI Spring Sympo. on Human Interac- dilemmas. tion with Autonomous Systems in Complex Environments, 2003. [He and Boyd-Graber, 2016] He He and Jordan L. Boyd- Graber. Opponent modeling in deep reinforcement learn- ing. In Proceedings of International Conference on Ma- chine Learning, pages 1804 -- 1813, 2016. [Hernandez-Leal and Kaisers, 2017] Pablo Hernandez-Leal and Michael Kaisers. Towards a fast detection of oppo- nents in repeated stochastic games. In Proceedings of In- ternational Conference on Autonomous Agents and Multi- agent Systems, pages 239 -- 257, 2017. [Hernandez-Leal et al., 2016] Pablo Hernandez-Leal, Matthew E. Taylor, Benjamin Rosman, Luis Enrique Sucar, and Enrique Munoz de Cote. Identifying and tracking switching, non-stationary opponents: A bayesian Interaction without Prior approach. Coordination, Papers from the 2016 AAAI Workshop, 2016. In Multiagent [Hernandez-Leal et al., 2017a] Pablo Hernandez-Leal, Michael Kaisers, Tim Baarslag, and Enrique Munoz de Cote. A survey of learning in multiagent environments: Dealing with non-stationarity. CoRR, abs/1707.09183, 2017. [Hernandez-Leal et al., 2017b] Pablo Hernandez-Leal, Yusen Zhan, Matthew E. Taylor, Luis Enrique Sucar, and Enrique Munoz de Cote. Efficiently detecting switches against non-stationary opponents. Autonomous Agents and Multi-Agent Systems, 31(4):767 -- 789, 2017. [Littman, 1994] Michael L. Littman. Markov games as a framework for multi-agent reinforcement learning. In Pro- ceedings of International Conference on Machine Learn- ing, pages 157 -- 163, 1994. [Mnih et al., 2015] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Belle- mare, Alex Graves, Martin A. Riedmiller, Andreas Fid- jeland, Georg Ostrovski, Stig Petersen, Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Kumaran, Daan Wierstra, Shane Legg, and Demis Has- sabis. Human-level control through deep reinforcement learning. Nature, 518(7540):529 -- 533, 2015. [Powers and Shoham, 2005] Rob Powers and Yoav Shoham. Learning against opponents with bounded memory. In Proceedings of International Joint Conference on Artifi- cial Intelligence, pages 817 -- 822, 2005. [Rosman et al., 2016] Benjamin Rosman, Majd Hawasly, and Subramanian Ramamoorthy. Bayesian policy reuse. Machine Learning, 104(1):99 -- 127, 2016. [Rudin, 1964] Walter Rudin. Principles of mathematical analysis, volume 3. 1964. [Shoham and Leyton-Brown, 2009] Yoav Shoham and Kevin Leyton-Brown. Multiagent Systems - Algorithmic, Game-Theoretic, and Logical Foundations. Cambridge University Press, 2009. [v. Neumann, 1928] J. v. Neumann. theorie der gesellschaftsspiele. Mathematische Annalen, 100(1):295 -- 320, 1928. Zur [Wunder et al., 2012] Michael Wunder, John Robert Yaros, Michael Kaisers, and Michael L. Littman. A framework for modeling population strategies by depth of reason- In Proceedings of International Conference on Au- ing. tonomous Agents and Multiagent Systems-Volume 2, pages 947 -- 954, 2012. [Yang et al., 2019] Tianpei Yang, Jianye Hao, Zhaopeng Meng, Yan Zheng, Chongjie Zhang, and Ze Zheng. Bayes- tomop: A fast detection and best response algorithm to- wards sophisticated opponents. In Proceedings of Interna- tional Conference on Autonomous Agents and MultiAgent Systems, pages 2282 -- 2284, 2019. [Zheng et al., 2018] Yan Zheng, Zhaopeng Meng, Jianye Hao, Zongzhang Zhang, Tianpei Yang, and Changjie Fan. A deep bayesian policy reuse approach against non- stationary agents. In Advances in Neural Information Pro- cessing Systems, pages 954 -- 964, 2018. Supplementary Materials Theorem 2 (Optimality on Strategy Detection) If the oppo- nent plays a strategy from the known policy library, Bayes- ToMoP can detect the strategy w.p.1 and selects an optimal response policy accordingly. Proof 1 Suppose the opponent strategy is j ∈ J at step t, the belief βt(j) is updated as follows: (cid:80) βt+1(j) = P (σtj, π)βt(j) (cid:48)∈J P (σtj (cid:48) j , π)βt(j (cid:48) ) where the signal σt is the average reward of policy π against the opponent strategy over last l episodes and can approxi- mate the expected payoff. Since Bayes-ToMoP already learned a best response against each opponent strategy in the offline phase, given the expected payoff against an opponent's strategy, there al- ways exists a corresponding best-response policy π ∈ Π, that makes the inequality P (σtj, π) ≥ P (σtj , π) establish for (cid:48) all ∀j (cid:48) ∈ J , i.e., the probability of receiving the expected payoff is larger than or equal to others. The equality holds when one policy can beat two or more types of opponent strategies. Besides, since βt(j) is bounded (0 ≤ βt(j) ≤ 1) and monotonically increasing (βt+1(j) − βt(j) > 0) which is deduced by the above-mentioned equation, based on the monotone convergence theorem [Rudin, 1964], we can eas- ily know the limit of sequence βt(j) exists. Thus, if we limit the two sides of the above Equation, the following equation establishes, P(σtj, π) = (cid:80) P(σtj(cid:48), π)βt(j(cid:48)), if f ∀j(cid:48) (cid:54)= j, βt(j) = 1, βt(j(cid:48)) = 0 j(cid:48)∈J So that Bayes-ToMoP can detect the strategy w.p.1 and selects an optimal response policy accordingly. As guaranteed by Theorem 2, Bayes-ToMoP behaves opti- mally when the opponent uses a strategy from the known pol- icy library. If the opponent is using a previously unseen strat- egy, following the new opponent strategy detection heuris- tic, this phenomenon can be exactly detected when the win- ning rate of Bayes-ToMoP during a fixed length of episodes is lower than the accepted threshold. Then Bayes-ToMoP be- gins to learn an optimal response policy. Network Architecture All experiments use the same parameter settings: c1 = 0.3, λ = 0.7, δ = 0.7 (experimentally selected). For deep Bayes-ToMoP1, we consider the DNN input which consists of different dimensions of environment information: for ex- ample, states in soccer includes coordinates of two agents and the ball possession. DQN we used has two fully-connected hidden layers both with 20 hidden units, the output layer is a fully-connected linear layer with a single output for each valid action. We train DQN each step with mini-batches of size 32 randomly sampled from a replay buffer of one mil- lion transitions. Parameters are copies to the target network every 500 episodes. The learning rate of DQN is 10−3 and the discount factor is 0.9. All results are averaged over 1000 runs. (a) RPS (b) Tabular soccer Figure 7: Different approaches against an opponent OT oM oP0 on different games. Performance against Different Opponents Rock-paper-scissors (RPS) [v. Neumann, 1928; Shoham and Leyton-Brown, 2009] is a two-player stateless game in which 0100200300Episodes-101Average rewardsBayes-ToMoP1BPRs0100200300400500Episodes-101Average rewardsBayes-ToMoP1BPRsquickly render additional performance gains marginal. The average adjustment counts stabilize around 2.8 when l > 35. We hypothesize that it is because the dynamic changes of the winning rate over a relatively small length of memory may be caused by noise thus resulting in inaccurate opponent type de- tection. As the increase of the memory length, the judgment about the opponent's types is more precise. However, as the memory length exceeds a certain threshold, the winning rate estimation is already accurate enough and thus the advantage of further increasing the memory length diminishes. Finally, the influence of threshold δ against opponent Ons is shown in Figure 12. We note that the average adjust- ment counts decrease as δ increases, but the decrease degree gradually stabilizes when δ is larger than 0.7. With the in- crease of the value of δ, the winning rate decreases to δ more quickly when the opponent switches its policy. Thus, Bayes- ToMoP1 detects the switching of the opponent's strategies more quickly. Similar with the results in Figure 11, as the value of δ exceeds a certain threshold, the advantage based on this heuristic diminishes. Figure 11: The impact of mem- ory length l. Figure 12: The impact of thresh- old δ. (a) RPS (b) Tabular soccer Figure 9: Different approaches against an opponent OT oM oP0−s on different games. two players simultaneously choose one of the three possi- ble actions 'rock' (R), 'paper' (P), or 'scissors' (S). If both choose the same action, the game ends in a tie. Otherwise, the player who chooses R wins against the one that chooses S, S wins against P, and P wins against R. Results on RPS and soccer games are shown as follows (Figure 7, 8, 9). (a) RPS (b) Tabular soccer Figure 8: Different approaches against an opponent Ons on different games. Bayes-ToMoP1 under Self-play Results (Figure 10) are the same as Bayes-ToMoP0 under self-play due to the similar reasons, thus it is worth inves- tigating higher order of Bayes-ToMoP to handle more kinds of opponents. Figure 10: Bayes-ToMoP1 under self-play. The Influence of Key Parameters In this section, we analyze the influence of key parameters on the performance of Bayes-ToMoP, e.g., the threshold δ and the memory length l. Figure 11 depicts the impact of the memory length l on the average adjustment counts of Bayes-ToMoP1 before tak- ing advantage of OT oM oP0. We observe a diminishing re- turn phenomenon: the average adjustment counts decrease quickly as the initial increase in the memory length, but 0200400600Episodes-101Average rewardsBayes-ToMoP1BPRs0200400600Episodes-101Average rewardsBayes-ToMoP1BPRs0200400600Episodes-101Average rewardsBayes-ToMoP1BPRs0200400600Episodes-101Average rewardsBayes-ToMoP1BPRs05001000Episodes00.51Average rewards10203040The length of memory l2.62.833.23.43.63.8Average adjustment counts0.50.60.70.80.9The value of the threshold 050100150200250Average adjustment counts
1602.03623	1	1602	2016-02-11T06:20:00	Dynamic Group Behaviors for Interactive Crowd Simulation	[ "cs.MA" ]	We present a new algorithm to simulate dynamic group behaviors for interactive multi-agent crowd simulation. Our approach is general and makes no assumption about the environment, shape, or size of the groups. We use the least effort principle to perform coherent group navigation and present efficient inter-group and intra-group maintenance techniques. We extend the reciprocal collision avoidance scheme to perform agent-group and group-group collision avoidance that can generate collision-free as well as coherent and trajectories. The additional overhead of dynamic group simulation is relatively small. We highlight its interactive performance on complex scenarios with hundreds of agents and compare the trajectory behaviors with real-world videos.	cs.MA	cs	Dynamic Group Behaviors for Interactive Crowd Simulation Liang He1 and Jia Pan2 and Sahil Narang1 and Wenping Wang2 and Dinesh Manocha1 ∗† Figure 1: Our method can generate dynamic group behaviors along with coherent and collision free navigation. We highlight the performance in a street-crossing scenario, where different groups are shown with different colors. Our approach can automatically adapt to the environment and the number, shape, and size of the groups can change dynamically. Abstract 6 1 0 2 b e F 1 1 ] A M . s c [ 1 v 3 2 6 3 0 . 2 0 6 1 : v i X r a We present a new algorithm to simulate dynamic group behaviors for interactive multi-agent crowd simulation. Our approach is general and makes no assumption about the environment, shape, or size of the groups. We use the least effort principle to perform coherent group navigation and present efficient inter-group and intra- group maintenance techniques. We extend the recipro- cal collision avoidance scheme to perform agent-group and group-group collision avoidance that can generate collision-free as well as coherent and trajectories. The additional overhead of dynamic group simulation is rel- atively small. We highlight its interactive performance on complex scenarios with hundreds of agents and com- pare the trajectory behaviors with real-world videos. 1 Introduction The problem of simulating the trajectories and be- haviors of a large number of human-like agents fre- quently arises in computer graphics, virtual reality, and computer-aided design. It includes generation of pedes- trian movements in a shared space and the collaboration between the agents governed by social norms and inter- actions. The resulting crowd simulation algorithms are used to generate plausible simulations for games and animation, as well as accurately predicting the crowd flow and patterns in architectural models and urban en- vironments. One of the main challenges is modeling differ- ent behaviors corresponding to navigation, collision- avoidance, and social interactions that lead to self- organization and emergent phenomena in crowds. Prior research and observations in sociology and behavioral psychology have suggested that real-world crowds are composed of (social) groups. The group is a meso-level concept and is composed of two or more agents that share similar goals, over a short or long period of time, and exhibit similar movements or behaviors. In many instances, up to 70% of observed pedestrians are walk- ing in such groups [1, 2]. As a result, it is important to model the group dynamics that includes intra-group and inter-group interactions within a crowd. In this paper, we address the problem of simulating the group behaviors that are similar to those observed in real-world scenarios. In crowds, small groups are dy- namically formed as some of the agents move towards their goals and generate behaviors such as aggregation, dispersion, following the leader, etc. As the individual agents respond to a situation (e.g. panic or evacuation), the dynamic behaviors can result in splitting a large group or new groups being formed. Such group behav- iors are frequently observed in public places, sporting events, street-crossing, cluttered areas when the pedes- trians tend to avoid the obstacles, etc. Furthermore, the geometric shape of the group and the size of these groups may change. Some earlier observations have suggested that group sizes different according to a Pois- son distribution [3]. Prior work on modeling group behaviors is mostly limited to cohesive movements or spatial group struc- tures. The simplest algorithms cluster the agents into a fixed number of groups and the size of each group re- mains fixed (i.e. static grouping). They are unable to model the changing shape or size of the group, splitting of a large group into sub-groups or merging of smaller groups into a large group. Furthermore, in some sce- narios an agent may switch from one group to the other group in close proximity. It is important to model such dynamic behavior in arbitrary environments. Main Results: We present a novel algorithm to gener- ate dynamic group behaviors using multi-agent crowd simulation. Our approach is general and makes no assumptions about the number, size, or shape of the 1 groups. We use spatial clustering techniques to generate group assignments that take into account the positions and velocities of the agent. Our group-level navigation algorithm is based on the principle of least effort that tends to maintain the group relationships as each agent proceeds towards its goal position. We present efficient inter-group and intra-group level techniques to perform coherent and collision-free navigation. The group shape and sizes are automatically updated as new agents are assigned to the group or when some agents leave the group. We extend the agent-agent reciprocal collision avoid- ance algorithm [4] to perform agent-group and group- group reciprocal collision avoidance. We formulate the velocity obstacle of the group in terms of the convex hull of the current agent positions and use that to per- form conservative collision avoidance. Our approach is used to compute the new preferred velocity for each agent that not only avoids collisions with other agents and obstacles, but also performs coherent group navi- gation. This makes it possible to handle high-density crowds as well as arbitrarily shaped groups. The overall approach has been implemented and we highlight its performance on many complex bench- marks with dynamic group behaviors. The additional overhead of group computation and maintenance is rela- tively small and our approach takes a few milli-seconds per frame on scenarios with hundreds of agents. Our formulation can generate smooth and coherent group- level trajectories and we demonstrate the benefits over prior methods based on agent-based or meso-scale al- gorithms. We compare the trajectory behaviors gener- ated by our algorithm with real-world crowd videos by extracting the pedestrian trajectories. Overall, our ap- proach offers the following benefits: 1. Our approach is general and makes no assumption about the environment, size or shape of the groups. 2. We present an efficient algorithm for agent-group and group-group collision avoidance by extending the reciprocal velocity obstacle formulation. 3. Our approach can generate dynamic group behav- iors in terms of formation, merging, splitting, and re-assignment. 4. We observe plausible group behaviors and smooth trajectories, similar to those observed in real-world crowds. The rest of the paper is organized as follows. We briefly survey prior work in crowd simulation and group behaviors in Section 2. We introduce the notation and give an overview of our approach in Section 3. The overall algorithm is described in Section 4, and we high- light its performance in Section 5. 2 Related Work In this section, we give a brief overview of prior work on crowd simulation and group behaviors. 2.1 Crowd Simulation There is extensive work on modeling the behavior of crowds. These include multi-agent simulation tech- niques for computing collision-free trajectories and navigation based on social forces [5], rule-based meth- ods [6, 7], geometric optimization [4, 8, 9], vision-based steering [10], cognitive methods [11], personality mod- els [12], etc. Other class of simulation algorithms are based on data-driven methods [13, 14] and estimating the simulation parameters based on real-world crowd data [15, 16]. The macroscopic simulation algorithms compute fields for pedestrians to follow based on con- tinuum flows [17] or fluid models [18] and are mainly used for high-density crowds. 2.2 Group Behavior Simulation Group behaviors have been studied in computer graph- ics [19, 20, 21, 22], robotics [23], pedestrian dynam- ics [2], and social psychology [24]. Prior techniques have been mainly used to simulate static or fixed- sized groups and perform group-based collision avoid- ance [25, 26, 27]. However, none of these methods can efficiently simulate dynamic groups of varying sizes in arbitrary environments. Golas et al. [28] have pro- posed a hybrid approach that combines microscopic and macroscopic methods, and generates grouping behav- iors by taking into account long range trajectory pre- dictions. However, this approach cannot generate stable grouping behavior, and long range prediction can be ex- pensive. Recently, a distributed following strategy [29] has been proposed for dynamic behaviors, but is lim- ited to scenes with a few agents and cannot simulate ar- bitrary merging and splitting behaviors or handle large number of groups. 3 Overview In this section, we introduce our notation and give an overview of our approach. 3.1 Dynamic Grouping Behavior Our approach is designed for multi-agent crowd simula- tion algorithms. We assume that during each step of the simulation, each agent in the crowd has an intermediate goal position that is used to compute its preferred ve- locity. This goal position can change over the course of the simulation. The notion of dynamic grouping is mo- tivated by real-world crowd observations. Many stud- ies have highlighted the importance of group dynamics in the context of modeling the interactions between the 2 velocity vG during τ: V Oτ aG = {v∃t ∈ [0, τ ] such that pa + (v − vG)t ∈ CH(G) ⊕ D(0, ra)}, (1) where D(0, ra) is a disc centered at the origin with radius ra, and CH(G) is convex hull of the set of agents constituting the group G. The convex hull pro- vides a conservative bound that can guarantee collision free navigation. This equation implies that if agent a chooses a velocity outside the velocity obstacle V OaG, it will not collide with group G within the time window τ. Intuitively, the velocity obstacle can be geometrically constructed as a cone with apex in pa and its sides tan- gent to CH(G) expanded by the radius ra of the agent a, which is then translated by vG, as shown in Figure 3. From the geometric interpretation, we can observe that the convex hull CH(G) need not be computed explic- itly, instead the velocity obstacle can be fully defined by the extreme agents in the radial directions of the group, as observed from pa. We denote the most "clockwise" a and the most "counterclockwise" agent as agent as er a are used to compute a and el a. These two agents er el collision free trajectory of agent a. Figure 3: Agent-group Velocity Obstacle: The velocity ob- aG for agent a induced by a group G of agents. The stacle V Oτ group G contains six agents and its convex hull is the dashed line. If G only contains a single agent, V OaG reduces to the traditional velocity obstacle between two agents. The black agents el a are two most extreme agents in the group G. The angle θ represents the steering angle required by agent a to avoid the group of agents G. a and er Figure 2: Algorithm pipeline: We show the various compo- nents of our algorithm for dynamic group behaviors. agents and with the objects in the environment [30]. The number of such groups or the size of each group (i.e. number of agents) can change during the course of the simulation. The dynamic grouping behavior within a crowd is classified based on how the agents are dynamically clus- tered into groups. Given a set of n independent agents sharing a (2D) environment consisting of obstacles, we automatically compute these groups using spatial and temporal clustering algorithms. In particular, given the current position pa and velocity va of an agent a, we need to cluster all agents into a set of groups {Gi} according to a pairwise similarity metric defined over the agents. It is possible that some agent may not be- long to any group and is treated as an isolated agent. The specific group assigned to an agent a is denoted as Ga ∈ {Gi}. We also compute the velocity of a group G as the average velocity of all agents belonging to G, and is denoted as vG. During the simulation, the number of agents in a group G may change or or may maintain the group formation. For example, nearby agents with sim- ilar goals and similar directions of motion will merge into a group and maintain the group by following one after another. A large group may split into several sub- groups while facing an obstacle or other groups, and these sub-groups may merge into one group after pass- ing the obstacle. As two groups come close to each other, it is possible that agents may switch from one group to the other (i.e. reassign groups for an agent). As a result, it is important to support such group be- haviors corresponding to formation, merging, splitting, reassignment, etc. 3.2 Agent-Group Velocity Obstacle For collision avoidance between the agents, we use the concept of velocity obstacles [4]. In order to perform collision avoidance between groups, we use the notion of velocity obstacle V OaG for one agent a induced by a group G. Given the velocity of the group vG, V OaG can be defined as the set of agent a's velocities va that will result in a collision with G at some point within time window τ assuming that the group G maintains its 3.3 Our Approach Our goal is to generate realistic dynamic grouping be- haviors for pedestrians. We assign the agents to differ- ent group during each frame and compute its trajectories by taking into account group dynamics. In cluttered ar- eas, the agents tend to be assigned to a large group, and in open areas the agents tend to be well spread out and may not be assigned to any group. As a result, we need the capabilities to support such dynamic merging and splitting behaviors. 3 sense positionsand velocities ofother robotsgroup formation& updatemerge & splitgroup maintenance& navigationgroup-groupcollision avoidanceagent-agentcollision avoidanceapply velocity torobot's actuators /update positionaGvGvGVOτaGvaelaeraθFurthermore, our approach is motivated by the prin- ciple of least effort [31] that has been used for comput- ing the agent trajectories in prior crowd simulation al- gorithms. An agent aligns itself with a group such that the resulting motion is governed by effort minimization. In particular, given the preferred velocity for each agent a, we tend to compute the actual velocity that tends to minimize the effort required by the entire group Ga to avoid the obstacles. In order to support dynamic groups, our approach supports the following computations: Group formation: We use spatial clustering algorithm to generate the initial group assignment for each agent in the crowd. The isolated agents are not assigned to any group. Group maintenance and navigation: Our approach tends to maintain these groups as long as possible dur- ing the navigation. All the agents belonging to a group exhibit coherent trajectories and behaviors. We per- form inter-group and intra-group computations to gen- erate such behaviors. At the inter-group level, each group needs to perform high-level coherent trajectory computation to avoid collisions with other groups and obstacles. The collision avoidance policy is chosen in a manner that if all agents in the same group consis- tently make the same choice, the entire group tends to avoid other groups altogether. At the intra-group level, each agent inside a group (except the group leader) will choose one fellow agent from the same group, and follow it to make progress towards the goal. If each agent in the group follows this policy, our approach doesn't need to explicitly check for agent-agent colli- sions within a group. Group update: The group assignments are updated and the number of agents belonging to a group may change. A key component for trajectory computation is an ef- ficient group-group collision avoidance algorithm. In our approach, this is achieved by first avoiding the colli- sions between the group leader of each group and other groups, and then determining a suitable preferred ve- locity for other non-leader agents. In particular, we use OCRA-based agent-group collision avoidance tech- nique [4] to compute the velocity for the group leader. All the other agents in the same group will compute their velocity according to the following policy. The new adapted preferred velocity for each agent is used by the agent-agent OCRA algorithm to compute the ac- tual velocity for each agent by taking into account all the constraints. The preferred velocity is chosen such that it guides the agent towards its goal position. The various components of our approach are shown in Fig- ure 2. (a) group formation (b) group maintenance (c) group split (d) group re-merge (e) leave a group Figure 4: Dynamic group behaviors during the navigation. (a) The agents first are clustered into groups according to their position and velocities. Each agent will has its individual goal, e.g., agent ab, c's goals are ga, gb, gc (please see (e)), re- spectively. (b) After a while, two groups will run into each other, but both groups will maintain their constitution during the navigation. (c) For collision avoidance, one group (marked by the red dashed line) is split into two groups. (d) After these two groups pass through each other, the split groups merge back into a single group. (e) If one agent in a group can ap- proach its goal easily, it will choose to leave the group and navigate alone. 4 Multi-agent Simulation In this section, we present our multi-agent simulation algorithm that can simulate dynamic grouping behav- iors. 4.1 Group Formation We use spatial clustering algorithm to compute the ini- tial group assignment for each agent. This assignment is based on the positions and velocities of all agents. The clustering criteria is based on the following crite- 4 agabcgbgacbgbagacbgbacgagcbgbagccgabgbaria. Given a pair of agents, a and b, they belong to the same group if the following conditions hold: • the position pa of agent a and the position pb of agent b are within a predefined distance p, and • the velocity va of agent a and the velocity vb of agent b are within a predefined threshold v. The transitive closure of this relation uniquely classified each cluster into groups, and can be formally described as (a ∼ b) ≡ ((cid:107)pb − pa(cid:107) < p ∧ (cid:107)vb − va(cid:107) < v), where ∼ is the binary operator defining whether two agents would be grouped together. Given this criteria for grouping, we use a greedy algorithm to compute these groups {Gi} in O(nk) time, where n are the num- ber of agents in the crowd and k is the number of groups in the partition. In particular, we iteratively check each agent whether it can be grouped into any existing groups according to the ∼ relationship. If an agent is not as- signed to any group, it is treat as a single or isolated agent during that frame. 4.2 Group Maintenance and Navigation One key point in simulating the group behavior for a crowd is how to maintain the groups based on colli- sion avoidance constraints during the navigation. We achieve the group maintenance by using a two-level ap- proach: the inter-group level makes sure that the entire group will avoid other groups as a whole, and the intra- group level ensures that all the agents belonging to a group do not collide with each other. 4.2.1 Inter-Group Level In most multi-agent simulation algorithms, each agent independently computes its current velocity for colli- sion avoidance. However, such navigation algorithms may not be able to maintain the group-like coherent mo- tion. This is because each agent may choose different extreme agents (as shown in Figure 3) from the same group to avoid collision, due to their difference in po- sitions and velocities relative to the obstacle group. In- stead, we would like that each agent in the same group as a should select the identical side (all el or er) while bypassing one group G. For this purpose, we first estimate the effort required for agent a to bypass one obstacle group G as Ea = (va − vG) × (pa − pG) · n, (2) where vG and pG are the average velocity and posi- tion of the group G respectively, and n is the normal of the 2D plane. As shown in Figure 3, this effort mea- surement is the sine function with the steering angle θ required by the agent to avoid with the obstacle group. 5 puted as E = (cid:80) (cid:40) Then the total effort for the entire group Ga can be com- b∈Ga Eb, and the bypassing side (for navigation) is computed as: s = r (right) l (left) if E < 0 otherwise. (3) In other words, each agent would choose the same by- passing side which has a smaller effort for collision avoidance. The solution of Equation 3 provides an ini- tial direction of motion for each agent. In this way, the group Ga will avoid the group G as a whole. Figure 5: Group-group collision avoidance: Our approach can compute collision-free as well as coherent trajectories for agents in each group. When the group Ga needs to avoid a set of different groups {Gi}, it first randomly selects one group Gi that may collide with it, and then computes the bypassing side s and the extreme agent es i for it. Next, it repeat- edly checks whether there are any other groups that may collide with it on the side s. If yes and suppose that par- ticular group is Gj, then it will choose to bypass Gj also from the side s and the corresponding extreme agent is j. If not, then the iteration stops and the extreme agent es is computed. One example for this process is illustrated in Figure 5. Suppose we are computing the bypassing side and extreme agent for group G1 which first chooses to avoid G2. It decides to bypass from the left side and 2. Since both G4 and the corresponding extreme agent el 2, we need to further G3 are both to the left side of el check for collision avoidance. Lets assume that we se- lect group G3 during the next step. To be consistent with the decision of avoiding G2, we continue to by- 3 as the pass group G3 from the left side, and choose el extreme agent. Since there is no more groups to the left 3, the iterative process stops. In this way, group G1 of el will bypass group G3 from the left side, as shown by the red trajectory in Figure 5. The trajectories for other groups can be computed in a similar manner. 4.2.2 Intra-Group Level After computing the bypassing side for the entire group, we can achieve coherent navigation within a group. G1G2G3G4el4el2el3However, this may not be sufficient to avoid the reas- signment, i.e., the exchange of agents between different groups. First, the bypassing decision in the inter-group level depends on the extreme agents of a group, which is computed based on the group's convex hull (see Equa- tion 1 and Figure 3). If the convex hulls of different groups overlap with each other, some agents in the same group may be isolated by the agents from other groups. In some other cases, the group needs to deform its shape (e.g., from a circle shape into a line shape) in order to maintain the group coherence for navigation a cluttered environment, and thus bypassing other groups from the same side may not be sufficient. To reliably avoid group reassignment, we need to keep agents connected during the navigation. In or- der to simulate this trajectory behavior, we use the dy- namic following strategy. In particular, we let each agent dynamically follow some other agents in the same group whenever possible. In this way, the members in a group will move along the same local path and will have the minimal risk for group reassignment or collisions with other agents. To achieve this behavior, we first need to decide whether one agent should be a leader or a follower in the group, and if it is a fol- lower, we need to determine whom it should follow. Suppose the group G chooses to bypass another group from the side s, then G's member agents all have es – the extreme agent in the obstacle group on the side s – as the temporary goal gG. We choose the leader of group G as the member that is closest to es, i.e., leader = arg mina∈G (cid:107)pa − gG(cid:107). All other members would be treated as the followers. If an agent a is a follower, we choose its following target as follows. First, we find all agents b in the group that satisfy (cid:107)pb − gG(cid:107) < (cid:107)pa − gG(cid:107), and the set of all qualified agents is denoted as F . In order to com- pute a stable connected group, we choose a's follow- ing target as the one in F that is closest to a. This is because if b is too far away from a then when tried to follow b, the group shape may change considerably, which makes it difficult to perform group-level collision avoidance. Formally, a's following target is selected as b∗ = arg minb∈F,b(cid:54)=a (cid:107)pb − pa(cid:107). 4.3 Group Update The group update or reassignment happens under two situations. The first situation is while the agents are in the open area and can easily approach their goals. In this case, the group bypassing and dynamic following strate- gies are usually sub-optimal for an individual agent's trajectory, even though they are beneficial for the over- all navigation. As a result, the notion of being able to stop following at the suitable time will help improve the performance of multi-agent navigation system. We perform this step by checking whether the original pre- ferred velocity vpref will result in making the agent col- lide with any other agents. If not, the agent will detach from the group and uses the discrete agent local navi- gation algorithm based on ORCA to move towards its goal. The second situation arises when the current group setting is not able to compute a collision-free velocity for the navigation. This is mainly because the original groups have deformed too much during the navigation, and their shapes are become quite non-convex. Our so- lution is to perform re-clustering over the entire crowd, to generate a group partition that can better describe the current dynamic behavior of the pedestrian crowd. 4.4 Collision Avoidance Besides the high-level grouping behaviors, we also need to make sure that there is no collision between the indi- vidual agents in the crowd. However, prior agent-agent collision avoidance schemes such as ORCA [4] or so- cial forces [5] may not maintain the group assignment. Some recent methods [25] extend the traditional agent- agent velocity obstacle by considering each group as a super-agent. However, they assume the group shape and size is should be fixed during the navigation, and thus requires all agents in the group must always choose the same velocity. As a result, simple extensions of velocity obstacle may not be able to find a feasible solution and does not work in cluttered environments where group deformation and/or reassignment are necessary for col- lision avoidance. Instead, we use a two-level to keep grouping behavior and make sure safe navigation simul- taneously. 4.4.1 Group-Group Collision Avoidance For the group-group collision avoidance, we leverage the result from the following strategy in Section 4.2.2. Given the leader agent a of one group Ga, we first com- pute the adapted preferred velocity of a that can avoid all the other groups. In order to avoid the collision with other groups within time τ, the agent a should choose the actual adapted velocity vadapt that is outside the union of the velocity obstacles with respect to each of the groups but is also closest to the preferred velocity, i.e., a vadapt a = v /∈(cid:83) arg min G∈{Gi}−Ga V Oτ aG (cid:107)v − vpref a (cid:107), (4) where V Oτ duced by the group G, as defined in Equation 1. aG is the velocity obstacle for agent a in- Once the leader's new preferred velocity is computed, we can calculate the preferred velocity for all other agents in the group Ga iteratively. In particular, we first compute all the agents {b} that follow the leader a, and the new preferred velocity vadapt is set by pro- jecting the old preferred velocity along the direction pb − pa. Once the adapted preferred velocity is com- puted for each agent in {b}, we continue to find the new b 6 5.1 Real-World Scenarios and Validation We compare the crowd simulation results using our dy- namic group behavior generation algorithm and prior approaches on scenarios inspired by real-world crowd videos. We extract the trajectories of the agents in the real-world video using a pedestrian tracking algo- rithm [32]. For each crowd simulation algorithm, the number of agents and their initial positions and goal po- sitions are assigned according to the pedestrian track- ing results. Given the initial and goal positions, we compare the trajectories of the pedestrians generated by each algorithm and compare them with those in the real videos in Figure 10. Figures 7 and 8 show the key frame for simulation sequences generated using different ap- proaches. We can observe that the simulation results using our dynamic group generation algorithm is most similar to the real world pedestrians in terms of trajec- tory behaviors. In terms of quantitative comparison, we evaluate the behavior of real pedestrians with that of simulated crowds by checking: 1. Compare the running time and number of colli- sions that occurred during the navigation from the initial to the goal positions, as shown in Table 2. 2. Compare the trajectories extracted using the track- ing algorithm (i.e. the ground truth) for some of the agents with the trajectories computed by different multi-agent simulation algorithms. In the first benchmark, agents are passing through a crosswalk as shown in Figure 1. During this simu- lation, agents automatically aggregate into groups and perform group-level collision avoidance. In this bench- mark, the total time taken by different crowd simulation approaches is almost similar. However, our dynamic group behavior approach result in fewer collisions be- tween the agents during navigation. Moreover, the tra- jectories generated using our algorithm have a better match with the ground truth data, as shown in Figure 7. This is due to the fact that ORCA and meso-scale simu- lation algorithms need more space to perform collision avoidance and therefore the agents are more spread out. In the second benchmark, agents are moving in a hallway inside the building, which represents a tight space. In this simulation, each agent's initial position and direction of movement is computed based on the real-world trajectories. Our approach can compute the navigation trajectories with a few collisions with coher- ent grouping behaviors, similar to real-world videos. In contrast, the agents in ORCA and meso-scale simula- tion algorithms take more time to move from the initial to the goal positions due to the tight spaces. Moreover, the trajectories computed by our algorithm are smoother and there is high co-relation with the ground truth data, i.e. the extracted trajectories. The third benchmark corresponds to a cluttered envi- ronment where the agents need to go through the hall- Figure 6: Group-group collision avoidance: a is the leader and the dashed lines illustrate the following relationship. The black vectors are the input preferred velocities, and the red vectors are the new preferred velocities computed by our al- gorithm. They tend to avoid collisions with the other agents and used for coherent group navigation. velocity for the followers. This iterative process con- tinues until we compute the new preferred velocities for all agents in the group (see Figure 6). 4.4.2 Agent-Agent Collision Avoidance The new preferred velocity computed is used as the in- put to the ORCA agent-agent collision avoidance algo- rithm [4] that finally computes the current velocity for each agent. The ORCA algorithm ensures the agent avoids collisions with nearby individual agents. The agent need only avoid pairwise collisions with imme- diately neighboring agents. This computation is per- formed independently for each agent. 5 Implementation mance and Perfor- In this section we describe our implementation and highlight the performance of our algorithm on different benchmarks. We compare our result with the group- ing behaviors generated by two state-of-the-art crowd simulators: agent-agent collision avoidance algorithm OCRA [4] and a group-based meso-scale navigation ap- proach [26]. We use five benchmarks to evaluate our algorithms and three of them are designed from real- world videos, and we compare the movement trajecto- ries generated by different approaches; and two other synthetic benchmarks, where we also compare the run- ning time and the number of collisions between the agents during the simulation. We have implemented our algorithms in C++ on an Intel Core i7 CPU running at 3.30GHz with 16GB of RAM and running Windows 7. All the timing results are generated on a single core. In practice, our 7 avprefavadapta(a) real-world video frame (b) ORCA (c) meso-scale (d) our method Figure 7: For benchmark 1, we compare the group behavior generated by our algorithm (d) on a real-world scenario (a). As compared to ORCA (b) and meso-scale (c), our approach can generate smoother and coherent trajectories. (a) real-world video frame (b) ORCA (c) meso-scale (d) our method Figure 8: Comparison between the key frame for simulation sequences generated using different approaches on benchmark 3. way, as shown in Figure 9. Both RVO and meso-scale methods are not able to compute collision-free naviga- tion as the crowd density is high. Instead, our method automatically enables the agents to move in groups and compute collision-free trajectories. We also observe that the trajectories computed using our algorithm have a better match with the ground truth data. 5.2 Other Benchmarks We also generated some synthetic scenes to further eval- uate the performance of our dynamic group behavior generation algorithm. In the fourth benchmark, agents are randomly placed in the scenario. Our approach au- tomatically cluster them into groups and generates co- herent trajectories. Furthermore, it results in fewer col- lisions and smoother trajectories. The fifth benchmark corresponds to adding several static obstacles in the en- vironment corresponding to the fourth benchmark. Our method can compute the paths to the goal position for each agent. On the other hand, the agents get stuck and pushed away from the goal position within ORCA and meso-scale simulation 6 Limitations, Conclusions and Future Work We present a novel multi-agent navigation algorithm that can automatically generate dynamical grouping be- haviors. Our approach is general and makes no assump- tion about the size or shape of the group, and can dy- namically adapt to the environment. Moreover, it results in smooth and coherent navigation behaviors as com- pared to prior multi-agent reciprocal collision avoid- ance algorithms. Furthermore, the agent's tend to avoid congestion based on group's follow-the-leader trajec- tory computation behavior, which is similar to human behaviors observed in real-world behaviors. We demon- strate its performance on complex benchmarks with a few hundred agents and show that the trajectories gen- erated by our algorithm are similar to those observed in real-world behaviors and exhibit similar group behav- iors. Unlike prior group behavior simulation schemes, our approach is adaptive and can model the dynamic behaviors of the agent in response to the environment. Our approach has some limitations. It is currently de- signed for homogeneous agents and the clustering algo- rithm only takes into account the position and velocity of each agent. We don't account for agents with varying personalities or how they respond to the environment ef- fects or situations or the psychological component cor- responding to the concept of personal space that varies along different cultures or the social norms. Our recip- rocal group-group collision avoidance algorithm can be conservative as it is implicitly based on the convex hull or extreme agents. There are many avenues for future work. In addi- tion to overcoming these limitations, we would like to evaluate its performance in complex scenarios with tens of thousands of agents (e.g. sporting events or religious gatherings). We would like to further validate its perfor- mance using other metrics, such as comparing with the collective behaviors and fundamental diagrams of real- world crowds. Finally, we would like to combine with macroscopic techniques to simulate very dense crowds. References [1] J. S. Coleman and J. James, "The equilibrium size distribution of freely-forming groups," Sociome- try, pp. 36–45, 1961. 8 Benchmark 1 #agents 49 #groups 8 Benchmark 2 #agents 97 #groups 9 Benchmark 3 #agents 185 #groups 6 Benchmark 4 #agents 184 #groups 11 Benchmark 5 #agents 151 #groups 9 Table 1: Number of agents and maximum number of groups in each benchmark. The number of groups change during the simulation. Figure 9: Key frames for the simulation sequence generated by our method on benchmark 3. (a) real-world video frame (b) ORCA (c) meso-scale (d) our method (e) real-world video frame (f) ORCA (g) meso-scale (h) our method (i) real-world video frame (j) ORCA (k) meso-scale (l) our method Figure 10: For all the three real-world benchmarks (the 1st column), we compare the trajectory behaviors generated by our algorithm (4th column, each color represents a group). As compared to ORCA (2nd column) and meso-scale (3rd column), our approach can generate smoother and coherent trajectories. Benchmark 1 Benchmark 2 Benchmark 3 Benchmark 4 Benchmark 5 Method ORCA Meso-scale Our (dynamic grouping) tpf 3.2 4.7 4.8 #steps 162 17.1 161 #colls 78 53 2 tpf 3.3 4.9 4.6 #steps 363 475 221 #colls 103 111 6 tpf 4.2 8.9 8.1 #steps 5000+ 5000+ 253 #colls 500+ 500+ 3 tpf 8.7 9.6 9.4 #steps 209 212 203 #colls 257 262 3 tpf 9.2 12.9 10.4 #steps 5000+ 5000+ 387 #colls 500+ 500+ 2 Table 2: The comparison between our approach and previous methods (ORCA, meso-scale) on five benchmarks. We report the average running time per frame (tpf) in milli-second, the average number of simulation time steps taken for each agent to reach the goal position (#steps) and the average number of pairwise collisions (#colls). These collisions can occur when the conservative collision avoidance schemes can't compute a feasible solution. In some case, the agents in the ORCA or meso-scale algorithms get stuck resulting in a high number of collisions. Even after 5000 simulation they have not reached the goal positions. We observe these behaviors with ORCA and meso-scale algorithms on Benchmark 3 and Benchmark 5. 9 [2] A. Gorrini, S. Bandini, and M. Sarvi, "Group dy- namics in pedestrian crowds: Estimating prox- emic behavior," Transportation Research Record: Journal of the Transportation Research Board, no. 2421, pp. 51–56, 2014. [3] J. James, "The distribution of free-forming small group size." American Sociological Review, 1953. [4] J. van den Berg, S. Guy, M. Lin, and D. Manocha, "Reciprocal n-body collision avoidance," in Robotics Research, ser. Springer Tracts in Ad- vanced Robotics, C. Pradalier, R. Siegwart, and G. Hirzinger, Eds. Springer Berlin Heidelberg, 2011, vol. 70, pp. 3–19. [5] D. Helbing and P. Molnar, "Social force model for pedestrian dynamics," Physical review E, 1995. [6] C. W. Reynolds, "Flocks, herds and schools: A distributed behavioral model," in SIGGRAPH, 1987, pp. 25–34. in large motion databases," in Symposium on Interactive 3D Graphics and Games, 2013, pp. 19–28. [Online]. Available: http://doi.acm.org/10. 1145/2448196.2448199 [15] D. Wolinski, S. J. Guy, A.-H. Olivier, M. C. Lin, D. Manocha, and J. Pettr´e, "Parameter estimation and comparative evaluation of crowd simulations," in Eurographics, 2014. [16] G. Berseth, M. Kapadia, B. Haworth, and P. Faloutsos, "Steerfit: Automated parameter fit- ting for steering algorithms," in Symposium on Computer Animation, 2014. [17] A. Treuille, S. Cooper, and Z. Popovi´c, "Contin- uum crowds," in SIGGRAPH, 2006, pp. 1160– 1168. [18] R. Narain, A. Golas, S. Curtis, and M. C. Lin, "Ag- gregate dynamics for dense crowd simulation," ACM Transactions on Graphics, vol. 28, no. 5, pp. 122:1–122:8, 2009. [7] N. Pelechano, J. M. Allbeck, and N. I. Badler, "Controlling individual agents in high-density crowd simulation," in Symposium on Computer animation, 2007, pp. 99–108. [19] K. H. Lee, M. G. Choi, Q. Hong, and J. Lee, "Group behavior from video: a data-driven ap- proach to crowd simulation," in Symposium on Computer Animation, 2007, pp. 109–118. [8] J. Pettr´e, J. Ondrej, A.-H. Olivier, A. Cretual, and S. Donikian, "Experiment-based modeling, simu- lation and validation of interactions between vir- tual walkers," in Symposium on Computer Anima- tion, 2009, pp. 189–198. [9] 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. [10] J. Ondrej, J. Pettr´e, A.-H. Olivier, and S. Donikian, "A synthetic-vision based steering approach for crowd simulation," ACM Transactions on Graph- ics, vol. 29, no. 4, pp. 123:1–123:9, 2010. [11] Q. Yu and D. Terzopoulos, "A decision network framework for the behavioral animation of virtual humans," in Symposium on Computer animation, 2007, pp. 119–128. [12] F. Durupinar, N. Pelechano, J. Allbeck, U. Gu anddu andkbay, and N. Badler, "How the ocean personality model affects the perception of crowds," Computer Graphics and Applications, vol. 31, no. 3, pp. 22 –31, 2011. [13] A. Lerner, Y. Chrysanthou, A. Shamir, and D. Cohen-Or, "Data driven evaluation of crowds," in Motion in Games, 2009, pp. 75–83. [14] M. Kapadia, I.-k. Chiang, T. Thomas, N. I. Badler, and J. T. Kider, Jr., "Efficient motion retrieval [20] S. I. Park, F. Quek, and Y. Cao, "Modeling social groups in crowds using common ground theory," in Winter Simulation Conference. Winter Simu- lation Conference, 2012, p. 113. [21] S. Curtis, J. Snape, and D. Manocha, "Way por- tals: Efficient multi-agent navigation with line- segment goals," in Symposium on Interactive 3D Graphics and Games, 2012, pp. 15–22. [22] S. Huerre, J. Lee, M. Lin, and C. O'Sullivan, "Simulating believable crowd and group behav- iors," in ACM SIGGRAPH ASIA 2010 Courses, 2010. [23] A. Krontiris, S. Louis, and K. Bekris, "Multi- level formation roadmaps for collision-free dy- namic shape changes with non-holonomic teams," in IEEE International Conference on Robotics and Automation, 2012, pp. 1570–1575. [24] E. S. Knowles, "Boundaries around group interac- tion: The effect of group size and member status on boundary permeability." Journal of Personal- ity and Social Psychology, vol. 26, no. 3, p. 327, 1973. [25] V. G. Santos and L. Chaimowicz, "Cohesion and segregation in swarm navigation," Robotica, vol. 32, pp. 209–223, 3 2014. [26] L. He and J. van den Berg, "Meso-scale planning for multi-agent navigation," in IEEE International Conference on Robotics and Automation, 2013, pp. 2839–2844. 10 [27] I. Karamouzas and S. Guy, "Prioritized group navigation with formation velocity obstacles," in IEEE International Conference on Robotics and Automation, 2015, pp. 5983–5989. [28] A. Golas, R. Narain, S. Curtis, and M. C. Lin, "Hybrid long-range collision avoidance for crowd simulation," IEEE Transactions on Visualization and Computer Graphics, vol. 20, no. 7, pp. 1022– 1034, 2014. [29] L. He, J. Pan, W. Wang, and D. Manocha, "Prox- emic group behaviors using reciprocal multi-agent navigation," Department of Computer Science, UNC Chapel Hill, Tech. Rep., 2015. [30] S. Reicher, "The psychology of crowd dynamics," Blackwell handbook of social psychology: Group processes, pp. 182–208, 2001. [31] S. J. Guy, J. Chhugani, S. Curtis, P. Dubey, M. Lin, and D. Manocha, "Pledestrians: a least-effort ap- proach to crowd simulation," in Symposium on computer animation, 2010, pp. 119–128. [32] A. Bera, S. Kim, and D. Manocha, "Efficient tra- jectory extraction and parameter learning for data- driven crowd simulation," in Graphics Interface, 2015, pp. 65–72. 11
1704.05905	1	1704	2017-04-19T19:25:55	A Coalition Formation Algorithm for Multi-Robot Task Allocation in Large-Scale Natural Disasters	[ "cs.MA", "cs.RO" ]	In large-scale natural disasters, humans are likely to fail when they attempt to reach high-risk sites or act in search and rescue operations. Robots, however, outdo their counterparts in surviving the hazards and handling the search and rescue missions due to their multiple and diverse sensing and actuation capabilities. The dynamic formation of optimal coalition of these heterogeneous robots for cost efficiency is very challenging and research in the area is gaining more and more attention. In this paper, we propose a novel heuristic. Since the population of robots in large-scale disaster settings is very large, we rely on Quantum Multi-Objective Particle Swarm Optimization (QMOPSO). The problem is modeled as a multi-objective optimization problem. Simulations with different test cases and metrics, and comparison with other algorithms such as NSGA-II and SPEA-II are carried out. The experimental results show that the proposed algorithm outperforms the existing algorithms not only in terms of convergence but also in terms of diversity and processing time.	cs.MA	cs	This paper has been accepted for presentation in IEEE IWCMC 2017 – The 13th International Wireless Communications and Mobile computing Conference (IWCMC) to be held in Valencia, Spain; June 26-30 June, 2017. This is an author copy. The respective Copyrights are with IEEE A Coalition Formation Algorithm for Multi-Robot Task Allocation in Large-Scale Natural Disasters Carla Mouradian¥, Jagruti Sahoo, Roch H. Glitho¥, Monique J. Morrow€, and Paul A. Polakos£ ¥Concordia University, Montreal, Canada, South Carolina State University, Orangeburg, South Carolina, USA € CISCO Systems, Zurich, Switzerland, [email protected], [email protected], [email protected] {mmorrow, ppolakos}@cisco.com £CISCO Systems, New Jersey, USA Abstract-In large-scale natural disasters, humans are likely to fail when they attempt to reach high-risk sites or act in search and rescue operations. Robots, however, outdo their counterparts in surviving the hazards and handling the search and rescue missions due to their multiple and diverse sensing and actuation capabilities. The dynamic formation of optimal coalition of these heterogeneous robots for cost efficiency is very challenging and research in the area is gaining more and more attention. In this paper, we propose a novel heuristic. Since the population of robots in large-scale disaster settings is very large, we rely on Quantum Multi-Objective Particle Swarm Optimization (QMOPSO). The problem is modeled as a multi-objective optimization problem. Simulations with different test cases and metrics, and comparison with other algorithms such as NSGA-II and SPEA-II are carried out. The experimental results show that the proposed algorithm outperforms the existing algorithms not only in terms of convergence but also in terms of diversity and processing time. Keywords- Coalition Formation, Multi-robot, Particle Swarm Optimization, Task Allocation I. INTRODUCTION In large-scale disasters scenarios, the primary goal is to find and rescue victims as quickly as possible. Conventional methods involving human rescuers and dogs have well-known limitations. The rescue teams often fail to reach the sites because of a fire break out, collapsed buildings, and closed roads when disaster occur. In contrast, the use of robots brings several advantages. Robots can, for instance, move quickly and find victims more accurately than their human counterparts. The use of robots for search and rescue mission was first witnessed during the rescue operations at the World Trade Center in New York City on September 2001, with the use of CRASAR (Center for Robot-Assisted Search And Rescue) rescue robots [1]. Robots have different capabilities. Forming dynamically an optimal coalition of robots with a required set of capabilities remains very challenging. It is even more challenging in large- scale disaster scenarios. This is due to the very large number of robots [2] needed in a single coalition in order to cover the whole disaster area and satisfy the search and rescue task requirements. The problem can be modeled as a multi-objective optimization problem where the time, the cost, and the number of robots in a coalition are minimized iteratively and simultaneously. In this paper, we propose a coalition formation algorithm to solve it. We address the ST-MR-IA (Single-Task Multi-Robots class of Multi-Robot Task taxonomy Instantaneous-Assignment) Allocation (MRTA) problems following the the tasks, and presented in [3]. In the problem at hand, each robot is capable of executing one task at a time and each task needs to be assigned to a robot coalition. Also, the available information about the robots, the environment permits only instantaneous allocation of tasks to robots, without planning for future allocations [3]. Our proposed algorithm is based on Quantum Multi-Objective Particle Swarm Optimization (QMOPSO). QPSO is a discrete version of PSO to solve optimization problems with binary-valued solution elements [4]. PSO is one of the many options for coalition formation. Simulated Annealing (SA), Genetic Algorithm (GA), and Column-Generation (CG) are other examples. PSO is chosen because of its effectiveness in solving a wide range of applications [5]. It has the ability to find optimal or near-optimal solutions for large-space problems in a short time compared to other heuristics [6]. The goal of the proposed algorithm is to ensure that the optimal coalition of robots is selected with the required capabilities for each task. The proposed algorithm consists of filtering method, the QMOPSO approach, and ranking method. Filtering is used to choose the best robots for the execution of the QMOPSO algorithm and to make the robots that have not been selected available for other requests. In addition, location constraints regarding the capability distribution of the robots are taken into consideration. For instance, some tasks require that the combination of a given sensor and actuator should reside on the same robot or on different robots. This is necessary to ensure proper execution of search and rescue task. The rest of the paper is organized as follows: Section II discusses related work. In section III, the problem is formulated. Section IV presents the proposed algorithm. Validation is done in Section V and we conclude in Section VI. II. RELATED WORK Forming an effective coalition has become an interesting research area. It has been widely studied over years. The problem is NP-hard [7]. Reference [8] proposes a modified version of Shehory and Kraus's algorithm. Authors in [9], propose an algorithm based on GA. Their major drawback is that they optimize only one objective (i.e., the overall utility and the coalition value respectively). However, there are other important objectives that need to be optimized, e.g., the time needed to perform a task. Reference [10] proposes an algorithm which aims at maximizing the number of completed tasks and the system efficiency. Two multi-objective algorithms are This paper has been accepted for presentation in IEEE IWCMC 2017 – The 13th International Wireless Communications and Mobile computing Conference (IWCMC) to be held in Valencia, Spain; June 26-30 June, 2017. This is an author copy. The respective Copyrights are with IEEE introduced: a Non-Dominated Sorting Genetic Algorithm (NSGA-II) and a Strength Pareto Evolutionary Algorithm (SPEA-II). Unfortunately, factors such as the minimizing the number of robots in a coalition is not considered. In [11], authors assign Unmanned Aerial Vehicles (UAVs) to search and prosecute missions. However, reducing the cost of UAV deployment is not considered. Authors in [12] propose PSO- based multiple objectives algorithm. The PSO is designed for continuous domains. We consider a discrete problem domain for the robots, which drives the need for designing new PSO. In [13], authors propose an auction-based approach for task allocation. However, their algorithm is evaluated over small scale i.e., 10 and 50 robots. Additionally, none of the works presented above discusses any filtering or ranking methods. In our work, we present ranking and filtering methods for the best solution among Pareto-optimal solutions, to optimize the processing time of the algorithm. In [14], authors propose an ant-colony based algorithm. They consider fix number of robots for each task. However, it is not efficient to fix the number of robots required for each task since robots have different capabilities and different capability distribution. Authors in [15], propose an algorithm based on dynamic ANT coalition technique. However, the performance of the algorithm is not evaluated with a large number of robots. In [16], Rauniyar and Muhuri modify the standard GA. They proposed adaptive Random Immigrants Genetic Algorithm (aRIGA) and adaptive Elitism-based (aEIGA). However, the proposed algorithm is single-objective. Immigrants Genetic Algorithm III. PROBLEM FORMULATION We consider an infrastructure composed of n robots: 𝑅 = {𝑅1, … , 𝑅𝑖, … , 𝑅𝑛} (1) where n is significantly large and hence the infrastructure can support search and rescue task in large-scale disasters. Each of these robots has two vectors of capabilities: sensing (e.g., cameras, sensors) and actuating (e.g., arms, wheels) capabilities. It is assumed that each capability is a real non- negative value and indicates the quantity of sensors/actuators owned by the robots. For robot 𝑅𝑖, the sensing and actuating capability vectors are: 𝑆𝑅 𝑖 = {𝑠1 𝑖 } (2) 𝐴𝑅𝑖 = {𝑎1 𝑖 , … , 𝑎𝑑 𝑖 , … , 𝑠𝑟 𝑖 } (3) where r and d are the number of possible sensing and actuating capabilities respectively. A robot can be in three states: Idle, Allocated, and Busy. The idle state is when the robot does not perform any tasks, the allocated is when the robot is locked with the algorithm running on it, and the busy state is when the robot is performing a task. The infrastructure can perform m tasks assigned to it: 𝑇 = {𝑇1, … , 𝑇𝑗, … , 𝑇𝑚} (4) Each task 𝑇𝑗 is composed of p sub-tasks: 𝑗} (5) 𝑍𝑇𝑗 = {𝑧1 It is assumed that the sub-tasks are executed independently and 𝑗, … , 𝑧𝑘 𝑗, … , 𝑧𝑝 that each robot is a member of only one sub-task. Each sub-task requires a specific set of sensing and/or a set of actuating capabilities to start. 𝑗 We represent the capability requirements of each sub-task 𝑧𝑘 by two vectors, sensing requirements and actuating requirements, as: 𝑗𝑘, … , 𝑠𝑟 , the capability requirement vectors are: 𝑗𝑘} (6) 𝐴𝑧𝑘 𝑗 For sub-task 𝑧𝑘 𝑗𝑘} (7) 𝑗 = {𝑠1 𝑆𝑧𝑘 Then, the capability requirement vectors for the task 𝑇𝑗 is the sum of the capability requirement vectors of the sub-tasks constituting the task 𝑇𝑗: 𝑗𝑘, … , 𝑎𝑑 𝑗 = {𝑎1 𝑝 𝑘=1 𝑗 𝐴𝑧𝑘 𝑝 𝑘=1 𝑗 𝑆𝑧𝑘 (8) 𝐴𝑇𝑗 = ∑ 𝑆𝑇𝑗 = ∑ (9) Some of the sub-tasks of task 𝑇𝑗 are tied by locational constraints regarding the capability distribution of the robots while others may be executed without any locational constraints. This is necessary in order to ensure proper execution of the sub-tasks. According to [8], there are two types of locational constraints; combination of sensors and actuators should reside on the same robot or combination of sensors and actuators should reside on different robots. The locational constraints can be represented as Constraints Satisfaction Problem (CSP) [8]. CSP consists of three components:  The set of variables, that is the required sensor and actuators for the task 𝑋 = {𝑥1, … , 𝑥𝑗, … , 𝑥𝑘} (10) where 𝑋 = {𝑠1, 𝑠2, … , 𝑠𝑛, 𝑎1, 𝑎2, … , 𝑎𝑛}  The set of values for each variable, that is the available robots possessing the required capabilities for each variable For variable 𝑥𝑖, the set of values is: 𝑉𝑥𝑖 = {𝑅𝑗, … , 𝑅𝑛} (11)  The set of constraints between different variables 𝐶 = {𝐶1, … 𝐶𝑖, … , 𝐶𝑛} (12) where each 𝐶𝑖 is one of the following types: 𝑥𝑖 ≠ 𝑥𝑗, 𝑥𝑖 = 𝑥𝑗. The goal in CSP is to assign a value for each variable such that the constraints are satisfied. A coalition m (𝐶𝐿𝑇𝑚) for any task has two vectors of capabilities: sensing 𝑆𝐶𝐿𝑇𝑚 and actuating 𝐴𝐶𝐿𝑇𝑚, while each is the sum of the capabilities owned by the robots in that coalition: 𝑆𝐶𝐿𝑇𝑚 = ∑𝑅𝑖Є𝐶𝐿𝑇𝑚 𝑆𝑅 𝑖 (13) 𝐴𝐶𝐿𝑇𝑚 = ∑𝑅𝑖Є𝐶𝐿𝑇𝑚 𝐴𝑅 𝑖 (14) 𝐶𝐿𝑇𝑚 can perform task 𝑇𝑗 only if: 1. The vector of its capabilities satisfy the following: 𝑆𝐶𝐿𝑇𝑚 ≥ 𝑆𝑇𝑗 And / or 𝐴𝐶𝐿𝑇𝑚 ≥ 𝐴𝑇𝑗 (15) 2. And its members meet the locational constraints. It is assumed that a coalition can work on a single task at a time and that each robot (𝑅𝑖) is a member of one coalition at a time. Where i = [1, n]. 2 = ∅ (16) The objective of this problem is to find a coalition of robots that minimizes simultaneously the deployment cost of the robots, the time needed to perform a task by the robots belonging 1 ⋂ 𝐶𝐿𝑇𝑅𝑖 𝐶𝐿𝑇𝑅𝑖 This paper has been accepted for presentation in IEEE IWCMC 2017 – The 13th International Wireless Communications and Mobile computing Conference (IWCMC) to be held in Valencia, Spain; June 26-30 June, 2017. This is an author copy. The respective Copyrights are with IEEE 𝑖 and then, according to the a random number for each 𝑣(𝑡)𝑗 condition in (19) and (20), the discrete particle vector is initialized. 𝑝(𝑡)𝑖 = [𝑝(𝑡)1 𝑖 ] (18) where n is the size of the problem, i.e., the total number of 𝑖 , … , 𝑝(𝑡)𝑗 𝑖 , … , 𝑝(𝑡)𝑛 robots. 𝑖 > 𝑣(𝑡)𝑗 𝑖 𝑝(𝑡)𝑗 If 𝑟𝑎𝑛𝑑𝑗 Otherwise 𝑝(𝑡)𝑗 𝑖 = 1 (19) 𝑖 = 0 (20) First, the initial population is evaluated by calculating the values of three objective functions for each particle. The particles that represent non-dominated solutions are stored in a repository (REP). Each particle keeps track of its best local position, which is the best solution obtained by this particle so far (𝑃𝑖 𝑙𝑜𝑐𝑎𝑙𝐵𝑒𝑠𝑡). At each iteration, the algorithm selects 𝑃𝑔𝑙𝑜𝑏𝑎𝑙𝐵𝑒𝑠𝑡 that denotes the best position achieved so far by any particle in the population. It is selected by ranking the solutions in REP and choosing the one with the highest rank. Also, the velocity equation is updated according to equation (21) and the particle vector is updated in the same way in equations (17) to (20). 𝑉(𝑡 + 1) = 𝑤 × 𝑉(𝑡) + 𝑐1 × 𝑉𝑙𝑜𝑐𝑎𝑙𝑏𝑒𝑠𝑡(𝑡) + 𝑐2 × 𝑉𝑔𝑙𝑜𝑏𝑎𝑙𝑏𝑒𝑠𝑡(𝑡) (21) to that coalition, and the number of the robots in a coalition in the same amount of time. IV. THE PROPOSED ALGORITHM FOR COALITION FORMATION We propose an algorithm for coalition formation problem. Our optimization problem includes constraints, including the task capability requirements and locational constraints as described in section III. A particle is said to be feasible if it satisfies both constraints and infeasible otherwise. Given a particle (i.e., candidate coalition), checking the task capability requirements is quite straightforward. However, the locational constraints are checked by solving CSP, which can be solved with either the brute-force technique or a better technique such as backtracking. Handling constraints during QMOPSO execution is described in the next section (section IV. B). A. Coalition Formation Algorithm for Multi-Robot System The Pseudocode for the coalition formation algorithm for multi-robot systems is given in Algorithm 1. The detailed description is as follows: The set of inputs for the algorithm are n (the maximum number of robots allowed in a group), Time (the maximum time period to complete a given task), Cost (the cost Filtering_Rule, Task_Requirements (the required sensors and actuators for a given task), Locational Constraints (the capability distributions for sub-tasks constituting a given task), and Criteria_Importance (defining the weights to rank the Pareto-optimal solutions based on more than one criterion - i.e., objectives in our case). customer agrees on), the The algorithm starts with filtering the robots based on the Filtering_Rule. In this function, if the battery level of the robots is lower than the Filtering_Rule, they are excluded from the next steps. It then applies the QMOPSO-based algorithm. Multi- objective problems generate a set of non-dominated or Pareto- optimal solutions. The solutions are ranked after excluding the solutions that exceed the time, the cost, the number threshold, and the infeasible solutions. Promethee II ranking [17] is applied, which is a multi-criteria ranking method with lots of success due to its mathematical properties and its user- friendliness. In this method, the Pareto-optimal solutions are compared pairwise. The difference between the evaluations of two Pareto-optimal solutions over each criterion is considered. The criteria in our case are the objectives (i.e., time, cost, and number of robots). The Pareto-optimal solutions are ranked using the Criteria-Importance/weight of the objectives. The highest rank solution denotes the best robot coalition. B. QMOPSO Algorithm The Pseudocode of QMOPSO is given in Algorithm 2. The algorithm first initializes the particles. A particle is defined based on the quantum bit. Two vectors are initialized:  Quantum particle vector 𝑉(𝑡)𝑖, which is the velocity for particle i and is initialized to random values between [0,1]: 𝑉(𝑡)𝑖 = [𝑣(𝑡)1 𝑖 ] (17)  Discrete particle vector 𝑝(𝑡)𝑖, which is initialized by initializing 𝑖 , … , 𝑣(𝑡)𝑗 𝑖 , … , 𝑣(𝑡)𝑛 This paper has been accepted for presentation in IEEE IWCMC 2017 – The 13th International Wireless Communications and Mobile computing Conference (IWCMC) to be held in Valencia, Spain; June 26-30 June, 2017. This is an author copy. The respective Copyrights are with IEEE TABLE I. ALGORITHM EVALUATION PARAMETERS Value Parameter General Population size Problem size (number of robots) Maximum number of iterations 𝛼, 𝛽 𝑤, 𝑐1, 𝑐2 Threshold for filtering Number of objectives Number of sub-tasks Criteria_Importance NSGA-II and SPEA-II Tournament size Pool size for tournament selection Mutation probability Crossover probability Distribution index for crossover Distribution index for mutation 100, 200 10-10000 100 0.3, 0.7 0.25, 0.25, 0.5 40% 3 3 time 2 Population number / 2 10% 90% 20 20 No. of Robots 1000 5000 10000 𝑉𝑙𝑜𝑐𝑎𝑙𝑏𝑒𝑠𝑡(𝑡) = 𝛼 × 𝑝𝑙𝑜𝑐𝑎𝑙𝑏𝑒𝑠𝑡(𝑡) + 𝛽 × (1 − 𝑝𝑙𝑜𝑐𝑎𝑙𝑏𝑒𝑠𝑡(𝑡)) (22) 𝑉𝑔𝑙𝑜𝑏𝑎𝑙𝑏𝑒𝑠𝑡(𝑡) = 𝛼 × 𝑝𝑔𝑙𝑜𝑏𝑎𝑙𝑏𝑒𝑠𝑡(𝑡) + 𝛽 × (1 − 𝑝𝑔𝑙𝑜𝑏𝑎𝑙𝑏𝑒𝑠𝑡(𝑡)) (23) where 𝛼 + 𝛽 = 1 , 𝛽 < 1, 0 < 𝛼. 𝛼 and 𝛽 are control parameters, 𝑤 represents the degree of belief on oneself, 𝑐1 is the local maximum, and 𝑐2 is the global maximum. The first part of the equation (21) indicates the interia of the previous probability. The second part is called cognition and represents the local exploration probability. The third part is the social part that indicates the cooperation among all quantum particles. 𝑃𝑖 𝑙𝑜𝑐𝑎𝑙𝑏𝑒𝑠𝑡 is updated by applying Pareto dominance. If the current position is dominated by the one in the memory, the one in the memory is kept; otherwise, the one in the memory is replaced by the current position. To update the REP, if the REP is empty, the current non- dominated particle is added to the REP; otherwise, the two particles are compared as follows: If both are feasible, Pareto- dominancy is applied; if one is feasible and the other infeasible, the feasible dominates; if both are infeasible, the one with the highest degree of constraint satisfactions is selected. We define a particle's feasibility degree as the degree of constraint satisfactions. A task 𝑇𝑗 is considered to have U capability requirements and M locational constraints. The capability requirements and locational constraints are considered as 𝑇𝑟𝑒𝑞 and C respectively. Then a particle is feasible if it satisfies 𝑇𝑟𝑒𝑞 and C, and it is infeasible otherwise. We determine a particle's feasibility degree as the weighted sum of feasibility degree with respect to 𝑇𝑟𝑒𝑞 and C. If a particle satisfies u capability requirements and satisfies m locational constraints, then the particle's feasibility degree with respect to 𝑇𝑟𝑒𝑞 and C are expressed as: 𝑠𝑎𝑡 𝑇𝑟𝑒𝑞 = 𝑢/𝑈 (24) 𝑠𝑎𝑡 𝐶 = 𝑚/𝑀 (25) A particle's feasibility degree can now be calculated as: 𝑃𝑓𝑒𝑎𝑠 = 𝑠𝑎𝑡 𝑇𝑟𝑒𝑞 ∗ 𝑊𝑇 + 𝑠𝑎𝑡 𝐶 ∗ 𝑊𝐶 (26) where 𝑊𝑇 and 𝑊𝐶 are the weights chosen such that: TABLE II. ERROR RATIO, SET COVERAGE AND SPACING (10 ROBOTS, Algorithm QMOPSO NSGA-II SPEA-II Error Rate 0.6 0.8 0.33 POPULATION SIZE=200) Set Coverage QMOPSO NSGA-II SPEA-II - 0.66 1 0.3 - - 0 - - Spacing 31.63 51.45 67.14 TABLE III. ERROR RATE & SET COVERAGE (POPULATION SIZE=100) Error Ratio Set Coverage QMOPSO NSGA- II 0.32 0.31 0.26 0.23 0.20 0.11 SPEA- II 0.30 0.41 0.2 (QMOPSO, NSGA-II) 0.72 0.9 1 (NSGA-II, QMOPSO) 0 0.1 0 (QMOPSO, SPEA-II) 0.8 1 0.93 (SPEA-II, QMOPSO) 0.1 0 0 TABLE IV. SPACING (POPULATION SIZE=100) No. of Robots QMOPSO Spacing NSGA-II 1000 5000 10000 18.23 16.11 8.24 23.11 35.61 19.23 SPEA-II 40.36 37.22 28.19 𝑊𝑇 + 𝑊𝐶 = 1, 0 ≤ 𝑊𝑇, 𝑊𝐶 ≤ 1 (27) If a particle is feasible, then the feasibility degree is 1. V. PERFORMANCE EVALUATION In order to evaluate the algorithm, we have performed our experiments with different problem and population sizes. In each experiment, the speed, the cost, the position, the battery level of each robot, and the position of the target - which is the disaster location – have been randomly generated. All the robots are in the idle state at the beginning of each experiment. We have compared our algorithm with two well-known heuristic-based algorithms: NSGA-II and SPEA-II [10]. All algorithms have been implemented in Matlab. Table I shows the evaluation parameters along with their values. A. Performance Metrics We measured the performance of three optimization algorithms in terms of convergence, diversity, and processing time. Convergence shows the solution's accuracy (i.e., the ability to produce good quality solutions and their ability to converge to true Pareto-solutions). We have used error ratio [12] and set coverage as the convergence metrics. Diversity, on the other hand, shows the spread of solutions. We have used spacing [12] as the diversity metric. The metrics are defined as follows: 1. Error Ratio, as the percentage of non-dominated solutions that are not part of a reference Pareto-set: When the true Pareto-set is known, it is used as the reference Pareto-set. When it is not known, the reference Pareto-set is obtained by combining Pareto-sets of all algorithms and applying non-dominancy. 2. Set Coverage (SC (A, B)), given two sets, is the percentage of non-dominated solutions in set B covered (i.e., dominated) by those in set A: If SC (A, B) > SC (B, A), then A is relatively better than B. A is absolutely better than B when SC (A, B) = 1 and SC (B, A) = 0. 3. Spacing, as the standard deviation of distances of non- This paper has been accepted for presentation in IEEE IWCMC 2017 – The 13th International Wireless Communications and Mobile computing Conference (IWCMC) to be held in Valencia, Spain; June 26-30 June, 2017. This is an author copy. The respective Copyrights are with IEEE (b) (c) Fig. 1. Non-dominated fronts obtained at different iteration for problem size 5000 and population size 200 (a) Fig. 2. Processing time with different problem sizes (Population size=100) Fig. 3. The effect of feasibility check on average repository updating time Fig. 4. The time needed for the filtering function in QMOPSO dominated solutions from their closest neighbors. 4. Processing Time of Algorithm (PT) (Sec), as the time needed for the algorithm to run and select the most efficient coalition. 5. Filtering Time (sec), as the time needed for filtering the robots in QMOPSO. 6. Repository Update Time (sec), as the time needed by QMOPSO to update the repository at each iteration. It includes the delay incurred by the constraint handling method. B. Results and Discussions Test case 1 - convergence and diversity: Table II shows the error ratio, the set coverage, and the spacing metrics of QMOPSO, NSGA-II, and SPEA-II for a small-scale problem (i.e.., 10 robots). We have also generated the true Pareto-optimal solutions by the enumerated search method. For the three algorithms, we have used a population size of 200. The error ratio of QMOPSO is higher than that of SPEA-II but lower than that of NSGA-II. The set coverage metric shows that 30% of solutions in NSGA-II are covered by QMOPSO and 66% of QMOPSO are covered by NSGA-II. Hence, NSGA-II is relatively better than QMOPSO. Since NSGA-II and SPEA-II do not cover each other, their relative dominance cannot be concluded. Overall, we observe that NSGA- II performs better than QMOPSO in terms of convergence. However, when it comes to diversity, QMOPSO outperforms NSGA-II and SPEA-II. This is concluded from the lowest spacing value in case of QMOPSO, which indicates a good distribution of solutions. Table III shows the error ratio for large-size problems (e.g., 1000, 5000, and 10000 robots). We observe that for any problem size, QMOPSO outperforms NSGA-II and SPEA-II. In fact, it achieves the lowest error ratio for the largest problem size (10000 robots). It shows a better convergence of QMOPSO for large-scale problems. SPEA-II has the highest error ratio for 5000 robots. Table III also shows the set coverage metric. As observed, when the problem size is 1000, QMOPSO is relatively better than both NSGA-II and SPEA-II. For a problem size of 5000, QMOPSO is absolutely better than SPEA-II as all solutions of SPEA-II are dominated by those of QMOPSO and none of QMOPSO solutions is dominated by those in SPEA-II. QMOPSO for a problem size of 5000 is relatively better than NSGA-II. However, for a problem size of 10000, QMOPSO is relatively better and absolutely better than SPEA-II and NSGA-II respectively. Overall, SC results show that QMOPSO produces a better solution than NSGA-II and SPEA-II. Table IV shows the spacing metric for three algorithms. We observe that QMOPSO attains the lowest value of spacing for any problem size, thereby achieving the highest diversity and even distribution of solutions. The diversity of NSGA-II lies between QMOPSO and SPEA-II. Fig. 1 (a)-(c) shows the non-dominated-fronts obtained at some iterations for a problem size of 5000. We have found that with an iteration increase, the solutions in QMOPSO evolve more quickly than those in NSGA-II and SPEA-II. It shows the ability of QMOPSO to explore the search space more efficiently than others. Test case 2 - processing time of the algorithms with a various number of robots: We compare the PT of our algorithm with NSGA-II and SPEA-II. The three algorithms are implemented and applied in the same environment, with the same This paper has been accepted for presentation in IEEE IWCMC 2017 – The 13th International Wireless Communications and Mobile computing Conference (IWCMC) to be held in Valencia, Spain; June 26-30 June, 2017. This is an author copy. The respective Copyrights are with IEEE number of robots, task requirements, and robot capabilities. The size of the population is 100. Fig. 2 shows the processing time of the three algorithms with a various number of robots. For the QMOPSO algorithm, we consider the PT with and without the filtering method. We notice that the PT decreases when the filtering method is used. This is because filtering method reduces the number of robots on which the algorithm runs. On the other hand, without filtering, the PT of the algorithm increases with an increase in the number of robots. The rationale behind this is the fact that the higher number of robots results in a higher dimension of the particle. As an important observation, the PT without these methods is still smaller than that of NSGA-II and SPEA-II; this is due to the simple mathematical operations of QMOPSO compared to other algorithms. In QMOPSO, the velocity equation is the sole equation updated at each iteration. Test case 3 - repository update time: We have considered two types of constraints: Task requirements and location constraints. For the task requirements, we have considered 6 requirements (𝑠1, 𝑠2, 𝑠3, 𝑎1, 𝑎2, 𝑎3) with random number of units for each. For the locational constraints, we have represented the problem using CSP as described in section III and we have considered three locational constraints ( 𝑠1 = 𝑎1, 𝑠2 = 𝑎2, 𝑠3 = 𝑎3). A simplified method is used to calculate the satisfaction degree of a particle/coalition for the task requirements and locational constraints. We calculate the effect of our proposed method to solve the two constraints (task requirements and locational constraints) on the average repository updating time. The results in Fig. 3 demonstrate the time needed to perform the feasibility check versus the overall repository update time, considering different numbers in a population. As we notice, the time needed for our proposed feasibility checking method is negligible compared to the total time for updating the repository. Test case 4 - filtering time: We have also calculated the time needed to perform the filtering function compared to the overall processing time of QMOPSO. Fig. 4 shows that the filtering time is negligible compared to the overall processing time of the algorithm. The time needed for filtering does not introduce additional overhead on the algorithm processing time. Since this method excludes some robots using battery levels, it ensures that the remaining robots have sufficient battery to accomplish the task. Since it does not affect the processing time of the algorithm, the overall efficiency is achieved. VI. CONCLUSION In this paper, we have proposed a coalition formation algorithm for multi-robot systems. To show the effectiveness of our algorithm, we have conducted extensive simulation experiments and compared our algorithm with other existing algorithms. The results demonstrate that the proposed algorithm cannot only improve the solution, but it also has a significantly short processing time. They also show that the filtering and the repository updating mechanisms do not add overhead on the processing time. It is also observed that QMOPSO achieves higher diversity, the lowest error rate, and produces better solution compared to NSGA-II and SPEA-II for large problem sizes. ACKNOWLEDGMENT The work is partially supported by CISCO systems grant CG-576719 and by the Canadian National Science and Engineering Research Council (NSERC) discovery program. REFERENCES [1] CRASAR, "World Trade Center 911 Disaster « Center for Robot- Assisted Search and Rescue (CRASAR) at Texas A&M University," 2001. [2] H. Kitano et al., "RoboCup Rescue: search and rescue in large-scale disasters as a domain for autonomous agents research," in 1999 IEEE International Conference on Systems, Man, and Cybernetics, 1999. IEEE SMC '99 Conference Proceedings, 1999, vol. 6, pp. 739–743. [3] B. P. Gerkey and M. J. Matarić, "A Formal Analysis and Taxonomy of Task Allocation in Multi-Robot Systems," Int. J. Robot. Res., vol. 23, no. 9, pp. 939–954, Sep. 2004. J. Kennedy and R. C. Eberhart, "A discrete binary version of the particle swarm algorithm," in, 1997 IEEE International Conference on Systems, Man, and Cybernetics, 1997. Computational Cybernetics and Simulation, 1997, vol. 5, pp. 4104–4108. [4] [5] S. Pandey, L. Wu, S. M. Guru, and R. Buyya, "A Particle Swarm Optimization-Based Heuristic for Scheduling Workflow Applications in Cloud Computing Environments," in 2010 24th IEEE International Conference on Advanced Information Networking and Applications, 2010, pp. 400–407. [6] G. Zhang, X. Shao, P. Li, and L. Gao, "An effective hybrid particle swarm optimization algorithm for multi-objective flexible job-shop scheduling problem," Comput. Ind. Eng., vol. 56, no. 4, pp. 1309–1318, May 2009. [7] T. Sandholm, K. Larson, M. Andersson, O. Shehory, and F. Tohmé, "Coalition Structure Generation with Worst Case Guarantees," Artif. Intell., vol. 111, no. 1–2, pp. 209–238, Jul. 1999. [8] L. Vig and J. A. Adams, "Multi-Robot Coalition Formation," IEEE Trans. Robot., vol. 22, no. 4, pp. 637–649, 2006. [9] H.-Y. Liu and J.-F. Chen, "Multi-Robot Cooperation Coalition Formation Based on Genetic Algorithm," in 2006 International Conference on Machine Learning and Cybernetics, 2006, pp. 85–88. [10] M. Agarwal, N. Kumar, and L. Vig, "Non-Additive Multi-Objective Robot Coalition Formation," Expert Syst. Appl., vol. 41, no. 8, pp. 3736– 3747, Jun. 2014. [11] J. G. Manathara, P. B. Sujit, and R. W. Beard, "Multiple UAV Coalitions for a Search and Prosecute Mission," J. Intell. Robot. Syst., vol. 62, no. 1, pp. 125–158, Apr. 2011. [12] C. A. C. Coello, G. T. Pulido, and M. S. Lechuga, "Handling Multiple Objectives with Particle Swarm Optimization," IEEE Trans. Evol. Comput., vol. 8, no. 3, pp. 256–279, 2004. [13] T. C. Service, S. D. Sen, and J. A. Adams, "A simultaneous descending auction for task allocation," in 2014 IEEE International Conference on Systems, Man and Cybernetics (SMC), 2014, pp. 379–384. [14] B. Qian and H. H. Cheng, "A mobile agent-based coalition formation system for multi-robot systems," in 2016 12th IEEE/ASME International Conference on Mechatronic and Embedded Systems and Applications (MESA), 2016, pp. 1–6. [15] P. M and G. R. Suresh, "Coalition formation and Task Allocation of multiple autonomous robots," in 2015 3rd International Conference on Signal Processing, Communication and Networking (ICSCN), pp. 1–5. [16] A. Rauniyar and P. K. Muhuri, "Multi-robot coalition formation problem: Task allocation with adaptive immigrants based genetic algorithms," in 2016 IEEE International Conference on Systems, Man, and Cybernetics (SMC), 2016, pp. 000137–000142. [17] J. Figueira, S. Greco, and M. Ehrgott, Multiple Criteria Decision Analysis: State of the Art Surveys. Springer, 2005.
1711.08319	1	1711	2017-11-16T01:31:36	Systems, Actors and Agents: Operation in a multicomponent environment	[ "cs.MA", "cs.AI" ]	Multi-agent approach has become popular in computer science and technology. However, the conventional models of multi-agent and multicomponent systems implicitly or explicitly assume existence of absolute time or even do not include time in the set of defining parameters. At the same time, it is proved theoretically and validated experimentally that there are different times and time scales in a variety of real systems - physical, chemical, biological, social, informational, etc. Thus, the goal of this work is construction of a multi-agent multicomponent system models with concurrency of processes and diversity of actions. To achieve this goal, a mathematical system actor model is elaborated and its properties are studied.	cs.MA	cs	Systems, Actors and Agents: Operation in a multicomponent environment Mark Burgin University of California, Los Angeles 405 Hilgard Ave. Los Angeles, CA 90095 Abstract. In this paper, we further develop multi-agent approach by creating new types of system models. The problem is that conventional models of multi-agent and multicomponent systems implicitly or explicitly assume existence of the absolute time or even do not include time in the set of defining parameters. However, it is rationalized theoretically and validated experimentally that there are different times and time scales in a variety of real systems – physical, chemical, biological, social, etc. Thus, the goal of this paper is construction of multi-agent and multicomponent system models with concurrency of processes and diversity of actions. To achieve this goal, a mathematically based system actor model is elaborated and its properties are studied. Keywords: time, system, actor, agent, action, process, interaction, environment 1. Introduction Multi-agent approach is becoming more and more popular in the area of computing, networking, artificial intelligence, robotics, distributed control, resource management, collaborative decision support systems, data mining, etc. (Buşoniu, et al, 2010; Shoham and Leyton-Brown, 2008; Vlassis, 2007; Weiss, 1999). Usually, it is assumed that a multi-agent system is a group of autonomous, interacting entities sharing a common environment, which they perceive with sensors and upon which they act with actuators. Time is a critically important characteristic of any real-life system. However, not all features of time are adequately presented in the conventional multi-agent models. If we analyze existing approaches and directions in the multi-agent approach, we can see that in all dynamic models of multi-agent systems, either time is implicitly induced by actions of agents and system states or it is explicitly assumed that unique time exists for the whole system. An archetypal 1 example of this situation is the absolute Newtonian time in the physical universe, which is innate for the entire classical physics. However, relativity theory and various experiments disproved this assumption bringing forth the concept of local time (cf., for example, (Einstein, et al, 1923)). The system theory of time extends this principle much further (Burgin, 1992; 2002). Other researchers also advocated existence of different times or different time scales in their theories (cf., for example, (Prigogine, 1980; Barwise and Seligman, 1997)). Besides, as Norbert Wiener (1961) writes, one of the most famous philosophers of the 20th century Bergson lays special emphasis on the distinction between the reversible time of physics, in which nothing new happens, and the irreversible time of evolution and biology, in which there is always something new (Bergson, 1910). In spite of this, time in general systems theory is similar to time in classical physics, namely, either all models of systems in general systems theory are still based on the principle of absolute (global) time or time is not explicitly expressed in these models. At the same time, there are many systems, in which it is unfeasible to introduce and preserve global time. For instance, it is proved that clock synchronization becomes impossible under definite conditions (Lamport, 1984; Dolev, et al, 1986; Fischer, et al, 1985; Attiya and Ellen, 2014).This precludes introduction of global time. As a result, in some systems, only local time (local time scale) can be treated in a consistent way. In addition, there are systems, in which global time (global time scale) coexists with a variety of local times (local time scales). Moreover, often these different times and time scales cannot be synchronized. All systems with these properties, which we call concurrent systems, cannot be portrayed by conventional models in general systems theory. The goal of this paper is to construct more advanced than utilized now models of multi-agent distributed systems using the concept of local time, which exists and can be different in distinct components and parts of real systems according to the system theory of time (Burgin, 1992; 2002). These models provide descriptions and tools for exploration not only of classical systems with one global time but also of relativistic and concurrent systems, which can multiplicities of time. It is interesting that absence of global time results in nonexistence of global states in a multicomponent system due to the concurrent functioning of the components and parts. As a result, time becomes multidimensional and demands specific unconventional mathematical structures for its representation. 2 In addition, exploring and modeling systems with a variety of independent and incoherent local times (local time scales), we come to the concepts of an observer, observation, synchronization and coordination of times and actions, which have not been studied in general systems theory. For instance, there are systems, in which it is possible to synchronize local time in different components and establish in such a way, global (absolute) time and a global time scale. However, conventional models from general systems theory give only a partial picture of these systems and do not allow exploration of synchronization. Note that synchronization plays a pivotal role in diversity of systems from organisms of people (cf., for example, (Winfree, 1987)) to computer networks (cf., for example, (Mills, 1991; Burgin, et al, 2016)) to distributed databases (cf., for example, (Lindsay, et al, 1979)) to physics and metrology (cf., for example, (Boixo, et al, 2006)) to human-computer interaction (cf., for example, (Burgin, et al, 2001)) to service systems (cf., for example, (Marzullo, 1983)). As Birman (2005) writes, clock synchronization is a necessary and critical part in most distributed systems. To develop our model employing different structures of local time and local time scales, we use methods and approaches developed for concurrent processes in computer science and information theory. On the first stage of our research, we construct a kinetic system model by fundamentally advancing and further developing the Actor Model originally constructed for distributed computations (Hewitt, et al, 1973). Here we expand the scope of this model from computational systems and processes to general systems making it applicable for any system comprised of interacting subsystems, e.g., for organization, society, group of people or a computer network. We call the basic component of the System Actor Model (SAM) constructed in this paper an actor although the conventional research typically uses the term agent. The reason for this is that according to the common usage, an agent is a system (an actor) who/that acts on behalf of another system (actor). Besides, in political science and economics, an agent is a person or entity able to make decisions and take actions on behalf of, or that impact, another person or entity called the principal (Rees, 1985; Eisenhardt, 1989). This methodology allows treating agents as specific actors, who/which act on behalf other actors - principals. As a result, it is possible to represent any multi-agent multicomponent system by the System Actor Model but not every actor system can be represented by a multi-agent multicomponent system. There are many situations, especially, in society, when this difference between agents and free actors is very important. Taking into account that actors can be software systems, we see that 3 software agents are a very special but essentially important case of actors. It is possible to compare relation between actors and agents with the relation between a function and a computable, e.g., recursive, function. The System Actor Model is more flexible than agent models. For instance, agents are usually treated as autonomous systems perceiving with sensors and acting with actuators (Buşoniu, et al, 2010; Vlassis, 2007; Weiss, 1999). At the same time, actors can be directed or controlled by other actors. Some of actors can be without sensors and/or actuators. For instance, in problems of resource management, identifying each resource with an actor can make available a helpful, detailed perspective on the system while each of them might not have sensors and/or actuators and could be controlled and managed by a central authority. This paper is structured in the following way. In Section 2, which goes after Introduction, we describe and explore the computational actor model. In Section 3, we construct and explore the system actor model, for which the computational actor model becomes a very special case. Besides, we go much further in comparison with the computational actor model by elaborating mathematical models of actors and environments where these actors function. This allows us to obtain many properties of actions, events, actors and their systems by rigorous mathematical techniques. One of the main targets of this work is to construct mathematical tools for exploration of collaborative and multi-agent systems, social network analysis and developing computational qualitative methods for data mining and digital humanities. 2. Actor model in computer science The Computational Actor Model (CAM) and its methodology were developed in the theory of computation to provide constructive and theoretical tool for modeling, analyzing and organizing concurrent digital computations (Hewitt, et al, 1973; Hewitt, 2012). In CAM, actors are interpreted as computing devices or computational processes. We will call them computational actors. To make the model uniform, the concept of a messenger, which is also a computational actor, is used instead of the concept of a message. An arbitrary event in the model is the receipt of a messenger, which impersonates a message, by the target (recipient) computational actor. In CAM, computational actors perform computations based on information about other computational actors and asynchronously communicate using their addresses for sending and receiving messages. Additionally, computational actors can make decisions about their actions and 4 behavior, create other computational actors, and determine how to react to the received messages. It is possible to treat all these actions as events in the space of computational actors although this not done in the original Computational Actor Model described in (Hewitt, et al, 1973). Computational actors are described by two groups of axioms - structural axioms and operational axioms. Structural axioms determine that the local storage of a computational actor may contain addresses of other computational actors such that satisfy one of the following conditions: 1. The addresses were provided when the computational actor was created. 2. The addresses have been received in messages. 3. The addresses were installed in computational actors created by the given computational actor. Operational axioms determine what a computational actor can do. Namely, a computational actor can: 1. Create more computational actors 2. Send messages to other computational actors 3. Receive messages from other computational actors Hewitt explains that CAM is rooted in physics while other theoretical models of computation are based on mathematics and/or logic (Hewitt, 2007). As a result, CAM has many properties similar to properties of physical models, especially, in quantum physics and relativistic physics. For instance, detailed observation of the arrival order of the messages for a computational actor affects the results of actor's behavior and can even increase indeterminacy. According to CAM, the performance of a computational actor is exactly defined when it receives a message while at other times, it may be indeterminate. Note that in reality, existence of nondeterministic models of computation, such as nondeterministic Turing machines (cf., for example, (Burgin, 2005)), shows that in some cases, the performance of a computing system or process cannot be exactly defined. An important feature of CAM is that it can model systems that cannot be represented by the deterministic Turing's model while the latter is a special case of CAM. As Milner wrote (Milner, 1993), Hewitt had explained that a value, an operator on values, and a process could all be computational actors. Taking into account that computational actors can be interpreted as software systems, we can see that software agents are a very special but essentially important case of computational actors. The relation between computational actors and software agents is similar to 5 the relation between the concept of a number and the concept of a rational number. As we know, there are numbers that are not rational. Being very useful for concurrent computations, CAM has very limited applications beyond computer science. That is why, taking the concept of an actor in all its generality and building a mathematical representation of a system actor, for which a computational actor is a very particular case, we extend CAM far beyond the area of computers, computer networks and computations. 3. Actor model in systems theory The basic concept in the System Actor Model (SAM) is the concept of an actor or more exactly, of a system actor, which, in particular, can be a computational actor. In what follows, we mostly call system actors simply by the name actor when it does not cause ambiguity. Informally, a system actor is a system that functions in some environment interacting with other systems. It means that System Actor Model developed in this paper is inherently dynamic. This notion of an actor is more formally described in the following way. Definition 3.1. Taking a system E of interacting systems {Rk ; k ∈ K}, which have the lower rank than E, we call the systems Rk actors and treat them as actors in E, while E is called the environment of each of the actors Rk. Note that in contrast to the Computational Actor Model where computational actors are processes or operators, a system actor can be (or can be interpreted as) an arbitrary system or an element/component of an arbitrary system, e.g., people, social networks, living beings, cells of living beings, molecules, artificial systems, such as computers or computer networks, processes and/or imaginary systems, such as heroes of novels or movies. Besides, computational actors can perform only three types of actions – create new actors, send messages and receive messages (Hewitt, 2007). In comparison with these limited abilities, system actors, in general, can perform any actions. Possible actions are described by the axioms that determine the environment of system actors. Although some authors call such systems by the name agent (cf., for example, (Doyle, 1983; Minsky, 1986)), it is more reasonable to call them actors because according to the common usage, an agent is a system (an actor) who/that acts on behalf of another system (actor). In addition, in political science and economics, an agent is a person or entity able to make decisions and take actions on behalf of another person or entity called the principal (Rees, 1985; Eisenhardt, 1989). 6 To build a mathematical model of an environment with actors, we construct a mathematical model (description) of an actor and an environment. Note that there is no similar mathematical model (description) in the computational Actor Model. A formal actor (system actor representation) A is described by a name and five structural components - three sets called set components of the actor A and two functions (or relations) called functional components of the actor A. Namely, we have the following structure A = (RelA , ActA , TrnA ; ReactA , ProactA) A is a name of the actor Three sets (set components) are: • RelA is the set of properties and relations of the actor A (usually only relations in the environment E are considered) • ActA is the set of possible actions of the actor A • TrnA is the set of possible actions aimed at the actor A Two functions (functional components), which are multivalued in the general case, are: The reaction function (reaction relation) shows responses of the actor A to actions on A ReactA: TrnA → ActA The proaction function (proaction relation) shows actions of the actor A instigated by properties and relations of A ProactA: RelA → ActA Reactions and proactions determine behavior of the actor. It is possible to consider the following example of a tentative proaction. Example 3.1. If an actor B is a friend of an actor A, then A is doing something good for B. The next example shows a prescribed proactions. Example 3.2. If an actor B is a friend of an actor A, then A always accepts messengers (massages) sent by B. As an example of reactions, we can consider the following situation. Example 3.3. The action aimed at an actor A is an e-mail from an actor B. The reaction of A is the response to this e-mail. Relations between an actor and data structures or knowledge structures, which may also be represented as actors, can represent the memory of the actor. Then self-actions can change this 7 memory performing computation, making decisions and deliberating subsequent actions. Note that it is possible to represent relations by properties and properties by relations (Burgin, 1985; 1990). Parts and elements of actor's components have their modalities described below. First, in this description of an actor A, it is useful to make a distinction between actualized parts (elements) and tentative parts (elements) of actor's components. For instance, some relations of A exist while others are only possible. Then the former relations are actualized while the latter are tentative. In a similar way, some actions have been performed or/and are performed while others are only possible. Then the former actions are actualized while the latter are tentative. Second, if an actor has a knowledge system, then it is useful to make a distinction between acknowledged parts (elements) and implicit parts (elements) of actor's components. For instance, an actor A can know about some of its relations and do not know about others. Then the former relations are acknowledged while the latter are implicit. Usually the components of an actor satisfy some restrictions. For instance, if an actor A is an automaton that does not give any output, e.g., if A is an accepting finite automaton, then all action of A are self-actions. In a formal setting, restrictions are described by axioms. Properties, relations and actions have various properties including temporal properties. For instance, a singular action is performed at one moment of time, while performance of a regular action always demands some interval of time. In the theory of computational automata, all actions are singular (Burgin, 2005). An important relation in this model is acquaintance. Namely, each actor A has a list of names (addresses) of forward acquaintances FAcq(A) and a list of names (addresses) of backward acquaintances BAcq(A). These lists regulate communication of the actor A. Namely, the actor A can send messages (messengers) only to forward acquaintances from FAcq(A) and can receive messages (messengers) from only backward acquaintances from BAcq(A). In particular, an actor (a system) can get feedback only from its backward acquaintances and can send feedback only to its forward acquaintances. This assumption is formalized by the following axioms. Let SMes(A, B) denotes the action of sending a messenger (a message) by an actor A to an actor B, ⇒ denotes implication, ◊ denotes modal value "possible" and ¬◊ denotes modal value "impossible". For instance, ◊ SMes(A, C) means that the actor A can send messages to the actor C. Axiom SM. a) ∀ A, C (C ∈ FAcq(A) ⇒ ◊ SMes(A, C)). 8 b) ∀ A, C (C ∉ FAcq(A) ⇒ ¬◊ SMes(A, C)). Informally, Axiom SMa means that the actor A can send messages (messengers) to any of its forward acquaintances. Axiom SMb means that the actor A cannot send messages (messengers) to any actor that (who) is not its forward acquaintance. Proposition 3.1. If Axiom SM is true, then ∀ A, C (C ∈ FAcq(A) ⇔ ◊ SMes(A, C)) Proof. By Axiom SMa, we have ∀ A, C (C ∈ FAcq(A) ⇒ ◊ SMes(A, C)) Thus, we have to prove only ∀ A, C (◊ SMes(A, C) ⇒ C ∈ FAcq(A)) Let us assume that the actor A can send messages to some actor C, i.e., ◊ SMes(A, C), but C does not belong to the forward acquaintances of A, i.e., C ∉ FAcq(A). However, by Axiom SMb, we have ¬◊ SMes(A, C)) and by principle of the Excluded Middle, our assumption is incorrect. Thus, we have ∀ A, C (◊ SMes(A, C) ⇒ C ∈ FAcq(A)) Proposition is proved. Let RMes(C, A) denotes the action of receiving a messenger (a message) by an actor A from an actor C. Axiom RM. a) ∀ A, C (C ∈ BAcq(A) ⇒ ◊ RMes(C, A)). b) ∀ A, C (C ∉ BAcq(A) ⇒ ¬◊ RMes(C, A)). Informally, Axiom RMa means that the actor A can receive messages (messengers) from any of its backward acquaintances. Axiom RMb means that the actor A cannot receive messages (messengers) from any actor that (who) is not its backward acquaintance. Proposition 3.2. If Axiom RM is true, then ∀ A, C (C ∈ BAcq(A) ⇔ ◊ Mes(C, A)) Proof is similar to the proof of Proposition 3.1. Note that C ∈ FAcq(A) does not always mean that A ∈ BAcq(C). Indeed, it is possible that an actor A can send messages to an actor C but C cannot receive messages from A. The following axiom for the environment E remedies this situation. Connectivity Axiom CA. ∀ A, C ∈ E (C ∈ FAcq(A) ⇔ A ∈ BAcq(C)). 9 Informally, it means that an actor A can receive messages (messengers) from an actor B if and only if B can send messages (messengers) to A. Acquaintances that belong to both lists FAcq(A) and BAcq(A) are called friends. We denote this set by In many cases (but not always), lists FAcq(A) and BAcq(A) coincide. In this case, they also Fr(A) = FAcq(A) ∩ BAcq(A)). coincide with the list Fr(A). Let us assume that Axioms CA, SM and RM are true. Proposition 3.3. ∀ A, B (B ∈ Fr(A) ⇒ A ∈ Fr(B)) Proof. The formula B ∈ Fr(A) means that B ∈ FAcq(A) and B ∈ BAcq(A). By Axiom CA, we have A ∈ BAcq(A) and A ∈ FAcq(A) Thus, A ∈ Fr(B). Proposition is proved. Proposition 3.3 allows proving the following result. Proposition 3.4. If in the environment E, all acquaintances are friends, then E satisfies Axiom CA. In the process of actor functioning, the lists of acquaintances and friends can change. There are five basic types of actor relations: • • • • Inner relations are relations between parts and elements of the actor A. For instance, if an actor A is an organization, then relations between members of this organization are inner relations of A. Internal relations are relations between the actor A and its parts and elements. For instance, if an actor A is an organization, then the relation "a member H of A receives salary from A" is an internal relation of A. Outer relations are relations of the actor A to other actors, their parts, elements and the environment. For instance, if actors A and B are organizations, then cooperation between A and B is an outer relation of A. Intermediate relations are relations of parts and elements of the actor A to other actors, their parts, elements and the environment. For instance, if an actor A is an organization, 10 then any relation between a member H of A and an actor K who is not a member of A is an intermediate relation of A. • External relations are relations of other actors, in which the actor A is included. For instance, if actors are companies, then "to be a supplier" is an external relation of A when A is a supplier for another company. Note that it is possible to consider actions, reactions and proactions as relations. However, it is more efficient to treat these structures separately making emphasis on the functionality and dynamics. According to the theory of autopoiesis developed by Maturana and Varela (1973), relations and properties play a crucial role for autopoietic systems, which can be described briefly as self- producing devices, or a self-generating systems with the ability to reproduce themselves recursively. Relations and properties of a system determine the structure of this system (Burgin, 2012). Indeed, autopoietic systems are structure-determined systems according to the principle of structural determinism, which states that the potential behavior of the system depends on its structure (Maturana, 1997). It means that all actions of actors representing autopoietic systems are functions of relations and properties. Observing actions in the real world, we see that there are different types, classes, groups and kinds of actions. Let us consider some of them. Temporal characteristics of actions determines three groups of reactions and proactions: − Sharp immediate reaction (proaction) of A starts immediately after the beginning of the corresponding action on A (immediately after the property or relation becomes overt). − Reserved immediate reaction (proaction) of A starts when the corresponding action on A ends (when the corresponding property or relation becomes comprehensible). − Delayed reaction (proaction) of A is performed when some time passes after the corresponding action on A (when some time passes after the corresponding property or relation becomes comprehensible). Definitions imply the following results. Proposition 3.5. If an action a is not immediate, then a and any sharp immediate reaction to a are parallel in time. Proposition 3.6. An action and a reserved immediate reaction to it are strictly sequential in time. Proposition 3.7. An action and a delayed immediate reaction to it are sequential in time. 11 There are other temporal relations between separate actions and events. Definition 3.2. a) Temporal independence of events (actions) E1 and E2 means autonomy of their occurrence, i.e., either E1 can take place before E2 or E2 can take place before E1 or they can take place at the same time. b) Two events (actions) are temporally dependent if they are not are temporally independent. For instance, events in two disconnected computing systems are temporally independent. Note that disconnectedness means that these computers are not connected not only to one another but also to another system, for example, to the Internet. However, temporal independence does not prohibit simultaneous occurrence or coincidence of actions and events. Proposition 3.8. Temporal dependence is a transitive relation. Another important concept is temporal incomparability. Definition 3.3. a) Temporal incomparability of events (actions) E1 and E2 means that it is not known whether they happen at the same time or which of them happens before the other. b) Two events (actions) are temporally comparable if they are not are temporally incomparable. For instance, events in two disconnected computers, which are not observed by the same observer, are temporally incomparable. Proposition 3.9. Temporal comparability is a transitive relation. Temporal independence and incomparability are related to concurrency. Definition 3.4. Concurrency of two or more events or actions means their temporal independence and/or temporal incomparability, or in other words, that time of their happening is independent and sometimes incomparable. As temporal independence allows simultaneous occurrence or coincidence, the introduced concept of concurrency comprises other interpretations of this term. Concurrency is intrinsically related to such properties of events and actions as parallelism and sequentiality. Definition 3.5. Two or more events or actions are parallel if their time intervals intersect (moments of their occurring coincide when they have zero duration, i.e., they are momentary). For instance, when people read and understand some text, these actions are usually parallel but not always strictly parallel. Note that independence of events allows them to be parallel. It implies that some parallel events can also be concurrent. 12 Proposition 3.10. If a momentary event (action) E1 is parallel to a momentary event (action) E2 and the event (action) E2 is parallel to a momentary event (action) E3 , then all three events (actions) are parallel. If the events are not momentary, then this result is not always true. For instance, let us consider events E1 , E2 and E3 such that E1 starts at time 0 and ends at time 3, E2 starts at time 2 and ends at time 5, and E3 starts at time 4 and ends at time 7. Then the event E1 is parallel to the event E2 and the event E2 is parallel to the event E3 , but the event E1 is not parallel to the event E3 . However, for interval events (actions), i.e., events (actions) with interval duration, it is possible to prove a result similar to Proposition 3.9. Proposition 3.11. If an interval event (action) E1 is parallel to an interval event (action) E2 , the event (action) E2 is parallel to an interval event (action) E3 and the event (action) E1 is parallel to the event E3, then all three events (actions) are parallel. However, if the events are neither interval nor momentary, then this result is not always true. For instance, let us consider events E1 , E2 and E3 such that E1 starts at time 0 and ends at time 3, E2 starts at time 2 and ends at time 5, and E3 starts at time 0 and continues to time 1, then restarts at time 4 and ends at time 7. Then the event E1 is parallel to the event E2 and the event E2 is parallel to the event E3 , the event E1 is parallel to the event E3 but all three events are not parallel. Definition 3.6. Two or more events or actions are strictly parallel if their beginning and end coincide and they go (take place) in the same time. For instance, when the user switches her computer on (the first event), the computer starts working (the second event, which is strictly parallel to the first event). Proposition 3.12. If an event (action) E1 is strictly parallel to an event (action) E2 and the event (action) E2 is strictly parallel to an event (action) E3 , then the event (action) E1 is strictly parallel to the event (action) E3 . Remark 3.1. For parallel events (actions), this result is not always true. Definition 3.7. a) Two events or actions are sequential if one of them, say E2 , starts after the other, say E1 , ends. b) In this case, the event (action) E2 is called subsequent to the event (action) E1 . For instance, reception of information is subsequent to sending this information but usually it is not strictly subsequent. Proposition 3.13. The relation between events and actions to be sequential is transitive. 13 Another important relation between events and actions is to be strictly sequential. Definition 3.8. a) Two events or actions are strictly sequential if one of them, say E2 , starts exactly at the moment the other, say E1 , ends. b) In this case, the event (action) E2 is called strictly subsequent to the event (action) E1 . In the theory of finite automata, it is assumed that starting from the second transition, each transition of the automaton is strictly subsequent to the previous transition (Burgin, 2005). Proposition 3.14. If an event (action) E1 is strictly subsequent to an event (action) E2 and the event (action) E2 has positive duration and is strictly subsequent to an event (action) E3 , then event (action) E1 is not strictly subsequent to the event (action) E3 . There are also structural characteristics of actions. One of them is direction. Direction of actions determines three groups of actions: • An external action of an actor is directed at other actors (cf. Figure 2). • An internal action or a self-action of an actor is directed at the same actor and usually results in self-transformation (cf. Figure 1). • A combined action of an actor is directed both at other actors and at the same actor(cf. Figure 3). A Figure 1. A self-action is an action of an agent on itself. A Figure 2. An external action is directed at other actors A Figure 3. A combined action goes inside and outside. 14 Example 3.4. Reception of information is an example of a self-action. Example 3.5. Computation performed by a system actor and any computational operation are examples of a self-action. Example 3.6. Decision-making of a system actor is an example of a self-action. Example 3.7. Sending information is an example of an external action. Example 3.8. Working an inductive Turing machine transforms the content of its working register and from time to time, sends this content to the output register (Burgin, 2005). The action of the machine when it is doing both operations at the same time is a combined action. Another structural characteristic of actions is modality, which determines the status of actions in the environment. There are three modalities of actions – positive, negative and neutral – and each of them contains four classes. Positive modalities of actions: − Possible actions − Tolerable actions − Permitted actions − Performed actions Negative modalities of actions: − Impossible actions − Intolerable actions − Prohibited actions − Not performed (but possible/permitted) actions Neutral modalities of actions: − Unknown actions − Unidentified actions − Unspecified actions − Indefinite actions There are definite relations between modalities of actions. Proposition 3.15. a) Any unknown action is unidentified. b) Any unidentified action is unspecified. c) Any performed action is possible. 15 d) Any unknown possible and permitted action is not performed. Structural characteristics of actions show that there are simple actions and compound actions, which are compositions of other actions. Compositions of actions are constructed using operations with actions. For instance, performing one action after another gives us the sequential composition of these actions. If an action a is a composition of actions a1 , a2 , a3 , …, an , for example, a = ω(a1 , a2 , a3 , …, an) where ω is an n–ary operation with actions, then any action ai (i = 1, 2, 3, … , n) is included in or is a part of the action a. It is denoted by ai ⊆ a. Informally, the relation b ⊆ a means that performance of the action a includes performance of the action b. Proposition 3.16. For any actions a, b and c, relations a ⊆ b and b ⊆ c imply the relation a ⊆ c. Indeed, as a composition of compositions of actions is a composition of actions, relations a ⊆ b and b ⊆ c imply the relation a ⊆ c. It means that the relation "to be a part" or "to be included" is transitive. Composition preserves direction of actions. Proposition 3.17. A composition of internal (external or combined) actions of the same actor is an internal (external or combined) action. Organization of actions determines three groups of actions: − Direct actions does not include additional operations (actions) − Mediated actions include additional operations (actions or processes), for example, such as computation, meditation, contemplation or actions of other actors − Void actions or inactions Not to perform an action is also an action. It is a void action. All other actions are proper actions. It is possible to build the system Actor Model (SAM) with one void action or with different void actions. It is possible to give a more precise description of actor's behavior when SAM allows different void actions. In this case, we have the following definition. Definition 3.9. Not to perform an action a is the inaction ¬a. For instance, when a person A is standing near the river and doing nothing seeing a person B is drowning, this is a negative void action. When the Allies did nothing to prevent Hitler from seizing Austria and a part of Czechoslovakia, it was also a negative void action. 16 At the same time, there are positive void actions. For instance, when a person does not steal, it is a positive void action. The concept of inaction or non-action plays an important role in Taoism because one of its central principles is the Principle of non-action (Wu wei in Chinese). Wu wei from the Tao Te Ching literally means non-action or non-doing and is connected to the paradox weiwuwei: "action without action" (Kirkland, 2004; Klaus, 2009). Let us consider some properties of void actions. Proposition 3.18. ¬¬a = a. Informally, it means that when non-doing of action a is not performed, then action a is performed. In essence, this is a version of the Principle of Excluded Middle because the proof of Proposition 3.18 uses this Principle and it is possible to consider systems of actors for which this assertion is not true. Common sense tells us that independently in what way you compose non-doing, it will always be non-doing. We formalize this impression in the following axiom. Emptiness Axiom EA. If a1 , a2 , a3 , …, an are actions and ω is an n–ary operation with actions, then ω(¬a1 , ¬a2 , ¬a3 , …, ¬an) = ¬ω(a1 , a2 , a3 , …, an) Axiom EA implies the following result. Proposition 3.19. A composition of inactions is an inaction. However, in general, Axiom EA is not always valid and a composition of inactions can be a proper action. For instance, let us consider the binary composition L(x, y), which combines two actions inferring the third action when only three actions can be performed. To provide an example of this situation, we can take the situation when a person can only either run (action a) or walk (action b) or stand (action c). Then combining two inactions ¬a (not running) and ¬b (not walking), we have L(a, b) = c, which is a proper action. Proposition 3.20. If a ⊆ b, then ¬b ⊆ ¬a. Indeed, if an action b includes an action a, then the absence of a implies and thus, includes, the absence of b. It is useful to consider the total inaction TIA when simply nothing is done. Proposition 3.21. For any action a, we have ¬a ⊆ TIA . Definitions imply the following result. 17 Proposition 3.22. A composition of non-void (proper) actions is a mediated action. There other important types of actions. A primitive action is a direct action that depends only on the input actions of other actors in the case of reactions or only properties and relations in the case of proactions. An automatic action is a direct action that depends both on actions of other actors and on properties/relations. Note that an inaction also can be primitive or automatic. Proposition 3.23. When an action a is primitive (automatic), the inaction ¬a is also primitive (automatic). Automatic actions allow unification of reactions and proactions in one (multivalued in a general case) function of combined actions CombactA: TrnA × RelA → ActA In this context, the function ReactA is a restriction of the function CombactA when the action on A is void and the function ProactA is a restriction of the function CombactA when the property/relation is void. This gives us the following result. Proposition 3.24. Any primitive action is an automatic action. Different types of actions spawn different types of actors. Definition 3.10. A behaviorally primitive actor A has only primitive actions. For instance, finite automata with one state are behaviorally primitive actors because their actions depend only on the input. Definition 3.11. A behaviorally automatic actor A has only automatic actions. For instance, finite automata are behaviorally primitive actors because their actions depend on both the input and inner state. Proposition 3.24 implies the following result. Corollary 3.1. Any behaviorally primitive actor is a behaviorally automatic actor. There are various relations between actors. Definition 3.12. Two actors are identical if they have the same structural components. For instance, in contemporary industry, identical copies of many devices, such as vehicles, planes, computers and cell phones, are produced. In the system Actor Model, all these copies are represented by identical actors. Lemma 3.1. Identity is an equivalence relation in sets of actors. 18 It is possible to find identical actors in many areas. One of them is theory and technology of information processing. Thus, there are models of computational systems, which contain many (sometimes, infinite) identical computing elements. Examples are cellular automata, artificial neural networks and iterative arrays. For instance, a cellular automaton is a system of identical finite automata called cells, which form a net and interact with one another. A cellular automaton is determined by the following parameters (Burgin, 2005): 1. The space organization of the cells. In the majority of cellular automata, cells organized in a simple rectangular grid (mostly it is a one-dimensional string of cells and a two- or three- dimensional grid of cells), but in some cases, other arrangements, such as a honeycomb or Fibonacci trees. a. The topology of the cellular automaton is determined by the type of the cell neighborhood, which consists of other cells that interact with this cell. In a grid, these are normally the cells physically closest to the cell in question. For instance, if each cell has only two neighbors (right and left), it defines linear topology. Such cellular automata are called linear or one-dimensional. It is possible to consider linear automata with the neighborhood of some radius r > 1. When there are four cells (upper, below, right, and left), the CA has two-dimensional rectangular topology. Such cellular automata are called planar or two-dimensional. 2. The dynamics of a cellular automaton, which determines by what rules cells exchange information with each other. Traditionally, only rectangular organization of the cells and their neighborhoods has been considered for cellular automata. Recently, researchers have begun studies of cellular automata in the hyperbolic plane or on a Fibonacci tree (Margenstern, 2002). It is proved that such automata are more efficient than traditional cellular automata in the Euclidean plane. This higher efficiency is a result of a better topology in cellular automata in the hyperbolic plane. According to the system Actor Model, each element of a cellular automaton is an actor and its actions consist of computing and communicating operations. Looking at computer technology, we see that from the perspective of a manufacturer, products, e.g., computers, of the same type are identical. Another important relation between actors is dynamic equivalence. 19 Definition 3.13. Two actors are dynamically equivalent if they have the same action components. When it is necessary to solve the same problem for different input data, it is possible to use equivalent actors to this in a parallel or concurrent mode. This is often done in multiprocessor computers where identical processors perform necessary computations. Lemma 3.2. Dynamic equivalence is an equivalence relation in sets of actors. Identity of actors is a stronger relation than dynamic equivalence. Lemma 3.3. Identical actors are dynamically equivalent. Dynamic equivalence determines similarities in actor's behavior. Proposition 3.25. An actor without actions is dynamically equivalent to an actor that has only void actions. Proposition 3.26. An actor A dynamically equivalent to a behaviorally primitive actor B is behaviorally primitive. Proposition 3.27. An actor A dynamically equivalent to a behaviorally automatic actor B is behaviorally automatic. Another important relation between actors is homology. Definition 3.14. Two actors A and B are homological if all their corresponding structural components are isomorphic. For instance, for homological actors A and B, there are isomorphisms between RelA and RelB , between ReactA and ReactB , and between ProactA and ProactB . Example 3.9. Let us consider two deterministic finite automata A and B. They have the same set of states and the same set of start and final states. The first has the alphabet {0, 1} and the second the alphabet {a, b}. Besides, all transitions of A produced by input 0 are the same as all transitions of B produced by input a and all transitions of A produced by input 1 are the same as all transitions of B produced by input b. Then these automata are homological actors. Lemma 3.4. Homology is an equivalence relation in sets of actors. Identity of actors is a stronger relation than homology. Lemma 3.5. Identical actors are homological. Let us assume that isomorphisms between ReactA and ReactB and between ProactA and ProactB . preserves primitive actions. Then we have the following result. 20 Proposition 3.28. An actor A homological to a behaviorally primitive actor B is behaviorally primitive. Let us assume that isomorphisms between ReactA and ReactB and between ProactA and ProactB . preserves automatic actions. Then we have the following result. Proposition 3.29. An actor A homological to a behaviorally automatic actor B is behaviorally automatic. A weaker type of relations is dynamic homology Definition 3.15. Two actors A and B are dynamically homological if all their corresponding action components are isomorphic. Lemma 3.6. Dynamic homology is an equivalence relation in sets of actors. Dynamic equivalence of actors is a stronger relation than dynamic homology. Lemma 3.7. Dynamically equivalent actors are dynamically homological. Let us assume that isomorphisms between ReactA and ReactB and between ProactA and ProactB . preserves primitive actions. Then we have the following result. Proposition 3.30. An actor A dynamically homological to a behaviorally primitive actor B is behaviorally primitive. Let us assume that isomorphisms between ReactA and ReactB and between ProactA and ProactB . preserves automatic actions. Then we have the following result. Proposition 3.31. An actor A dynamically homological to a behaviorally automatic actor B is behaviorally automatic. According to their structure, we discern four classes of actors: − A structurally prime actor A does not have components or parts. − A structurally primitive actor A does not have components or parts, which are also actors. − A structurally composite actor A has parts and/or components. − A structurally compound actor A has parts and/or components, which are also actors. In the actor's structure elements are also treated as parts. The scale of observation defines what actors are prime. Thus, to be a prime actor depends on the scale of observation/treatment. For instance, in the observation scale of society, people are primitive actors. At the same time, in the observation scale of biology, people are composite actors. The scale of modeling defines what actors are primitive. Thus, to be a primitive actor depends on the scale of modeling /representation. For instance, in the modeling scale of society, it is natural to 21 represent people as primitive actors. At the same time, in the modeling scale of biology, it is natural to represent people as compound actors. It is possible to develop a scale (ranging) of actors and deal with parts and components of a actor in this scale. Namely, an actor A that is a part/component of another actor B has lower range than B. The system (environment) E can be a model of a real system R, which can be physical, mental or structural. The system R is called a modeled domain of E. In general, one environment E can model different domains. Let us consider a modeled domain R of an environment E. Proposition 3.32. If R is the modeled domain of environment E and a subdomain P of R is a modeled domain of D, then there is an injection of the set of all actors from D into the set of all actors from E. It is possible to introduce the following axiom Modeling Axiom MA. Any object in the modeled domain R is modeled by an actor in E. If Pythagoras asserted "Everything is a number," the Modeling Axiom states "Everything and everybody is an actor." The computational Actor Model that satisfies the Modeling Axiom is called the universe of CAM (Agha, 1986). Let us consider an actor A with the inner structure Q. Proposition 3.33. If the Modeling Axiom is valid for an environment E and its modeled domain R, then: (a) Any structurally primitive actor is structurally prime. (b) Any structurally composite actor is structurally compound. Corollary 3.2. If the Modeling Axiom is valid for an environment E and its modeled domain R, then there are only structurally primitive and structurally compound actors in E. Definition 3.16. A primary actor A is not a part or component of other actors. According to their communication, we discern five classes of actors – closed, inactive, generative, undemanding and open actors. Definition 3.17. A closed actor A does not send and receive messengers (messages). The concept of a closed actor allows treating almost anything, for example, tables, chairs, mountains, rivers, words, sounds, etc. as actors. Definition 3.18. An inactive actor A does not send messengers (messages). 22 For instance, a sleeping woman does not send messengers (messages). Another example of an inactive actor is a receptor such as an automaton, which accepts input but gives no output (Burgin, 2005). Definitions imply the following result. Lemma 3.8. Any closed actor A is inactive. The dual concept to an inactive actor is a non-receptive actor. Definition 3.19. A non-receptive actor A does not receive messengers (messages). An example of a non-receptive actor is a generator, i.e., such as an automaton, which does not accept input but gives output (Burgin, 2005). Another example of a non-receptive actor is a black hole (Thorne, 1994; Davies, 1995). Definitions imply the following results. Lemma 3.9. Any closed actor A is non-receptive. It means that the property "to be closed" is stronger than the property "to be non-receptive." Lemma 3.10. A non-receptive and inactive actor A is closed. Opposite to closed actor are open actors. Definition 3.20. An open actor A sends and receives messengers (messages). Definitions imply the following results. Lemma 3.11. Any open actor A is active. It means that the property "to be open" is stronger than the property "to be active." Lemma 3.12. A receptive and active actor A is open. It is possible to distinguish actor by messages they send. Definition 3.21. An undemanding actor A does not send requesting messengers (requests). Definitions imply the following results. Lemma 3.13. Any inactive actor A is undemanding. Lemmas 3.9 and 3.13 imply the following result. Corollary 3.3. Any closed actor A is undemanding. It is possible to develop a scale (ranging) of actors and deal with parts and components of a primary actor in this scale. Because an actor is functioning in some environment, it is also practical to use an extended actor representation, which includes relevant characteristics of the environment. 23 An extended actor representation consists of two names, three sets and four functions (or relations) (A, E) = (RelA , ActA , TrnA ; ReactA , ProactA , VReactA , VProactA) A is a name of the actor. C is a name of the actor's environment. Three sets are: • RelA is the set of properties of A and relations of A to other actors and the environment • ActA is the set of possible actions of A • TrnA is the set of possible actions on A Four functions (multivalued in the general case) are: The reaction function shows responses of A to actions on A ReactA: TrnA → ActA Proactions show actions on A instigated by properties and relations of A ProactA: RelA → ActA For instance, if B is a friend of A, then A is doing something good for B. The virtual reaction function shows responses of A to all possible actions VReactA: ActpE → ActA Here ActpE is the set of all possible actions in E. The virtual proaction function shows actions on A instigated by all properties and relations, which exist in E VProactA: RelpC → ActA Here RelpC is the set of all possible properties and relations in E. Definitions imply the following results. Lemma 3.14. ReactA is a restriction of VReactA . Lemma 3.15. ProactA is a restriction of VProactA . In the System Actor Model, we also have a mathematical model of an environment. An environment representation is described by a name, two sets and two functions (or relations) E = (RelpE , ActpE , TrnE ; EReactE , EProactE) A is a name of the actor Two sets are: • RelpE is the set of all possible properties and relations in E. 24 • ActpE is the set of all possible actions in E. Two functions (multivalued in the general case) are: EReactions show all possible responses to actions in E EReactE: TrnE → ActE EProactions show all possible actions instigated by properties and relations in E EProactE: RelE → ActE Note that the systems Rk in the environment E can have different ranks. For instance, in society, actors include separate individuals, organizations, countries, and so on. Definition 3.22. a) If an actor A is a proper subsystem of an actor B, then the rank of A is lower than the rank of B. b) If actors A and B consist of elements of the same rank, then the rank of A is equal to the rank of B. By definition, the environment E has the highest rank in the system Actor Model. Proposition 3.34. Elements, parts and components of an actor A have lower rank than A. 4. Conclusion We have built a mathematical model of multicomponent interactive systems, which is called the System Actor Model and based on the formal structure of actors functioning in a multifarious convoluted environment. Different properties of such systems represented by an environment with actors have been obtained. Actions and events are analyzed in this context and different classes of events and actions are explicated and studied. Actors are also classified according to their traits. In addition, we elaborated a mathematical model of the environment. One of the main targets of this work is to construct mathematical tools for exploration of social systems. To conclude, we formulate some open problems for the System Actor Model. The first cluster of problems is related to actions. Problem 1. Formalize and study results of actions. Problem 2. Formalize and study consequences of actions. Problem 3. Formalize and study causes of actions. Problem 4. Formalize and study in more detail structural, temporal and spatial characteristics of actions. 25 The second cluster of problems is related to actors. Problem 5. Formalize and study tasks of actors. Problem 6. Formalize and study obligations of actors. Problem 7. Formalize and study norms of actors. Problem 8. Formalize and study values of actors. The third cluster of problems is related to concepts of agents and oracles, which are connected to the concept of actors. Problem 9. Formalize and study relations between agents and actors. Problem 10. Formalize and study relations between oracles and actors. References 1. Agha, G. A. (1986) ACTORS: A Model of Concurrent Computation in Distributed Systems, The MIT Press Series in Artificial Intelligence, The MIT Press, Cambridge, Massachusetts 2. Attiya H. and Ellen, F. Impossibility Results for Distributed Computing, Morgan & Claypool Publishers, May 1, 2014 3. Barwise, J. and Seligman, J. (1997) Information Flow: The Logic of Distributed Systems, Cambridge Tracts in Theoretical Computer Science 44, Cambridge University Press 4. Bergson, H. Time and Free Will: An Essay on the Immediate Data of Consciousness, George Allen and Unwin, London, 1910 5. Birman, K.P. (2005) Clock Synchronization and Synchronous Systems, in Reliable Distributed Systems, Springer New York, pp. 493-508 6. Boixo, S., Caves, C.M., Datta, A. and Shaji, A. (2006) On decoherence in quantum clock synchronization, Laser Physics, v. 16, issue 11, pp. 1525-1532 7. Burgin, M. (1985) Abstract theory of properties, in Non-classical Logics, Institute of Philosophy, Moscow, pp. 109-118 (in Russian) 8. Burgin, M. (1990) Abstract Theory of Properties and Sociological Scaling, in Expert Evaluation in Sociological Studies, Kiev, pp. 243-264 (in Russian) 9. Burgin, M. A System Approach to the Concept of Time, Philosophical and Sociological Thought. - 1992. – No. 8 (in Russian and Ukrainian) 10. Burgin M. Time as a Factor of Science Development, Science and Science of Science, 1997, No. 1/2, pp. 45-59 11. Burgin, M. Elements of the System Theory of Time, LANL, Preprint in Physics 0207055, 2002, 21 p. (electronic edition: http://arXiv.org) 12. Burgin, M. Super-recursive Algorithms, Springer, New York/Heidelberg/Berlin, 2005 13. Burgin, M. (2012) Structural Reality, Nova Science Publishers, New York 26 14. Burgin, M., Karplus, W. and Liu, D. The Problem of Time Scales in Computer Visualization, in "Computational Science", Lecture Notes in Computer Science, v. 2074, part II, 2001, pp.728-737 15. Burgin, M., Mikkilineni, R. and Morana, G. Intelligent organisation of semantic networks, DIME network architecture and grid automata, International Journal of Embedded Systems, v. 8, No. 4, 2016, pp. 352-366 16. Buşoniu, L., Babuška, R. and De Schutter, B. Multi-agent reinforcement learning: An overview, in Innovations in Multi-Agent Systems and Applications (D. Srinivasan and L.C. Jain, eds.), v. 310 of Studies in Computational Intelligence, Berlin, Germany: Springer, pp. 183–221, 2010. 17. Davies, P. About Time, Simon & Schuster, New York/London/Tokyo, 1995 18. Dolev, D., Halpern, J.Y. and Strong, H. R. On the Possibility and Impossibility of Achieving Clock Synchronization, Journal of Computer and System Sciences, v. 32, 230-250 (1986) 19. Einstein, A., Lorentz, H.A., Weil, H., and Minkowski, H. The Principle of Relativity, Dover, 1923 20. Fischer, M.J., Lynch, N.A. and Paterson, M.S. Impossibility of Distributed Consensus with One Faulty Process, Journal of the Association for Computing Machinery, v. 32, No. 2, April 1985, pp. 374-382 21. Hewitt, C. (2007) What is Commitment? Physical, Organizational, and Social, Lecture Notes in Artificial Intelligence (Vázquez-Salceda, J. and Pablo Noriega, P. EdsS), Springer 22. Hewitt, C. (2012) What is computation? Actor Model versus Turing's Model, in A Computable Universe, Understanding Computation & Exploring Nature as Computation (H. Zenil, Ed.) World Scientific Publishing Company/Imperial College Press 23. Hewitt, C., Bishop, P. and Steiger, R. (1973) A Universal Modular Actor Formalism for Artificial Intelligence, IJCAI'73 Proceedings of the 3rd international joint conference on Artificial intelligence, Morgan Kaufmann Publishers Inc., San Francisco, CA, pp. 235-245 24. Kirkland, R. Taoism: The Enduring Tradition, Routledge, London/New York, 2004 25. Klaus, H. The Tao of Wisdom. Laozi – Daodejing. Chinese-English-German. Hochschulverlag, Aachen, 2009 26. Lamport, L. Unsolved problems, solved problems, and non-problems in concurrency , in Proceedings of the 3rd ACM Conference on Principles of Distributed Computing, 1984, pp. l-l 1 27. Lindsay, B. G., Selinger, P. G., Galtieri, C., Gray, J. N., Lorie, R. A., Price, T. G., Putzolu, F., Traiger, I. L., and Wade, B. W. Notes on distributed databases. IBM Res. Rep. RJ2571, IBM Research Division, San Jose, Calif., 1979. 28. Marzullo, K. Loosely-Coupled Distributed Services: A Distributed Time System, Ph.D. dissertation, Stanford University, 1983 29. Maturana, H. (1997) Metadesign, in Articulos y Conferences "Diez Años de Post- Racionalismo en Chile", Instituto de Terapia Cognitiva Web, Santiago http://www.inteco.cl/articulos/006/doc_ing1.htm last visited 9/6/2017 27 30. Maturana, H. and Varela, F. (1998) The Tree of Knowledge: The Biological Roots of Human Understanding, Shambhala, Boston 31. Mills, D.L. (1991) Internet time synchronization: the Network Time Protocol. IEEE Trans. Communications COM-39, v. 10, pp. 1482-1493 32. Milner, R. (1993) Elements of Interaction, Communications of the ACM, v. 36 No. 1, pp. 78- 89 33. Prigogine, I. From being to Becoming: Time and Complexity in Physical Systems, San Francisco, 1980 34. Shoham, Y. and Leyton-Brown, K. Multiagent Systems: Algorithmic, Game Theoretic and Logical Foundations, Cambridge University Press, Cambridge, 2008 35. Thorne, K.S. Black Holes and Time Warps, Norton, New York, 1994 36. Vlassis, N. A Concise Introduction to Multiagent Systems and Distributed Artificial Intelligence, Synthesis Lectures in Artificial Intelligence and Machine Learning, Morgan & Claypool Publishers, 2007 37. Weiss, G. (ed.) Multiagent Systems: A Modern Approach to Distributed Artificial Intelligence, MIT Press, New York/London, 1999 38. Wiener, N. Cybernetics, or Control and Communication in the Animal and the Machine, MIT Press and Wiley, New York/London, 1961 39. Winfree, A.T. The Timing of Biological Clocks, Scientific American Library, New York, 1987 28
1908.09658	1	1908	2019-08-26T13:08:49	Dynamic Term-Modal Logic for Epistemic Social Network Dynamics (Extended Version)	[ "cs.MA", "cs.LO", "cs.SI" ]	Logics for social networks have been studied in recent literature. This paper presents a framework based on dynamic term-modal logic (DTML), a quantified variant of dynamic epistemic logic (DEL). In contrast with DEL where it is commonly known to whom agent names refer, DTML can represent dynamics with uncertainty about agent identity. We exemplify dynamics where such uncertainty and de re/de dicto distinctions are key to social network epistemics. Technically, we show that DTML semantics can represent a popular class of hybrid logic epistemic social network models. We also show that DTML can encode previously discussed dynamics for which finding a complete logic was left open. As complete reduction axioms systems exist for DTML, this yields a complete system for the dynamics in question.	cs.MA	cs	Dynamic Term-Modal Logic for Epistemic Social Network Dynamics (Extended Version) Andrés Occhipinti Liberman1 and Rasmus K. Rendsvig2 1 DTU Compute [email protected] 2 Center for Information and Bubble Studies, University of Copenhagen [email protected] Abstract. Logics for social networks have been studied in recent liter- ature. This paper presents a framework based on dynamic term-modal logic (DTML), a quantified variant of dynamic epistemic logic (DEL). In contrast with DEL where it is commonly known to whom agent names refer, DTML can represent dynamics with uncertainty about agent iden- tity. We exemplify dynamics where such uncertainty and de re/de dicto distinctions are key to social network epistemics. Technically, we show that DTML semantics can represent a popular class of hybrid logic epis- temic social network models. We also show that DTML can encode pre- viously discussed dynamics for which finding a complete logic was left open. As complete reduction axioms systems exist for DTML, this yields a complete system for the dynamics in question. Keywords: social networks, term-modal logic, dynamic epistemic logic 1 Introduction Over recent years, several papers have been dedicated to logical studies of social networks, their epistemics and dynamics [2, 10 -- 14, 18 -- 22, 24, 25]. The purpose of this literature typically is to define and investigate some social dynamics with respect to e.g. long-term stabilization or other properties, or to introduce formal logics that capture some social dynamics, or both. This paper illustrates how dynamic term-modal logic (DTML, [1]) may be used for the second purpose. In general, term-modal logics are first-order modal logics where the index of modal operators are first-order terms. I.e., the operators double as predicates to the effect that e.g. ∃xKxN (x, a) is a formula -- read, in this paper, as "there there exists an agent that knows of itself that it is a social network neighbor of a". The dynamic term-modal logic of [1] extends term- modal logic with suitably generalized action models that can effectuate both factual changes (e.g. to the network structure) as well as epistemic changes. For all the DTML action model encodable dynamics, [1] presents a general sound and complete reduction axiom-based logic in the style of dynamic epistemic logic (DEL, [3, 4]). Hence, whenever an epistemic social network dynamics is encodable using DTML, completeness follows. With this in mind, the main goal of this paper is to introduce and illustrate DTML as a formalism for representing epistemic social network dynamics, and to show how it may be used to obtain completeness results. To this end, the paper progresses as follows. Sec. 2 sketches some common themes in the logical literature on social networks before introducing DTML and its application to epistemic social networks. Sec. 2 contains the bulk of the paper, with numerous examples of both static DTML models and action models. The examples are both meant to showcase the scope of DTML and to explain the more non-standard technical details involved in calculating updated models. In Sec. 3, we turn to technical results, where it is shown that DTML may encode popular static hybrid logical models of epistemic networks, as well as the dynamics of [12], for which finding a complete logic was left open. Sec. 4 contains final remarks. 2 Models and Languages for Epistemic Social Networks To situate DTML in the logical literature on social networks, we cannot but describe the literature in broad terms. We omit both focus, formal details and main results of the individual contributions in favor of a broad perspective. That said, then all relevant literature in one way or other concern social networks. In general, a social network is a graph (A, N ) where A is a set of agents and N ⊆ A × A is represents a social relation, e.g., being friends on some social media platform. Depending on interpretation, N may be assumed irreflexive and symmetric. Social networks may be augmented with assignments of atomic properties to agents, representing e.g. behaviors, opinions or beliefs. One set of papers investigates such models and their dynamics using fully propositional static languages [13, 20, 24, 25]. A second set of papers combines social networks with a semantically repre- sented epistemic dimension in the style of epistemic logic. In these works, the fundamental structure of interest is (akin to) a tuple (A, W, {Nw}w∈W , ∼) with agents A and worlds W , with each world w associated with a network Nw ⊆ A×A, and ∼: A → P(W ×W ) associating each agent with an indistinguishability (equivalence) relation ∼a. Call such a tuple an epistemic network structure. The existing work on epistemic network structures may be organized in terms of the static languages they work with: propositional modal logic [2,14] or hybrid logic [9 -- 12, 18, 19, 21, 22]. In the former, the social network is described using designated atomic propositions (e.g., Nab for 'b is a neighbor of a'). To produce a model, an epistemic network structure is augmented with a propositional val- uation V : P → P(W ). Semantically, Nab is then true at w iff (a, b) ∈ Nw. Knowledge is expressed using operators {Ka}a∈A as in standard epistemic logic with Ka the Kripke modality for ∼a. In the hybrid case, the network is instead described using modal operators. The hybrid languages typically include a set of agent nominals N om (agent names), atoms P and indexical modal operators K and N , read "I know that" and "all my neighbors". Some papers additionally include state nominals, hy- brid operators (@x, ↓x) and/or universal modalities U ("for all agents"). A hy- brid network model is an epistemic network structures extended with two assignments: a nominal assignment g : N om → A that names agents, and a two-dimensional hybrid valuation V : P → P(W × A), where (w, a) ∈ V (p) represents that the indexical proposition p holds of agent a at w. The satis- faction relation is relative to both an epistemic alternative w and an agent a, where the noteworthy clause are: M, w, a = p iff (w, a) ∈ V (p); M, w, a = Kϕ iff M, v, a = ϕ for every v ∼a w; and M, w, a = N ϕ iff M, w, b = ϕ for every b such that Nw(a, b). With these semantics, formulas are read indexically. E.g. KN p reads as "I know that all my neighbors are p". In relation to these two language types, the term-modal approach of this pa- per lies closer to the former: By including a binary 'neighbor of' relation symbol N in the signature of a term-modal language, the social network component of models is described non-modally. This straightforwardly allows expressing e.g. that that all agents know all their neighbors (∀x∀y(N (x, y) → Kx(N (x, y))) or that an agent has de re vs. de dicto knowledge of someone being a neighbor (∃xKaN (a, x) vs. Ka∃xN (a, x)). Moreover, hybrid languages can be translated into DTML, in such a way that hybrid formulas such as @ap ("agent a has prop- erty p") become equivalent to P (a), if a is the name of a. 2.1 Term-Modal Logic and Epistemic Network Structures In general, term-modal languages may be based on any first-order signature, by for the purposes of representing social networks and factual properties of agents, we limit attention to the following:3 . Definition 1. A signature is a tuple Σ = (V, C, P, N, =) with V a countably infinite set of variables, C and P countable sets of constants and unary predicates, . = for identity. The terms of Σ are T := V ∪ C. N a binary relation symbol and With t1, t2 ∈ T, x ∈ V and P ∈ P, the language L(Σ) is given by ϕ := P (t1) N (t1, t2) (t1 . = t2) ¬ϕ ϕ ∧ ϕ Ktϕ ∀xϕ Standard Boolean connectives, ⊤, ∃ and Kt are defined per usual. With ϕ ∈ L, t ∈ T, x ∈ V, the result of replacing all occurrences of x in ϕ with t is denoted ϕ(x 7→ t). Formulas from the first three clauses are called atoms; if an atom contains no variables, it is ground. Throughout, a, b, etc. are used for constants and the relation symbol N de- notes a social network. The reading of N (t1, t2) depends on application. Ktϕ is a term-indexed epistemic operator which read as "agent t knows that ϕ". L(Σ) neither enforces nor requires a fixed-size agent set A, in contrast with standard epistemic languages, where the set of operators is given by reference to A. Hence the same language may be used to describe networks of varying size. 3 The defined are special cases of the setting in [1], which allows general signatures and non-agent terms. [1] also reviews the term-modal literature. To interpret L(Σ), we use constant-domain models (the same number of agents in each world) with non-rigid constants (names, like predicates and re- lations, may change extension between worlds; this allows for uncertainty about agent identity). See Figs. 1 and 2 for examples of such models. w : b v : b u : b a c a, b, c a c a, b, c a c Fig. 1. Example 1, pt. 1 (Server Error). Three agents a,b and c work in a company with a hierarchical command structure, −→: a is the direct boss of b, who is the direct boss of c. The server has thrown an error after both b and c tampered with it. Either w) the server failed spontaneously, v) b made a mistake (marked by gray) or u) c made a mistake. Lines represent indistinguishability with reflexive and transitive links omitted. There is no uncertainty about the hierarchy, but nobody knows why the server failed. In fact, c made a mistake: the actual world has a thick outline. Definition 2. An L(Σ)-model is a tuple M = (A, W, ∼, I) where A is a non- empty domain of agents, W is a non-empty set of worlds, ∼ : A → P(W ×W ) assigns to each agent a ∈ A an equivalence relation on W denoted ∼a, and I is an interpretation satisfying, for all w ∈ W , 1. for c ∈ C, I(c, w) ∈ A; 2. for P ∈ P, I(P, w) ⊆ A; 3. I(N, w) ⊆ A × A. A pointed model is a pair (M, w) with w ∈ W called the actual world. A variable valuation of Σ over M is a map g : V → A. The valuation identical to g except mapping x to a is denoted g[x 7→ a]. The extension of the term t ∈ T at w in M under g is JtKI,g w = I(t, w) for t ∈ C. w = g(t) for t ∈ V and JtKI,g Given the inclusion of N in the signature Σ, each L(Σ)-model embeds an epis- temic network structure (A, W, (∼a)a∈A, (I(N, w))w∈W ). Formulas are evaluated over pointed models using a direct combination of first-order and modal semantics: Definition 3. Let Σ, M and g be given. The satisfaction of formulas of L(Σ) is given recursively by w ∈ I(P, w), for P ∈ P. w , Jt2KI,g w ) ∈ I(N, w). w = Jt2KI,g w . M, w (cid:15)g P (t1) iff Jt1KI,g M, w (cid:15)g N (t1, t2) iff (Jt1KI,g . = t2) iff Jt1KI,g M, w (cid:15)g (t1 M, w (cid:15)g ¬ϕ iff not M, w (cid:15)g ϕ. M, w (cid:15)g ϕ ∧ ψ iff M, w (cid:15)g ϕ and M, w (cid:15)g ψ. M, w =g ∀xϕ iff M, w =g[x7→a] ϕ for all a ∈ A. M, w (cid:15)g Ktϕ iff M, w′ (cid:15)g ϕ for all w′ such that w ∼JtKI,g w′. w 2.2 Knowing Who and Knowledge De Dicto and De Re First-order modal languages can represent propositional attitudes de dicto (about the statement) and de re (about the thing) in principled manners. For example, Ka∃xP (x) is a de dicto statement: knowledge is expressed about the proposi- tion that a P -thing exists. In contrast, ∃xKaP (x) is a de re statement: it is expressed that of some thing x, that x is known to be a P -thing. In general, de re statements are stronger than de dicto statements. The difference has been ap- preciated in epistemic logic since Hintikka's seminal [16], where he argues that . = b) expresses that a knows who b is (see Fig. 2). Semantically, the ∃xKa(x formula ensures that the constant b refers to the same individual in all a's epis- temic alternatives (i.e., b is locally rigid ). Both de dicto and de re statements may partially be expressed in propositional languages (e.g. de dicto Ka(pb ∨ pc) vs. de re Kapb ∨ Kapc; see [2] for such a usage), but not in a principled manner: the required formulas will depend on the specific circumstances. w1 : t 7→ t, b 7→ b, a 7→ h w2 : t 7→ b, b 7→ t, a 7→ h t h t h c b i b i c c t h t h b i b i c c c c c w3 : t 7→ t, b 7→ b, a 7→ i w4 : t 7→ b, b 7→ t, a 7→ i Fig. 2. Example 2, pt.1 (Knowing Who). Two thieves, t and b, hide in a building with hostages h and i. Outside, a cop, c, waits. To communicate safely, the thieves use code names 'Tokyo' and 'Berlin' for each other and 'The Asset' for the specially valuable hostage h. Agents t, b, h and i all know whom the code names denote (the names are rigid for them), but the cop does not. The code names are t for t, b for b and a for h. Known by all, h and i are in fact called h and i. The thief network ( -- ) is assumed symmetric and transitive. The case is modeled using four worlds, identical up to code name denotation, (shown by 7→). E.g., in the actual world is w1, t names t, but . in w4, it names b. Hence the cop does not know who Tokyo is: M, w1 (cid:15)g ¬∃xKc(x = t). 2.3 Dynamics: Action Models and Product Update To code operations on static models, we use a a variant of DEL-style action models, adapted to term-modal logic (see Fig. 3). They include (adapted versions of) preconditions specifying when an event is executable ( [3, 4]), postconditions describing the factual effects of events ( [5, 7, 15]) as well as edge-conditions representing how an agent's observation of an action depends on the agent's circumstances ( [6]) -- for example their position in a network, cf. Fig. 3. Edge- conditions are non-standard and deserve a remark. With E the set of events, edge-conditions are assigned by a map Q. For each edge (e, e′) ∈ E × E, Q(e, e′) is a formula with a single free variable x⋆. Given a model M , an agent i cannot distinguish e from e′ iff the edge-condition Q(e, e′) is true in M when the free variable x⋆ is mapped to i. Intuitively, if the situation described by the edge- condition is true for i, the way in which i is observing the action does not allow her to tell whether e or e′ is taking place. See Figure 4 for an example. See [1] for a comparison of this approach to that of [6] and the term-modal action models of [17]. ϕ := ∃xN (x, x⋆) 1 : ¬∃xM (x) 2 : ϕ ⊤ M (b) ⊤ 3 : ϕ M (c) ⊤ 4 : ϕ ∃xM (x) ⊤ Fig. 3. Example 1, pt. 2 (Edge-Conditions: Announcement to Subgroup). To learn what happened to the server, the top boss a requests its log file. The log holds one of four pieces of information: 1) Nobody made a mistake, 2) b made a mistake (M ), 3) c made a mistake or 4) somebody made a mistake. Each box represents one of these events: top lines are preconditions, bottom lines postconditions (⊤ means no factual change). In fact, the log rats on c. N denotes the hierarchy. The log is send only to the top boss: the others cannot see its content. This is represented by the edge-condition ϕ: If you, x⋆, have a boss, then you cannot tell 1) from 2) nor 2) from 3) etc. For unillustrated edges, Q(e, e) = (x⋆ . = x⋆) and Q(e, e′) = ϕ when e 6= e′. For simplicity, we here only define action models that take pre-, post, and edge-conditions in the static language L(Σ). However, dynamic conditions are needed for completeness; we refer to [1] for details. Definition 4. An action model for L(Σ) is a tuple ∆ = (E, Q, pre, post) where ✄ E is a non-empty, finite set of events. ✄ Q : (E × E) → L(Σ) where each edge-condition Q(e, e′) has exactly one free variable x⋆. ✄ pre : E → L(Σ) where each precondition pre(e) has no free variables. ✄ post : E → (GroundAtoms(L(Σ)) → L(Σ)) assigns to each e ∈ E a post- condition for each ground atom. To preserve the meaning of equality, let post(e)(t . = t) = ⊤ for all e ∈ E. With no general restrictions on Q, to ensure that all agents' indistinguishability relations continue to be equivalence relations after updating, Q must be cho- sen with care. Throughout, we assume Q(e, e) = (x⋆ . = x⋆) for all e ∈ E. To update, product update may be altered to fit the edge-condition term-modal set- ting as below. Fig. 4 illustrates the product update of Figs. 1 with 3. The use of postconditions is illustrated in Figs. 7 and 8. Definition 5. Let M = (A, W, ∼, I) and ∆ = (E, Q, pre, post) be given. The product update of M and ∆ is the model M ⊗ ∆ = (A′, W ′, ∼′, I ′) where 1. A′ = A 2. W ′ = {(w, e) ∈ W × E : (M, w) (cid:15)g pre(e)} for any g, 3. (w, e) ∼′ 4. I ′(c, (w, e)) = I(c, w) for all c ∈ C, and i (w′, e′) iff w ∼i w′ and M, w (cid:15)g[x⋆7→i] Q(e, e′), I ′(X, (w, e)) = (I(X, w) ∪ X +(w)) \ X −(w), for X = {P, N }, P ∈ P, where: w : (M, w) (cid:15)g post(e)(P (t))}; w : (M, w) 6(cid:15)g post(e)(P (t))}; P +(w) :={JtKI,v P −(w) :={JtKI,v N +(w) :={(Jt1KI,v N −(w) :={(Jt1KI,v w , Jt2KI,v w , Jt2KI,v w ) : (M, w) (cid:15)g post(e)(N (t1, t2))}; w ) : (M, w) 6(cid:15)g post(e)(N (t1, t2))} If (M, w) = pre(e), then (A, e) is applicable to (M, w), and the product update of the two is the pointed model (M ⊗ ∆, (w, e)). Else it is undefined. w1 : b a c v2 : b, c v4 : b, c a a b b b, c c c u3 : b, c u4 : a, b, c a a b b b, c c c Fig. 4. Example 1, pt. 3 (Product Update: Edge-Conditions). The product update of Fig. 1 and Fig. 3. After checking the logs, the boss has learned that c made a mistake, while both b and c are now both uncertain about this, as well as about the boss' information. Worlds are named using by the world-event pair they represent: w1 is the child of w and 1, etc. The pair w2 is not a world: w did not satisfy the precondition of 1. We have w1 ∼′ b v2 as w ∼b v and M, w (cid:15)g[x⋆ 7→a] Q(1, 2) -- as M, w (cid:15)g ∃xN (x, b). Likewise, v2 ∼′ a v2 follows as . M, w (cid:15)g ¬∃xN (x, b), but v4 ∼′ = a). The same reason, reflexive loops are preserved. The boss now knows that c made a mistake: KaM (c). b w1 as v ∼b w and M, v (cid:15)g ∃xN (x, b). That w1 6∼′ a u4 as M, v (cid:15)g (a 2.4 Announcements De Dicto and De Re With de dicto and de re statements expressible in DTML, they may be used to define principled announcements, as exemplified in Fig. 5 and 6. The action models are applicable to any DTML model for a signature that includes the constant a and the predicate M , irrespective of the size of the set of agents. This level of general applicability is not mirrored in standard DEL action models. e : Ka∃xM (x) ⊤ v2e : v4e : a a b b b, c c c u3e : b, c u4e : a, b, c a a b b b, c c c Fig. 5. Example 1, pt. 3 (De Dicto Announcement). The boss breaks the news from the log to b and c piecemeal. Left: First, a makes a de dicto announcement: a knows that somebody made a mistake. Right: The effect on Fig. 4. Only w1 does not survive. In u3e, everybody knows de dicto that somebody messed up: ∀xKx∃yM (y). The boss also knows de re, i.e., knows who: u3e (cid:15)g ∃xKaM (x), as u3e (cid:15)g[x7→c] KaM (x). The employees do not know that a knows de re: u3e (cid:15)g ∀x(∃yN (y, x) → Kx¬∃zKaM (z)) -- since v4e (cid:15)g M (x) iff g(x) = b, but then u4e 6(cid:15)g M (x). I.e., there is no one object to serve as valuation for x such that v4e and u4e satisfy M (x) simulta- neously). The employees are held in suspense! σ : ∃xKaM (x) ⊤ v2eσ : b u3eσ : b a c b, c a c Fig. 6. Example 1, pt. 4 (De Re Announcement). Following a dramatic pause, the boss reveals a stronger piece of information: the boss knows who messed up. This de re announcement is on the left, with Q(e, e) = (x⋆ = x⋆); its result on Fig. 5 (Right) on the right. In u3eσ, everybody knows that a has de re knowledge: ∀xKx∃yKaM (y), but b and c still only have de dicto knowledge: ∀x((x = b ∨ x = c) → Kx∃yM (y) ∧ ¬∃zKxM (z)). 2.5 Postconditions and Network Change Action models with postconditions allows DTML to represent changes to the social network. Such changes may be combined with the general functionality of action models such that some agents may know what changes occur while others remain in the dark. Fig. 7 provides a simple example, including the details calculating the updated network. Fig. 8 presents an example of how de re/de dicto knowledge affects what is learned by a publicly observed network change. † : ⊤ v2eσ† : u3eσ† : b b, c b N (a, b), N (b, c) 7→ ⊥, N (a, c) 7→ ⊤ a c a c Fig. 7. Example 1, pt. 5 (Getting Fired). The employees are dying to know who messed up the server. But the boss just proclaims: 'b, you are fired! c, you are pro- moted!' Left: Action with three instructions for factual change: post(†)(N (a, b)) = ⊥, post(†)(N (b, c)) = ⊥ and post(†)(N (a, c)) = ⊤ (illustrated by 7→). Else post = id. As u3eσ 6(cid:15) ⊥, the first two instructions entail that (a, b), (b, c) ∈ N −(u3eσ), while the lat- ter implies that (a, c) ∈ N +(u3eσ). Right: The network is updated to I ′(N, u3eσ†) = (I(N, u3eσ) ∪ N +(u3eσ))\N −(u3eσ) = ({(a, b), (b, c)} ∪ {(a, c)})\{(a, b), (b, c)} = {(a, b)}. In u3eσ†, neither b nor c know who made the mistake. Unrepresented, a thinks that only bad superiors let their employees make mistakes. e : ⊤ N (·, a), N (a, ·) 7→ ∃xN (·, x) w1e : t 7→ t, b 7→ b, a 7→ h w2e : t 7→ b, b 7→ t, a 7→ h t h t h c b i b i c c c c t h t h b i b i c c c w3e : t 7→ t, b 7→ b, a 7→ i w4e : t 7→ b, b 7→ t, a 7→ i Fig. 8. Example 2, pt.2 (Becoming Criminal) Left: The thieves convince The Asset to cooperate with them, in exchange for stolen goods. For simplicity, assume that the action of a joining the thief network is noticed by everyone. We model this with the action model, with post(e)(N (·, a)) = ∃xN (·, x) and post(e)(N (a, ·)) = ∃xN (x, ·) for · ∈ {t, b, a, h, i, c}. Informally, these say: "If you are a member of the network, then a becomes your neighbor". Right: The effect of event e on Fig. 2: The network has changed in all worlds, but differently. E.g., in w1, we had ¬N (b, a); in (w1, e), we have N (b, a) as (b, h) ∈ N +((w1, e)) since w1 (cid:15)g post(e)(N (b, a)) -- i.e., ∃xN (b, x). Now all thieves and hostages know the new network, as they know whom a refers to. E.g.: Tokyo knows all her neighbors, (w1, e) (cid:15)g ∀x(N (t, x) → KtN (t, x)). The cop only learns that . some hostage has joined the network, but can't tell whom: (w1, e) (cid:15)g Kc∃x(x 6 = b ∧ N (t, x)) but (w1, e) 6(cid:15)g ∃xKc(x 6 . = b ∧ N (t, x)). . = t ∧ x 6 . = t ∧ x 6 2.6 Learning Who Allowing for the possibility of non-rigid names has the consequence that pub- lic announcements of atomic propositions may differ in informational content depending on the epistemic state of the listener. This can be exploited by the thieves of Example 2 to enforce a form of privacy -- as code names should. The notion of privacy involved is orthogonal to the notion of privacy modeled in DEL using private announcements. Though the message is public in the standard sense of everyone being aware of it and its content, as it involves non-rigid names, its epistemic effects are not the same for all agents. This is in contrast with standard public announcements, which yield the same information to everyone. σ : a . = h ⊤ w1eσ : t 7→ t, b 7→ b, a 7→ h w2eσ : t 7→ b, b 7→ t, a 7→ h t h b i c c t h b i c Fig. 9. Example 2, pt.4 (Revealing the Asset) In the model in Fig. 8 (Right), even a public announcement of N (t, a) would not inform the cop about who joined the network. To know who joined the network, the cop must learn who The Asset is. As the cop knows who h is, learning that h is The Asset suffices. Left: The event model . σ for the public announcement that a = h, revealing the identity of The Asset. Right: The product update of Fig. 8 (Right) and event σ. The cop now knows the structure of the network, as a result of the removal of w3e and w4e. 3 Embedding Dynamic Social Network Logics in DTML This section examines relations between the hybrid network models and their languages to DTML. As hybrid languages corresponds to fragments of first-order logic with equality (FOL=), which term-modal logic extends, it stands to reason that the hybrid languages and models mentioned in Sec. 2 may be embedded in term-modal logic. A precise statement and a proof sketch follows below. Turning to dynamics, things are more complicated. [22] presents a very flexible hybrid framework expressing network dynamics using General Dynamic Dynamic Logic (GDDL, [23]). We leave general characterizations of equi-expressive fragments of GDDL and DTML as open question, but remark that all GDDL action-examples of [22] may be emulated using DTML action models, and in many cases via fairly simple ones. More thoroughly, we show that the logic of Knowledge, Diffusion and Learning (KDL, [12]) has a complete and decidable system, a question left open in [12]. This is shown by encoding KDL in DTML. 3.1 Embedding Static Languages and Models The static hybrid languages of [9 -- 12,19,21,22] are all sub-languages of L(P, N om), defined and translated into DTML below. [18] also includes state nominals, which our results do not cover. L(P, N om) is read indexically, as described in Sec. 2. Definition 6. With p ∈ P and x ∈ N om, the language L(P, N om) is given by ϕ := p ¬ϕ ϕ ∧ ϕ @xϕ Kϕ N ϕ U ϕ Denote the fragments without U and @x by L−U (P, N om) and L−@(P, N om). Hybrid logics may be translated into FOL=; our translation resembles that of [8]. We identify agent nominals with first-order variables, translate the modal operator N to the relation symbol N (·, ·), and relativize the interpretation of the indexical K to the nominal/variable x by using the term-indexed operator Kx. Formally, the translation is defined as follows. Definition 7. Let Σn(P, N om) = (V, C, P, N, =) be the signature with V = N om, C = {a1, . . . , an} and P = P . Translations Tx, Ty both mapping L(P, N om) to L(Σn(P, N om)) are defined by mutual recursion. It is assumed that two nominals x and y are given which do not occur in the formulas to be translated. For p ∈ P and i ∈ N om, define Tx by: Tx(p) = p(x) Tx(i) = x =i Tx(ϕ ∧ ψ) = Tx(ϕ) ∧ Tx(ψ) Tx(¬ϕ) = ¬Tx(ϕ) Tx(@iϕ) = Tx(ϕ)(x 7→ i) Tx(N ϕ) = ∀y(N (x, y) → Ty(ϕ)) Tx(Kϕ) = KxTx(ϕ) Tx(U ϕ) = ∀xTx(ϕ) The translation Ty is obtained by exchanging x and y in Tx. To show the translation truth-preserving, we embed the class of hybrid net- work models into a class of term-modal models: Definition 8. Let M = (A, W, (Nw)w∈W , ∼, g, V ) be a hybrid network model for L(P, N om). Then the TML image of M is the L(Σn(P, N om)) TML model T(M ) = (A, W, ∼, I) sharing A, W and ∼ with M and with I given by 1. ∀c ∈ C, ∀w, v ∈ W, ∀a, b ∈ A, (I(c, w) = a and w ∼b v ⇒ I(c, v) = a) 2. I(p, w) = {a : (w, a) ∈ V (p)} 3. I(N, w) = {(a, b) ∈ A × A : (a, b) ∈ Nw} The model T(M ) has the same agents, worlds and epistemic relations as M . The interpretation 1. encodes weak rigidity : if (w, v) ∈ Sa∈A ∼a, then any constant denotes the same in w and v, emulating the rigid names of hybrid network models; 2. ensures predicates are true of the same agents at the same worlds, and 3. ensures the same agents are networked in the same worlds. With the translations Tx, Ty and the embedding T, it may be shown that DTML can fully code the static semantics of L(P, N om) hybrid network logics: Proposition 1. Let M = (A, W, (Nw)w∈W , ∼, g, V ) be a hybrid network model. Then for all ϕ ∈ L(P, N om), M, w, g(•) = ϕ iff T(M ), w =g T•(ϕ), • = x, y. 3.2 KDL Dynamic Transformations and Learning Updates in DTML We show that KDL [12] dynamics may be embedded in DTML, for finite agent sets (as assumed in [12]). Given Prop. 1, we argue that each KDL model transformer is representable by a DTML action model and that the dynamic KDL language is truth-preservingly translatable into a DTML sublanguage. The logic of the class of KDL models is, up to language translations, the logic of its corresponding class of DTML models. We show that the logic of this class of DTML models can be completely axiomatized, and the resulting system is decidable. Thus, by embedding KDL in DTML, we find a complete system for the former. In KDL4, agents are described by feature propositions reading "for feature f, I have value z". With F a countable set of features and Zf a finite set of possible values of f ∈ F, the set of feature propositions is FP = {(f + z) : f ∈ F, z ∈ Zf}. The static language of [12] is then L−U (FP, N om). The dynamic language LKDL extends L−U (FP, N om) with dynamic modalities [d] and [ℓ] for dynamic trans- formations d and learning updates ℓ: ϕ ::= (f + z) i ¬ϕ ϕ ∧ ϕ @iϕ N ϕ Kϕ [d]ϕ [ℓ]ϕ A dynamic transformation d changes feature values of agents: each is a pair d = (Φ, post) where Φ ⊆ LKDL is a non-empty finite set of pairwise inconsistent formulas and post : Φ × F → (Zn ∪ {⋆}) is a KDL post-condition. Encoded by post(ϕ, f) = x is the instruction: if (w, a) (cid:15) ϕ, then after d, set f to value x at (w, a), if x ∈ Zn; if x = ⋆, f is unchanged. A learning update cuts accessibility relations: the update with finite ℓ ⊆ LKDL keeps a ∼a link between worlds w and v iff, for all ϕ ∈ ℓ, (w, b) (cid:15) ϕ ⇔ (v, b) (cid:15) ϕ for all neighbors b of a. 4 Notation here is equivalent but different to fit better with the rest of this paper. Definition 9. Given a KDL model M = (A, W, (Nw)w∈W , ∼, g, V ), the model reached after applying d is M d = (Ad, W d, (N d w)w∈W , ∼d, gd, V d) where only V d is different, and is defined as follows: (w, a) ∈ V d(f + z) iff (a) post(ϕ, f) = x for some ϕ ∈ Φ such that M, w, a = ϕ, where x 6= ⋆; or (b) condition (a) does not hold and (w, a) ∈ V (f + z). Definition 10. A learning update is a finite set of formulas ℓ ⊆ LKDL. Given a KDL model M = (A, W, (Nw)w∈W , (∼a)a∈A, g, V ), the model after ℓ is M ℓ = (A, W, (Nw)w∈W , (∼′ w ∼′ a v iff w ∼a v and ∀b ∈ A(Nw(a, b) ⇒ ∀ϕ ∈ ℓ(M, w, b = ϕ iff M, v, b = ϕ)) a)a∈A, g, V ) where: Let D and L be the sets of dynamic transformations and learning updates. The result of applying † ∈ D ∪ L to M is denoted M †, and the [†] modality has semantics M, w, a = [†]ϕ iff M †w, a = ϕ. As we show below, for every † ∈ D ∪ L, there is a pointed DTML action model ∆† with identical effects. As KDL operations may involve formulas with [†]-modalities, we must use DTML action models that allow [∆, e]-modalities in their conditions, and translate LKDL into the general DTML language that results, denoted L(Σn(FP, N om)+[∆]).5 This language is interpreted over DTML models with standard action model semantics: (M, w) (cid:15)g [∆, e]ϕ iff M ⊗ ∆, (w, e) (cid:15) ϕ. We define now the action models ∆†. For a dynamic transformation d ∈ D, [11] provide reduction axioms showing d's instructions statically encodable in LKDL). The reduction axiom for atoms is as follows: [d]f + z ↔   _ ϕ∈Φ:post(ϕ,f)=z,z∈Zf ϕ  ∨  ¬  _ ϕ∈Φ:post(ϕ,f)=z,z∈Zf ϕ  ∧ f + z  As d changes atomic truth values under a definable instruction, its effects may be simulated by an action model with a matching post-condition (i.e., the translation of the definable instruction). More specifically, the action model ∆d is defined as follows. Definition 11. For dynamic transformation d = (Φ, post), the action model ∆d = (E, Q, pre, post) is defined by E = {ed}, Q(ed, ed) = pre(ed) = ⊤ and for each constant a, post(e)(Tx(f + z)(x 7→ a)) = Tx _ ϕ∈Φ:post(ϕ,f)=z,z∈Zf ϕ∈Φ:post(ϕ,f)=z,z∈Zf ϕ  ∨  ¬      For a learning update ℓ ∈ L, ∆ℓ has events eX, eY for any consistent subsets X, Y of {ϕ(c), ¬ϕ(c) : ϕ ∈ ℓ, c ∈ C} with edge-condition Q(eX , eY ) satisfied for agents for whom all neighbors agree on X and Y . Unsatisfied edge-conditions thereby capture the link cutting mechanism of ℓ. The detailed definition of ∆ℓ is as follows. ϕ  ∧ f + z    (x 7→ a) _ 5 Defined using double recursion as standard; see [1] for details. Definition 12. Let ℓ = {ϕ1, . . . , ϕm} be a learning update. Let Tx(ℓ) := {Tx(ϕi) i = 1, . . . , n} and let Gℓ := {Tx(ϕ)(x 7→ a) Tx(ϕ) ∈ Tx(ℓ), a ∈ C} be the ground- ing of Tx(ℓ) obtained by replacing each free occurrence of x in Tx(ϕ) for each possible constant a ∈ C. Define a Gℓ-valuation as a function val : Gℓ → {0, 1} and let Vℓ be the set of all such valuations. Definition 13. Let ℓ be a learning update. The corresponding DTML action model ∆ℓ = (Eℓ, Qℓ, preℓ, postℓ) is defined by letting ✄ Eℓ = {eval val ∈ Vℓ}, ✄ preℓ(eval) = V{ϕ val(ϕ) = 1} ∪ {¬ϕ val(ϕ) = 0} ✄ Qℓ(eval, eval) = ⊤ ✄ Qℓ(eval, eval′ ) = V{a∈C∃ϕ∈ℓ s.t. val(Tx(ϕ)(x7→a))6=val′(Tx(ϕ)(x7→a))} ¬N (x⋆, a), for any two distinct events eval, eval′ ✄ postℓ(e) = id for all e ∈ Eℓ Note that the signature Σn(F P, AN om) is defined to have finitely many con- stants C = {a1, . . . , an}, and hence both E, the preconditions and the edge- conditions in ∆ℓ are finite, as required. The action model ∆ℓ works as follows. Each event eval corresponds to one way the agents can be with respect to Gℓ, as indicated by val. The edge conditions control how links get cut. Two worlds (w, eval) and (v, eval′ ) in the updated model will keep a link for the agent named a, if any disagreement between val and val′ does not concern a neighbor of a. Or, equivalently, if all neighbors of a are identical with respect to Gℓ. Precisely this condition is encoded in Q(eval, eval′ ). To formally state that the dynamics of † ∈ D ∪ L are simulated by ∆†, the following clauses are added to translation T•, for • = x, y: T•([d]ϕ) = [∆d, ed]T•(ϕ), T•([ℓ]ϕ) = ^ e∈Eℓ (preℓ(e) → [∆ℓ, e]T•(ϕ)) where (∆†, e†) is an action model implementations of † ∈ D ∪ L. Then KDL statics and dynamics can be shown performable in DTML: Proposition 2. For any finite agent hybrid network model M with nominal valuation g and ϕ ∈ LKDL: M, w, g(•) = ϕ iff T(M ), w =g T•(ϕ), for • = x, y. Proof. By induction on ϕ. We include the cases for the dynamic modalities. Let ϕ = [d]ψ, where d = (Φ, post). We need to show that M, w, g(x) = [d]ψ iff T(M ), w =g [∆d, ed]Tx(ψ) (the case for Ty is analogous). Note that M, w, g(x) = [d]ψ iff M d, w, g(x) = ψ iff (by i.h.) T(M d), w =g Tx(ψ). We will show that T(M d) and T(M ) ⊗ ∆d satisfy the same formulas. To prove this, we will show that there is a bounded morphism linking these two models (it is straightforward to show that term-modal formulas are preserved when this is the case, as in the propositional modal setting). Define b : T(W d) → T(W ∆d ) by w 7→ (w, ed). We show that b is a bounded morphism. 1. w and (w, ed) satisfy the same basic formulas: T(M d), w =g Tx(f + z) iff (i.h.) M d, w, g(x) = f + z iff M, w, g(x) = [d]f + z iff (reduction axiom for [d]f + z) iff (i.h., where we let g(x) = a for some a ∈ A named a) M, w, g(x) = (cid:16)Wϕ∈Φ:post(ϕ,f)=z,z∈Zf ϕ(cid:17) ∨ (cid:16)¬(cid:16)Wϕ∈Φ:post(ϕ,f)=z,z∈Zf T(M ), w =g Tx((cid:16)Wϕ∈Φ:post(ϕ,f)=z,z∈Zf (cid:16)¬(cid:16)Wϕ∈Φ:post(ϕ,f)=z,z∈Zf ϕ(cid:17) ∧ f + z(cid:17) ϕ(cid:17) ∨ ϕ(cid:17) ∧ f + z(cid:17))(x 7→ a) iff (by definition of ∆d) T(M ), w =g post(e)(Tx(f + z)(x 7→ a)) iff T(M ) ⊗ ∆d, (w, ed) =g Tx(f + z)(x 7→ a) iff (since g(x) = a and a is named a) T(M ) ⊗ ∆d, (w, ed) =g f + z. 2. if (w, v) ∈ T(∼d a) then ((w, ed), (v, ed)) ∈ T(∼∆d a ): (w, v) ∈ T(∼d T(∼∆d a) iff (w, v) ∈∼d a iff (w, v) ∈∼a iff (w, v) ∈ T(∼a) iff (w, v) ∈ a ) (since ∆d does not change the accessibility relations). 3. if ((w, ed), (v′, ed)) ∈ T(∼∆d a ) then there is v such that (w, v) ∈ T(∼d a) and b(v) = (v′, ed): Reasoning as in step 2, ((w, ed), (v′, ed)) ∈ T(∼∆d b(v′) = (v′, ed). a ) iff (w, v′) ∈ T(∼d a), and Hence, b is a bounded morphism, and T(M d) and T(M )⊗∆d satisfy the same formulas. Thus, M, w, g(x) = [d]ψ iff M d, w, g(x) = ψ iff (by i.h.) T(M d), w =g Tx(ψ) iff (bounded morphism) T(M ) ⊗ ∆d, (w, ed) =g Tx(ψ) iff T(M ), w = Tx([d]ψ). Next, let ϕ = [ℓ]ψ. We need to show that M, w, g(x) = [ℓ]ψ iff T(M ), w =g ^ e∈Eℓ (preℓ(e) → [∆ℓ, e]Tx(ψ)) (the case for Ty is analogous). Note that M, w, g(x) = [ℓ]ψ iff M ℓ, w, g(x) = ψ iff (by i.h.) T(M ℓ), w =g Tx(ψ). As in the previous case, we will show that T(M ℓ) and T(M ) ⊗ ∆ℓ satisfy the same formulas by defining a bounded morphism linking the two. Note that the preconditions in ∆ℓ are pairwise inconsistent and jointly exhaustive, since each precondition corresponds to one way of assigning truth values to the formulas in Gℓ. Hence, for each w ∈ T(W ), there is exactly one event eval such that T(M ), w = preℓ(eval). Define b : T(W ℓ) → T(W ∆ℓ ) by w 7→ (w, eval). We show that b is a bounded morphism. 1. w and (w, eval) satisfy the same basic formulas: This is clear from the fact that learning updates do not change the acces- sibility relations. T(M ℓ), w =g Tx(f + z) iff (i.h.) M ℓ, w, g(x) = f + z iff M, w, g(x) = f + z iff (i.h.) T(M ), w =g Tx(f + z) iff T(M )⊗∆ℓ, (w, eval) =g Tx(f + z). 2. if (w, v) ∈ T(∼ℓ a) then ((w, eval), (v, eval′ )) ∈ T(∼∆d a ): a) iff w ∼ℓ As T(M ) is weakly rigid, each agent has the same name in each equivalence class [w]∼a of ∼a. In what follows, we let the name of any agent o ∈ A in worlds of [w]∼a be o. Now, (w, v) ∈ T(∼ℓ iff w ∼a v and ∀b ∈ A(Nwab ⇒ ∀ϕ ∈ ℓ(M, w, b = ϕ iff M, v, b = ϕ)) iff (contrapositive) w ∼a v and ∀b ∈ A(∃ϕ ∈ ℓ((M, w, b = ϕ and M, v, b = ¬ϕ) or (M, w, b = ¬ϕ and M, v, b = ϕ)) ⇒ ¬Nwab) iff (by i.h.) (w, v) ∈ T(∼a) and (by def. of T(M ) ⊗ ∆ℓ) T(M ), w =g pre(eval) and T(M ), v =g pre(eval′ for all b ∈ C: if there is a ϕ ∈ ℓ such that ) for some val, val′ ∈ Vℓ, and a v (cid:0)T(M ), w =g Tx(ϕ)(x 7→ b) and T(M ), v =g Tx(¬ϕ)(x 7→ b)(cid:1) or (cid:0)T(M ), w =g Tx(¬ϕ)((x 7→ b)) and M, v =g Tx(ϕ)(x 7→ b)(cid:1) then T(M ), w =g ¬N (a, b) iff (w, v) ∈ T(∼a) and T(M ), w =g pre(eval) and T(M ), v =g pre(eval′ some val, val′ ∈ Vℓ and (by def. of ∆ℓ) T(M ), w =g[x⋆7→a] Q(eval, eval′ ((w, eval), (v, eval′ )) ∈ T(∼∆ℓ ) for ) iff a ). )) ∈ T(∼∆ℓ 3. if ((w, eval), (v′, eval′ and b(v) = (v′, eval′ ): Reasoning as in step 2, ((w, eval), (v′, eval′ and b(v′) = (v′, eval′ ). a ) then there is v such that (w, v) ∈ T(∼ℓ a) )) ∈ T(∼∆d a ) iff (w, v′) ∈ T(∼ℓ a), Hence, b is a bounded morphism, and T(M ℓ) and T(M ) ⊗ ∆ℓ satisfy the same formulas. Thus, M, w, g(x) = [ℓ]ψ iff M d, w, g(x) = ψ iff (by i.h.) T(M ℓ), w =g Tx(ψ) iff (bounded morphism) for the unique event eval such that T(M ), w =g preℓ(eval), we have T(M )⊗∆ℓ, (w, eval) =g Tx(ψ) iff T(M ), w =g Ve∈Eℓ(preℓ(e) → [∆ℓ, e]Tx(ϕ) iff T(M ) = Tx([ℓ]ψ). This completes the proof. With Prop. 2 embedding KDL in DTML, it remains to show that there is a complete and decidable system for the image of KDL. Up to translation, such a logic is then a logic for the class of KDL models. To state the result, denote the TML image of the class of n-agent KDL models by T(KDLn). We now define a set of formulas, Fn, which can be shown to characterise the class T(KDLn). Definition 14. Let Fn ⊆ L(Σn(FP, N om)+[∆]) be the logic extending the term- modal S5 logic with the reduction axioms for action models (∆†, e†), † ∈ D ∪ L (defined in [1]), as well as the following static axioms: 1. There are n agents and they are all named: Namedn := ∃x1, ..., xn( i,j≤n,i6=j  ^   ^ i,j≤n,i6=j y = xi   ∧  ) xi 6= xj   ∧ ∀y  _   ∧  ^ i≤n i≤n ci 6= cj xi = ci 2. Weak rigidity (Def. 8): Rign := ^ c∈C ∀x((c = x) → ∀y(Ky(c = x))) 3. The neighbour relation is irreflexive and symmetric: Neigh := ∀x∀y(¬N (x, x) ∧ (N (x, y) ↔ N (y, x))) 4. Agents know their neighbors: KnowNeigh := ∀x∀y(N (x, y) ↔ KxN (x, y)) We then obtain the result: Proposition 3. Fn statically characterizes T(KDLn). Proof. By model-checking of the formulas in Fn. Which we can use to state completeness: Theorem 1. For any n ∈ N, the logic Fn is sound, strongly complete and de- cidable w.r.t. T(KDLn). Proof (sketch). By Prop. 3, Fn statically characterizes T(KDLn). The result then follows from three results from [1]: 1. Any extension of the term-modal logic K with axioms A is strongly complete with respect to the class of frames charac- terized by A, and 2. If A characterizes a class with finitely many agents, then the logic is also decidable, and 3. any dynamic DTML formula is provably equivalent to a static DTML formula using reduction axioms. Thus, since Fn characterizes T(KDLn), which is a class with finitely many agents, and all dynamic axioms in Fn are probably equivalent to static DTML, it follows that K + Fn is strongly complete and decidable with respect to T(KDLn). 4 Final Remarks This paper has showcased DTML as a framework for modeling social networks, their epistemics and dynamics, including examples in which uncertainty about name reference and de dicto/de re distinctions are key to modelling information flow and network change correctly. It was shown that DTML may encode the popular hybrid logical models of epistemic networks, and that DTML may be used to obtain completeness for an open-question dynamics through emulation. We are very interested in learning how DTML relates to GDDL with respect to the encodable dynamics. We have been able to emulate the updates used in the examples of [22], but the general question is open. Further, the statics of frameworks that describe networks using propositional logic [2, 14] must be DTML encodable, and we expect the name about their updates, where reduc- tion axioms exist. This raises two questions: if we can show this by a general results instead of piecemeal, and whether principled DTML action models exist for classes of updates. E.g., the threshold update of [2] gives an agent's property P if a given fraction of neighbors are P ; for a fixed agent set, this is DTML encod- able by using the reduction axioms of [2] to provide pre- and postconditions. For a principled update, however, seemingly we need a generalized quantifier (e.g., a Rescher quantifier). If so, the general update form is not DTML encodable. Classification results like these would add valuable insights on network logics. References 1. A. Achen, A. O. Liberman, and R. K. Rendsvig. Dynamic Term-Modal Logics for Epistemic Planning. under review. arXiv:1906.06047, 2019. 2. A. Baltag and Z. Christoff. Dynamic Epistemic Logics of Diffusion and Prediction in Social Networks. Studia Logica, 2018. 3. A. Baltag and L. S. Moss. Logics for Epistemic Programs. Synthese, 139(2):165 -- 224, 2004. 4. A. Baltag, L. S. Moss, and S. S. Solecki. The Logic of Public Announcements, Common Knowledge, and Private Suspicions. In TARK '98, p. 43 -- 56. 1998. 5. J. van Benthem, J. van Eijck, and B. Kooi. Logics of communication and change. Information and Computation, 204(11):1620 -- 1662, 2006. 6. T. Bolander. Seeing is Believing: Formalising False-Belief Tasks in Dynamic Epis- temic Logic. ECSI 2014, volume 1283, p. 87 -- 107. 7. T. Bolander and M. B. Andersen. Epistemic planning for single- and multi-agent systems. Journal of Applied Non-Classical Logics, 21(1):9 -- 34, 2011. 8. T. Brauner. Hybrid logic and its proof-theory. Springer, 2011. 9. Z. Christoff. Dynamic Logics of Networks. PhD thesis, U. of Amsterdam, 2016. 10. Z. Christoff and J. U. Hansen. A two-tiered formalization of social influence. LORI 2013, volume 8196 of LNCS, 68 -- 81. Springer, 2013. 11. Z. Christoff and J. U. Hansen. A Logic for Diffusion in Social Networks. Journal of Applied Logic, 13:48 -- 77, 2015. 12. Z. Christoff, J. U. Hansen, and C. Proietti. Reflecting on social influence in net- works. Journal of Logic, Language and Information, 25:299 -- 333, 2016. 13. Z. Christoff and P. Naumov. Diffusion in social networks with recalcitrant agents. Journal of Logic and Computation, 29(1):53 -- 70, 12 2018. 14. Z. Christoff and R. K. Rendsvig. Dynamic Logics for Threshold Models and their Epistemic Extension. Proc. of ELISIEM, 2014. 15. H. van Ditmarsch and B. Kooi. Semantic Results for Ontic and Epistemic Change. In Logic and the Foundations of Game and Decision Theory (LOFT 7), Texts in Logic and Games, Vol. 3, pages 87 -- 117. Amsterdam University Press, 2008. 16. J. Hintikka. Knowledge and Belief: An Introduction to the Logic of the Two No- tions. College Publications, 2nd, 2005 edition, 1962. 17. B. Kooi. Dynamic term-modal logic. In LORI 2007, Texts in Computer Science 8, p. 173 -- 185, 2007. 18. L. Zhen and J. Seligman. A Logical Model of the Dynamics of Peer Pressure. Electronic Notes in Theoretical Computer Science, 278(0):275 -- 288, 2011. 19. F. Liu, J. Seligman, and P. Girard. Logical Dynamics of Belief Change in the Community. Synthese, 191(11):2403 -- 2431, 2014. 20. R. K. Rendsvig. Diffusion, Influence and Best-Response Dynamics in Net- In ESSLLI 2014 Student Session, p. 63 -- 75. works: An Action Model Approach. arXiv:1708.01477, 2014. 21. J. Seligman, F. Liu, and P. Girard. Logic in the Community. In Logic and Its Applications, p. 178 -- 188. Springer, 2011. 22. J. Seligman, F. Liu, and P. Girard. Facebook and the epistemic logic of friendship. In TARK 2013, p. 229 -- 238, 2013. 23. J. Seligman, F. Liu, and P. Girard. General Dynamic Dynamic Logic. In Advances in Modal Logic, vol. 9, p. 239 -- 260. Springer, 2012. 24. S. Smets and F. R. Valezquez-Quesada. How to Make Friends: A Logical Approach to Social Group Creation. In LORI 2017, LNCS vol. 10455, p. 377 -- 390, 2017. 25. S. Smets and F. R. Valezquez-Quesada. In Dynamic Logic. New Trends and Ap- plications (DALI 2017), LNCS vol. 10669, p. 171 -- 184. Springer, 2018.
0912.3984	1	0912	2009-12-20T05:25:13	Multi-Agent Model using Secure Multi-Party Computing in e-Governance	[ "cs.MA" ]	Information management and retrieval of all the citizen occurs in almost all the public service functions. Electronic Government system is an emerging trend in India through which efforts are made to strive maximum safety and security. Various solutions for this have been proposed like Shibboleth, Public Key Infrastructure, Smart Cards and Light Weight Directory Access Protocols. Still, none of these guarantee 100 percent security. Efforts are being made to provide common national identity solution to various diverse Government identity cards. In this paper, we discuss issues related to these solutions.	cs.MA	cs	JOURNAL OF COMPUTING, VOLUME 1, ISSUE 1, DECEMBER 2009, ISSN: 2151-9617 HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/ Multi-Agent Model using Secure Multi-Party Computing in e-Governance Dr. Durgesh Kumar M ishra, Sam iksha Shuk la 195 Abstract— Information management and retrieval of all the citizen occurs in almost all the public service functions. Electronic Government system is an emerging trend in India through which efforts are made to strive maximum safety and security. Vari- ous solutions for this have been proposed like Shibboleth, Public Key Infrastructure, Smart Cards and Light Weight Directory Access Protocols [1]. Still, none of these guarantee 100% security. Efforts are being made to provide common national identity solution to various diverse Government identity cards. In this paper, we discuss issues related to these solutions. Index Terms: Secure multi-party computation, Security, Privacy, e-Governance, Multi-Agent —————————— (cid:139) —————————— 1 INTRODUCTION Digitalization of all the Government activities and offices has given rise to the concept of e‐Governance which is vital to the IT revolution. It has to be done keeping in mind the large diverse cultural background. Fast progress in network technologies and development in data mining and distributed data applications have encouraged the computerization in even Government sector, which was till now untouched. Privacy Preserving data mining, secure multiparty com‐ putation, cryptography, randomization and anonymiza‐ tion can be the suggestive security mechanisms [8] that needs to be enforced for all the Government information systems. Data of such Government offices needs to be kept extremely secure and confidential with proper ac‐ cessing and authorization checks of individual employee. With an e‐Governance System, we try to provide a com‐ mon information system that can maintain and serve the common information requirements of various diverse public and private sectors. This involves large data trans‐ fer among various organizations and hence, citizen pri‐ vacy is at stake, if not properly checked and administered. Special attention needs to be provided for such web‐ based applications. E‐governance, as a concept, can be a great technological step ahead. It can be implemented using privacy preserving data mining [1, 2]. 2 Security Issues in e-Governance Systems It is the key requirement of an e‐Governance system with an interoperable secure infrastructure to meet the current and future needs. Such sub‐system must act in coordina tion with its horizontally similar other sub‐systems pro‐ viding and getting information from varying sources [2]. Success of such a concept lies entirely in the hands of the ———————————————— • Dr. Durgesh Kumar Mishra is with the Department of Computer Engi- neering Acropolis Institute of Technology and Research, Indore, India. • Samiksha Shukla is with the Department of Computer Engineering, Christ University, Bangalore. citizens, their trust and confidence. Although, prone to security thefts, such a system can greatly introduce a re‐ volutionary change in the public and private sector or‐ ganizations. The basic security requirements include [1]: • Client authentication of the message and content ve‐ rification. • Sender and receiver authentication. • No information, messages and data leaks. • 100%availability and reliability. • Confidentiality of the messages and overall system working. The solutions to these any many more security issues can be to adopt existing and contemporary technology that ensures safety throughout the communication process. This can be done in combination with LDAP, PKI, PKI Smart Cards, SSO (Single Sign On), Web Single Sign On, Shibboleth etc… [1] The combination of these methods can help design efficient and secure e‐governance archi‐ tecture. 3 Privacy Issues in e-Governance Systems Privacy of the information in a web‐based e‐governance system is equally important issue that cannot be com‐ promised as the information of citizens on web is a valu‐ able resource which needs its privacy and confidentiality. This data may be for person identification details, his family details, or any other personal information which is to be kept private. Also, as Government to citizen interac‐ tion increases, there is a need to safeguard the Govern‐ ment activities and data from information thefts [2]. In today’s era, when personal information has become an extremely valuable resource, it is necessary to enforce security and legal restrictions to such hacking activities and data thefts. For example, with the available personal information on web, whole profile of a person can be created, and de‐ pending on the confidentiality of hacked data, it can re‐ JOURNAL OF COMPUTING, VOLUME 1, ISSUE 1, DECEMBER 2009, ISSN: 2151-9617 HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/ 196 sult in any kind of loss, which can be financial loss, pass‐ word theft, account information loss, address and tax de‐ tails stolen and so on. If bank information is lost, it can result to online banking theft. Jurisdiction and a separate legal Department which can enforce this code of conduct of e‐Governance need to be setup, to keep an eye on such malicious activities [1]. A web security Board should be setup to check and control the online thefts and other malicious activities that cause great information loss and thefts. Till now, in India we haven’t yet been able to implement such strict web regula‐ tory mechanism, but it has been successfully worked out in US [1]. As per the US privacy Legislation, there is a privacy cer‐ tificate that is signed by citizens at the time they fill such privacy sensitive information in which they define the authentication and access levels which must be enforced on there data. Later, whenever this information is re‐ quested for web, first the certificate is checked for the va‐ lidity of the access, and only if it is valid, it is sent to web, else access is denied. Similarly, there is other Government acts as well that define and control the privacy rights of the individuals [1]. 4 Security Solutions to e-Governance A single safety mechanism cannot suffice the security so‐ lution, but the complete communication must be must be kept safe using a combination of several security mecha‐ nisms. Some of the security methods are: 4.1 Public Key Infrastructure: It provides strong authe tication and secure communication to the entities involved. It is based on asymmetric keys and digital cer‐ tificates to enable public key cryptography. It consists of a trusted third party called the certificate authority which binds public key to the entities involved. Certificate Au‐ thority maintains a table containing entries for each entity along with its public key and other entries. This infra‐ structure has various benefits like cost effective, interop‐ erable, and consistent [1]. 4.2 Smart Cards/National Identity Cards (NIC): It is hardware based cryptographic mechanism in which a card reader with the desired functionality and it imple‐ ments the authentication and security mechanism. It too stores the digital certificates, private and public keys oth‐ er entity related information. It performs the entire task with minimal human intervention. Additionally, these can also be used as electronic identification mechanism. Its benefits are convenient, strong authentication mechanism which is a one time investment [1]. Alternatively, NICs can also be employed with PKI to provide similar functionality. 4.3 LDAP (Light Weight Directory Access Proto- col): It is an Internet Standard Protocol in which directo‐ ries are organized according to X.500 data model. It can be used to issue, revoke (CRLs) and organize PKI certifi‐ cates by means of directories [2]. 4.4 SSO (Single Sign On): This mechanism allows user to sign in once and then make use of multiple re‐ sources with a single sign in [2]. 4.5 Web Single Sign on (SSO): It allows surfing across or within organizational boundaries. By this, au‐ thorized decisions about a site for a particular entity per‐ taining to a particular entity can be made [2]. Shibboleth: It is a Security Assertion Mark up 4.6 Language. Its key concepts include Federated Admini‐ stration, Access Control Based on Attributes and Activity Management of Privacy. A collaboration of these above mentioned security architecture can give evolutionary Horizontal Infrastructure [2]. 5 Privacy Solutions to e-Governance Privacy is essentially an important right of each individ‐ ual and must be protected by regulatory policies or legis‐ lations. It cannot be ensured without lawful acts and legal amendments. India, although has been a Democratic na‐ tion in which every Government activities have been fully transparent and well known, but now in this Internet Age, privacy of an individual is a major concern and has to be made through restricted access. Some of the privacy measures can be as under: 5.1 Legal framework for Privacy Enforcement of Individual: Prior this revolution, Government must frame out a strong set of principles and rules to work out the whole e‐governance infrastructure. Some legal framework like US Policy Act of 1974 [2], the Policy Act of Senator, the Social Security Number Misuse Prevention Act, the Notification of Risk to Personal Data Act etc... must be prepared so that there are unambiguous and transparent rules for smooth running of the system across web. These rules can be similar to our fundamental rights and duties, which suggests that rights abide duties. Such acts must ensure that every citizen can insert, delete, modify, and view only his own record without any permission to ac‐ cess other records [2]. They must be able to determine what all records pertaining to them have been collected, allow purposeful access of record and so on. Also, the similar restrictions must be made for Government em‐ ployees maintaining the record so that they may not pur‐ posely or mistakenly manipulate any record. Large fines and appropriate punishments for the hackers and mali‐ cious elements must also be defined so that such activities can be avoided or rather prohibited completely [2]. 5.2 Privacy Enhancing Technologies: Privacy Rules enforcement is as important as its initial definition. Without its enforcement, everything shall be just a matter of trust of citizens which always very from person to per‐ son. Privacy Enhancing Policies are defined for this pur‐ pose to reinforce the policy rules from time to time. Sev‐ eral technologies exist that can address the privacy poli‐ cies of the Government, Prominent among which are as under: JOURNAL OF COMPUTING, VOLUME 1, ISSUE 1, DECEMBER 2009, ISSN: 2151-9617 HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/ 197 5.3 Privacy Language: Specification e‐ The Government Act of 2002 requires Federal Agencies to put in place the privacy protection of information collected electronically [2]. The World Web Consortium Standard, the Platform for Privacy Preferences or P3P, is a formal language for pri‐ vacy communication to customers. It is a kind of specifi‐ cation that covers all the access and manipulation permis‐ sions and authorization levels. Each transaction on web, before being executed is checked for validity with the help of these specifications, and if these transactions are found valid they are executed or else aborted as invalid [2]. Another example of such a language which provides ma‐ chine enforceable policies is IBM’s Enterprise Privacy Au‐ thorization Language [EPAL]. It is an XML based Privacy Specification Language [2]. Privacy during Data Mining: In order to avoid 5.4 false profiles, hacking and information thefts during the information retrievals, it is required. Such methods may include data modification, randomization, encryption, selective transformation, perturbation etc… [2]. Another solution to all e‐Governance problems is Secure Multiparty Computation which shall be discussed in the next section. 5.5 Privacy Preserving Databases: Several ad‐ vanced databases provide all the privacy related con‐ straints to be specified on the data. Such mechanisms may include strong authentication, single sign on, LBAC poli‐ cies, encryption schemes for databases, virtual private databases, role based authentication, Hippocratic data‐ bases etc… These databases allow mining process to take place only if they don’t impose any privacy and security threats, and thus, ensures safe working [2]. 5.6 Transactional Privacy: One way of achieving trans‐ actional privacy is by encryption. The transactional data is supplied in encrypted form and once the transaction is complete, it is again encrypted. So, even if the hacker hacks the data, he gets nothing but only garbage string. Alternatively, a direct encrypted data connection can also be established in which encryption is managed automati‐ cally [2]. Besides this, a special program can also be used that runs in background to safeguard all the ongoing transactions. An Administrator can also be employed which take cares of all the authorization checks and transactional validity [1]. 5.7 Statistical Data Protection: Selective disclosure of statistical data is another safe approach in which statis‐ tical estimates are disclosed publicly without revealing the information or identity of an individual. The various ways in which such statistical Disclosure Control [SDC] can be implemented are query restriction, data perturba‐ tion, output perturbation. Similar to SDC are some other database protection techniques like tabular data protec‐ tion, dynamic databases, microdata protection, data ano‐ nymization and anonymized data analysis, use of privacy brokers for proper privacy tracking and many more [2]. 6 Secure Multiparty Computation [SMC] as a Comprehensive Solution to Privacy and Security Issues of e-Governance The importance of data in transactional and computa‐ tional environment cannot be overlooked as its loss can cause great chaos and problems [7]. When thinking of e‐ Governance infrastructure in a web based environment, we have to take into consideration some potential threats in that environment and their solutions or avoidance me‐ chanisms [9]. Internet is a vast distributed architecture. Therefore, transactions and computations need to be de‐ signed with this architecture in mind. Secure Multiparty Computation provides a safe and efficient distributed computing environment in which no safety requirements of individuals are compromised [10]. Instead, it provides such an environment in which the privacy and confiden‐ tiality of the system as well as of the system is promised in polynomial time complexity and no extra overheads. At present, it is one of the unbeatable solutions to secu‐ rity. The aim of a secure multiparty computation task is for the participating parties to securely compute some function of their distributed and private inputs [10]. In SMC, sev‐ eral parties or workstation who wish to compute some results or carry out certain transaction, send their inputs to some trusted third party [TTP] which then computes the result and then announces the results publicly [10]. In this scenario, each party learns nothing more than their own inputs and the final results. Also, the main require‐ ments of a transaction, i.e. Correctness, independence of inputs/outputs, fairness and privacy are guaranteed [8]. 7 Multi-agent Secure Multiparty Computa- tion using Arithmetic Cryptography Agent may be defined as a TTP that does the computation for the distributed parties. A multi‐agent scenario basi‐ cally consists of multiple agents that carry out computa‐ tion, both in ideal and adversial manner [5]. For each task, arbitrary numbers of agents are employed, and the com‐ putations are made. A result is said to be correct if at least n/3 agents return the same result [5]. Moreover, arithmetic cryptography helps to secure data/ information leaks as the whole data on the channel and at the agents are full proof encrypted and thus, cannot be interpreted [12]. Even if hacked, it is of no use. Any of the protocol can be used by these agents for the computations as long as they yield the final results correct. The Architecture of Multi‐ agent SMC is given in fig.1. In this architecture, Decision makers are the trustworthy components that check the agent’s availability and trust and accordingly allocate tasks. The data channels are en‐ crypted data channels in which data travels in encrypted manner. Parties fragment the data and then send it along the encrypted data channels to the decision makers for whatever task like, some mining task, statistical analysis, mathematical computation or else. These fragments are sent to the decision makers which provide some interme‐ JOURNAL OF COMPUTING, VOLUME 1, ISSUE 1, DECEMBER 2009, ISSN: 2151-9617 HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/ 198 diate conclusions according to the available data and re‐ quested operation. These intermediate conclusions are then forwarded to the agents that make the final conclud‐ ing remarks and the result from the majority of agents is taken to be correct. Fig. 1 Architecture of ‐agent Secure Multiparty Computation As the data from parties is encrypted, the chances of be‐ ing hacked are very less and moreover, decision makers contain the agent’s past transactional performance and it selects the new agent as per its performance. The correct‐ ness of result comes from the strength of the crypto‐ graphic algorithms used. Special arithmetic encryption algorithms are designed for the purpose which provides reliable encryption. Agents are also not allowed to de‐ crypt the data, but they apply the computations on same encrypted data. Thus the threat of agents being unreliable is also completely removed. Moreover, if some agent be‐ comes malicious and does false computation or gives wrong results, it can be caught immediately, as we are using multiple agents, so the results from their majority will still be correct. Thus, the above mentioned architec‐ ture confirms all the requirements of secure e‐governance and provides all the provisions under one architecture, which, as we have seen above are found very distributed through other mechanisms. We have to use their combi‐ nations to achieve desired features, whereas, here we get all of it under one unified architecture. 8 Features of the Multi-agent Secure Multi- party Computation The protocol as we discussed above meets the safety re‐ quirements of our e‐Governance System through its vari‐ ous features which are as under: Security: The security is guaranteed by its three 8.1 level architecture. Data is sent over encrypted and does not move collectively; instead it is fragmented so that no one can get the complete data. Also, the neither the deci‐ sion makers nor the agents untrustworthiness can affect the security of the system as they have just mere sequence of encrypted bits which gives no sense and is useless. 8.2 Privacy: The decision makers helps in maintain‐ ing privacy as the data packets cannot be identified as to which party they belong. They provides an abstraction in between the agents and parties, which otherwise would have been linked. The agents selected for computation are also not revealed by the decision makers. If in worst case, if they even disclose the agent, no one can affect the data as it if encrypted and also agents don’t know its decryp‐ tion mechanism. 8.3 Correctness: The correctness comes from the fact correctness of the computation procedure that has been used which is completely robust. It always guaran‐ tees correct results. Also, as the result is taken by the hon‐ est majority and not by a single agent, there is no possibil‐ ity for incorrect results. 9 Performance Analysis of the Multi-agent Secure Multiparty Computation The data is forwarded to decision makers in encrypted form. Therefore, additional encryption mechanism needs to be implemented. Let the total number of decision mak‐ ers be Dn. Since, the data is first fragmented and they for‐ warded, hence, the probability of hacking the data of nth party with r data fragments is, Pr (n) data = 1/r (1) This probability of data hacking increases as the number of packets r increases as shown in fig. 2. Also, since there are m decision makers, then the prob‐ ability of decision maker to be corrupt is, Pm (n) decision maker =1/m (2) Therefore, the probability of the decision maker to be‐ come bias and make wrong decisions can be observed as shown in fig. 3. Also, the numbers of agents are p; hence the probability of selecting a wrong agent by a biased decision maker is given by, P (n) wrong agent = (1⁄m) + (1/p)‐(1/ (m×p)) Although, the number of agents and decision makers may vary, but the decision makers are always less than the agents. Therefore, the probability of this situation also decreases as the number of agents increases, and it even becomes better if we increase the number of decision makers also with increase in number of agents. 10 Conclusion Thus, from the above analysis, we can observe that secure multiparty computation can be an effective solution to e‐ Governance issues and helps to solve is major issues without much complexity. JOURNAL OF COMPUTING, VOLUME 1, ISSUE 1, DECEMBER 2009, ISSN: 2151-9617 HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/ 199 Probability of packet loss as the number of Packets inc reases 8. Durgesh Kumar Mishra and Manohar Chandwani, “A Zero Hacking Protocol for Secure Multiparty Computation using Multiple TTP”, in the proceeding of Tencon’08, 19‐21 Nov. 2008, pp:1‐6. 9. Y.Lindell and B. Pinkas, “Privacy Preserving Data Mining”. In advances in Cryptography‐CRYPTO‐2000, pp 36‐54, Springer‐ Verlag, August 24 2000. 10. O. Goldreich, “Secure Multiparty Computation”, September 1998 [Working draft] Online available on: http://www.wisdom.weizmann.ac.il/~oded/pp.html. 11. Vassilios S. Verykios, Elisa Bertino, Igor Nai Fovino, Loredana Parsiliti Provenza, Yucel Saygin, Yannis Theodoridis, “State‐of‐ The‐Art in Privacy Preserving Data Mining”, SIGMOD Record, Vol. 33, No. 1, March 2004. 12. Durgesh Kumar Mishra; Manohar Chandwani,”Arithmetic cryptography protocol for secure multi‐party computation”, in the proceeding of SECON’07, 22‐25 March 2007 pp:22 – 22. Authors Profile Dr. Durgesh Kumar Mishra Professor (CSE) and Dean (R&D), Secretary IEEE MP-Subsection, Acropolis Institute of Technology and Research, Indore, MP, India, Ph - +91 9826047547, +91-731-4730038 Biography: Dr. Durgesh Kumar Mishra has received M.Tech. degree in Computer Science from DAVV, Indore in 1994 and PhD degree in Computer Engineering in 2008. Presently he is working as Profes- sor (CSE) and Dean (R&D) in Acropolis Institute of Technology and Research, Indore, MP, India. He is having around 20 Yrs of teaching experience and more than 5 Yrs of research experience. He has completed his research work with Dr. M. Chandwani, Director, IET-DAVV Indore, MP, India in Secure Multi- Party Computa- tion. He has published more than 60 pa- pers in refereed International/National Journal and Conference in- cluding IEEE, ACM etc. He is a senior member of IEEE and Secre- tary of IEEE MP-Subsection under the Bombay Section, India. Dr. Mishra has delivered his tutorials in IEEE International conferences in India as well as other countries also. He is also the programme committee member of several International conferences. He visited and delivered his invited talk in Taiwan, Bangladesh, USA, UK etc in Secure Multi-Party Computation of Information Security. He is an author of one book also. He is also the reviewer of tree International Journal of Information Security. He is a Chief Editor of Journal of Technology and Engineering Sciences. He has been a consultant to industries and Government organization like Sale tax and Labor Department of Government of Madhya Pradesh, India. Samiksha Shukla Asst. Professor (CSE), Christ University, Banglore, India. Biography: Samiksha Shukla has received M.Tech. degree in Com- puter Science from DAVV, Indore in 2005. Presently she is working as Asst. Professor (CSE) Christ University, Banglore, India.. She is having around 04 Yrs of teaching experience. She is doing her re- search work with Dr. Durgesh Kumar Mishra. t e k c a P f o y t i l i b a b o r P s s o L 1.2 1 0.8 0.6 0.4 0.2 0 1 2 3 4 5 6 7 8 Numbe r of Pa cke ts 9 10 Fig. 2. Probability of packet loss with increase in number of packets. Probability of dec is ion makers to become biased g n i e b f o y t i l i b a b o r P d e s a i b 1.2 1 0.8 0.6 0.4 0.2 0 1 9 10 8 7 6 5 4 3 2 Numbe r of De cision Ma ke rs Fig. 3 Probability of the decision makers to become biased with in- crease in number of Decision Makers 2. REFERENCES 1. Claudio Biancalana and Francesco Saverio Profiti “Security and Privacy Preserving Data in e‐Government Integration”, http://www.esiig2.it , Department of Computer Science and Au‐ tomation, Roma Tre University, Rome, Italy. Jaijit Bhattacharya,” Privacy Technology for E‐Governance”, Department of Management Studiesn Indian Institute of Tech‐ nology, Delhi, India, pp: 114‐124. 3. Tu Bao Hu,”Privacy Preserving Data Mining and E‐Commerce and E‐Governance”, School of Knowledge Science, Japan Ad‐ vanced Institute of Science and Technology and IOIT, Vietnam‐ ese Academy of Science and Technology. Arry Brandt, Lois Delcambrie, Sharon Dawes, Howard Brad‐ sher‐Fredrick “Being Successful in Digital Government Project”, Birds of a Feather Session, University of Wisconsin‐Madison, 2004. Jacques Calmet, Regine Endsuleit, Pierre Maret “A Multi Agent Model for Secure and Scalable E‐Business Transactions”, http://www.avalon.ira.uka.de. Jacques Calmet, Regine Endsuleit, “An Agent Framework for Legal Validations of E‐Transactions”, University of Karlsruhe, Germany. 7. Rachet GreenStadt, “Privatizing Constraint Optimization”, Harvard University. 4. 5. 6.
1903.03053	1	1903	2019-03-07T17:17:23	A Privacy-preserving Disaggregation Algorithm for Non-intrusive Management of Flexible Energy	[ "cs.MA", "math.OC" ]	We consider a resource allocation problem involving a large number of agents with individual constraints subject to privacy, and a central operator whose objective is to optimizing a global, possibly non-convex, cost while satisfying the agents'c onstraints. We focus on the practical case of the management of energy consumption flexibilities by the operator of a microgrid. This paper provides a privacy-preserving algorithm that does compute the optimal allocation of resources, avoiding each agent to reveal her private information (constraints and individual solution profile) neither to the central operator nor to a third party. Our method relies on an aggregation procedure: we maintain a global allocation of resources, and gradually disaggregate this allocation to enforce the satisfaction of private contraints, by a protocol involving the generation of polyhedral cuts and secure multiparty computations (SMC). To obtain these cuts, we use an alternate projections method \`a la Von Neumann, which is implemented locally by each agent, preserving her privacy needs. Our theoretical and numerical results show that the method scales well as the number of agents gets large, and thus can be used to solve the allocation problem in high dimension, while addressing privacy issues.	cs.MA	cs	A Privacy-preserving Disaggregation Algorithm for Non-intrusive Management of Flexible Energy Paulin Jacquot, Olivier Beaude, Pascal Benchimol, Stéphane Gaubert, Nadia Oudjane 9 1 0 2 r a M 7 ] A M . s c [ 1 v 3 5 0 3 0 . 3 0 9 1 : v i X r a Abstract -- We consider a resource allocation problem involv- ing a large number of agents with individual constraints subject to privacy, and a central operator whose objective is to opti- mizing a global, possibly non-convex, cost while satisfying the agents' constraints. We focus on the practical case of the man- agement of energy consumption flexibilities by the operator of a microgrid. This paper provides a privacy-preserving algorithm that does compute the optimal allocation of resources, avoiding each agent to reveal her private information (constraints and individual solution profile) neither to the central operator nor to a third party. Our method relies on an aggregation procedure: we maintain a global allocation of resources, and gradually disaggregate this allocation to enforce the satisfaction of private contraints, by a protocol involving the generation of polyhedral cuts and secure multiparty computations (SMC). To obtain these cuts, we use an alternate projections method à la Von Neumann, which is implemented locally by each agent, preserving her privacy needs. Our theoretical and numerical results show that the method scales well as the number of agents gets large, and thus can be used to solve the allocation problem in high dimension, while addressing privacy issues. I. INTRODUCTION Motivation. Consider an operator of an electricity mi- crogrid optimizing the joint production schedules of renew- able and thermal power plants in order to satisfy, at each time period, the consumption constraints of its consumers. To optimize the costs and the renewables integration, this operator relies on demand response techniques, that is, taking advantage of the flexibilities of some of the consumers electric appliances -- those which can be controlled without impacting the consumer's confort, as electric vehicles or water heaters [1]. However, for privacy reasons, consumers are not willing to provide neither their consumption con- straints nor their consumption profiles to a central operator or any third party, as this information could be used to induce private information such as their presence at home. The global problem of the operator is to find an allocation of power (aggregate consumption) p = (pt)t at each time period (resource) t ∈ T , such that p ∈ P (feasibility constraints of power allocation, induced by the power plants constraints). Besides, this aggregate allocation has to match an individual comsumption profile xn = (xn,t)t∈T for each of the consumer (agent) n ∈ N considered. The problem can P. Jacquot, O. Beaude, P. Benchimol and N. Oudjane are with EDF R&D, OSIRIS, Palaiseau, France. P. Jacquot and S. Gaubert are with Inria Saclay and CMAP, Ecole polytechnique, Palaiseau, France. [email protected] be written as follows: min x∈RN×T, p∈P f (p) xn ∈ Xn, ∀n ∈ N xn,t = pt, ∀t ∈ T , (cid:88) n∈N (1a) (1b) (1c) The (aggregate) allocation p can be made public, that is, revealed to all agents. However, the individual constraint set Xn and individual profiles xn constitute private information of agent n, and should not be reavealed to the operator or any third party. The approach adopted in our paper is to deal with the problem (1) as two kinds of interdependent subproblems. The firsts are optimal resource allocation problems, or mas- ter problems, minp∈P (s) f (p) which consists in finding an aggregate allocation over T resources (P (s) ⊂ P ⊂ RT ). The second kind is problems of finding the disaggregation of a given aggregate allocation p, that is, to find an individual profile xn for each agent (consumer) n satisfying her individ- ual constraint (1b), such that the aggregate of the profiles is the optimal allocation (1c) determined in a master problem. Aside from the example above, ressource allocation prob- lems (optimizing common resources shared by multiple agents) find many applications in energy [1, 2], logistics [3], distributed computing [4], health care [5] and telecom- munications [6]. In these applications, several entities or agents (e.g. consumers, stores, tasks) share a common re- source (energy, products, CPU time, broadband) which has a global cost for the system. For large systems composed of multiple agents, the dimension of the overall problem can be prohibitive and one can rely on decomposition and distributed approaches [7 -- 9] to answer to this issue. Besides, agents' individual constraints are often subject to privacy issues [10]. These considerations have paved the way to the development of privacy-preserving, or non-intrusive methods and algorithms, e.g. [11, 12]. In this work, we consider that each agent has a global demand constraint (e.g. energy demand or product quantity), which confers to the disaggregation problem the particular structure of a transportation polytope [13]: the sum over the agents is fixed by the aggregate solution p, while the sum over the T resources are fixed by the agent global demand constraint. Besides, individual constraints can also include minimal and maximal levels on each resource, as for instance electricity consumers require, through their appliances, a minimal and maximal power at each time period. Main Results. The major contribution of the paper is to provide a non-intrusive and distributed algorithm (Algo- rithm 4) that computes an aggregated resource allocation p, optimal solution of the -- possibly nonconvex -- optimization problem (1), along with feasible individual profiles x for agents, without revealing the individual constraints of each agent to a third party, either another agent or a central operator. The algorithm solves iteratively intances of master problem minp∈P (s) f (p) by constructing successive approx- imations P (s) ⊂ P of the aggregate feasible set of (1) for which a disaggregation exists, by adding a new constraint on p to P (s), before solving the next master problem. To identify whether or not disaggregation is feasible and to add a new constraint in the latter case, our algorithm relies on the alternating projections method (APM) [14, 15] for finding a point in the intersection of convex sets. Here, we consider the two following sets: on the one hand, the affine space defined by the aggregation to a given resource profile, and on the other hand, the set defined by all agents individual constraints (demands and bounds). As the latter is defined as a Cartesian product of each agent's feasibility set, APM can operate in a distributed fashion. The sequence constructed by the APM converges to a single point if the intersection of the convex sets is nonempty, and it converges to a periodic orbit of length 2 otherwise. Our key result is the following: if the APM converges to a periodic orbit, meaning that the disaggregation is not feasible, we construct from this orbit a polyhedral cut, i.e. a linear inequality satisfied by all feasible solutions p of the global problem (1), but violated from the current resource allocation (Thm. 4). Adding this cut to the master problem, we can recompute a new resource allocation and repeat this procedure until disaggregation is possible. Another major result stated in this paper is the explicit upper bound on the convergence speed of APM in our framework (Thm. 2), which is obtained by spectral graph theory methods, exploiting also geometric properties of transportation polytopes. This explicit speed shows a linear impact of the number of agents, which is a strong argument for the applicability of the method in large distributed systems. Related Work. A standard approach to solve resource allocation problems in a distributed way is to use a La- grangian (dual) decomposition technique [8, 16, 17]. Those techniques are generally used to decompose a large problems into several subproblems of small dimension. They may also be implemented in a way which preserve privacy (see Remark 2 in Sec. IV). However, Lagrangian decomposition methods are based on strong duality property, requiring global convexity hypothesis which are not satisfied in many practical problems (e.g. MILP, see Sec. V). On the contrary, our method can be used when the master allocation problem is not convex. In [2], the authors study a disaggregation problem similar to the one considered in this paper. Their results concern zonotopic sets, which is different from the structure we described in Sec. II. The APM has been the subject of several works in itself [15, 18, 19]. The authors of [20] provide general results on the convergence rate of APM for semi-algebraic sets. They show that the convergence is geometric for polyhedra. However, it is generally hard to compute explicitly the geometric convergence rate of APM, as this requires to bound the singular values of certain matrices arising from the polyhedral constraints. In [21], the authors provide an explicit convergence rate for APM on a class of polyhedra arising in submodular optimization. The sets they consider differ from the present transportation polytopes. Structure. In Sec. II, we describe the master resource allocation problem and formulate the associated disaggre- gation problem. In Sec. III, we focus on the APM and state our main results. In Sec. IV, we apply these results to describe a non-intrusive version of APM (NI-APM) that is used to describe our non-intrusive algorithm for computing an optimal resource allocation. Finally, in Sec. V, we provide a concrete numerical example based on a MILP to model the management of a local electricity system (microgrid), and study numerically the influence of the number of agents on the time needed for convergence of our algorithm. Notation. Vectors and matrices are denoted by bold fonts, v(cid:62) denotes the transpose of v, 1K denotes the vector (1 . . . 1)(cid:62) of size K, U([a, b]) stands for the uniform dis- tribution on [a, b]. We use (cid:107)x(cid:107)2 to denote the Frobenius n,t, and PC(.) to denote the Euclidean norm (cid:107)x(cid:107)2 =(cid:80) n,t x2 projection on a convex set C. II. MASTER PROBLEM AND DISAGGREGATION STRUCTURE In this work, we suppose an operator wishes to determine an allocation of resources, represented by a T -dimensional vector p, in order to minimize a global cost function f, for instance, an electricity power economic dispatch (or the allocation of different types of merchandise in warehouses in logistics applications) subject to a set of constraints described by a feasibility set P. This problem can be nonconvex either because of nonconvex costs f or because of a nonconvex feasible set P (see Sec. V). In the proposed method, the operator will consider master problems of the form: min p∈RT f (p) s.t. p ∈ P (s) , (2a) (2b) where the set P (s) ⊂ P is an aggregate approximation of disaggregation constraints. Indeed, the resource allocation p has to be shared between N agents (e.g. consumers). Each agent has a global demand (total energy needed) En and some lower and upper bounds on each of the resource t ∈ T . The admissible set of profiles of agent n is therefore: 1T =En and ∀t, xn,t≤xn,t≤xn,t}. (3) Xn The disaggregation problem consists in finding individual profiles x = (xn,t)n,t ∈ RN T of a given aggregated allocation p such that xn is feasible for each agent n: def= {xn∈RT x(cid:62) n FIND x ∈ Yp ∩ X where Yp def= {y ∈ RN Ty(cid:62)1N = p} and X def= (cid:89) n∈N (4) Xn . Following (4), the disaggregated profile refers to x, while the aggregated profile refers to the allocation p. Problem (4) may not always be feasible. Some neces- sary conditions for a disaggreagation to exist, obtained by summing the individual constraints on N , are the following aggregated constraints: p(cid:62)1T = E(cid:62)1N and x(cid:62)1N ≤ p ≤ x(cid:62)1N . (5) However, (II) are not sufficient conditions, as shown in Fig. 1 where the problem (4) is represented as a flow or circulation problem from source nodes t ∈ T to sink nodes n ∈ N . (cid:88) Indeed, with this circulation representation of the disaggre- gation problem (4), an immediate consequence of Hoffmann theorem [22, Thm. 3.18][23] is the following characterization of the disaggregation feasibility, which involves an exponen- tial number of inequalities: Theorem 1. The disaggregation problem (4) is feasible (i.e. X ∩ Yp (cid:54)= ∅) iff for any Tin ⊂ T ,Nin ⊂ N : xn,t + xn,t − (cid:88) pt ≤ (cid:88) t∈Tin,n /∈Nin t /∈Tin,n∈Nin t /∈Tin The inequality (6) has a simple interpretation: the residual demand (the left hand side composed of demand and exports minus production) in Tin ∪ Nin cannot exceed the import capacity (right hand side of the inequality). One can see that, in the example of Fig. 1, inequality (6) does not hold when using the cut composed of the dashed nodes p1 and E1. (cid:88) n /∈Nin En. (6) p1 = 0 p2 = 3 E1 = 2 E2 = 0.5 E3 = 0.5 Fig. 1. Example of disaggregation structure (T = 2, N = 3), with x = 0 and x := 1. Although the aggregate constraints (II) are satisfied, the disaggregation (4) of p is not feasible in this example (see Thm. 1). There are two main reasons for which solving (1) is harder than solving (2) and (4) separately: i) the dimension of (1) can be huge, as the number of agents N can be really important, for instance in the example of individual consumers; ii) also, and this is the main motivation of this work, the information related to (xn)n, (xn)n and (En)n might not be available to the centralized operator in charge of optimizing resources p, as this information may be confidential and kept by each agent n, not willing to reveal it to any third party. In the next sections, we provide a method that addresses those two issues, by considering subproblems (2) and (4) in- dependently and iteratively, and exploiting the decomposable structure of problem (4). III. ALTERNATE PROJECTION METHOD (APM) A. Convergence of APM on Transportation Polytopes In this section, we consider a fixed aggregated profile p and present the Von Neumann Alternate Projections Method (APM) [14] which solves the problem Eq. (4) of finding a point in the intersection X ∩ Yp. In the remaining, we will often ommit p and just write Y to denote Yp. The key idea of the method proposed in this paper is to use results of APM to generate a cut in the form of (6) and to add it as a new constraint in the master problem (2) to "improve" the aggregated profile p for the next iteration. As described in Algorithm 1, APM can be used to decompose (4) and only involves local operations. Algorithm 1 Alternate Projections Method (APM) Require: Start with y(0), k = 0 , εcvg, a norm (cid:107).(cid:107) on RN T 1: repeat 2: 3: 4: x(k+1) ← PX (y(k)) y(k+1) ← PY (x(k+1)) k ← k + 1 5: until(cid:13)(cid:13)y(k) − y(k−1)(cid:13)(cid:13) < εcvg The convergence of Algorithm 1 is proved by Thm. 2: Theorem 2 ([15]). Let X and Y be two convex sets with X bounded, and let (x(k))k and (y(k))k be the two infinite sequences generated by Algorithm 1 with εcvg = 0. Then there exists x∞ ∈ X and y∞ ∈ Y such that: k→∞ y∞; k→∞ x∞ , y(k) −→ (7b) In particular, if X ∩ Y (cid:54)= ∅, then (x(k))k and (y(k))k converge to a same point x∞ ∈ X ∩ Y. x∈X ,y∈Y (cid:107)x − y(cid:107)2 . x(k) −→ (cid:107)x∞ − y∞(cid:107)2 = min (7a) If disaggregation is not feasible, Thm. 2 states that APM will "converge" to an orbit (x∞, y∞) of period 2. The convergence rate of APM has been the subject of several works [18, 20], and it strongly depends on the structure of the sets on which the projections are done: for instance, if the sets are polyhedral, [20, Prop. 4.2] shows that the convergence is geometric. However, there are very few cases in which an explicit upper bound on the convergence rate has been proved. In our case, we are able to obtain such a bound, as shown in the following theorem: Theorem 3. For the sets X and Y defined in (3-4), the two subsequences of alternate projections converge at a geometric rate to x∞ ∈ X , y∞ ∈ Y, with: (cid:107)x(k)− x∞(cid:107)2≤2(cid:107)x(0)− x∞(cid:107)2 × ρk (cid:0)N (T + 1)2(T − 1)(cid:1) < 1 , def= 1 − 1 N T 4 where ρN T Same inequalities hold for the convergence of y(k) to y∞. Proof. Appendix II provides a sketch of the proof. Thm. 3 shows that the APM is efficient in our case of bounded transport polytopes. It shows that the number of iterations for a given accuracy grows linearly in the number of agents N. n Xn, so that the projection (13) can be computed by N projections As stated in (4), the set X is a Cartesian product(cid:81) on (Xn)n, which can be executed in parallel. Now, instead of solving the quadratic program by standard interior point methods and due to its particular structure, we can use the algorithm of Brucker [24], which has a complexity in O(T ). On the other hand, PY (.) is a projection on an affine space, and the solution can be obtained explicitly as: ∀n, t, yn,t = xn,t + νt and ν = 1 N (p − x(cid:62)1N ) . (8) B. Generation of a cut from APM iterates Our key result is the following: in the case where APM converges to a periodic orbit (x∞, y∞) with x∞ (cid:54)= y∞ (see Thm. 2), we obtain from (x∞, y∞) an inequality (6) that is violated by p: Theorem 4. For the sets X and Y defined in (3-4) and if X ∩ Y = ∅ , the following sets given by the limit orbit (x∞, y∞) defined in Thm. 2: n,t} n∈N x∞ t /∈T0 (cid:88) def= {tpt >(cid:80) def= {n En −(cid:80) En−(cid:88) define a Hoffman cut of form (6) violated by p, that is: xn,t < 0} xn,t < 0 . (10) (cid:88) T0 N0 (9b) (9a) pt + t∈T0 n∈N0 t /∈T0,n∈N0 This cut can be reformulated in terms of 1(cid:62) t∈T0,n /∈N0 t∈T0 xn,t −(cid:80) xn,t− (cid:88) (cid:88) (cid:88) t∈T0 n∈N (cid:88) t∈T0 AT0 < pt with AT0 def= N x∞ as: x∞ n,t. (11) Proof. Appendix I gives the sketch of the proof of Thm. 4. The complete proofs will be given elsewhere. One can see that, intuitively, N0 is the subset associated to T0 that minimizes the right hand side of (6). Note that Thm. 4 gives an alternative constructive proof of Hoffman circulation's theorem (Thm. 1) in the case of a bipartite graph of the form of Fig. 1. Moreover, in the case where the disaggregation problem (4) is not feasible, the negation of equation (11) provides a new valid constraint as a condition for the existence of a disaggreagated profile of p. This constraint can be used in the master problem (2) to update the vector of resources p for the next iteration. This constraint only involves the aggregate information 1(cid:62) N x∞ on the users profile. To make the process fully non-intrusive, we explain in Sec. IV-A how the operator can compute this constraint without making the agents reveal their profiles (x∞ n )n∈N . IV. NON-INTRUSIVE PROJECTIONS AND COMPUTATION OF DISAGGREGATED OPTIMAL RESOURCES A. Non-Intrusive Alternate Projections Method (NI-APM) Because of the particular structure of the problem, the projections in APM can be computed separately by the operator and the agents. The projection PY is made by the operator, which only requires to know p and the aggregate profile x(cid:62)1N according to from (8). The projection PX n Xn is executed in parallel by each agent: n computes PXn which only needs her private information En and xn, xn. However, in the way APM is described on X = (cid:81) in Algorithm 1, the operator and the agents still need to exchange the iterates x(k), y(k) at each step. To avoid the transmission of agents' profiles to the operator, we use a secure multiparty computation (SMC) technique (see [25]) which enables the operator to obtain the aggregate profile S(k) := 1(cid:62) N x(k) in a non-intrusive manner, as described in Algorithm 2. The main idea of SMC is that, instead of sending her profile xn, agent n splits xn,t for each t into N random parts (sn,t,m)m, according to an uniform distribution and summing to xn,t (Lines 2-3). Thus, each part sn,t,m taken individually does not reveal any information on xn nor on Xn, and can be sent to agent m. Once all exchanges of parts are completed (Line 5), and n has herself received the parts from other agents, agent n computes a new aggregate quantity σn (Line 7), which does not contain either any information about any of the agents, and sends it to the operator (Line 8). The operator can finally compute the quantity S = x(cid:62)1N = σ(cid:62)1N . n∈N xn Algorithm 2 SMC of Aggregate (SMCA)(cid:80) def= xn,t −(cid:80)N−1 Require: Each agent has a profile (xn)n∈N 1: for each agent n ∈ N do Draw ∀t, (sn,t,m)N−1 m=1∈U([0, A]N−1) 2: and set ∀t, sn,t,N 3: Send (sn,t,m)t∈T to agent m ∈ N 4: 5: done 6: for each agent n ∈ N do 7: 8: 9: done Compute ∀t, σn,t =(cid:80) 10: Operator computes S =(cid:80) Send (σn,t)t∈T to operator m∈N sm,t,n n∈N σn m=1 sn,t,m Remark 1. As σn, and sn are random by construction, an eavesdropper aiming to learn the profile xn of n has no choice but to intercept all the communications of n to all other agents (to learn (sn,t,m)m(cid:54)=n and (sm,t,n)m(cid:54)=n) and to the operator (to learn σn). To increase the confidentiality of the procedure, one could use any encryption scheme (such as RSA [26]) for all communications involved in Algorithm 2. We can use this non-intrusive computation of aggregate S in APM to obtain a non-intrusive algorithm NI-APM (Algorithm 3) in which agents do not reveal neither their profiles nor their constraints to the operator. One can see that x and y computed in Lines 3 and 8 in Algorithm 3 correspond to the projections computed in the original APM Algorithm 1. In Algorithm 3, the operator obtains the aggregate profile S(k) (Line 5), computes and sends the corrections ν(k) to all agents (Line 6). Then, each n = PY (x(k) agent can compute locally the projection y(k) n ) by applying the correction ν(k) (Line 8). N (p− 1(cid:62) N x∞). Thm. 4 ∞ def= {t ∈ T 0 < ν∞ t }. uses this limit value through T0 Yet, from APM, one can only access to ν(k) and thus to Using (8), we get ν(k) → ν∞ def= 1 n n ← PXn (y(k−1) x(k) for each agent n ∈ N do ) Algorithm 3 Non-intrusive APM (NI-APM) Require: Start with y(0), k =0, εcvg, εdis, norm (cid:107).(cid:107) on RN T 1: repeat 2: 3: 4: 5: 6: 7: 8: 9: 10: done Operator obtains S(k) ←SMCA(x(k)) (cf Algo.2) N (p − S(k)) ∈ RT to agents N and sends ν(k) := 1 for each agent n ∈ N do n ← x(k) done k ← k + 1 11: until (cid:13)(cid:13)x(k) − x(k−1)(cid:13)(cid:13) < εcvg 12: if (cid:13)(cid:13)x(k) − y(k)(cid:13)(cid:13) ≤ εdis then Compute y(k) n + ν(k) (cid:46) found a εdis-solution of the disaggregation problem Each agent adopts profile x(k) n return DISAG ← TRUE (cid:46) have to find a valid constraint violated by p t } Operator computes T0 ← {t ∈ T 3 2 Bεcvg < ν(k) 1(cid:62) N σ(k) Operator computes AT0 def= (cid:80) t if AT0 −(cid:80) t∈T0 pt < 0 then t∈T0 return DISAG ← FALSE, AT0 Return to Line 1 with εcvg ← εcvg/2 else (cid:46) need to run APM with higher precision 13: 14: 15: else 16: 17: 18: 19: 20: 21: 22: 23: end end the approximation T0, computed on Line 16), where B is a pre-defined constant. However, we show that for εcvg small enough and a well-chosen value of B, we obtain T0 = T ∞ 0 , so that we get the termination result: Proposition 1. For B > (1 − ρN T )−1, Algorithm 3 termi- nates in finite time. Thm. 3. In the case where(cid:13)(cid:13)x(k) − y(k)(cid:13)(cid:13) ≤ εdis, Algorithm 3 The termination of the loop Lines 1-11 is ensured by terminates. Otherwise, if (cid:107)x∞ − y∞(cid:107) > εdis, then Algo- rithm 3 terminates (i.e. Line 18 is True and a new cut is found) as soon as Bεcvg < min , where ν def= min{ν∞ > 0} and with (cid:107).(cid:107) = (cid:107).(cid:107)2. The complete proof is ommited here. (cid:110)(cid:107)x∞−y∞(cid:107)1 , 2 5 ν (cid:111) √ 2 N t In practice, we can start with a large εcvg to obtain the first constraints while avoiding useless computation, and then half εcvg if needed (Line 21) until the termination condition holds. Remark 2. SMC techniques could also be used to implement non-intrusive Lagrangian decomposition methods. However, these methods rely on a convexity hypothesis that we do not need in the proposed method. B. Non-intrusive Disaggregation of Optimal Allocation In this section, we describe a method to compute a solution of the global problem (1), is, an optimal resource allocation p for which a disaggregation exists, along with an associated disaggregated profile xn for each agent n. that This computation is done in a non-intrusive manner: the operator in charge of p does not have access neither to the bounding constraints x and x of the agents nor to the agents disaggregated profile x, as detailed in Algorithm 4 below. cs p Compute p(s) = arg minp∈P (s) DISAG ← NI-APM(p(s)) if DISAG then Algorithm 4 Non-intrusive Optimal Disaggregation Require: s = 0 , P (0) = P ; DISAG = FALSE 1: while Not DISAG do 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: done P (s+1) ← P (s) ∩ {p(cid:80) Operator adopts p(s) Obtain T (s) from NI-APM(p(s)) pt ≤ A(s)T0 end s ← s + 1 , A(s)T0 t∈T (s) else } 0 0 to the resource problem (2). This constraint Algorithm 4 iteratively calls NI-APM (Algorithm 3) and in case disaggregation is not possible (Line 6), a new constraint is added (Line 8), obtained from the quantity AT0 defined in (11), is an inequality on p and thus does not reveal significant individual information to the operator. The algorithm stops when disaggregation is possible (Line 4). The termination of Algorithm 4 is ensured by the following property and the form of the constraints added (10): Proposition 2. Algorithm 4 stops after a finite number of iterations, as at most 2T constraints (Line 8) can be added to the master problem (Line 2). Although there exist some instances with an exponential number of independent constraints, this does not jeopardize the proposed method: in practice, the algorithm stops after a very small number of constraints added (see the example of Sec. V). Intuitively, we will only add constraints "support- ing" the optimal allocation p. Thus, Algorithm 4 is a method which enables the operator to compute a resource allocation p and the N agents to adopt profiles (xn)n, such that (x, p) solves the global problem (1), and the method ensures that both: 1) the information relative to each agent constraints (upper bounds xn, lower bounds xn, demand En); 2) the final disaggregated profile xn (as well as the iterates (x(k))k and (y(k))k in NI-APM) are kept confidential by agent n and can not be induced by a third party (either the operator or any other agent m (cid:54)= n). V. APPLICATION TO MANAGEMENT OF A MICROGRID We apply the proposed method to solve a nonconvex distributed problem in the energy field. We consider a micro- grid [27] composed of N electricity consumers with flexible appliances (such as electric vehicles or water heaters), a photovoltaic (PV) power plant and a conventional generator. A. Mixed Integer Problem Formulation The operator responsible of the microgrid aims at satisfy- ing the demand constraints of consumers over a set of time periods T = {1, . . . , T}, while minimizing the energy cost for the community. We have the following characteristics: (pPV • the PV plant generates a nondispatchable power profile t )t∈T at marginal cost zero; • the conventional generator has a starting cost C ST, min- imal and maximal power production pg, pg, and piecewise- linear and continuous generation cost function pg (cid:55)→ f (pg): f (pg) = αk + ckpg, if pg ∈ Ik def= [θk−1, θk[, k = 1 . . . K, def= pg; where θ0 • each agent n ∈ N has some flexible appliances which require a global energy demand En on T , and has consump- tion constraints on the total household consumption, on each time period t ∈ T , that are formulated with xn, xn. These parameters are confidential because they could for instance contain some information on agent n habits. The master problem (2) can be written as a MILP (12): def= 0 and θK (cid:17) α1bON t + ckpg kt + C STbST t (12a) (cid:16) (cid:88) k,t, ∀t ∈ T k=1 pg 1,t ≤ θ1, ∀t ∈ T t∈T p,pg,(pg min k),(bk),bON,bST t =(cid:80)K (cid:88) k t ... , ∀t ∈ T t ≤ pgbON (12b) (12c) (12d) t−1, ∀t ∈ {2, . . . , T} 2,t ≤ b1,t(θ2 − θ1), ∀t ∈ T K,t ≤ bK−1,t(θK − θK−1), ∀t ∈ T t − bON pg b1,tθ1 ≤ pg b2,t(θ2 − θ1) ≤ pg ... 0 ≤ pg t ≥ bON bST t ≤ pg pgbON t , b1,t, . . . , bK−1,t ∈ {0, 1}, ∀t ∈ T bON , bST t p ≤ pPV + pg p(cid:62)1T = E(cid:62)1N x(cid:62)1N ≤ p ≤ x(cid:62)1N . In this formulation (12b-12f) are a mixed integer formu- lation of the generation cost function f: one can show that k,t ≥ θk for the boolean variable bk,t is equal to one iff pg each k ∈ {1, . . . , K − 1}. Note that only α1 appears in (12a) because of the continuity assumption on f. (12e) (12f) (12g) (12h) (12i) (12j) (12k) (12l) Constraints (12g-12h) ensure the on/off and starting con- straints of the power plant, (12j) ensures that the power allo- cated to consumption is not above the total production, and (12k-12l) are the aggregated feasibility conditions already referred to in (II). Note that more complex and realistic MILP models exist for power plants (e.g. [28]), but with the same structure than (12). The nonconvexity of (12) comes from the existence of starting costs and on minimal the power constraint, which makes necessary to use boolean state variables bST, bON. B. Parameters with random parameters for each value of N. A scaling def= N/20 is applied on parameters to ensure factor κN that production capacity is large enough to meet consumers demand. The parameters are chosen as follows: • T = 24 (hours of a day); (cid:105) • production costs: K = 3 , θ = [0, 70, 100, 300]κN , c = [0.2, 0.4, 0.5], pg =50κN , pg =300κN , α1 =4 and C ST = 15; • photovoltaic: pPV κN t = for t ∈ {6, . . . , 20}, pPV • for consumption parameters, we used xn,t ∼ U([0, 10]), xn,t ∼ U([0, 5]) + xn,t and En ∼ U([1(cid:62) T xn]), so that individual feasibility (Xn (cid:54)= ∅) is ensured. t = 0 otherwise (see Fig. 3); 50(1−cos( (t−6)2π )+U([0, 10]) T xn, 1(cid:62) (cid:104) 16 N = # master pb. # projs. 24 193.6 9506.9 25 194.1 15366.7 26 225.5 24319.3 27 210.9 26537.5 28 194.0 26646.4 NUMBER OF SUBPROBLEMS SOLVED (AVERAGE ON 100 INSTANCES) TABLE I Fig. 2. Total number of computed projections for different values of N. We observe that the number of agents N has a sublinear impact on the total number of projections needed. C. A limited impact of the number N of agents t xn,t (to avoid the defined by (cid:107)x(cid:107) = maxn∈N(cid:80) We implement Algorithm 4 using Python 3.5. The MILP (12) is solved using Cplex Studio 12.6 and Pyomo inter- face. Simulations are run on a single core of a cluster at 3GHz. For the convergence criteria (see Lines 11 and 12 of √ Algorithm 3), we use εdis = 0.01 with the operator norm N factor in the convergence criteria appearing with (cid:107).(cid:107)2), and starts with εcvg = 0.1. The largest instances took around 10 minutes to be solved in this configuration and without parallel implementation. As the CPU time needed depend on the cluster load, it is not a reliable indicator of the influence on N on the complexity of the problems. Moreover, one advantage of the proposed method is that the projections in APM can be computed locally by each agent in parallel, which could not be implemented here for practical reasons. Tab. I gives two robust indicators of the influence of N on the problem complexity: the number of master problems solved and the total number of projections computed, on average over the hundred instance for each value of N: We simulate the problem described above for different values of N ∈ {24,25, 26,27,28} and one hundred instances • one observes that the number of master problems solved (MILP (12) ), which corresponds to the number of constraints or "cuts" added to the master problem, remains almost constant when N increases; • in all instances, this number is way below the upper bound of 224 > 1, 6 × 107 possible constraints (see proof of Prop. 2), which suggests that only a polynomial number of constraints are added in practice; • the average total number of projections computed for each instance (total number of iterations of the while loop of Algorithm 3, Line 1 over all calls of APM in the instance) increases in a sublinear way, as illustrated in Fig. 2, which is even better that one could expect from the upper bound given in Thm. 3. Fig. 3. Example of optimal resource allocation (power production p = pPV + pg) in the example of Sec. IV, with N = 20 agents. Exploiting consumption flexibilities, the consumption is higher during the PV production periods. VI. CONCLUSION We provided a non-intrusive algorithm that enables to compute an optimal resource allocation, solution of a -- possibly nonconvex -- optimization problem, and affect to each agent an individual profile satisfying a global demand and lower and upper bounds constraints. Our method uses local projections and works in a distributed fashion. Hence, it en- sures that the problem is not affected by the high dimension relative to the large number of agents, and that it is privacy- preserving, as agents do not need to reveal any information on their constraints or their individual profile to a third party. Several extensions and generalizations can be considered. First, we could generalize the abstract circulation problem on the bipartite graph depicted in Fig. 1 to an arbitrary network, where the set of nodes is partitioned in K parts defining K sets on which we could make alternating projections. Second, our method relies on the particular structure obtained from the form of constraints. Although these kind of constraints are widely used to model many practical situations, it would also be useful to obtain similar results for arbitrary (or polyhedral) agents constraints. Last, a deeper complexity analysis, with a thinner upper bound on the maximal number of constraints (cuts) added in the algorithm (see Prop. 2 and Tab. I) would constitute interesting results. REFERENCES [1] P. Jacquot, O. Beaude, S. Gaubert, and N. Oudjane, "Analysis and implementation of an hourly billing mechanism for demand response management," IEEE Trans. Smart Grid, 2018. [2] F. L. Müller, J. Szabó, O. Sundström, and J. Lygeros, "Aggregation and disaggregation of energetic flexibility from distributed energy resources," IEEE Trans. Smart Grid, 2017. [3] K. K. Lai, K. Lam, and W. K. Chan, "Shipping container logistics and allocation," J. Oper. Res. Soc., vol. 46, no. 6, pp. 687 -- 697, 1995. [4] P.-Y. R. Ma et al., "A task allocation model for distributed computing systems," IEEE Trans. Computers, vol. 100, no. 1, pp. 41 -- 47, 1982. [5] A. Rais and A. Viana, "Operations research in healthcare: a survey," Int. Trans. Oper. Res., vol. 18, no. 1, pp. 1 -- 31, 2011. [6] M. Zulhasnine, C. Huang, and A. Srinivasan, "Efficient resource allo- cation for device-to-device communication underlaying lte network," in WiMob, 2010 IEEE 6th Int. Conference. IEEE, 2010, pp. 368 -- 375. [7] D. P. Bertsekas and J. N. Tsitsiklis, Parallel and distributed computa- tion: numerical methods. Prentice hall Englewood Cliffs, NJ, 1989, vol. 23. [8] D. P. Palomar and M. Chiang, "A tutorial on decomposition methods for network utility maximization," IEEE J. Sel. Areas Commun., vol. 24, no. 8, pp. 1439 -- 1451, 2006. [9] L. Xiao and S. Boyd, "Optimal scaling of a gradient method for distributed resource allocation," J. Optim. Theory. Appl., vol. 129, no. 3, pp. 469 -- 488, 2006. [10] B. A. Huberman, E. Adar, and L. R. Fine, "Valuating privacy," IEEE security & privacy, vol. 3, no. 5, pp. 22 -- 25, 2005. [11] A. Zoha, A. Gluhak, M. A. Imran, and S. Rajasegarar, "Non-intrusive load monitoring approaches for disaggregated energy sensing: A survey," Sensors, vol. 12, no. 12, pp. 16 838 -- 16 866, 2012. [12] G. Jagannathan, K. Pillaipakkamnatt, and R. N. Wright, "A new privacy-preserving distributed k-clustering algorithm," in Proc. of the 2006 SIAM Int. Conf. on Data Mining. SIAM, 2006, pp. 494 -- 498. [13] E. D. Bolker, "Transportation polytopes," Journal of Combinatorial Theory, Series B, vol. 13, no. 3, pp. 251 -- 262, 1972. [14] J. Von Neumann, Functional operators: Measures and integrals. Princeton University Press, 1950, vol. 1. [15] L. Gubin, B. Polyak, and E. Raik, "The method of projections for finding the common point of convex sets," USSR Comput. Math. & Math. Phys., vol. 7, no. 6, pp. 1 -- 24, 1967. [16] L. Xiao, M. Johansson, and S. P. Boyd, "Simultaneous routing and resource allocation via dual decomposition," IEEE Trans. Comm., vol. 52, no. 7, pp. 1136 -- 1144, 2004. [17] K. Seong, M. Mohseni, and J. M. Cioffi, "Optimal resource allocation for ofdma downlink systems," in Information Theory, 2006 IEEE Int. Sym. IEEE, 2006, pp. 1394 -- 1398. [18] H. H. Bauschke and J. M. Borwein, "On the convergence of von neumann's alternating projection algorithm for two sets," Set-Valued Analysis, vol. 1, no. 2, pp. 185 -- 212, 1993. [19] H. H. Bauschke, J. Chen, and X. Wang, "A bregman projection method for approximating fixed points of quasi-bregman nonexpansive mappings," Applicable Analysis, vol. 94, no. 1, pp. 75 -- 84, 2015. [20] J. M. Borwein, G. Li, and L. Yao, "Analysis of the convergence rate for the cyclic projection algorithm applied to basic semialgebraic convex sets," SIAM J. Optim., vol. 24, no. 1, pp. 498 -- 527, 2014. [21] R. Nishihara, S. Jegelka, and M. I. Jordan, "On the convergence rate of decomposable submodular function minimization," in Advances in Neural Information Processing Systems, 2014, pp. 640 -- 648. [22] W. J. Cook, W. Cunningham, W. Pulleyblank, and A. Schrijver, Combinatorial optimization. Springer, 2009. [23] A. J. Hoffman, "Some recent applications of the theory of linear inequalities to extremal combinatorial analysis," in Proc. of Symposia on Applied Mathematics, 1960, pp. 113 -- 127. [24] P. Brucker, "An O(n) algorithm for quadratic knapsack problems," Oper. Res. Lett., vol. 3, no. 3, pp. 163 -- 166, 1984. [25] A. C. Yao, "How to generate and exchange secrets," in 27th Annual Symp. Found. of Comp. Sci. (SFCS), Oct 1986, pp. 162 -- 167. [26] R. L. Rivest, A. Shamir, and L. Adleman, "A method for obtaining digital signatures and public-key cryptosystems," Communications of the ACM, vol. 21, no. 2, pp. 120 -- 126, 1978. [27] F. Katiraei, R. Iravani, N. Hatziargyriou, and A. Dimeas, "Microgrids management," IEEE power and energy magazine, vol. 6, no. 3, 2008. [28] M. Carrión and J. M. Arroyo, "A computationally efficient mixed- integer linear formulation for the thermal unit commitment problem," IEEE Trans. Pow. Sys., vol. 21, no. 3, pp. 1371 -- 1378, 2006. APPENDIX I PROOF OF PROP. 3 To show Thm. 4, we formulate the projections PX and PY as the solutions of the constrained quadratic programs: (cid:107)x − y(cid:107)2 2 1 min x∈RN T 2 x1T = E x ≤ x ≤ x and: (cid:107)x − y(cid:107)2 2 1 min y∈RN T 2 y(cid:62)1N = p (13a) (13b) (13c) (14a) (14b) (λ) (µ, µ) (ν) , ; n,t = xn,t n,t = xn,t ; 0 are nonempty. n,t ≥ xn,t and x∞ where λ, µ, µ, ν are the Lagrangian multipliers associated to the constraints. Although there is no such explicit char- acterization of the solution of (13) as the one (8) given for (14), we can obtain the following properties: Proposition 3. Suppose that X ∩ Y = ∅ and consider the sets T0 and N0 given by (9). Then we have the following: (i) ∀t ∈ T0,∀n /∈ N0, y∞ (ii) ∀n ∈ N0, λn < 0 ; (iii) ∀t /∈ T0,∀n ∈ N0, x∞ (iv) the sets T0,T c The proof of Prop. 3 relies on the KKT optimality condi- tions associated to (13) and (14). Then we use Prop. 3 and the convergence condition to prove Thm. 4: xn,t −(cid:88) xn,t − (cid:88) (cid:33) (cid:32)(cid:88) (cid:88) t∈T0 −ν∞ (cid:12)(cid:12)(cid:33) = −(cid:107)x∞ − y∞(cid:107)1 where the last equality comes from(cid:80) (cid:80) n∈N (xE n,t − xn,t − (cid:88) xP n,t) = 0. The compact form (11) also follow from Prop. 3: (cid:88) (cid:88) (cid:88) 0 , N0 and N c (cid:88) n,t − (cid:88) (cid:12)(cid:12)x∞ (cid:32)(cid:88) (cid:32) −(cid:88) (cid:88) (cid:88) n∈N n,t − y∞ (cid:88) (cid:88) (cid:88) (cid:88) AT0(x) = t∈T0,n /∈N0 t /∈T0,n∈N0 (cid:33) n∈N0 n∈N0 En + En + x∞ y∞ < 0 , t∈T0 t∈T0 t∈T0 t∈T0 n∈N n∈N n∈N xn,t pt = = = n,t n,t 2 t t∈T0,n /∈N0 x∞ n,t + t /∈T0,n∈N0 x∞ n,t = x∞ n,t . = n∈N0,t∈T0 t∈T0 n∈N t∈T0,n /∈N0 APPENDIX II PROOF OF THM. 2 For this analysis, we use the space RN T = RT ×···×RT , where the (n−1)T +1 to nT coordinates correspond to agent n, for 1 ≤ n ≤ N. We make use of the following results: Lemma 1 ([21]). For APM on polyhedra X and Y, the sequences (x(k)k and (y(k)k converge at a geometric rate, where the rate is bounded by the maximal value of the square of the cosine of the Friedrichs angle cF (U, V ) between a face U of X and a face V of Y, where cF (U, V ) is given by: cF (U, V ) = sup{uT v (cid:107)u(cid:107) ≤ 1,(cid:107)v(cid:107) ≤ 1 u ∈ U ∩ (U ∩ V )⊥, v ∈ V ∩ (U ∩ V )⊥}. Lemma 2 ([21]). Let A and B be matrices with or- thonormal rows and with equal numbers of columns and Λsv(AB(cid:62)) the set of singular values of AB(cid:62). Then if Λsv(AB(cid:62)) = {1}, then cF (Ker(A), Ker(B)) = 0. Other- wise, cF (Ker(A), Ker(B)) = maxλ<1{λ ∈ Λsv(AB(cid:62))}. N N −1 −1 √ √ 1T} with A def= In our case, the polyhedra Y is an affine subspace Y = J1,N ⊗ IT , {x ∈ RN T Ax = where ⊗ denotes the Kronecker product. The matrix A has orthonormal rows and the direction of Y is Ker(A). Describing the faces X is more complex. We have a polyhedral description of X , and the faces of X are subsets of the collection of affine subspaces indexed by (T n,T n)n ⊂ T N (with T ∩ T = ∅): (cid:111) A(T n,T n)n (cid:110) def= ∀t∈ T n, xn,t = xn,t, and ∀t ∈ T n, xn,t = xn,t (x)nt ∀n, x(cid:62) 1T = En and n . (cid:33) (cid:32)(cid:88) (cid:88) ((cid:80) T −1 def= √ and the matrix B has b def= (cid:80) The associated linear subspace is given by Ker(B), where the IN ⊗ J1,T , N first rows of B are given by [B][N ] n card(Tn) more rows, where def= T n ∪ T n, corresponding to the saturated inequalities def= card(Tn). Then, renormalizing B, we can show that the double product S := (AB(cid:62))(A(cid:62)B), of size T × T is given by: Tn (xn,t = xn,t or xn,t). We denote by Kn n n n k,(cid:96) + 1 N 1 N S def= 1≤t≤T 1t∈Tn ) Et,t. 1{k,(cid:96)}⊂T c T − Kn n and P =def= IT − S. As P can be Denote ¯T def= ∪nT c written as a block diagonal matrix P = diag(P ¯T , 0 ¯T c ), we can restrict ourselves to the subspace Vect(et)t∈ ¯T to find the least positive eigenvalue of P , that we denote by λ1. Consider the weighted graph G = ( ¯T ,E) whose vertices are the time periods ¯T and edge (k, (cid:96)) has weight Sk,(cid:96) = (if this quantity is zero, then there is no (cid:96)(cid:54)=k −Pk,(cid:96) = Using the Laplacian property and Cauchy-Schwartz, one (cid:80) edge between k and (cid:96)). One can show that (cid:80) Pkk, which shows that P is the Laplacian matrix of G. shows that for any u ⊥ 1: 1{k,(cid:96)}⊂T c T−Kn n 1 N n u(cid:62)P u ≥ min k,(cid:96)∈(s∗-t∗) (−Pk,(cid:96)) (ut∗−us∗ )2 ds∗ ,t∗ ≥ 4T(cid:107)u(cid:107)2 N (T +1)2(T−1)2 2 where ut∗ := maxt ut, us∗ := mint ut and ds∗,t∗ is the the distance between s∗ and t∗ in G, and (s∗-t∗) a path from s∗ to t∗. As 1 is an eigenvector of P associated to λ0 = 0, from the minmax theorem, we get λ1(P ) ≥ N (T +1)2(T−1) := 1−ρN T and the greatest singular value lower than one of BA(cid:62)AB(cid:62) is ρN T and then, applying the preceding Lemmas of [21], we obtain the result stated in Thm. 3 4
0802.1393	1	0802	2008-02-11T08:55:46	Les Agents comme des interpr\'eteurs Scheme : Sp\'ecification dynamique par la communication	[ "cs.MA", "cs.AI" ]	We proposed in previous papers an extension and an implementation of the STROBE model, which regards the Agents as Scheme interpreters. These Agents are able to interpret messages in a dedicated environment including an interpreter that learns from the current conversation therefore representing evolving meta-level Agent's knowledge. When the Agent's interpreter is a nondeterministic one, the dialogues may consist of subsequent refinements of specifications in the form of constraint sets. The paper presents a worked out example of dynamic service generation - such as necessary on Grids - by exploiting STROBE Agents equipped with a nondeterministic interpreter. It shows how enabling dynamic specification of a problem. Then it illustrates how these principles could be effective for other applications. Details of the implementation are not provided here, but are available.	cs.MA	cs	Les Agents comme des interpréteurs Scheme : Spécification dynamique par la communication Agents as Scheme Interpreters: Enabling Dynamic Specification by Communicating Clement Jonquet Stefano A. Cerri LIRMM - Université Montpellier II 161, rue Ada 34392 Montpellier Cedex 5 - France {cerri, jonquet}@lirmm.fr Résumé Nous avons proposé dans de précédents papiers une extension et une implémentation du modèle STROBE, qui considère les Agents comme des interpréteurs Scheme. Ces Agents sont capables d 'interpréter des messages dans des environnements donnés incluant un interpréteur qui apprend de la conversation et donc qui représente l’évolution de sa connaissance au niveau méta. Quand ces interpréteurs sont non déterministes, le dialogue consiste à raffiner les spécifications d’un problème par des ensembles de contraintes. Ce papier présente un exemple de génération dynamique de service – tels qu’ils sont nécessaires sur le GRID – exploitant des Agents STROBE équipés d’un interpréteur non déterministe. Il montre comment réaliser la spécification dynamique d’un problème. Puis il illustre comment ces principes peuvent être intéressants pour d’autres applications. Les détails de l’implémentation ne sont pas fournis ici mais sont disponibles. Mots Clef interprétation dynamique de Communication Agent, message, déterministe, STROBE, non évaluation spécification dynamique, contraintes, dialogue. Abstract We proposed in previous papers an extension and an implementation of the STROBE model, which regards the Agents as Scheme interpreters. These Agents are able to interpret messages in a dedicated environment including an interpreter that learns from the current conversation evolving meta-level Agent representing therefore knowledge. When the Agent interpreter is a nondeterministic one, the dialogues may consist of subsequent refinements of specifications in the form of constraint sets. The paper presents a worked out example of dynamic service generation – such as necessary on Grid – by exploiting STROBE Agents equipped w i th a nondeterministic interpreter. It shows how enabling dynamic specification of a problem. Then, it illustrates how these principles could be effective for other applications. Details of the implementation are not provided here, but are available. Keywords Agent communication, message dynamic interpretation, nondeterministic interpretation, STROBE model, ACL, dynamic specification, constraints, dialogue. 1. Introduction La transmission du savoir est quelque chose d 'essentiel pou r tou tes les soc ié tés humaines c'est elle qui assure l'évolution et l'adaptation des ces sociétés à travers le temps. Nous ne pouvons imaginer où en serait l'homme s'il réapprenait à chaque génération à tailler des silex ou à contrôler le feu. Mais le problème ne se pose pas car les êtres humains possèdent une faculté d 'apprentissage et d'adaptation qui n'est ni prévisible ni mesurable. Il n'en est pas de même pour les entités informatiques. En effet, assurer la transmission du savoir, l'apprentissage et l'adaptabilité des sociétés d 'Agents est un véritable sujet qui promet encore de longues années de recherche. Nous tentons d'amener, ici, une petite pierre à cet énorme édifice en proposant un modèle d 'apprentissage de connaissances basé sur la communication. Notre travail 1 est le résultat de la mise en commun de deux domaines : l’interprétation des langages et la communication Agent. Notre idée est de profiter de l'apprentissage comme effet secondaire de la communication. En effet, le but de l'éducation est de faire changer d'état son interlocuteur. Ce changement se fait après l'évaluation (1 e r domaine) des nouveaux éléments apportés par la communication (2 ème domaine). Nous proposons dans cet article un modèle permettant de réaliser ce changement. Plus précisément, notre travail es t basé sur le modèle STROBE [4], qui considère les Agents comme des interpréteurs Scheme. Ces Agents sont capables d 'interpréter les messages d 'une conversation dans un environnement donné, incluant un interpréteur, dédié à la conversation courante. Nous allons montrer comment, grâce à la communication, les variables stockées dans ces environnements, et en particulier les interpréteurs, peuvent être modifiés dynamiquement. Ainsi, considérant les environnements d’un Agent comme sa connaissance, nous montrero ns comment il apprend plus qu'une simple information, en modifiant, sa façon de voir ces informations et en devenant capable d'en intégrer de nouvelles. Nous illustrerons ce que nous appelons apprentissage au méta-niveau par une expérimentation de dialogue de type « professeur-élève » . Nous pensons que considérer les Agents comme des interpréteurs est un point de vue très intéressant et nous montrerons comment cela peut être effectif pour des domaines tels que le Web, le GRID, la génération de service dynamique. Ce dernier intérêt constitue le noyau de cet article où nous illustrerons le potentiel d’Agents considérés comme des interpréteurs non déterministes dans un scéna r io de type e-commerce. Le but de l’article étant de convaincre que sur le Web, l’approche classique du génie logiciel qui consiste à spécifier un problème puis le coder peut être remplacé par une approche qui alterne spécification et exécution basée sur le dialogue. Le but est de réaliser ces deux étapes en même temps, soit de permettre la s pécification dynamique d’un problème. La suite de l’article est organisée de la manière suivante : la partie 2 propose un aperçu des problématiques associées aux notions de communication et d’apprentissage dans les SMA. Nous tentons également de définir un scénario « idéal » pour l'évolution des sociétés d'Agents et pour la communication dans l'avenir. Les parties 3 et 4 présentent notre modèle qui consiste à considérer les Agents comme des interpréteurs Scheme et qui leur fournit un ensemble de couples (environnement interpréteur) pour la représentation des autres. La partie 5 présente brièvement notre architecture Scheme et illustre l’apprentissage au méta-niveau dans un exemple jouet de t y p e « professeur – éleve ». Puis, la partie 6 illustre la spécification dynamique, après avoir succinctement introduit le concept d’interpréteur non déterministe. Finalement, la partie 7 met en avant les intérêts et extensions de ces principes pour d’autres domaines. 2 2. Communication et apprentissage dans les SMA La simple mi se en commun de plusieurs Agents ne suffit pas à former un SMA (Systèmes Multi-Agents), c 'est le fait que ces Agents communiquent qui le permet. C'est grâce à la communication (directe ou non) que sont possibles la coopération et la coordination entre Agent [8]. Définir et modéliser la communication a toujours été difficile. Aujourd'hui, il existe de nombreux modèles et langages de communication mais pouvons nous dire qu'ils sont adaptés au monde Agent ou à de nouvelles formes de communication comme celles que propose le Web ? Il ne s 'agit pas de prendre les langages traditionnels de communication ou même les paradigmes actuels de programmation et de les adapter au Web et aux Agents. Il s'agit de développer de nouvelles architectures et de nouveaux langages conçus pour le Web et les Agents. En effet, les langages traditionnels sont ficelés et il est souvent très difficile de les faire évoluer. Pour être efficace, la communication doit être intrinsèque à un langage. Par exemple, en ce qui concerne l'apprentissage, il ne suffit pas simplement d'apprendre au niveau donnée ou au niveau contrôle , il faut aussi apprendre au niveau interpréteur. Nous pouvons faire ressortir quelques pré -requis aux langages de communication Agent : Si nous considérons le fait qu'une communication a des effets sur les interlocuteurs (acte perlocutoire), alors il nous faut obligatoirement considérer que les Agents peuvent changer de but ou de point de vue au milieu de cette communication. Ils doivent donc être autonomes et s'adapter pendant la communication [4]. La partie 5 explique brièvement comment notre architecture Scheme nous permet cela. Il faut aussi considérer que des Agents peuvent interagir entre eux ou avec des humains suivant les mêmes principes [12] [6]. Ce qui compte c'est la représentation qu'un Agent se fait de son interlocuteur. La partie 6 montrera comment le concept d’Environnement Cognitifs [3] peut nous aider. Historiquement, les SMA étaient construits avec un langage de communication intégré fonctionnant de manière ad -hoc. Aujourd'hui, la communauté SMA tend à fournir des Agent Communication Languages (ACL) applicables à un maximum d'interactions entre Agents. En effet, fournir un ACL fort d'un point de vue sémantique donne un gros avantage pour la création et l'évolu tion d'un SMA [7]. Ces ACL sont basés sur la théorie des actes de langage1. Traditionnellement, les messages KQML ou FIPA -ACL fournissent un élément qui correspond à l'ontologie utilisée dans la communication. Cela permet de rendre les ACL indépendants de n'importe quel vocabulaire et donne à l'interlocuteur un moyen d'établir la correspondance concept / signification des éléments du contenu d 'un message. Dorénavant un ACL ad-hoc n’es t 1 Modèle issu de la philosophie du langage [2] et [15]. plus spécifié pour l'incorporer à un SMA, mais une ontologie est constru ite et passée en paramètre des messages. Nous proposons dans ce papier une alternative à cet état de fait. Les ACL reçoivent souvent la critique du manque de performatif. Notre première expérimentation propose un exemple de solution à ce problème. Elle illustre une technique pour diffuser des nouveaux performatifs dans un SMA, par apprentissage direct d 'Agent à Agent. langages, de nombreux modèles de Face à ces communication ont été proposés. Une alternative sur la façon de considérer les Agents est présentée dans [11]. Entre autres, STROBE s'intéresse à de nombreux principes importants pour une communication et se base sur les trois primitives Scheme : STReam, OBject, et Environment. Il met en avant des points comme la représentation de l'interlocuteur, la conservation de l'historique d'une conversation, l'apprentissage issu de la communication, etc. Nous reviendrons souvent aux propositions de STROBE car le modèle proposé ici s'en inspire. Imaginons maintenant un scénario « idéal » de ce que pourrait être un SMA dans quelques années. Il considère toutes les entités du Web comme des Agents d 'une même société qui peuvent communiquer les uns avec les autres naturellement et se transmettre des connaissances. Cette société pouvant s'étendre sans aucune limite. Da ns ce SMA, chaque Agent est initialisé avec un minimum requis de connaissances (pour interagir) ainsi qu'avec une spécialité qui le caractérise et qu'il peut transmettre aux autres. Cet Agent possède un ensemble d 'interpréteurs qui représentent sa connaissance et son évolution dans le temps. Il apprend tout le reste au fur et à mesure de ses communications. Il apprend même à enseigner ! Pour chaque Agent avec qui il communique, il a une représentation spécifique de celui-ci, ce qui lui permet de tenir compte de ce qu 'il apprend tout en gardant éventuellement son comportement et ses croyances d'origine intactes. Bien sur, ces Agents profitent également des méthodes d’apprentissage classique, leur permettant d’analyser leurs environnements locaux (de représentations des autres) pour décider ou pas de modifier leur comportement (leur environnement global). D 'un point de vue coopération / coordination, un Agent peut demander à un autre d'interpréter pour lui tel ou tel programme et de lui renvoyer le résultat. Un Agen t peu t également transmettre à un autre un interpréteur qui lui permet d'effectuer sa tâche. Ces interpréteurs peuvent même être transmis avant une conversation, comme est transmise aujourd'hui une ontologie. Mais vu qu'aucun interpréteur n'évolue de la même manière, puisque deux conversations ne sont jamais identiques, cela donne à cette société d 'Agents une pluralité des connaissances incommensurable ! Son évolution devient totalement imprévisible et autonome. Dans cette société toutes les taches son t réalisées par le dialogue. L'intégration des nouveaux Agents se fait naturellement et petit à petit. Il n'est plus possible alors de prouver de manière théorique que tel ou tel Agent sait accomplir une tâche ; le seul 3 moyen est de regarder les solutions émergentes qui apparaissent lorsqu 'un problème se pose. Nous tentons dans cet article de proposer des idées pour rendre réalisable ce scénario encore utopiste aujourd'hui. Entre autres, nous allons voir comment un Agent qui possède plusieurs les modifier interpréteurs peut dynamiquement pour évoluer au fur et à mesure des conversa t ions . 3. Les Agents comme des interpréteurs Scheme Le modèle que nous proposons a pour caractéristique principale le fait qu'il considère les Agents comme des interpréteurs. En s’inspirant de la boucle classique d 'évaluation nous considérons les Agents comme des interpréteurs de type REPL (Read, Eval, Print, Listen). Lors d'une communication, chaque Agent exécute une boucle REPL qui sont imbriquées les unes dans les autres. Ce princip e est important car il permet de considérer les Agents comme des entités autonomes dont le comportement et les interactions sont dirigées par des procédures concrètes. Notre modèle utilise Scheme aussi bien pour le contenu des messages que pour leur représ entation. De cette façon nous pouvons utiliser le même interpréteur pour évaluer le message et son contenu. Exemple d'expression Scheme représentée par des messages : >(define x 2) (cid:243) >(assertion (define x 2)) :x (cid:243) :(ack x) (cid:243) >(request x) >x :2 (cid:243) :(answer 2) Ce point de vue est très intéressant car les Agents bénéficient de tous les avantages liés à Scheme pour la représentation de connaissances en particulier le modèle de mémoire que constitue l'environnement et le modèle de contrôle obtenu grâce aux procédures et aux continuations de première classe. Par exemple, STROBE propose une structure d'environnement conservant l'historique, basé sur les streams, c 'est à dire que les liaisons ne sont plus du type (var val) mais du type (var val1 … valn -1 valn). Cette structure devient accessible à des Agents représentés par des interpréteurs. Pour implémenter ce modèle nous avons écrit un méta- évaluateur Scheme (appelé méta-eval) ainsi qu’un méta- évaluateur non déterministe (appelé méta-ambeval) reconnaissant un certain langage (dans lequel doivent être écrit le contenu des messages d’une conversation) et nous y avons rajouté un module d 'interprétation des messages, la fonction ambevaluate-kqmlmsg. Ceci nous permet de bien expliciter les trois niveaux d’abstraction : donnée , contrôle , interpréteur. L’apprentissage au niveau donnée consiste à affecter des valeurs à des variables déjà existantes ou à définir de nouvelles données (exemple : (set! a 4) ou (define a 3)) ; Au niveau contrôle , consiste à définir de nouvelles fonctions par abstractions sur celles déjà existantes (exemple : (define (square x) (* x x))). Et enfin, l’apprentissage au niveau interpréteur ou méta-niveau consiste à faire évoluer directement l’interpréteur correspondant à l’Agent (exemple : rajo uter une forme spéciale). Notre travail propose un mécanisme d’apprentissage à ces trois niveaux et en particulier au troisième. En faisant évoluer son interpréteur2, un Agent apprend plus qu 'une simple information, il change complètement sa façon de percevoir ces informations. C'est la différence entre apprendre une donnée et apprendre à traiter une classe de données. 4. Représentation des autres Un des principes de STROBE est d’avoir un « modèle du partenaire », c’est la notion d’Environnement Cognitif. En effet, pour ce qui est de la représentation de l'interlocuteur, STROBE met en avant le fait qu'il faut avoir un modèle du partenaire pour pouvoir reconstruire son état interne. Ce modèle propose que chaque dialogue soit interprété dans une paire d'environnement : le premier privé, appartenant à l'Agent, et le deuxième représentant le modèle du partenaire courant. Notre travail exploite cette notion. En il se greffe dessus car, comme fait, le concept fournit aux Agents un d'Environnement Cognitif environnement global (ou privé) et plusieurs environnements locaux de représentation des autres, notre proposition fournit aux Agents non pas un interpréteur mais plusieurs interpréteurs dont un global (ou privé) et un pour chaque Agent dont ils ont une représentation. Ainsi, l 'évaluation des messages d 'une conversation se fait avec un interpréteur donné dans un environnement donné. Notre travail se greffe sur ce concept d 'environnement dans le sens où ces interpréteurs, pour être accessibles, doivent eux-mêmes être stockés dans ces environnements. Ainsi nos Agents possèdent les trois attributs suivants : - GlobalEnv leur environnement global. - GlobalInter leur interpréteur global. - Other = {(name, environment, interpreter)} un ensemble de triplet correspondant aux représentations des au t res . Leur environnement global est privé et ne change pas. C'est cet environnement qui est dupliqué3 lors de l'arrivée d 'une nouvelle conversation et c 'est son clone, stocké dans un élément de Other, qui est modifié au fur et à mesure de la conversation. La Figure 1 illustre ces représentations. 2 Nous verrons à la partie suivante qu’en fait nos Agents sont vus comme un ensemble d’interpréteur. 3 Pas forcément si nous considérons un Agent qui voudrait reprendre une conversation dans le contexte d'une autre déjà existante ou ayant existée (avec son environnement et son interpréteur). 4 Agent A GlobalEnvA GlobalInterA Other = {(BA, InterBA, EnvBA) (CA, InterCA, EnvCA)} Agent B GlobalEnvB GlobalInterB Other = {(AB, InterAB, EnvAB) (CB, InterCB, EnvCB)} Agent C GlobalEnvC GlobalInterC Other = {(BC, InterBC, EnvBC) (AC, InterAC, EnvAC)} Conversation s Figure 1 . Les attributs d’un Agent et ses représentations des autres. 5. Modification dynamique d’un interpréteur 5.1. Architecture Scheme En Scheme, la boucle classique d’interprétation REPL consiste à attendre que l’utilisateur tape une expression, lire cette expression, l’évaluer renvoyer le résultat et attendre l’entrée de l’expression suivante. Nous sommes alors classiquement au niveau donnée et contrôle . Le n iveau du dessus , interpréteur, n’est pas accessible directement. Pour y accéder il faut soit simplement avoir accès aux sources de cet interpréteur (codé dans un certain langage) et modifier ces sources. Soit d’une manière plus sophistiquée si l’interpréteur est réflexif et propose le mécanisme des reifying procedures, alors ce mécanisme peut être utilisé pour accéder au contexte d’exécution4 d’une fonction et le modifier [9]. L’interpréteur est alors dynamiquement modifiable. Nous utilisons un autre schéma. Au lieu de faire évaluer les expressions utilisateur par l’interpréteur courant (Scheme), nous utilisons celui-ci pour appeler un autre interpréteur (méta-eval), stocké lui dans une structure de donnée (un environnement de première classe que nous con trôlons) qui lui applique le bon interpréteur (méta- ambeval) sur l’expression utilisateur. Ainsi, les expressions utilisateur peuvent consister à modifier l’environnement et en particulier l’interpréteur stocké dans celui-ci. L’interpréteur évaluant l’expression utilisateur est donc accessible et modifiable. Notre idée est de transporter ce schéma dans nos Agents . Ainsi nous pouvons considérer les trois interpréteurs suivants caractérisés par deux procédures : 4 Le contexte d'exécution d'une expression est constitué de l’environnement (r) d’évaluation de cette expression (e) et de la continuation (k) qui correspond à la prochaine expression à évaluer avec le résultat. - Scheme : eval et apply - Méta-eval : evaluate et apply-procedure - Méta-ambeval : ambevaluate et ambapply- procedure5 Alors la boucle classique d’évaluation est (eval e r k).où e r k est le contexte d’exécution. La notre est (eval (apply-procedure ‘ambevaluate e r ks kf)). 5.2. Protocole de communication Les me ssages que nous cons idérons son t insp irés de la structure des messages KQML ou FIPA -ACL. Ils seront notés par la suite kqmlmsg, ils sont de la forme : (kqmlmsg performative sender receiver content). Pour que méta-ambeval puisse interpréter les messages kqmlmsg, il faut lui ajouter un test qui les reconnaît dans la fonction ambevaluate, ainsi qu'une fonction qui les traite en évaluant leur contenu, ambevaluate-kqmlmsg . Les performatifs essentiels sont : assertion, order, ack, executed ainsi que, nous le verrons , broadcast , qui fait l'objet de l'exemple expliqué plus loin. Lorsqu'un Agent reçoit un message indexé par un performatif qu'il ne connaît pas il l'indique de manière particulière : - Les messages assertion ont pour but de modifier le comportement ou les représentations de l'interlocuteur. Leurs réponses son t des messages de type accusé de réception ack signifiant un succès ou une erreur. - Les messages order demandent à l'interlocuteur d'exécuter une procédure. Le résultat est envoyé par celui-ci par un me ssage de type executed. - Les messages broadcast consistent à envoyer en contenu un couple (perform, content) qui signifie que l'interlocuteur doit envoyer un message avec comme performatif perform et comme contenu content à interlocuteurs en cours tous ses (diffusion). Il n'y a pas de réponse définie pour les messages broadcast . 5.3. : Exemple d’un dialogue « professeur - élève » Cet exemple jouet illustre les aspects techniques de notre travail, il montre que ce modèle est viable et peut réellement être impléme nté. L’expérimentation montre comment un Agent peut dynamiquement changer sa fonction d’interprétation des messages, donc une partie de son interpréteur. Pour cela nous utilisons l’architecture vue plus haut , chaque Agent possédant un ensemble d’instances de méta-ambeval. C’est un dialogue type « professeur-élève ». Un Agent teacher demande à un autre Agent student de diffuser (kqmlmsg de type lui un message broadcast) pour à tous ses correspondants mais le student ne connaît pas le performatif utilisé par le teacher. Par conséquent, le 5 Nous avons écrit méta-eval et méta-ambeval en s’inspirant respectivement de [9] et [1]. 5 teacher transmet au student deux messages (de type assertion et order) explicitant comment prendre en compte ce performatif. Il lui re -transmet alors son message d'origine et obtient alors satisfaction. Le dialogue exact de l'expérimentation est décrit par la Figure 3. Nous avons développé pour les besoins de notre expérimentation des Agents capables de communiquer, c 'est à dire de s 'échanger des messages les uns avec les a u t r e s e t d'y produire des réponses significatives (ceci suivant le protocole définit plus haut)6. Ils ne font rien lorsqu'ils ne communiquent pas et leur autonomie se caractérise par le fait qu'ils apprennent seuls. Ils ont comme attributs, name, globalEnv, globalIn ter, other, ainsi que deux files de messages (une en sortie et une en entrée), et une structure stockant les conversations courantes. Leur comportement consiste uniquement à appliquer la boucle REPL (F igu re 2). Notons que ces Agen ts ne possèden t aucune techn ique d’appren t issage classique, mais nous nous inscrivons là -de s su s dan s une logique globale de la communauté IA. LISTEN - Retirer un message de la file des messages en entrée - Si ce message est nul, renvoyer(pasdemessage ) - Sinon transmettre ce message à READ READ - Transmettre ce message à EVAL EVAL - Récupérer l’interpréteur et l’environnement dédié à la conversation - Evaluer ce message avec les éléments choisis - Transmettre via la continuation le résultat à PRINT PRINT - Si le résultat est une liste de messages envoyer le ou les messages - Sinon traiter la réponse - Revenir à LISTEN Figure 2 . La boucle REPL de nos Agents. A l'issue du traitement du dernier message le student a modifié sa fonction ambevaluate-kqmlmsg e t donc sa façon d’interpréter les messages. Le code correspondant à cette fonction dans son environnement dédié à cette conversation est changé. Il est donc maintenant apte à traiter les messages indexés par le performatif broadcast. Remarque (F igu re 3) : Le message learn-broadcast- code-msg indique comment générer le nouveau code en tenant compte de l’ancienne définition du student . Ce code est stocké dans la variable learn-broadcast- code. 6 Nos Agents sont en fait des objets programmés en Scheme d'après les techniques présentées par Normark dans Simulation of Object-Oriented and Mechanisms in Scheme [13] TEACHER STUDENT Voici la définition de la procédure square : (kqmlmsg 'assertion teacher student '(define (kqmlmsg 'ack student teacher '(.)) (square x) (* x x))) Ok, je connais maintenant cette procédure : Diffuse à tous tes correspondants : Désolé, je ne connais pas ce performatif : (kqmlmsg 'broadcast teacher student '(order (kqmlmsg 'answer student teacher ’’(no-such- (square 3))) performative broadcast)) Ok, voilà comment ajouter broadcast à la liste des performatifs que tu reconnais : Voilà le code que tu devras générer et ajouter à ta procédure Ok, j'ai rajouté ce code dans une variable de mon environnement : ambevaluate-kqmlmsg : (kqmlmsg 'assertion teacher student learn- (kqmlmsg 'ack student teacher '(.)) broadcast-code-msg) Exécute cette procédure : Ok, je viens de modifier mon évaluateur : ’(set! (kqmlmsg student teacher 'order (kqmlmsg 'executed student teacher '(.)) ambevaluate-kqmlmsg learn-broadcast-code))) Diffuse à tous tes correspondants : Ok, je diffuse… (kqmlmsg 'broadcast teacher student '(order (kqmlmsg 'order … '(square 3)) (square 3))) Figure 3 . Dialogue « professeur-élève » pour l’enseignement de broadcast interpréteur non 6. Spécification dynamique par évaluation non déterministe Avant de regarder comment utiliser des interpréteurs non déterministes pour la génération de service et la spécification dynamique, nous devons introduire la notion de d’interpréteur non déterministe telle qu’elle est présenté dans [1]. 6.1. Qu’est ce qu’un déterministe ? L’idée principale est qu’avec un langage non déterministe, une expression peut avoir plusieurs valeurs possibles. Une expression représente en fait un ensemble de « mondes » possibles, chacun déterminé par un ensemble de choix. Un programme peut avoir, en évaluation non déterministe, plusieurs exécutions différentes. L’évaluation non déterministe repose sur la forme spéciale amb. L’expression (amb exp1 exp2 … expn) retourne une des n expressions expi7. Par exemple l’expression : (list (amb 1 2 3) (amb 'a 'b)) peut avoir 6 valeurs : (1 a) (1 b) (2 a) (2 b) (3 a) (3 b) L’intérêt d’un tel évaluateur est qu’ensuite des fonctions peuvent appeler la forme amb en rajoutant des contraintes (sous forme de prédicat) sur les valeurs renvoyées par 7 L’expression (amb) sans argument correspond à un échec. amb. Ces contraintes s’expriment avec la forme spéciale require définie comme ceci : (define (require p) (if (not p) (amb)))8 L’évaluation d’une expression amb peut être vue comme l’exploration d’un arbre de solutions où le traitement de l’expression continue trouver une solution jusqu'à respectant toutes les contraintes et ceci tant que l’arbre complet n’a pas été parcouru. Lorsque la forme (amb) e s t évaluée cela correspond à une feuille de cet arbre, une autre branche est donc explorée. Pour implémenter un évaluateur non déterministe il faut bien sur gérer la forme spéciale amb, et il faut introduire le concept de failure con t inua t ion (kf) qui est la continuation appelé en cas d’échec dans une évaluation. Le contexte d’exécution devient donc (e r ks kf) où ks correspond à la continuation classique. Un peu comme en Prolog, nous pouvons demander l’ensemble des solutions d’une expression logique une à une avec un évaluateur non déterministe, la forme try-again permet de voir la prochaine évaluation à succès d’une fonction appelant amb. La boucle classique d’interprétation est modifiée en évaluation non déterministe pour tenir compte de ces retours en arrière (back t rack s). 8 The form (if cond exp) returns exp value if cond is true. Else it return no value. 6 Considérons la fonction (an-element-of list) qui renvoie la valeur d’un élément d’une liste donnée. Alors son éva lua t ion es t : > (an-element-of ’(a b c)) : b > try-again : a > try-again : c > try-again : no more values Avant tout, regardons comment est écrit le corps de la fonction an-element-of. La contrainte dans ce cas est que la liste ne soit pas vide. Lorsqu’elle est null? alors (amb) est évaluée et l’évaluateur non déterministe passe à la branche suivante de l’arbre des solutions : (define (an-element-of items) (require (not (null? items))) (amb (car items) (an-element-of (cdr items)))) Regardons maintenant un exemple de programme non déterministe un peu plus compliqué tiré de [1]. Il s’agit d’un problème type puzzle logique : “Baker, Cooper, Fletcher, Miller, and Smith live on different floors of an apartment house that contains only five floors. Baker does not live on the top floor. Cooper does not live on the bottom floor. Fletcher does not live on either the top or the bottom floor. Miller lives on a higher floor than does Cooper. Smith does not live on a floor adjacent to Fletcher's. Fletcher does not live on a floor adjacent to Cooper's. Where does everyone live?” Nous pouvons déterminer qui habite à chaque étage en énumérant toutes les possibilités et en leur appliquant les différentes contraintes ; Cela donne la fonction suivante : (define (multiple-dwelling) (let ((baker (amb 1 2 3 4 5)) (cooper (amb 1 2 3 4 5)) (fletcher (amb 1 2 3 4 5)) (miller (amb 1 2 3 4 5)) (smith (amb 1 2 3 4 5))) (require (distinct? (list baker cooper fletcher miller smith))) (require (not (= baker 5))) (require (not (= cooper 1))) (require (not (= fletcher 5))) (require (not (= fletcher 1))) (require (> miller cooper)) (require (not (= (abs (- smith fletcher)) 1))) (require (not (= (abs (- fletcher cooper)) 1))) (list (list 'baker baker) (list 'cooper cooper) (list 'fletcher fletcher) (list 'miller miller) (list 'smith smith)))) 7 L’évaluation de cette fonction (multiple-dwelling) renvoie, avec un évaluateur non déterministe : ((baker 3) (cooper 2) (fletcher 4) (miller 5) (smith 1)) 6.2. Spécification dynamique par la communication dans un scénario e -commerce Nos Agents sont considérés comme des interpréteurs Scheme donc ils peuvent très bien être vus comme des interpréteurs non déterministes ; C’est à dire reconnaître et traiter les formes amb, require, try-again… Ceci est, en effet, intéressant pour eux car ils pourraient alors résoudre des programmes comme ceux vus dans la section précédente. Mais ce n’est pas le seul intérêt. Pour faire le lien avec notre travail orienté communication Agent, le po in t le p lus in téressan t es t que nos Agen ts peuven t construire de tels programmes au fur et à mesure d’un dialogue et appliquer ensuite ces programmes pour donner une réponse ou accomplir une tâche pour un autre Agent. Les contraintes, définissant un programme non déterministe, peuvent être explicitées au fur et à mesure du dialogue en utilisant les outils présentés plus haut, de modification dynamique des fonctions et de la façon dont elles sont in terprétées. Considérons par exemple un dialogue type e-commerce, de recherche de billet de train comme celui présenté dans [11]. Un billet est caractérisé par une ville de départ, une ville de destination, un prix, une date. Un Agent SNCF est sollicité pa r un Agen t Cl ien t pour lui faire des propositions de billets. Le dialogue en situation réelle pourrait être : a. C l ien t : Je voudrais un billet de Montpellier à Paris. b . SNCF : Pour quand ? c. C l ien t : Demain avant 10H du matin. Pourriez vous me faire une proposition ? d . SNCF : Voilà, train 34170, départ demain 9H30, Montpellier en direction de Paris, 150€. e. C l ien t : Vous n’auriez pas à moins de 100 € ? f. SNCF : Voilà, train 34730, départ demain 8H41, Montpellier en direction de Paris, 95€. g . C l ien t : Une autre proposition s’il vous plait ? h . SNCF : Voilà, train 34392, départ demain 9H15, Montpellier en direction de Paris, 98€. i. C l ien t : Ok, celui-ci me va. Nous pouvons constater que les interactions a, b, c et e consistent à établir les contraintes sur la sélection du billet. Les interactions d, f et h sont des applications de la fonction de recherche de billet avec différentes contraintes. L’interaction g correspond à une demande du Cl ien t pour avoir une autre réponse, soit, pour explorer une autre branche de l’arbre des solu tions. La Figure 4 illustre la façon dont ce dialogue peut être traduit en expression Scheme pour être réalisé par nos Agents. L’Agent Cl ien t transmet ses requêtes sous forme de require et de try-again. CLIENT SNCF Je voudrais un billet de Montpellier à Paris (require (eq? depart montpellier)) (require (eq? dest paris)) Début de la construction de find-ticket : (define (find-ticket) (let ((depart(amb ens-ville)) (dest (amb ens-ville)) (prix (amb ens-prix)) (date (amb ens-date))) (require (not (eq? depart dest))) (require (eq? depart montpellier)) (require (eq? dest paris)) (list (list ’depart depart) (list ’destination dest) (list ’prix prix) (list ’date date)))) Pour quand ? Demain avant 10H du matin (require (< date demain10H)) Pourriez vous me faire une proposition ? (find-ticket) Modification de la fonction find-ticket en lui ajoutant la nouvelle contrainte. Puis exécution de cette fonction. ((depart montpellier) (destination paris) (prix 150) (date dem9H30)) Voilà, train 34170, départ demain 9H30, Montpellier en direction de Paris, 150€. Vous n’auriez pas à moins de 100 € ? idem. (require (< prix 100)) (find-ticket) ((depart montpellier) (destination paris) (prix 95) (date dem8H41)) Une autre proposition s’il vous plait ? (try-again) Voilà, train 34730, départ demain 8H41, Montpellier en direction de Paris, 95€. Execution de find-ticket : ((depart montpellier) (destination paris) (prix 98) (date dem9H15)) Voilà, train 34392, départ demain 9H15, Montpellier en direction de Paris, 98€ Ok, celui-ci me va Figure 4 . Dialogue entre l’Agent Client et l’Agent SNCF pour la recherche de billet L’Agent SNCF commence, au début de la conversation, la construction d’une nouvelle fonction find-ticket qu’il modifie et exécute au fur et à mesure de cette conversation. Ces modifications consistent à changer une valeur dans l’environnement de l’Agent SNCF représentant l’Agent C l ien t . Elles sont réalisées à l’aide des principes présentés précédemment. Cette idée est très intéres sante car c’est le dialogue qui construit le calcul à effectuer et non le contraire. C’est un scénario que nous pourrions retrouver dans de nombreuses applications e-commerce ou d’autres du même genre où des Agents doivent construire un programme solution. trouver une ensemble pour L’approche classique de construction d’un programme (la plus fréquemment utilisée dans le génie logiciel), qui consiste à spécifier un programme avant de le coder, est changée par une approche de spécification dynamique pendan t la construction d’un programme. C'est-à-dire que la spécification et la réalisation sont faites en même temps. Ce scénario ressemble et s’adapterait facilement à ceux envisagés sur le GRID [14] [5] où il s’agit de générer des services dynamiquement plutôt q ue de fournir une prestation prédéfinie et statique. 8 7. Intérêts et extensions de ces principes L’expérimentation de la 5èm e partie montre comment prendre en compte un nouveau performatif, donc comment modifier la fonction d'interprétation des messages d'un Agent. Mais les mêmes principes peuvent être utilisés pour modifier n'importe qu'elle partie de l'interpréteur d'un Agent. Par exemple nous aurions pu faire un exemple qui rajoute cond ou let* au langage reconnu par notre méta- ambeval. Voir même un Agent qui enseigne à un autre comment transformer son interpréteur en interpréteur paresseux en changeant ses fonctions ambevaluate et ambapply-procedure. Grâce à ce protocole, nos Agents possèdent en fait un ensemble d 'interpréteurs qui représentent ils effet, connais sances. En leurs correspondent aux sous-langages que nos Agen ts reconnaissent et donc à leurs facultés à effectuer une tâche. Comme vu dans le scénario « idéal » , les Agen ts peuvent effectuer des tâches pour d 'autres ou même interpréteurs 9 s'échanger leurs comme dans des architectures GRID où il est plus intéressant de déplacer les processus que les données. Leurs interpréteurs peuvent même être aussi transmis avant une conversation, plutôt que de transférer ontologie. Imaginons une société d 'Agents ou un SMA qui suit ce genre de protocole alors n'importe quel Agent peut apprendre quelque chose d 'un autre. Si nous construisons un Agent avec une connaissance et une spécialité le caractérisant alors cette spécialité peut se diffuser au fur et à mesure de ses communications. Pour le Web, ce genre de principe est très intéressant. Considérons un Agent qui joue le rôle de serveur pour une nouvelle application orientée Web et qui utilise une série de performatifs correspondant exactement à ce qu'il doit faire. S'il est construit avec le potentiel pour enseigner ces performatifs alors il s'intégrera très bien à une société d'Agents en les enseignants au début des communications qu’il peut avoir. En outre, pour faire le lien avec XML, nous pouvons considérer une DTD, un XML-Schema et surtout un programme XSL comme des « interpréteurs » de données XML. Le modèle présenté ici permet alors de faire évoluer dynamiquement ces « interpréteurs » et donc les documents XML associés donnant au Web un dynamisme et une adaptabilité inégalable. Cela peut être facilement mis en place considérant l’analogie entre Scheme et XML, vu que les documents XML sont des arbres et quoi de mieux que Scheme pour traiter des arbres. Dans la même idée de nombreux langages liant les S-expressio ns (Scheme) et XML apparaissent [10]. 9 Ceci s'inspire de l'idée énoncée dans [1] : « If we wish to discuss some aspect of a proposed modification to Lisp with another member of the Lisp community, we can supply an evaluator that embodies the change. The recipient can then experiment with the new evaluator and further send back comments as modifications. » Les principes présentés dans cet article peuvent être également utiles pour d’autres types de scénarios. Imaginons qu’au lieu de transférer la définition de la procédure square, l’Agent teacher transfère à l’Agent studen t la définition d’une procédure implémentant un algorithme optimisé de résolution d’un problème. Par exemple, la fonction calculant le nombre de Fibonacci en utilisant la mémoization (cette fonction, memo-fib, transforme le calcul exponentiel d’un nombre de Fibonacci en un calcul linéaire). Dans ce cas, l’Agent student , après avoir appris le performatif broadcas t , peut jouer le rôle de serveur de grille de calcul (Grid Computing) en procédant à une sélection des Agents participants à un lourd calcul utilisant par exemple memo-fib de la manière suivante : Il peut demander à tous ses correspondants de réaliser pour lui un calcul simple d’un nombre de Fibonacci et de lui renvoyer le résultat. En fonction de certain critère comme en particulier le temps de réponse ce t Agen t es t capab le de sélectionner un ensemble d’Agents et de leur demander de participer à un gros calcul, utilisant les nombres de Fibonacci. Il peut même enseigner aux autres Agents la bonne version de la fonction implémentant cet algorithme. Cette idée est effectivement particulièrement intéressante pour le Grid Computing mais peut être également utilisée dans des protocoles de communication type contract net . Conclusion Nous avons essayé de montrer dans ce papier une méthode d 'apprentissage pour les Agents cognitifs issue de la communication. Cet apprentissage peut se faire par communication simple (niveau donnée et contrôle), ou par modification interne de l'Agent (niveau interpréteur). Nous avons illustré cette idée par un couple d’exemples jouets. Dans le second, nous pouvons voir comment notre modèle convient tout particulièrement à des scénarios de type e-commerce en permettant la spécification dynamique de contraintes par interactions entre Agents. Si les Agents interprètent de façon dynami que les messages qu'ils reçoivent, ils deviennent adaptables et, peuvent sans aucune intervention extérieure, communiquer avec des entités qu'ils n'ont jamais rencontrées auparavant. En outre, comme leur interpréteur est modifié pour acquérir une connaissance, il peut l'être aussi pour apprendre à enseigner une connaissance. Dans ce cas le transfert du savoir devient exponentiel au fur et à mesure des communications. Ce papier n'a pas simplement pour vocation de proposer un artéfact de plus de programmation à ajouter aux Agents, mais l'idée est plutôt de montrer une technique, faite de manière simple et utilisable, d 'évolution autonome des Agents dans une société. 9 Workshop on Reflection and Meta-level architecture, Tokyo, Novembre 92. [10] Oleg Kiselyov and Kirill Lisovsky. XML, XPath, XSLT implementations as SXML, SXPath, and SXSLT. International Lisp Conference: ILC2002, Octobre 2002. [11] Daniele Maraschi and Stefano A. Cerri. Relations entre les technologies de l’apprentissage humain et les agents. In Aniorté, P., Gouarderes, S., and al. (eds), Cognitique vers une informatique plus cognitive et sociale , pages 87–98, 2003. [12] John McCarthy. Elephant 2000: A programming language based on speech acts. Stanford University, 1989. [13] Kurt Normark. Simulation of object-oriented and mechanisms in scheme. Technical report, Institute of Electronic Systems, Aalborg University, Denmark, 1991. [14] David De Roure, Nicholas Jennings, and Nigel Shadbolt. Research agenda for the Semantic Grid : A future e-science infrastructure. Technical report, Report commissioned for EPSRC/DTI Core e-Science Programme , University of Southampton, UK, 2001. [15] John Searle. Les actes de langages, essai de ph i losoph ie du langage . Herman Editeur, Paris, 1971. Annexe Le modèle présenté ici a été sujet à une petite implémentation, non compète encore à ce jour, mais d'ores et déjà fonctionnelle. Elle est disponible en ligne sur http://www.lirmm.fr/~jonquet. Elle fut réalisée avec MIT Scheme 7.7.1, norme R5RS. Vous y les trouverez expérimentations citées ici. Remerciements Ce travail a été réalisé en grande partie pendant le stage de DEA d’un des auteurs (CJ). Le support du projet européen LeGE-WG (Learning Grid Excellence Working Group) est amplement remercié. Bibliographie [1] Harold Abelson, Gerald J. Sussman, and Julie Sussman. Structure o f Computer Interpretation and Programs.MIT Press, Cambridge, Massachusetts, 2nd edition, 1996. [2] John L. Austin. Quand dire c’est faire. Editions du Seuil, Paris, 1970. [3] Stefano A. Cerri. Cognitive environments in the STROBE model. Presented at EuroAIED : the European Conference in Artificial Intelligence and Educa t ion , 1996. [4] Stefano A. Cerri. Shifting the focus from control to communication: the STReams OBjects environments model of communicating agents. In In Padget, J.A. (ed.) Collaboration between Human and Artificial Soc ie ties, New York: Springer-Verlag, Lecture Notes in Artificial Intelligence, pages 74–101, Berlin, Heidelberg, 1999. [5] Stefano A. Cerri. Human an artificial agents conversations on the grid. In Electronic Workshops in Computing (eWic), 1 st LeGE-WG Internationa l Workshop on Educational Models for Grid Based Services, Lausanne, Switzerland, September 2002. [6] Stefano A. Cerri, Jean Sallantin, Emmanuel Castro, and Daniele Maraschi. Steps towards C+C: A language for interactions. In AIMSA2000, New York: Springer- Verlag, Lecture Notes in Artificial Intelligence, pages 34–48, Berlin, Heidelberg, 2000. [7] Frank Dignum and Mark Greaves. Issues in agent communication: An introduction. Dignum, F., and Greaves, M. (Eds), Agent Communication, LNAI 1916 , pages 1–16, 2000. [8] Jacques Ferber. Les Systemes Multi-Agents, vers une intelligence collective . InterEditions, Paris, 1995. [9] Stanley Jefferson and Daniel P. Friedman. A Simple Reflective Interpreter. In IMSA’92 International 10
1906.11064	1	1906	2019-06-26T13:00:39	Reasoning about Hypothetical Agent Behaviours and their Parameters	[ "cs.MA" ]	Agents can achieve effective interaction with previously unknown other agents by maintaining beliefs over a set of hypothetical behaviours, or types, that these agents may have. A current limitation in this method is that it does not recognise parameters within type specifications, because types are viewed as blackbox mappings from interaction histories to probability distributions over actions. In this work, we propose a general method which allows an agent to reason about both the relative likelihood of types and the values of any bounded continuous parameters within types. The method maintains individual parameter estimates for each type and selectively updates the estimates for some types after each observation. We propose different methods for the selection of types and the estimation of parameter values. The proposed methods are evaluated in detailed experiments, showing that updating the parameter estimates of a single type after each observation can be sufficient to achieve good performance.	cs.MA	cs	Reasoning about Hypothetical Agent Behaviours and their Parameters Stefano V. Albrecht Department of Computer Science The University of Texas at Austin Austin, TX 78712, USA [email protected] Peter Stone Department of Computer Science The University of Texas at Austin Austin, TX 78712, USA [email protected] 9 1 0 2 n u J 6 2 ] A M . s c [ 1 v 4 6 0 1 1 . 6 0 9 1 : v i X r a ABSTRACT Agents can achieve effective interaction with previously un- known other agents by maintaining beliefs over a set of hy- pothetical behaviours, or types, that these agents may have. A current limitation in this method is that it does not recog- nise parameters within type specifications, because types are viewed as blackbox mappings from interaction histories to probability distributions over actions. In this work, we pro- pose a general method which allows an agent to reason about both the relative likelihood of types and the values of any bounded continuous parameters within types. The method maintains individual parameter estimates for each type and selectively updates the estimates for some types after each ob- servation. We propose different methods for the selection of types and the estimation of parameter values. The proposed methods are evaluated in detailed experiments, showing that updating the parameter estimates of a single type after each observation can be sufficient to achieve good performance. CCS Concepts •Computing methodologies → Multi-agent systems; Intelligent agents; Planning under uncertainty; Co- operation and coordination; Keywords Ad hoc teamwork; Agent types; Parameter learning 1. INTRODUCTION An important open problem in multi-agent systems is the design of autonomous agents that can quickly and effectively interact with other agents when there is no opportunity for prior coordination, such as shared world models and com- munication protocols [2, 11, 37]. Several works addressed this problem by proposing methods which utilise beliefs over a set of hypothetical behaviours for the other agents [1,4,8,9,16,36]. Behaviours in this approach are specified as types, which are blackbox mappings from interaction histories to probability distributions over actions. If the types are sufficiently repre- sentative of the true behaviours of other agents, then this method can lead to rapid adaptation and effective interaction in the absence of explicit prior coordination [3, 8]. Appears in: Proceedings of the 16th International Confer- ence on Autonomous Agents and Multiagent Systems (AA- MAS 2017), S. Das, E. Durfee, K. Larson, M. Winikoff (eds.), May 8 -- 12, 2017, Sao Paulo, Brazil. Copyright c(cid:13) 2017, International Foundation for Autonomous Agents and Multiagent Systems (www.ifaamas.org). All rights reserved. There is, however, a current limitation in this type-based method, which is that it does not recognise parameters within types. Complex behaviours often involve various continuous parameters which govern certain aspects of the behaviour. For example, reinforcement learning methods often use learning, discounting, and exploration rates [38]. If we were to use such a method as a type, we would have to instantiate its parameters to some fixed values. Thus, an agent that wants to account for n different parameter settings will have to reason about n instances of the same type whose only difference is in their parameter values. This, however, is very inefficient as it leads to redundancy in space (storing n copies of the type) and time (computing the outputs of n copies). Our goal in this work is to devise a method which allows an agent to reason about both the relative likelihood of types and the values of their parameters. To be useful in practice, this reasoning should be efficient and allow for any bounded continuous parameters, without a need for the user to specify maximum likelihood estimators for the individual parameters. We show that the problem of space redundancy is typically unavoidable because the internal state of a type may depend on both the history of observations and the parameter values. Regarding the time requirements, due to the blackbox nature of types, the only way to ascertain the effect of a specific parameter setting is to evaluate the type with that parameter setting. Thus, our goal is to minimise the number of type evaluations while achieving a useful and robust estimate of the type's true parameter setting. We propose a general method which maintains individual parameter estimates for each type and selectively updates the estimates for some types after each observation. We propose different methods for the selection of types and the estimation of parameter values. The proposed methods are evaluated in the level-based foraging domain [3], where they achieved substantial improvements in task completion rates compared to random estimates, while updating only a single parameter estimate in each time step. 2. MODEL & OBJECTIVE We consider an interaction process with two or more agents. The process starts at time t = 0. At time t, each agent i receives a signal st i and independently chooses an action at i from some countable set of actions Ai. The signal st i may encode information about the state of the environment, a private reward, etc. We leave the precise structure and dynamics of st i open. This process continues indefinitely or until some termination criterion is satisfied. The probability with which action at iH t i , θi, p), where H t i = (s0 P (at i is chosen is given by i) is agent i's history of i , ..., st observations, θi is i's type, and p = (p1, ..., pn) is a vector of continuous parameters in θj. Each parameter pk takes a ] ⊂ R. To fixed value from some bounded interval [pmin simplify the exposition, we assume that all types have the same number of parameters, but in general this need not be the case. Which type θj a parameter vector p belongs to is disambiguated from context. , pmax k k j ∈ st i for t > 0. The true type of j, denoted θ∗ We control a single agent, i, which reasons about the be- haviour of another agent, j. We assume that i knows j's action space Aj and that it can observe j's past actions, i.e. at−1 j , and its true parameter values, p∗, are unknown to i. However, i has access to a finite set of hypothetical types θj ∈ Θj, with j ∈ Θj. We furthermore assume that i has all information θ∗ relevant to j's decision making, so that H t j is a function of H t i and we can write P (at The goal in this work is to devise a method which allows agent i to reason about the relative likelihood of types θj ∈ Θj and the values of their parameters p, based only on agent j's observed actions. i , θj, p). jH t 3. MARKOVIAN PARAMETERS Types are often implemented as Markov chains, such that j and j, θj, p). j is then incorporated into , usually by aggregating the information the choice of action depends only on the current signal st a current internal state wt The information contained in st the next state wt+1 within a collection of variables inside the state. j of the type, i.e. P (at jst j, wt j For types which are realised in this way, it is important to note that the internal state of the type may depend on both the history of observations and the parameter values. To illustrate this, consider a simple Q-learning agent [40] which uses three parameters, α, γ, ∈ [0, 1]. Its internal state is defined by a matrix, Q, which is used to compute and store expected payoffs for state-action pairs. This matrix is updated at each time step as Q(s, a) ← (1 − α)Q(s, a) + α r + γ max (cid:48) a(cid:48) Q(s (cid:48) ) , a (cid:20) (cid:21) where s, a is the previous state-action pair, r is some reward, and s(cid:48) is the new state. Given a state s, the agent chooses an action in arg maxa Q(s, a) with probability 1 − , and a random action otherwise. In this example, the values of Q depend on the history of observed states and rewards and the values of α, γ. This dependence on parameter values has important con- sequences for space requirements. Suppose we use the Q- learning agent as a type θj and fix its parameter setting to some values p. Its internal state wt j, defined by Q, will depend on past observations and p. Now, if we change the param- eter setting to p(cid:48) (cid:54)= p at some time t, we have a potential jst j, θj, p(cid:48)) may not be equal to inconsistency in that P (at j , θj, p(cid:48)), since wt P (at j has thus far been updated using p. Therefore, to ensure correct probabilities, we may have j to conform to the new parameter setting p(cid:48). In to adjust wt general, this can be done by recomputing the internal state "from the ground up" using the new parameter setting. How- ever, more efficient methods may be possible depending on how the internal states are influenced by parameters. jH t j, wt We adopt the naming convention and say that parameters p of type θj are Markovian if θj's action probabilities are independent of past values of p given their current values, i.e. i ) and parameter Algorithm 1 Selective parameter estimation in types Given: type space Θj, initial belief P (θjH 0 estimate p0 for each type θj ∈ Θj Repeat for each t > 0: 1: Select a subset Φt ⊂ Θj for parameter updates 2: For each θj ∈ Φt: 3: 4: 5: Set pt = pt−1 for all θj (cid:54)∈ Φt 6: For each θj ∈ Θj, update belief: H t−1 Obtain new parameter estimate pt for θj If pt non-Markovian, adjust internal state of θj , θj, pt)P (θjH t−1 i ) ∝ P (at−1 P (θjH t ) j i i P (ajH t j , θj, pt, pt−1, ..., p0) = P (ajH t j , θj, pt) (1) where pτ are the parameter values at time τ . Hence, the parameters in the Q-learning example (specifically α, γ) are not Markovian since they directly influence the values of Q. 4. LEARNING PARAMETERS IN TYPES We propose a method whereby agent i maintains individual parameter estimates for each hypothetical type θj ∈ Θj and selectively updates the estimates after each observation. The method starts with an initial belief P (θjH 0 i ) which specifies the relative likelihood (probability) that agent j has type θj. In addition, for each type θj ∈ Θj, it maintains an initial parameter estimate p0 within the respective value bounds. Then, at each time t > 0, the method selects a subset of types Φt ⊂ Θj and obtains a new parameter estimate pt for each θj ∈ Φt. (Sections 4.1 and 4.2 propose methods for each of these operations.) If the parameters of a type θj ∈ Φt are non-Markovian, then the internal state of θj may have to be adjusted to conform to the new parameter estimate (cf. Section 3). The parameter estimates of types not in Φt remain unchanged. Given the estimate pt for type θj, the current belief is updated via i ) ∝ P (at−1 , θj, pt)P (θjH t−1 P (θjH t H t−1 (2) ) j i i and the method continues in this fashion (cf. Algorithm 1). The use of point estimates of parameters effectively allows us to use Algorithm 1 as a pre-routine on top of an existing implementation A of the type-based method (e.g. [3, 10]). That is, at each time t > 0, we first execute Lines 1-5 to set the parameter values for each type, after which Line 6 executes A to update the belief and perform the planning step. From the perspective of A, there is formally no difference in the types since their parameters were set externally. i j , θ∗ H t−1 j , pt) = 0 while P (at−1 However, using point estimates can also cause a potential problem in our setting: it may generally be the case that P (at−1 j , p∗) > 0.1 The latter can cause P (θ∗ i ) to prematurely converge to zero, even though we may learn the correct parameter values p∗ at a later time. To prevent this, we assume that for any θj ∈ Θj, if P (at−1 , θj, p) is positive for some p, then it is positive for all p. In practice, this can be ensured by using close-to-zero probabilities instead of zero probabilities. H t−1 H t−1 jH t , θ∗ j j i i 1As an example, consider the Q-learning agent from Section 3 and set t = 0, ∗ = .5, and at−1 (cid:54)∈ arg maxa Q(s, a). j 4.1 Selecting Types for Parameter Updates Since we do not know which type in Θj is the true type θ∗ j , the safe choice of Φt is to update the parameter estimates of all types in Θj. However, this is also the most costly choice in terms of computation time. On the other hand, we may minimise computation costs by updating parameter estimates only for some subset Φt ⊂ Θj, but this carries the risk that θ∗ j may not be included in Φt. In this sense, we view the choice of Φt as a decision problem which balances exploitation (i.e. choosing types which are in some sense expected to benefit the most from an update) and exploration. We propose two approaches to make this choice, which entail different notions of exploitation, exploration, and risk. 4.1.1 Posterior Selection The first approach is to select types which are believed to be most likely, with the expectation that one of them is the true type. Here, exploitation amounts to choosing types θj ∈ Θj which have maximum probability P (θjH t−1 ). However, depending on the observation history H t−1 i and parameter estimates p0, ..., pt−1, there is a risk that P (θjH t−1 ) assigns high probability to incorrect types θj (cid:54)= θ∗ j . This can lead to premature convergence of beliefs to incorrect types if we do not update the parameter estimates of the true type θ∗ j . Thus, exploration in this approach means choosing types which currently seem less likely than other types. To balance exploitation and exploration, we propose to sample Φt from the belief P (θjH t−1 ). i i i 4.1.2 Bandit Selection The second approach is to select types according to their expected change in parameter estimates after the new obser- vation is accounted for. This is predicated on the assumption that parameter estimates will converge, so that exploitation entails selecting types which are expected to make the largest leaps toward convergence. The risk is that the parameter estimates for some types, including the true type θ∗ j , may not change significantly until certain observations are made. Hence, exploration entails choosing types even if their pa- rameter estimates are not expected to change much. To balance exploitation and exploration, we can frame this approach as a multi-armed bandit problem [34]. In the general setting, there are k arms to choose from at each time step t, and each choice results in a reward rt drawn from an unknown distribution associated with the chosen arm. The goal is to choose arms so as to maximise the sum of rewards. In our case, the arms represent the types in Θj and we define the reward rt after updating the parameter estimate of type θj as the normalised L1 norm pt k − pt−1 k − pmin pmax n(cid:88) n(cid:88) rt = η η = (3) −1 , k . k k=1 k=1 Thus, rewards are in the range [0, 1], where a reward of 0 means no change in the parameter estimate and a reward of 1 represents maximum change. Several algorithms exist which solve this problem, subject to different assumptions regarding the distribution of rewards (e.g. [6,25]). In our case, the reward distributions of arms are independent but possibly changing over time (e.g. if estimates converge). Therefore, one should also consider algorithms designed for changing reward distributions (e.g. [7, 17]). Algorithm 2 Approximate Gradient Ascent Given: parameter estimate pt−1, degree d 1: Collect samples D = (p(l), f (p(l))) // e.g. uniform grid 2: Fit polynomial f of degree d to D 3: Compute gradient ∇ f (pt−1) and step size λt 4: Update estimate pt = pt−1 + λt ∇ f (pt−1) i , θj), represented as polynomial of deg. d Algorithm 3 Approximate Bayesian Updating Given: P (pH t−1 1: Fit f to f as in Algorithm 2 2: Compute polynomial product g = f · P (pH t−1 3: Collect samples D = (p(l), g(p(l))) // e.g. uniform grid 4: Fit new polynomial h of degree d to D 5: Compute integral I =(cid:82) pmax h(p) dp , θj) i 6: Set new belief P (pH t i , θj) = h/I 7: Extract estimate pt from P (pH t pmin i , θj) // e.g. sample 4.2 Estimating Parameter Values We propose three different methods for the estimation of parameter values pt of a type θj. For notational convenience, we define f (p) = P (at−1 . H t−1 , θj, p). j i 4.2.1 Approximate Gradient Ascent The idea in this method is to update parameter estimates by following the gradient of a type's action probabilities with respect to the parameter values. Formally, the estimate is updated as pt = pt−1 + λt ∇f (pt−1), where ∇f denotes the gradient of f and λt is some suitably chosen step size (e.g. constant or optimised via line search). This requires a representation of f which is differentiable in p and flexible enough to allow for a variety of shapes, including skewness and multi-modality. We can obtain such a representation by approximating f as a polynomial f of some specified degree d, fitted to a suitable set of samples (p(l), f (p(l))). For exam- ple, one could use a uniform grid over the parameter space that includes the boundary points. Algorithm 2 provides a summary of this method. We note that operations such as fitting and differentiation of multivariate polynomials can be costly, even in the ap- proximate case [19], whereas univariate polynomials can be processed very efficiently. To alleviate this, one may parti- tion parameters p1, ..., pn into clusters C1, C2, ... according to their degree of correlation in f (so that parameters from different clusters are independent or only weakly correlated; cf. [5]) and use separate polynomials for each cluster. If the resulting clusters are small, this can significantly reduce com- putational costs [12, 30]. However, care must be taken not to break important correlations between parameters, which may degrade the accuracy of parameter estimates. 4.2.2 Approximate Bayesian Updating Rather than using f to perform gradient-based updates, we can use f to perform Bayesian updates that retain infor- mation from past updates. In addition to the belief P (θjH t i ), agent i now also has a belief P (pH t i , θj) to quantify the rela- tive likelihood of parameter values p for θj. This new belief is represented as a polynomial of the same degree d as f . The (a) Prior belief (b) f (likelihood) (c) Posterior belief Figure 1: Approximate Bayesian Updating for a single parameter p ∈ [0, 1] with true value p∗ = 0.11. The polynomials have degree 4 and are fitted using 5 uniformly spaced points from the parameter space. i Bayesian update is then constructed as follows: After fitting f , we take the convolution (i.e. polynomial product) of P (pH t−1 , θj) and f , resulting in a polynomial g of degree greater than d. To restore the original representa- tion, we fit a new polynomial h of degree d to any suitably chosen set of sample points from the convolution g. Again, we could use a uniform discretisation of the parameter space. Finally, we compute the integral of h under the parameter space and divide the coefficients of h by the integral, to ob- tain the new belief P (pH t i , θj). This new belief can then be used to obtain a parameter estimate, e.g. by finding the max- imum of the polynomial or by sampling from the polynomial. Algorithm 3 provides a summary of this process and Figure 1 gives a graphical example. While the use of polynomials allows for great flexibility, it does not come without limitations: Polynomials suffer from known instability issues in extrapolation and interpolation. Extrapolation is not of concern here since we are confined to bounded parameter spaces. However, instability of interpola- tion can lead to negative values between fitted samples (cf. Figure 1b). While this poses no difficulty for the calculation of maxima and sampling, it does mean that the integral in the normalisation of h has to be "absolute", in that any area below the zero axis is assigned a positive sign. Moreover, due to the nature of approximate fitting and finite machine ac- curacy, care should be taken that the samples taken from g to construct h (cf. Figure 1c) are not negative in g, as oth- erwise negative minima may be propagated across updates, which can lead to further instabilities. 4.2.3 Exact Global Optimisation The previous methods rely on an approximation f of f to perform successive updates. An alternative approach is to reason directly with f . In addition to avoiding the potential inaccuracies caused by the approximations, this would allow for the detection of possible discontinuities in f (p) which cannot be represented by continuous polynomials. Specifically, the estimation of parameter values can be viewed as a global optimisation problem [22] in which the goal is to find a parameter setting pt with maximum prob- ability over the history of observations H t i . Formally, the optimisation problem is defined as follows: Find pt ∈ arg max p F (p) = P (aτ−1 j H τ−1 i , θj, p) (4) t(cid:89) τ =1 s.t. ∀k pk ∈ [pmin k , pmax k ] Since the evaluation of the objective function F for a given parameter setting p can be relatively costly, one would ideally solve this problem using an optimisation method that seeks to minimise the number of evaluations. Bayesian optimisation was specifically designed for such settings and has been shown to be effective for low-dimensional problems [29]. The idea is to use a Gaussian process [33] to represent uncertainty over the values of F . Each iteration of the method selects a new point p to evaluate, according to some tradeoff criterion for exploitation (choosing points which are expected to have high values) and exploration (minimising uncertainty). A crucial choice in this method is the form of the covariance function, which is used to measure similarity of points [35]. 5. EXPERIMENTAL EVALUATION We provide a detailed experimental evaluation of our meth- ods in the level-based foraging domain [3], which was intro- duced as a test domain for ad hoc teamwork [37]. 5.1 Domain Description The domain consists of a rectangular grid in which a team of agents must collaborate to collect a number of items in minimal time. The agents' ability to collect items is limited by skill levels: each agent and item has an individual level which is represented by a number in the range [0, 1]. A group of agents can collect an item if (i) they are located next to the item, (ii) they simultaneously choose the load action, and (iii) the sum of the agents' levels is at least as high as the item's level. Thus, in Figure 2, the two agents in the left half can jointly collect an item which individually they cannot collect. When an item is collected, it is removed from the grid and the team receives a reward of 1; in all other cases, the reward is 0 (timing will become relevant via a discount factor). In addition to the load action, each agent has 4 actions N, E, S, W, which move the agent into the corresponding direction if the target cell is empty and inside the grid. Ties are resolved by executing actions in random order. To enforce collaboration and keep this solvable, skill levels are chosen such that all agents have levels below the highest item level, and no item has a level greater than the sum of all agent levels. Furthermore, items are placed such that the Euclidean distance between each item is greater than 1, and no item is placed at any border of the grid. We extend this domain by adding view cones for agents, which are parameterised by a radius and angle. An agent's view cone determines which items and other agents it can see, as well as the certainty with which they are seen. The 00.20.40.60.81p00.20.40.60.81Belief densityP(pHt−1i,θj)p∗00.20.40.60.81p00.0020.0040.0060.0080.01f(p)foraging.H (true type)SamplesfromfFittedf00.20.40.60.81p00.20.40.60.81Belief densitySamplesfromgFittedhP(pHti,θj)p∗00.20.40.60.81p0123456Belief densityP(pHt−1i,θj)p∗00.20.40.60.81p-0.500.511.5f(p)SamplesfromfFittedf00.20.40.60.81p-101234Belief densitySamplesfromgFittedhP(pHti,θj)p∗00.20.40.60.81p00.511.52Belief densityP(pHt−1i,θj)p∗00.20.40.60.81p00.20.40.60.811.2f(p)SamplesfromfFittedf00.20.40.60.81p-10123Belief densitySamplesfromgFittedhP(pHti,θj)p∗00.20.40.60.81p00.20.40.60.81Belief densityP(pHt−1i,θj)p+00.20.40.60.81p00.0020.0040.0060.0080.01f(p)foraging.H (true type)SamplesfromfFittedf00.20.40.60.81p00.20.40.60.81Belief densitySamplesfromgFittedhP(pHti,θj)p+00.20.40.60.81p0123456Belief densityP(pHt−1i,θj)p+00.20.40.60.81p-0.500.511.5f(p)SamplesfromfFittedf00.20.40.60.81p-101234Belief densitySamplesfromgFittedhP(pHti,θj)p+00.20.40.60.81p00.511.52Belief densityP(pHt−1i,θj)p+00.20.40.60.81p00.20.40.60.811.2f(p)SamplesfromfFittedf00.20.40.60.81p-10123Belief densitySamplesfromgFittedhP(pHti,θj)p+00.20.40.60.81p00.20.40.60.81Belief densityP(pHt−1i,θj)p∗00.20.40.60.81p00.0020.0040.0060.0080.01f(p)foraging.H (true type)SamplesfromfFittedf00.20.40.60.81p00.20.40.60.81Belief densitySamplesfromgFittedhP(pHti,θj)p∗00.20.40.60.81p0123456Belief densityP(pHt−1i,θj)p∗00.20.40.60.81p-0.500.511.5f(p)SamplesfromfFittedf00.20.40.60.81p-101234Belief densitySamplesfromgFittedhP(pHti,θj)p∗00.20.40.60.81p00.511.52Belief densityP(pHt−1i,θj)p∗00.20.40.60.81p00.20.40.60.811.2f(p)SamplesfromfFittedf00.20.40.60.81p-10123Belief densitySamplesfromgFittedhP(pHti,θj)p∗if Mem (cid:54)= ∅ and Loc (cid:54)= Mem then Dest ← Mem (A, I) ← VisibleAgentsAndItems(Loc) Targ ← ChooseTarget(A, I) if Targ (cid:54)= ∅ then Algorithm 4 Template for foraging types Parameters: skill level p1, view radius p2, view angle p3 Initialise: destination memory Mem ← ∅ Repeat for each t: 1: // Select destination 2: Loc ← own location, Dest ← ∅ 3: 4: 5: else 6: 7: 8: 9: 10: Save destination in memory: Mem ← Dest 11: // Assign action probabilities 12: 13: 14: else 15: 16: 17: 18: 19: 20: Add probability 0.01 to each action and normalise Use A∗ [21] to find path from Loc to Dest Assign probability 1 to first move action in path if Dest is item and Loc is next to Dest then Assign probability 0.25 to each move action Assign probability 1 to load action if Dest = ∅ then Dest ← Targ else The internal state of the template is defined by a memory Mem for the current destination (x/y position) which the agent is trying to reach. Once the destination in Mem has been reached, the template chooses a new destination using the ChooseTarget routine. Thus, the contents of Mem is directly affected by the parameters, and we can classify them as non-Markovian (cf. Section 3). Finally, Line 20 in Algorithm 4 is a simple way of guaran- teeing the assumption that the set of actions with positive probability is invariant of the parameter values (see last para- graph before Section 4.1). 5.3 Experimental Setup We tested various configurations of Algorithm 1. For the selection of types for parameter updates (Φt), we tested updating all types in Θj, sampling a single type from Θj using the belief P (θjH t−1 ) (Section 4.1.1), and sampling a single type from Θj using a bandit algorithm (Section 4.1.2). A number of bandit algorithms were tried in preliminary experiments, including UCB1 [6], EEE [17], S [25], Exp3 [7], and Thompson sampling [39]. All reported results are based on UCB1, which achieved the best performance. i For the estimation of parameter values, we tested Approx- imate Gradient Ascent (AGA), Approximate Bayesian Up- dating (ABU), and Exact Global Optimisation (EGO). AGA and ABU used univariate polynomials of degree 4 for each parameter, which were fitted using 5 uniformly spaced points over the parameter space (as shown in Figure 1). AGA op- timised the step size λt in each update using backtracking line search (with the search parameters set to 0.5/0.5). ABU used uniform initial beliefs P (pH 0 i , θj) for each type θj ∈ Θj and generated parameter estimates by averaging over 10 sam- ples taken from P (pH t i , θj) (which we found to be more robust than taking the maximum). EGO was implemented Figure 2: Level-based foraging domain. Agents are marked by circles (blue is our agent) and items are marked by grey squares. Skill levels are shown inside agents and items. The dashed magenta lines show the other agents' view cones. latter is calculated as the percentage (measured in [0, 1]) with which the view cone overlaps with the grid cell occupied by an agent or item. Thus, the agent in the right half of Figure 2 can see two items, one with certainty 1 and another one with certainty ≈ 0.85. We assume that our agent can see the entire grid (cf. Section 2), hence it has no view cone. 5.2 Hypothetical Type Space The hypothetical type space Θj consists of four types which are all based on the template given in Algorithm 4. The tem- plate uses three parameters: p1 ∈ [0, 1] specifies the agent's skill level; p2 ∈ [.1, 1] specifies the agent's view radius as √ w2 + h2, where w and h are the width and height of the p2 grid; and p3 ∈ [.1, 1] specifies the view angle as p32π. The parameters p2, p3 are used in the VisibleAgentsAndItems routine, which returns two sets containing the visible agents and items with a view certainty of 0.1 or higher. The param- eter p1 is used in the ChooseTarget routine, which returns a specific target out of the visible agents and items. The four types in Θj differ from each other in their speci- j fication of the ChooseTarget routine: • θL1 • θL2 : if items visible, return furthest2 one; else, return ∅ : if items visible, return item with highest level below j own level, or item with highest level if none are below own level; else, return ∅ • θF 1 • θF 2 : if agents visible but no items visible, return furthest j agent; if agents and items visible, return item that furthest agent would choose if it had type θL1 ; else, return ∅ j : if agents visible but no items visible, return agent with j highest level above own level, or furthest agent if none are above own level; if agents and items visible, select agent as before and return item that this agent would choose if it had type θL2 ; else, return ∅ j j j and θL2 Intuitively, types θL1 can be viewed as leaders: they choose targets on their own and expect others to follow their lead. Conversely, types θF 1 can be viewed as followers: they assume other agents know best and attempt to follow their lead. The leader and follower types are further distinguished by whether they consider skill levels. and θF 2 j j 2We found that choosing the furthest item/agent penalises wrong parameter estimates more than choosing closest ones, since the latter is invariant to overestimation of view cone parameters. .48.20.12.24.20.86.60.91.68.43.11.60.39using Bayesian optimisation with the "expected improvement" search criterion [29] and squared exponential covariance with automatic relevance detection [33]. The number of points evaluated by EGO (cf. (4)) was limited to 10. All configurations used uniform initial beliefs P (θjH 0 i ) over the set Θj (specified in Section 5.2) and random initial parameter estimates for each θj ∈ Θj. In each time step, Monte Carlo Tree Search (MCTS), specifically UCT [26], was used to compute optimal actions with respect to the beliefs and types. Each rollout in the tree search used the current belief P (θjH t i ) to sample a type θj ∈ Θj which was used for the entire rollout. Each time step generated 300/500 rollouts in the 10x10/15x15 worlds, respectively (see below), which we found to be robust numbers. Each rollout was over a horizon of 100 time steps, and the rewards accumulated during a rollout were discounted with a factor of 0.95. Subtrees from previous time steps were reused to accelerate the tree search. The configurations were tested in two different sizes of the level-based foraging domain: a 10x10 world with 2 agents and 5 items, and a 15x15 world with 3 agents and 10 items (so our agent reasons about the types and parameters of two other agents). Each configuration was tested in the same sequence of 500 instances, which were generated as follows: First, we set random initial positions and skill levels for each agent and item, subject to the constraints noted in Section 5.1. Then, for each agent not under our control, we randomly selected its true type θ∗ j from the type space Θj and completed its parameter setting by choosing random values for the view cone parameters. Finally, for each θj ∈ Θj, we sampled random initial parameter estimates which were used by the tested configuration. Instances of the 10x10/15x15 world were run for a maximum of 100/150 time steps, respectively. We used two baselines to facilitate the comparison of our methods: Rnd, which used fixed random parameter values for each type, and Cor, which used the correct parameter values for the true type and fixed random parameter values for all other types (baselines did not update parameters). 5.4 Results Figure 3 shows the average number of time steps and the completion rates for each of the tested configurations and world sizes. The completion rate is the percentage of instances which were completed successfully (i.e. all items collected) within the given amount of time. The average time steps are for completed instances. To put the results into perspective, we will begin by discussing the results of the two baselines, Cor and Rnd. (In the following, all significance statements are based on paired t-tests with a 5% significance threshold.) The first observation is that there was only a small dif- ference between Cor and Rnd in their average number of time steps for completed instances, with margins of less than 10 time steps in both world sizes. This may seem surpris- ing, given that the random parameter settings used by Rnd can lead to significantly different predictions than the cor- rect settings. However, in instances which were completed by both baselines, we found that the MCTS planner was robust enough to "absorb" the differences, in that it often produced similar courses of actions despite the differences. On the other hand, there were substantial differences in the comple- tion rates of Cor and Rnd, dropping from 98% to 71% in the 10x10 world and 79% to 41% in the 15x15 world, respec- tively. We found that the random parameter settings used by Rnd often led to predictions that fooled Rnd into taking (a) 10x10 world, 2 agents, 5 items (b) 15x15 world, 3 agents, 10 items Figure 3: Time steps required in completed instances (means and standard deviations) and completion rates for the tested methods. Results are averaged over 500 instances in each world. Dashed lines mark the baseline performances, where Cor had lowest time steps and highest completion rates. the wrong actions without ever realising it, thus inducing an infinite cycle which the agent never escaped. This effect has been described previously as "critical type spaces" [4]. Given the means and standard deviations of time steps shown in Figure 3, one can see that simply increasing the maximum allowed time steps per instance would not significantly affect Rnd's ability to complete instances. 4 th, since Θj = 4) of the computation time. We now turn to a comparison of our proposed methods. Most notably, the results show that updating a single type in each time step achieved comparable performance to up- dating all types in each time step, albeit at only a fraction (approximately 1 Moreover, bandit selection significantly outperformed poste- rior selection in all tested configurations, except for EGO in the 10x10 world, where the two were equivalent. We found that this difference was due to the fact that posterior selec- tion tended to exploit more greedily than bandit selection, because the beliefs P (θjH t i ) often placed high probability on certain types early on in the interaction. In contrast, bandit selection was more exploratory because the rewards defined in Section 4.1.2 tended to be more uniform across types than beliefs. Given that the distributions underlying these rewards were not stationary, it is worth pointing out that bandit algorithms which were specifically designed for changing distributions (e.g. [7, 17]) did not perform better than those which assume stationary reward distributions.3 These results show that our approach of viewing the selec- tion of types as a decision problem, balancing exploitation and exploration, can be effective in practice. Regarding the different estimation methods, the results show a gradual improvement from AGA to ABU to EGO. AGA performed worst because the gradient update used in AGA did not retain information from past updates. Thus, its 3The analysis in [26] provides some insights into the performance of UCB1 for non-stationary ("drifting") reward distributions. AGAABUEGO20406080100Time stepsAllPostBanditCor/RndAGAABUEGO5060708090100Completion rateAGAABUEGO80100120140Time stepsAllPostBanditCor/RndAGAABUEGO304050607080Completion rate-1-0.500.51-1-0.500.51AllPostBanditCor/Rnd-1-0.500.51-1-0.500.51AllPostBanditCor/RndFigure 4: Average seconds (log-scale) needed per parameter update for one type. Measured in Matlab R2015b on a UNIX dual-core machine with 2.66 GHz per core. (a) Parameter p1 (skill level) estimates were dominated by the most recent observations, which often prevented convergence to good estimates. In ad- dition, AGA and ABU's ability to estimate parameters was hindered by the fact that they used individual polynomials for the parameters, thus ignoring possible parameter correla- tions at the benefit of reduced computation time. EGO, due to its ability to detect parameter correlations and discontinu- ities, achieved the best performance in our experiments. We note that the results shown for EGO are for a maximum of 10 evaluated points. We were able to drive its performance up by increasing the number of evaluated points, approach- ing the performance of the Cor baseline in both worlds. However, this performance came at a significant cost in com- putation time (cf. Figure 4): while AGA and ABU needed on average about 0.03 and 0.05 seconds per update, EGO needed about 1 (2.3) seconds per update when evaluating 10 (20) points, which increased slowly for longer histories. Thus, ABU provided the best tradeoff between task comple- tion and computation time. However, the time requirements of EGO may be reduced drastically by using a more efficient implementation of Bayesian optimisation, e.g. [28]. Figure 5 shows the mean error in the parameter estimates for the true type θ∗ j . The figure shows that AGA's estima- tion errors increased slowly over time. One reason for this was that f (i.e. the action probabilities of types with respect to parameters; cf. Section 4.2) was often multi-modal and hence non-convex, causing the gradient to point away from the true parameter values. Another reason was that f could change drastically between time steps, which in some cases had a similar "trapping" effect on the gradient. Nonethe- less, AGA still managed to produce good estimates in some of the instances. A different picture is shown for ABU: its mean errors dropped substantially after the first time step and remained stable after. This shows that ABU was able to effectively retain information from past updates, through its conjugate polynomial update. While EGO did also re- tain information from past observations, its estimates were less stable than ABU's estimates, often jumping radically between different values. This was a result of the search strat- egy used in Bayesian optimisation and the fact that it only evaluated 10 points in each update, which can cause it to find different solutions after each new observation. An in- teresting observation is that EGO seemed to differentiate between parameters, with substantially different mean errors for the individual parameters. This, too, was a result of its search strategy, which can concentrate on certain parame- ters if they lead to better solutions. Thus, p1 (the skill level) seemed to be less relevant than p2/p3 (the view cone param- eters). Given that ABU's mean error was substantially lower than EGO's mean error, it may be surprising that EGO still (b) Parameter p2 (view radius) (c) Parameter p3 (view angle) Figure 5: Mean error in parameter estimates for the true type θ∗ j in the 15x15 world (updating all types in each time step), averaged over 500 instances and both other agents. The error at time t is defined as the absolute difference p∗ k. Errors are shown for the first 15 and last time steps of an instance. k − pt outperformed ABU in completion rates. However, a closer inspection showed that EGO more often estimated the right combination of parameter values (i.e. it recognised correla- tions in parameters), which in many cases was crucial for the correct planning of actions. Finally, Figure 6 shows the evolution of beliefs in the 10x10 world (the same picture was obtained in the 15x15 world). The correct baseline Cor had a robust convergence to the true type, with an average final probability of 0.975 for the true type. In contrast, the random baseline Rnd converged in many cases to an incorrect type, with an average final probability of 0.314 for the true type. The corresponding probabilities produced by our methods were 0.313 for AGA, 0.401 for ABU, and 0.482 (0.574) for EGO with 10 (50) eval- uated points. Thus, AGA did not improve belief convergence over Rnd while ABU and EGO produced statistically sig- nificant improvements, albeit still a long way from Cor. By the end of an instance, all methods placed most of their be- lief mass on one type, with average maximum probabilities (over any type) in the 0.9x range. These numbers show that parameter estimates that deviate from the true values can have a significant impact on the evolution of beliefs. As our data show, convergence to the true type correlates with (and causes) higher completion rates. AGAABUEGO0.010.11Seconds (log)01234567891011121314endTime step0.200.250.300.350.40Mean errorAGAABUEGO01234567891011121314endTime step0.200.250.300.350.40Mean errorAGAABUEGO01234567891011121314endTime step0.200.250.300.350.40Mean errorAGAABUEGOmethods proposed in [15,36] are specific to parameters of the used distributions (e.g. Dirichlet). Classical methods for opponent modelling assume a fixed model structure (e.g. a decision tree or finite-state machine) and attempt to fit the model parameters based on observed actions (e.g. [10, 13, 27]). Because such models may involve many parameters, the learning process may need many obser- vations to produce useful fits. This is in contrast to type-based methods, in which types are blackbox functions and we only "fit" one probability for each type. The latter can lead to rapid adaptation, but may not be as flexible as classical methods. Here, too, our work can be viewed as a hybrid between the two approaches: in addition to fitting probabilities over types we now also fit parameters within types, giving them greater flexibility, but the number of such parameters is usually lower than that found in classical methods. Our proposed method is in part inspired by methods of selective inference in dynamic Bayesian networks [5]. In our work, we selectively choose types whose parameter values we wish to infer. However, the selection of types is viewed as a decision problem whereas the selective inference in [5] is predetermined by the structure of the network. 6.3 Conclusion & Outlook This work extends the type-based interaction method by allowing an agent to reason about both the relative likeli- hood of types and the values of any bounded continuous parameters within types. A key element in our approach to minimise computation costs is to perform selective updates of the types' parameter estimates after new observations are made. Moreover, our proposed methods for the estimation of parameter settings can be applied to any continuous param- eters in types, without requiring additional structure in type specifications. We evaluated our methods in detailed experi- ments, showing that they achieved substantial improvements in task completion rates compared to random estimates, while updating only a single parameter estimate in each time step. There are several potential directions for future research. Our experiments showed that parameter estimates can have a significant effect on the evolution of beliefs over types. However, we do not currently have a formal theory that char- acterises the interaction between parameter estimates and beliefs. Such a theory might have useful implications for the selection of types and the derivation of estimates. Further- more, our methods assume that we can observe (or derive) the chosen actions and observations of other agents. A use- ful generalisation of our work would be to also account for possible uncertainties in such observations, e.g. [32]. Finally, further enhancements of our methods could be made. For in- stance, another approach to select types for updates might be to estimate the impact that updating a particular type may have on our beliefs and future actions. However, such meth- ods can be computationally expensive, even in the myopic approximate case [15]. Acknowledgements: This work took place in the Learning Agents Research Group (LARG) at UT Austin. LARG research is sup- ported in part by NSF (CNS-1330072, CNS-1305287, IIS-1637736, IIS-1651089), ONR (21C184-01), AFOSR (FA9550-14-1-0087), Raytheon, Toyota, AT&T, and Lockheed Martin. Peter Stone serves on the Board of Directors of, Cogitai, Inc. The terms of this arrangement have been reviewed and approved by The University of Texas at Austin in accordance with its policy on objectivity in re- search. Stefano Albrecht is supported by a Feodor Lynen Research Fellowship from the Alexander von Humboldt Foundation. Figure 6: Average belief P (θ∗ j in the 10x10 world (updating all types in each time step). Probabil- ities are averaged over 500 instances and shown for the first 10 and last time steps of an instance. i ) for the true type θ∗ jH t 6. DISCUSSION 6.1 A Note on Belief Merging A central feature of keeping beliefs over a set of behaviours is a property called belief merging [23]. Under a condition of "absolute continuity", this property entails that the believed distribution over future play converges in a strong sense to the true distribution induced by the true behaviour. One may ask if this property also holds in our method, given that (2) may use different parameter estimates in each update. The simple answer to this question is no, because changing the parameter estimates means that the beliefs effectively refer to a different type space in the original result [23]. Would a method that uses distributions over parameter values rather than point estimates inherit the belief merging property? It can be shown that the answer here, too, is negative, and we provide an example below (we assume basic familiarity with the work of Kalai and Lehrer [23]): Suppose agent j can choose between two actions. Its true type, θ∗ j , is to choose action 1 with probability δ and action 2 with probability 1− δ. Assume that agent i knows θ∗ j but not the value of the δ parameter, and so maintains a continuous distribution over the interval [0, 1]. The probability measures µ and µ(cid:48) over play paths are induced in the usual way [23] from the true type and the distribution, respectively. Now, consider the set Ω consisting of all infinite play paths in which action 1 has limit frequency δ. We have µ(Ω) = 1, since θ∗ can only realise paths in Ω, but µ(cid:48)(Ω) = 0 due to the diffused distribution over δ. Thus, the absolute continuity condition is violated and belief merging does not materialise (absolute continuity is in fact necessary for belief merging [24]). j Nonetheless, it has been argued that absolute continuity and the resulting convergence (which implies accurate predic- tion of infinite play paths [23]) are too strong for practical applications [18, 24, 31]. It is easy to see that the ABU and EGO methods described in Section 4.2 would converge point- wise to the correct parameter value in the above example. 6.2 Related Work Several works proposed methods which maintain Bayesian beliefs over a set of possible behaviours or types [1, 9, 14, 15, 20, 36]. Some methods assume discrete (usually finite) type spaces [3, 9, 14] while others assume continuous type spaces [15,36]. Our work can be viewed as bridging these methods by doing both: we maintain beliefs over a finite set of types, and we allow each type to have continuous parameters. Moreover, our methods can deal with any parameterisation, while the 0123456789endTime step0.20.40.60.81ProbabilityAGAABUEGOCorRndREFERENCES [1] S. Albrecht, J. Crandall, and S. Ramamoorthy. Belief and truth in hypothesised behaviours. Artificial Intelligence, 235:63 -- 94, 2016. [2] S. Albrecht, S. Liemhetcharat, and P. Stone. Special issue on multiagent interaction without prior coordination: Guest editorial. Autonomous Agents and Multi-Agent Systems, 2016. [3] S. Albrecht and S. Ramamoorthy. A game-theoretic model and best-response learning method for ad hoc coordination in multiagent systems. Technical report, School of Informatics, The University of Edinburgh, 2013. [4] S. Albrecht and S. Ramamoorthy. On convergence and optimality of best-response learning with policy types in multiagent systems. In Proceedings of the 30th Conference on Uncertainty in Artificial Intelligence, pages 12 -- 21, 2014. [5] S. Albrecht and S. Ramamoorthy. Exploiting causality for selective belief filtering in dynamic Bayesian networks. Journal of Artificial Intelligence Research, 55:1135 -- 1178, 2016. [6] P. Auer, N. Cesa-Bianchi, and P. Fischer. Finite-time analysis of the multiarmed bandit problem. Machine Learning, 47(2-3):235 -- 256, 2002. [7] P. Auer, N. Cesa-Bianchi, Y. Freund, and R. Schapire. Gambling in a rigged casino: The adversarial multi-armed bandit problem. In Proceedings of the 36th Symposium on the Foundations of Computer Science, pages 322 -- 331, 1995. [8] S. Barrett and P. Stone. Cooperating with unknown teammates in complex domains: a robot soccer case study of ad hoc teamwork. In Proceedings of the 29th AAAI Conference on Artificial Intelligence, pages 2010 -- 2016, 2015. [9] S. Barrett, P. Stone, and S. Kraus. Empirical evaluation of ad hoc teamwork in the pursuit domain. In Proceedings of the 10th International Conference on Autonomous Agents and Multiagent Systems, pages 567 -- 574, 2011. [10] S. Barrett, P. Stone, S. Kraus, and A. Rosenfeld. Teamwork with limited knowledge of teammates. In Proceedings of the 27th AAAI Conference on Artificial Intelligence, pages 102 -- 108, 2013. [11] M. Bowling and P. McCracken. Coordination and adaptation in impromptu teams. In Proceedings of the 20th National Conference on Artificial Intelligence, pages 53 -- 58, 2005. [12] X. Boyen and D. Koller. Tractable inference for complex stochastic processes. In Proceedings of the 14th Conference on Uncertainty in Artificial Intelligence, pages 33 -- 42, 1998. [13] D. Carmel and S. Markovitch. Learning models of intelligent agents. In Proceedings of the 13th National Conference on Artificial Intelligence, pages 62 -- 67, 1996. [14] D. Carmel and S. Markovitch. Exploration strategies for model-based learning in multi-agent systems. Autonomous Agents and Multi-Agent Systems, 2(2):141 -- 172, 1999. [15] G. Chalkiadakis and C. Boutilier. Coordination in multiagent reinforcement learning: a Bayesian approach. In Proceedings of the 2nd International Conference on Autonomous Agents and Multiagent Systems, pages 709 -- 716, 2003. [16] M. Chandrasekaran, P. Doshi, Y. Zeng, and Y. Chen. Team behavior in interactive dynamic influence diagrams with applications to ad hoc teams. In Proceedings of the 13th International Conference on Autonomous Agents and Multiagent Systems, pages 1559 -- 1560, 2014. [17] D. de Farias and N. Megiddo. Exploration-exploitation tradeoffs for experts algorithms in reactive environments. In Advances in Neural Information Processing Systems 17, pages 409 -- 416, 2004. [18] P. Doshi and P. Gmytrasiewicz. On the difficulty of achieving equilibrium in interactive POMDPs. In Proceedings of the 21st National Conference on Artificial Intelligence, pages 1131 -- 1136, 2006. [19] B. Fu. Multivariate polynomial integration and differentiation are polynomial time inapproximable unless P = NP. In Lecture Notes in Computer Science, volume 7285, pages 182 -- 191. Springer, 2012. [20] P. Gmytrasiewicz and P. Doshi. A framework for sequential planning in multiagent settings. Journal of Artificial Intelligence Research, 24(1):49 -- 79, 2005. [21] P. Hart, N. Nilsson, and B. Raphael. A formal basis for the heuristic determination of minimum cost paths. In IEEE Transactions on Systems Science and Cybernetics, volume 4, pages 100 -- 107, July 1968. [22] R. Horst, P. Pardalos, and N. Thoai. Introduction to Global Optimization. Kluwer Academic Publishers, 2000. [23] E. Kalai and E. Lehrer. Rational learning leads to Nash equilibrium. Econometrica, 61(5):1019 -- 1045, 1993. [24] E. Kalai and E. Lehrer. Weak and strong merging of opinions. Journal of Mathematical Economics, 23:73 -- 86, 1994. [25] R. Karandikar, D. Mookherjee, D. Ray, and F. Vega-Redondo. Evolving aspirations and cooperation. Journal of Economic Theory, 80(2):292 -- 331, 1998. [26] L. Kocsis and C. Szepesv´ari. Bandit based Monte-Carlo planning. In Proceedings of the 17th European Conference on Machine Learning, pages 282 -- 293. Springer, 2006. [27] A. Ledezma, R. Aler, A. Sanchis, and D. Borrajo. Predicting opponent actions by observation. In RoboCup 2003: Robot Soccer World Cup VII, pages 286 -- 296. Springer, 2004. [28] R. Martinez-Cantin. BayesOpt: A Bayesian optimization library for nonlinear optimization, experimental design and bandits. Journal of Machine Learning Research, 15:3735 -- 3739, 2014. [29] J. Mockus. Bayesian approach to global optimization: theory and applications. Springer Science & Business Media, 2013. [30] K. Murphy and Y. Weiss. The factored frontier algorithm for approximate inference in DBNs. In Proceedings of the 17th Conference on Uncertainty in Artificial Intelligence, pages 378 -- 385, 2001. [31] J. Nachbar. Beliefs in repeated games. Econometrica, 73(2):459 -- 480, 2005. [32] A. Panella and P. Gmytrasiewicz. Interactive POMDPs with finite-state models of other agents. Autonomous Agents and Multi-Agent Systems, 2017. [33] C. Rasmussen and C. Williams. Gaussian Processes for Machine Learning. MIT Press, 2006. [34] H. Robbins. Some aspects of the sequential design of experiments. Bulletin of the American Mathematical Society, 58:527 -- 535, 1952. [35] J. Snoek, H. Larochelle, and R. Adams. Practical Bayesian optimization of machine learning algorithms. In Advances in Neural Information Processing Systems 25, pages 2951 -- 2959, 2012. [36] F. Southey, M. Bowling, B. Larson, C. Piccione, N. Burch, D. Billings, and C. Rayner. Bayes' bluff: opponent modelling in poker. In Proceedings of the 21st Conference on Uncertainty in Artificial Intelligence, pages 550 -- 558, 2005. [37] P. Stone, G. Kaminka, S. Kraus, and J. Rosenschein. Ad hoc autonomous agent teams: collaboration without pre-coordination. In Proceedings of the 24th AAAI Conference on Artificial Intelligence, pages 1504 -- 1509, 2010. [38] R. Sutton and A. Barto. Reinforcement Learning: An Introduction. The MIT Press, 1998. [39] W. Thompson. On the likelihood that one unknown probability exceeds another in view of the evidence of two samples. Biometrika, 25:285 -- 294, 1933. [40] C. Watkins and P. Dayan. Q-learning. Machine Learning, 8(3):279 -- 292, 1992.
1601.08154	1	1601	2016-01-29T15:29:21	JADE, TraSMAPI and SUMO: A tool-chain for simulating traffic light control	[ "cs.MA" ]	Increased stress, fuel consumption, air pollution, accidents and delays are some of the consequences of traffic congestion usually incurring in tremendous economic impacts, which society aims to remedy in order to leverage a sustainable development. Recently, unconventional means for modeling and controlling such complex traffic systems relying on multi-agent systems have arisen. This paper contributes to the understanding of such complex and highly dynamic systems by proposing an open-source tool-chain to implement multi-agent-based solutions in traffic and transportation. The proposed approach relies on two very popular tools in both domains, with focus on traffic light control. This tool-chain consists in combining JADE (Java Agent DEvelopment Framework), for the implementation of multi-agent systems, with SUMO (Simulation of Urban MObility), for the microscopic simulation of traffic interactions. TraSMAPI (Traffic Simulation Manager Application Programming Interface) is used to combine JADE and SUMO allowing communication between them. A demonstration of the concept is presented to illustrate the main features of this tool-chain, using Q-Learning as the reinforcement learning method for each traffic light agent in a simulated network. Results demonstrate the feasibility of the proposed framework as a practical means to experiment with different agent-based designs of intelligent transportation solutions.	cs.MA	cs	JADE, TraSMAPI and SUMO: A tool-chain for simulating traffic light control Tiago M. L. Azevedo Paulo J. M. de Araújo Rosaldo J. F. Rossetti Ana Paula C. Rocha Artificial Intelligence and Computer Science Lab Department of Informatics Engineering Faculty of Engineering, University of Porto, Portugal {tiago.manuel, paulo.araujo, rossetti, arocha}@fe.up.pt 6 1 0 2 n a J 9 2 ] A M . s c [ 1 v 4 5 1 8 0 . 1 0 6 1 : v i X r a ABSTRACT Increased stress, fuel consumption, air pollution, accidents and delays are some of the consequences of traffic congestion usually incurring in tremendous economic impacts, which society aims to remedy in order to leverage a sustainable development. Recently, unconventional means for modeling and controlling such complex traffic systems relying on multi- agent systems have arisen. This paper contributes to the understanding of such complex and highly dynamic systems by proposing an open-source tool-chain to implement multi- agent-based solutions in traffic and transportation. The proposed approach relies on two very popular tools in both domains, with focus on traffic light control. This tool-chain consists in combining JADE (Java Agent DEvelopment Framework), for the implementation of multi-agent systems, with SUMO (Simulation of Urban MObility), for the micro- scopic simulation of traffic interactions. TraSMAPI (Traffic Simulation Manager Application Programming Interface) is used to combine JADE and SUMO allowing communication between them. A demonstration of the concept is presented to illustrate the main features of this tool-chain, using Q- Learning as the reinforcement learning method for each traffic light agent in a simulated network. Results demonstrate the feasibility of the proposed framework as a practical means to experiment with different agent-based designs of intelligent transportation solutions. Categories and Subject Descriptors I.2 [Artificial Intelligence]: Miscellaneous; I.6 [Simulation and Modeling]: Miscellaneous General Terms Algorithms, Design, Experimentation, Verification Keywords MAS, traffic light, JADE, SUMO, TraSMAPI, Q-learning 1. INTRODUCTION Nowadays urban centers face the daily problem of traffic congestion, which in addition to the obvious confusion can create also other negative consequences. Increased stress, fuel consumption, air pollution, accidents and delays are some of these consequences, which society aims to remedy in order to leverage a sustainable development, while mitigating tremendous economic impacts. Solutions to this problem have evolved over time, more in an immediate response perspective than on a long-term resolution perspective. Initially, the approach was based on the construction of alternative routes with increased capacity. However, available money and territorial area ceased to exist for continuing implementation of this sort of solution. In parallel, traffic lights and roundabouts were introduced but the urban centers continued growth now are demanding more advanced and efficient alternative measures. The aim of the work described in this paper was to use a tool-chain that allows us to implement a multi-agent sys- tem (MAS) for traffic light control. Therefore, a multi- agent system approach was used to answer the daily problem of traffic congestion. This tool-chain consisted in integrat- ing JADE (Java Agent DEvelopment Framework) for con- trolling the multi-agent system to SUMO (Simulation of Urban MObility) for traffic simulation. TraSMAPI (Traffic Simulation Manager Application Programming Interface) was the middleware combining JADE and SUMO and allow- ing communication between both environments. For the sake of illustration, the implemented agents’ learning method was Q-Learning. As a motivation, just a few simulation tools truly sup- port the concept of agents and multi-agent systems in traffic simulation; MATSim-T [3, 4] and ITSUMO [9, 6] are good ex- amples to be mentioned. However, no standard of wide reach for the implementation of such tools actually exists. Indeed, alternative approaches would require either general purpose MAS-based simulators to be adapted to the specific domain of traffic and transportation, or the other way around with the adaptation of traffic simulators to be adapted so as to support the MAS-based models. With our approach, we ex- pect to benefit from both worlds on an integrated basis. Also, it is important to notice that although SUMO and ITSUMO are both open-source microscopic simulators and have a quite similar acronym, they are no related applications. ITSUMO is a Cellular-Automata-based simulator, whereas SUMO uses a continuous representation of space on road segments. Be- sides, ITSUMO explicitly consider the metaphor of agents, whereas SUMO can be considered a traditional microscopic simulator, where agents are not explicitly implemented. The expected contribution of this work, rather than imple- menting a new agent-based simulator from scratch, adapting or extending existing ones, is to devise an open-source tool- chain to implement MAS-T (MAS in traffic and transporta- tion) on the basis of two very popular tools in both domains. On the one hand, JADE supports the implementation of any MAS solution and, on the other hand, SUMO supports an appropriate representation of the traffic environment in which agents inhabit and perform their tasks. This paper will start to deeply describe the tools. The conceived model is detailed and instantiated in the proposed tool-chain. An experimental set-up is used to illustrate the proposed approach, followed by the discussion of preliminary results. After discussion on related works, conclusions are drawn as well as are further developments suggested. 2. A MAS-BASED TRAFFIC SIMULATION TOOL-CHAIN The MAS-based traffic simulation tool-chain used consisted in three main tools: JADE, SUMO and TraSMAPI. A multi-agent system based approach seems to be the appropriate way to represent the different traffic lights in a network. Consequently, it is necessary that a multi-agent system framework take care of the different agent behaviours, as it is the case in JADE. Next, a microscopic simulator is needed to take care of the traffic road dynamics, such as vehicles decisions. It should be noted that although it is necessary to have vehicles in order to test traffic light control, these vehicles do not need to be modeled as agents. It would be very computationally expensive to simulate a huge quantity of vehicles, each one with driver’s decision-making and other cognitive aspects and details. SUMO was the microscopic simulator chosen. Finally, as traffic lights are considered to be agents, it is necessary they communicate with the simulator. This is important so as to allow their traffic lights in the simulation to have the semaphore plans always updated and agents to perceive the network dynamics. This communication was made through TraSMAPI, consisting of an integration API implemented in Java. 2.1 JADE JADE is a framework completely developed in Java. It simplifies the implementation of multi-agent systems through a middleware that complies with the FIPA1 specifications and through a set of graphical tools that supports the de- bugging and deployment phases. The agent platform can be distributed across machines and the configuration can be controlled via a remote GUI [19]. The version used in this work was 4.3.0, released on March 2013. One advantage of using JADE to implement MAS is its ability to allow run-time visualisation and control of the interactions among agents in the application. As relevant features for this work, some can be pointed that are not directly connected to agents, that is, are independent of the applications: message transportation, codification and parsing of messages or lifetime of an agent, for instance. 2.2 SUMO SUMO is an open-source program (licenced under GPL2) 1Foundation of Intelligent Physical Agents, an organization that promotes agent-based technology and the interoperabil- ity of its standards with other technologies 2GNU General Public License, a free, copyleft license for for traffic simulation. Its simulation model is microscopic, that is, each vehicle is explicitly modeled, has its own route and moves individually over the network. It is mainly de- veloped by Institute of Transportation Systems, located at German Aerospace Center [12]. The version used in this work was 0.18.0, released on August 2013. Among other features, it allows the existence of differ- ent types of vehicles, roads with several lanes, traffic lights, graphical interface to view the network and the entities that are being simulated, and interoperability with other applica- tions at run-time through an API called TraCI. Moreover, the tool is considered to be fast, still allowing a version with- out a graphical interface where the simulation is accelerated putting aside visual concerns and overheads[12]. In Figure 1 it is possible to visualize the SUMO’s graphical interface with a running simulation. It is possible to point out almost all specified features: vehicles stopped at the traffic light as well as a long vehicle entering an intersection. Figure 1: SUMO working This tool was crucial in this work! First, it allows loading different maps (described in XML files) in order to test vari- ous scenarios with vehicles and traffic lights. Then, with the simulation itself there is no need to waste time implementing the dynamics of many vehicles and traffic lights, starting soon with the evaluation of algorithms. Finally, interoperability with other applications allows that each agent can be bound to an entity in SUMO, so that changes in the dynamics of traffic lights, for instance, can be visually seen in the SUMO’s graphic interface. 2.3 TraSMAPI TraSMAPI can be seen as a generic API for microscopic traffic that allows real-time communication between agents of urban traffic management (such as vehicles and traffic signals) and the environment created by various simulators. This tool was developed in LIACC (Artificial Intelligence and Computer Science Laboratory), University of Porto, having already been tested with two different simulators, including SUMO [20]. This API offers a higher abstraction level than most of mi- croscopic traffic simulators in such a way that the solution is independent from the microscopic simulator to use. Initially, this tool also aimed to gather relevant metrics/statistics and offer an integrated framework for developing multi-agent systems, as shown in Figure 2 [21]. As it can be seen, there were three main modules: a com- munication module with possibility of various microscopic software and other kinds of works 3.1 Concepts For the purpose of this work, a traffic light is defined as an intersection that has a semaphore plan, which is characterized by a sequence of phases. Each phase has a duration and a color scheme (green, yellow, flashing yellow and/or red), whose values correspond to every possible maneuver at the intersection. The execution of the phases sequence is called a cycle and has a period equal to the sum of the durations of the phases. In Figure 4 the intersection has six possible maneuvers, indicated by the arrows, which means that each phase has to specify a color for each maneuver (M1, ..., M6). The sequence of phases is guided by the phase number, and after the end of the sixth phase a 80 cycle duration is completed, following again phase 1. For each maneuver the traffic light may show the green color with symbol G, yellow with symbol y, flashing yellow with symbol g and red with symbol r. Figure 4: Example of a semaphore plan with illus- trative image for phase 5 3.2 Scenario Definition As a demonstration of the concept, it was used a grid (Manhattan-like) map (Figure 5) in order to make some experiments for traffic light control. A grid map is relatively simple to implement and where it is fairly possible to define consistent semaphore plans. The Q-learning algorithm was chosen as the learning method for the traffic light agents. Figure 2: TraSMAPI’s initial architecture simulators, the module generating statistics and the module for the MAS management. Presently, only the communica- tion module is functional and this is the module that interests to the scope of the presented work. 2.4 The tool-chain In order to achieve a tool-chain with the previous described tools, it was necessary to extend the TraSMAPI API, enabling to build an abstraction over a SUMO’s traffic light entity. Thus, it was necessary to implement the communication protocol regarding the methods of traffic light for value retrieval and state change, in TraCI [23]. Figure 3: Communication between JADE and SUMO using TraSMAPI for a traffic light The architecture described in Figure 3 shows how it is possible the existence of one or more traffic light agents. Each traffic light agent has a tie to the respective traffic light to be modeled in SUMO. This tie is supported by the TraSMAPI communication module that interacts with the SUMO API, TraCI. 3. EXPERIMENTAL SETUP A simple scenario for the sake of illustration is now de- scribed. Although the following scenario is simple and not intended to deeply discuss the appropriateness of implement- ing traffic control through agents, it illustrates well how our integrated framework could be practically used in this sort of experiments. Figure 5: The grid map where simulation took place Thus, these experiments consisted in performing four sim- ulations: one with traffic lights with fixed semaphore plans, one with traffic lights with fixed semaphore plans but with different durations for distinct day periods, another with traf- fic lights with Q-Learning taking into account the duration of the phases and another with traffic lights with Q-Learning taking into account the duration of the phase and the period of the day. The metrics that will be used to evaluate the results are described in Section 3.3. Therefore, to be a basis for comparison, each of the simulations had the same back- ground: the same vehicles leaving at the same time, from the same place and with the same route. Theoretical hour 00h00 07h30 09h00 18h00 20h00 Starting step Traffic 0 150 000 180 000 360 000 400 000 Low High Medium High Low Table 1: Traffic distribution during the day As SUMO’s time unit is step (step of execution), and as each step can last more or less a second, it was necessary to make a correlation between number of steps and the time in simulation. This correlation is necessary to implement time compression and allow for an entire day to be simulated correctly and much quicker than in the real-life duration. Thus, the approach taken was that 20000 steps correspond to 1 theoretical simulation hour. There are three traffic scenarios throughout the day: low traffic, medium traffic with a predominance of horizontal flows of vehicles, and heavy traffic. The distribution of traffic is performed according to Table 1. A manual approach was carried out for the definition of the green splitting for the phases in the simulations where Q-Learning was not used. In the specific case of the traffic lights with fixed semaphore plans but with different durations for distinct day periods, in the low traffic period faster green durations were used in opposition to the high traffic period where long green durations were used. Each simulation corresponded to a 4-day simulation. This way, at the end of each simulation, that is, when all vehicles arrived at their destination, metrics were generated. The tool-chain takes some time to add all desired vehicles at startup. This way, simulation time should not be such that would make the startup take longer than necessary. However, simulation time should be enough so traffic lights have time to learn. 4-day simulation seemed to be the best way for balancing these issues. It is also important to note that the insertion of network traffic was not made in a distributed manner again because of the slowness that would result with the startup of the tool- chain. Thus, two approaches have been considered for the four simulations: on the one hand, insertions with intervals of 7000 steps, and on the other hand, insertions with intervals of 10000 steps. In each of these intervals, the quantity of vehicles to add would depend on the period of day that the simulation was on. So, in reality, there were 8 simulations. 3.3 A Q-Learning traffic control It is important to be aware that the state representation has influence in the Q-Learning performance, in other words, it is only possible to learn something if it is relevant to the problem. In this sense, it is intended to use two relevant vari- ables: phase durations and period of the day. It is considered that phases initially with duration under 20 seconds will not suffer any variation and the other phases will have durations between 20 and 60 seconds, with a granularity of 5 seconds. Assuming that could exist two or three phases with variable durations for each semaphore plan, there are a total of 81 or 729 duration combinations, respectively. Possible actions are decrease, maintain or increase (-5, 0 or +5 seconds) each variable duration, which results in a Q-Table with 243 or 2187 pairs. Considering the period of the day these numbers would increase. The reward function consists of two portions: the own reward multiplied by 0, 5 and the weighted average (concern- ing distance of roads) of neighboring traffic lights rewards multiplied by 0, 5. These rewards are calculated using the average of vehicles in the vicinity of an intersection, during a complete cycle. In what concerns exploration, it is used a 0-greedy strategy. The learning rate was 50% as well as was the discount factor. In order to evaluate the learning process, the following metrics will be used: • Travel time and average waiting time in queues, that allow to check the individual performance of each vehi- cle; • Standard deviations of travel times and of average waiting times in queues, that allow to check the network traffic homogeneity, in other words to check whether vehicles will have a similar experience both in travel time and waiting time in queues; • Average of travel times and of average waiting times in queues, that allows to check the global network traffic performance. 3.4 A multi-agent system for traffic control System could be implemented using two agent models: an agent for each traffic light with a super coordinator agent, or an agent for each traffic light with distributed coordination. First model allows a greater process synchronization between agents, has a single point of failure for the entire system and has a computation volume highly concentrated in the coordi- nator. The second model can hardly obtain synchronization but yet in the event of a failure, this is not spread to the entire system, and computation is homogeneous. Therefore, system will be implemented using the second model in which agents are traffic lights. The architecture of each agent displayed in Figure 6 is based on a learning agent architecture [16, p. 54-57] but specified to the Q- Learning process. In this Figure, the presented behaviour does not include the initial phase in which the Q-Table is initialized and where each agent finds the neighbors (in Figure 6 represented as Agent n). There exist two types of communication between agents: reward requests and answers to reward requests. The former is implemented using the performative QUERY REF with content ”reward”, whereas the latter uses the performative INFORM REF with the reward itself in the content. Figure 7 is described a possible situation between agents, in which Agent2 is a neighbor of Agent1 and Agent3, and Agent1 and Agent3 are not neighbors. 4. PRELIMINARY RESULTS AND DISCUS- SION Figures 8 and 9 show, for each vehicle, the average waiting time in queues. For each vehicle, with intervals of 7000 steps there are greater peaks in the average waiting time in queues compared to intervals of 10000 steps. This is explained as the network gets easily saturated with fewer steps and vehicles wait longer in queues. Another fact is that, in individual terms, the average waiting time in queues does not vary a considerably with the types of semaphore plan. Figure 6: Traffic light agent architecture and behaviour Figure 8: The average waiting time in queues with intervals of 10000 steps Figure 9: The average waiting time in queues with intervals of 7000 steps Figure 11: Average of averages waiting time in queues steps. Even so, Q-Learning B plan has slightly better results. The peculiar result that semi-fixed plans induces lower waiting times in queues but longer travel times than Q- Learning B may be explained. A simple example where this makes sense is that while in Q-Learning B a vehicle can pass through a lot of green traffic lights (inducing lower travel times), in the few traffic lights that it has to wait, it waits a lot of time (inducing a greater average waiting time in queues). In the semi-fixed plan a vehicle may have to wait, in average, shorter in queues but as it stops in more traffic lights than in Q-Learning B, it takes longer to travel through. Finally, Figures 12 and 13 show the results of standard deviations. Figure 7: Interaction example between agents In what concerns the travel time for each vehicle, once again intervals of 7000 steps produce greater peaks, in other words, easily a vehicle takes longer to travel the same path. It is curious to verify that with a changing of the step intervals a vehicle can take longer or shorter in different plans. In other words, unlike the previous metric, there is not a better semaphore plan for the majority of the vehicles, and so a semaphore plan can give better individual results for some vehicles, but not for all vehicles. Figures 10 and 11 present metrics to a more global evalua- tion of the explored solutions. It is called Q-Learning A to the plan taking into account the duration of the phases and Q-Learning B to the plan taking into account the duration of the phases and period of the day. Figure 12: Standard deviation of travel time Figure 10: Average of travel times In these Figures a clear difference between the fixed and semi-fixed plan is shown: while the fixed plan presents the worst results, semi-fixed plans presents the best results, even compared to the Q-Learning plans. Even so, the Q-Learning B plan has better results than Q-Learning A, as it was expected. However, what matters the most for the driver is the total travel time. Looking at the Figures, the differences between plans are not big, mainly for the plans with intervals of 7000 Figure 13: Standard deviation of average waiting time in queues For the averages of waiting times in queues, the semi-fixed plan has the best network traffic homogeneity for intervals of 7000 steps and the second best for intervals of 10000 steps. Nevertheless, in general terms Q-Learning B can obtain more network traffic homogeneity. Passing to the total travel times, Q-Learning B can widely overcome the other plans, obtaining greater network traffic homogeneity, both for intervals of 7000 and 10000 steps. The network traffic homogeneity is an important factor for a driver, who intends to know that when he goes to his destination there is not a probability to take longer than it was expected. 5. RELATED WORKS The specific case of traffic lights is one of the areas where much has been researched for new solutions, from the design of intersections [13] (including physical layout and semaphore plans), to the definition of semaphore plans through statis- tical analysis. Current solutions try to answer the highly dynamic system [8, p. 343] using coordinated control. Sev- eral methodologies have been used such as genetic algorithms [18], fuzzy logic [2] and reinforcement learning [1]. To date, there are not many solutions for traffic that make full use of the intelligent agent concept. However, the multi- agent system approach has become recognized as a convenient approach for modelling and simulating complex systems [15]. Also, it has grown enormously not only applied to traffic but also to transportation in general terms [7]. In the last decade some microscopic simulators have been developed, such as MITSIM, Paramics, Aimsun, CORSIM and VisSim. However, none of these is strictly defined as agent-based simulation systems, even though they model vehicles in an object-oriented manner. Just a few simulation tools truly support the concept of agents and multi-agent systems in traffic simulation; MATSim-T [3, 4] and ITSUMO [9, 6] are good examples to be mentioned. Regarding this simulation tools some examples of multi- agent system approaches for traffic lights control can be seen in [11], [5], [14] and [10]. Simulators used in these works were Aimsun, ITSUMO, VisSim and ITSUMO, respectively. With MAS being recognized as a convenient approach, there must be a sufficiently general way to couple this ap- proach to such a huge quantity of microscopic simulators that exist now. The platform that integrates SUMO and JADE consists of an API intended to allow interoperability among simulators. The platform, coined TraSMAPI, is suf- ficiently general to allow other simulators to interact with MAS frameworks such JADE. A previous paper [20] reports on an experiment integrating ITSUMO and SUMO under TraSMAPI, thus demonstrating such an ability. In another study [22], external traffic controller agents operate over Aimsun-simulated scenarios through TraSMAPI. In this spe- cific work, we illustrate how non-agent-based simulators can be extended with TraSMAPI to support MAS-T assessment. There are certainly other options to simulate agent-based traffic and transportation, such as MATSim. Although such tools are open-source then allowing full customisation, the use of JADE over a traditional microscopic simulation tool is expected to promote greater flexibility in terms of agent architectures that can be implemented. In respect to the described tool-chain, a similar approach has already been proposed. In [17] it is possible to see the tool-chain JADE+TraSMAPI+SUMO. However, the goal of this work was focused on the vehicles itself instead of traffic lights. 6. CONCLUSIONS This paper explores the use of a specific tool-chain for the implementation of intelligent traffic light control. At the end, we have a tool that allows us to implement and test real MAS-based solutions in the domain of traffic and transportation, using commodity computers and open-source tools of wide reach. Q-Learning was used as the reinforcement learning method to illustrate the implementation of traffic light agents. The tool-chain resulting from the integration of JADE and SUMO through TraSMAPI is the main expected contribution of this paper. Nonetheless, many improvements can be identified for fu- ture work. This paper did not analyse other forms for traffic control. For example, there are solutions based on the sim- ple statistical analysis of traffic information and posterior adjustment according to such analytical procedures. This kind of solution can contrast with others as it can be highly dynamic and therefore can be applied to very specific scenar- ios. Another possible solution is the installation of sensors in each traffic light that, on a reactive way, can simply respond according to the number of waiting vehicles in the queue, needing neither great computation power nor the analysis of the traffic network, totally or partially. The tool-chain itself could be improved in some different possible ways, including scalability, robustness, and efficiency. Firstly, the increase of performance in information retrieval by decreasing time in communication between the agent and the simulator. SUMO, that is still in its very young stage, proved to be much slower than desired with a larger number of vehicles and constant information retrieval. Certainly this aspect will be improved in next versions of SUMO, but it is necessary to analyse who is to blame: Is TraCI too much slow? is TraSMAPI implemented well in what concerns performance issues? During simulations is the number of generated requests to TraCI greater than necessary? and so forth! On other hand, for the real simulated system implementation it would be necessary to develop a distributed system where each agent was executed in each machine. In this specific study we did not use JADE ability to dis- tribute agents over a computer network, as our main objective is to demonstrate how JADE and SUMO can be integrated through TraSMAPI. Nonetheless, larger networks will cer- tainly require more robust computational power, which can be achieved through an appropriate distribution of computa- tion across a computer network. The traffic network itself could also be improved: a more realistic map for simulation can give more relevant results. Maybe the multi-agent system used could not be the best for the proposed approach. An analysis of the best tool to use is certainly imperative. We intend to use the proposed framework to further in- vestigate traffic control strategies through more robust and complex signal agents. Contrary to the manual approach adopted to set up semaphore plans, tools such as Transit can be used to assist a more coherent definition of phases at each junction of the network. Finally, in terms of general results, it seems that Q-Learning taking into account the duration of the phases and the period of the day obtains better general results, even if they are not very significant. Nevertheless, it is necessary to perform these experiments in more real settings, not only in what concerns the network, but also in what concerns simulation. So, it would be possible to better conclude whether the Q-Learning implementation in traffic networks is an added value not only for drivers, but also for the system as a whole. 7. ACKNOWLEDGMENTS Authors greatly acknowledge invaluable contributions by Filipe Oliveira, who also worked directly on this project. We also thank Professor Eug´enio Oliveira and Dr. Henrique Lopes Cardoso for important suggestions and comments on the course of this work. 8. REFERENCES [1] I. Arel, C. Liu, T. Urbanik, and A. Kohls. Reinforcement learning-based multi-agent system for network traffic signal control. Intelligent Transport Systems, IET, 4(2):128–135, 2010. [2] E. Azimirad, N. Pariz, and M.-B. N. Sistani. A novel fuzzy model and control of single intersection at urban traffic network. Systems Journal, IEEE, 4(1):107–111, 2010. [3] M. Balmer, K. Meister, M. Rieser, K. Nagel, K. W. Axhausen, K. W. Axhausen, and K. W. Axhausen. Agent-based simulation of travel demand: Structure and computational performance of MATSim-T. ETH, Eidgenossische Technische Hochschule Zurich, IVT Institut fur Verkehrsplanung und Transportsysteme, 2008. [4] M. Balmer, M. Rieser, K. Meister, D. Charypar, N. Lefebvre, K. Nagel, and K. Axhausen. MATSim-T: Architecture and simulation times. Multi-agent systems for traffic and transportation engineering, pages 57–78, 2009. [5] A. L. Bazzan, D. de Oliveira, and B. C. da Silva. Learning in groups of traffic signals. Engineering Applications of Artificial Intelligence, 23(4):560–568, 2010. [6] A. L. Bazzan, M. d. B. do Amarante, T. Sommer, and A. J. Benavides. ITSUMO: an agent-based simulator for its applications. In Proc. of the 4th Workshop on Artificial Transportation Systems and Simulation. IEEE, 2010. [7] A. L. Bazzan and F. Klugl. A review on agent-based technology for traffic and transportation. The Knowledge Engineering Review, pages 1–29, 2013. [8] Y. K. Chin, W. Y. Kow, W. L. Khong, M. K. Tan, and K. T. K. Teo. Q-learning Traffic Signal Optimization within Multiple Intersections Traffic Network. In Computer Modeling and Simulation (EMS), 2012 Sixth UKSim/AMSS European Symposium on, pages 343–348. IEEE, 11 2012. [9] B. C. da Silva, A. L. Bazzan, G. K. Andriotti, F. Lopes, and D. de Oliveira. ITSUMO: an intelligent transportation system for urban mobility. In Innovative Internet Community Systems, pages 224–235. Springer, 2006. [10] D. de Oliveira, A. L. Bazzan, B. C. da Silva, E. W. Basso, L. Nunes, R. Rossetti, E. de Oliveira, R. da Silva, and L. Lamb. Reinforcement Learning based Control of Traffic Lights in Non-stationary Environments: A Case Study in a Microscopic Simulator. In EUMAS. Citeseer, 2006. [11] L. B. de Oliveira and E. Camponogara. Multi-agent model predictive control of signaling split in urban traffic networks. Transportation Research Part C: Emerging Technologies, 18(1):120–139, 2010. [12] German Aerospace Center, Institute of transportation Systems. SUMO at a Glance. http: //sumo-sim.org/userdoc/Sumo_at_a_Glance.html. Accessed: 2013-10-23. [13] P. Koonce, L. Rodegerdts, K. Lee, S. Quayle, S. Beaird, C. Braud, J. Bonneson, P. Tarnoff, and T. Urbanik. Traffic signal timing manual. Technical report, US Department of Transportation, 2008. [14] W. Lu, Y. Zhang, and Y. Xie. A multi-agent adaptive traffic signal control system using swarm intelligence and neuro-fuzzy reinforcement learning. In Integrated and Sustainable Transportation System (FISTS), 2011 IEEE Forum on, pages 233–238. IEEE, 2011. [15] L. J. Moya and A. Tolk. Towards a taxonomy of agents and multi-agent systems. In Proceedings of the 2007 spring simulation multiconference-Volume 2, pages 11–18. Society for Computer Simulation International, 2007. [16] S. Russell and P. Norvig. Artificial Intelligence: A Modern Approach. Prentice Hall series in artificial intelligence. Prentice Hall, 2010. [17] G. Soares, J. Macedo, Z. Kokkinogenis, and R. J. Rossetti. An integrated framework for multi-agent traffic simulation using sumo and jade. In SUMO2013, The first SUMO User Conference, May 15-17, 2013 - Berlin-Adlershof, Germany, pages 125–131. DLR - Institut fur Verkehrssystemtechnik, 2013. [18] F. Teklu, A. Sumalee, and D. Watling. A genetic algorithm approach for optimizing traffic control signals considering routing. Computer-Aided Civil and Infrastructure Engineering, 22(1):31–43, 2007. [19] Telecom Italia Lab. JADE description. http://jade.tilab.com/description-index.htm. Accessed: 2013-10-20. [20] I. J. Tim´oteo, M. R. Ara´ujo, R. J. Rossetti, and E. C. Oliveira. Using trasmapi for the assessment of multi-agent traffic management solutions. Progress in Artificial Intelligence, 1(2):157–164, 2012. [21] I. J. P. M. Tim´oteo, M. R. Ara´ujo, R. J. F. Rossetti, and E. C. Oliveira. TraSMAPI: An API oriented towards Multi-Agent Systems real-time interaction with multiple Traffic Simulators. In Intelligent Transportation Systems (ITSC), 2010 13th International IEEE Conference on, pages 1183–1188, 9 2010. [22] C. Vilarinho, G. Soares, J. Macedo, J. P. Tavares, and R. J. Rossetti. Capability-enhanced {AIMSUN} with real-time signal timing control. Procedia - Social and Behavioral Sciences, 111(0):262 – 271, 2014. Transportation: Can we do more with less resources? - 16th Meeting of the Euro Working Group on Transportation - Porto 2013. [23] A. Wegener, M. Pi´orkowski, M. Raya, H. Hellbruck, S. Fischer, and J.-P. Hubaux. TraCI: an interface for coupling road traffic and network simulators. In Proceedings of the 11th communications and networking simulation symposium, pages 155–163. ACM, 2008.
1808.04813	2	1808	2018-11-06T08:41:36	AMoDSim: An Efficient and Modular Simulation Framework for Autonomous Mobility on Demand	[ "cs.MA" ]	Urban transportation of next decade is expected to be disrupted by Autonomous Mobility on Demand (AMoD): AMoD providers will collect ride requests from users and will dispatch a fleet of autonomous vehicles to satisfy requests in the most efficient way. Differently from current ride sharing systems, in which driver behavior has a clear impact on the system, AMoD systems will be exclusively determined by the dispatching logic. As a consequence, a recent interest in the Operations Research and Computer Science communities has focused on this control logic. The new propositions and methodologies are generally evaluated via simulation. Unfortunately, there is no simulation platform that has emerged as reference, with the consequence that each author uses her own custom-made simulator, applicable only in her specific study, with no aim of generalization and without public release. This slows down the progress in the area as researchers cannot build on each other's work and cannot share, reproduce and verify the results. The goal of this paper is to present AMoDSim, an open-source simulation platform aimed to fill this gap and accelerate research in future ride sharing systems.	cs.MA	cs	AMoDSim: An Efficient and Modular Simulation Framework for Autonomous Mobility on Demand Andrea Di Maria1, Andrea Araldo2[0000−0002−5448−6646], Giovanni Morana1, and Antonella Di Stefano3 1 Aucta Cognitio R&D Labs, Catania 95123, Italy {adimaria,gmorana}@auctacognitio.net 2 R´eseaux et Services de T´el´ecommunications, Te´l´ecom SudParis, Evry 91011, France [email protected] 3 Universit´a di Catania, Catania 95125, Italy [email protected] Abstract. Urban transportation of next decade is expected to be dis- rupted by Autonomous Mobility on Demand (AMoD): AMoD providers will collect ride requests from users and will dispatch a fleet of au- tonomous vehicles to satisfy requests in the most efficient way. Differently from current ride sharing systems, in which driver behavior has a clear impact on the system, AMoD systems will be exclusively determined by the dispatching logic. As a consequence, a recent interest in the Opera- tions Research and Computer Science communities has focused on this control logic. The new propositions and methodologies are generally eval- uated via simulation. Unfortunately, there is no simulation platform that has emerged as reference, with the consequence that each author uses her own custom-made simulator, applicable only in her specific study, with no aim of generalization and without public release. This slows down the progress in the area as researchers cannot build on each other's work and cannot share, reproduce and verify the results. The goal of this paper is to present AMoDSim, an open-source simulation platform aimed to fill this gap and accelerate research in future ride sharing systems. Keywords: smart mobility · smart city · shared mobility · autonomous vehicles · simulation 1 Introduction Transportation is traversing a period of big transformations driven by Informa- tion and Communication Technology (ICT). For instance, the ubiquitous con- nectivity guaranteed by 3G and 4G has triggered the emergence of ride sharing services, e.g., Uber and Lyft, in which users reserve a ride through a smartphone app and service providers match them to a fleet of vehicles. Goldman Sachs quantifies the importance of these services by predicting a market of 285 billion dollars in 2030 [11]. In more and more cities, ride sharing services are also deter- mining a transformation of every-day life [12]. This revolution will become even 2 Di Maria et al. deeper when these services will be provided by Autonomous Vehicles (AVs). Au- tonomous Mobility on Demand (AMoD) services [8] will be very cheap for the users, since providers will not have to sustain the cost of labor of the drivers. One reason for the efficiency of these systems is that vehicles can be shared among many users. To do so, efficient and scalable algorithms are needed. While the Vehicle Routing Problem [4] has been studied from the 1950s, the success of ride sharing systems has lead to a renovated interest in this decade, where the problem has been specialized to the case of matching ride requests from passen- gers to available vehicles, while respecting some constraints on users' waiting and riding time. A particular focus has regarded the computation of condensed vehi- cle trips to properly aggregate many rides in order to minimize provider's costs while keeping the user quality of level acceptable. The request-vehicle matching problem has been shown to be NP hard [7]. Therefore, a vast literature has de- veloped to propose "good" heuristics with a reasonable computation time to be used in practice and has resorted to simulation to evaluate them. Unfortunately, up to now no reference simulation tool has emerged for this, which is shown by the fact that most of the authors have been forced to build from scratch their own case-specific simulator. The negative consequences are: -- Waste of time and effort, to create every time a simulator. -- Impossibility to build on the effort of past research. -- Difficulty for the community to reproduce and verify results. On the other side, there are few exceptions of complex transportation sim- ulation tools extended with models of ride sharing systems. However, they are not suitable for the researchers interested in the development of algorithms for ride-sharing, whom we target in this work. The reasons are: -- They require to specify scenarios with high level of realism, like economic indicators of the population and of the area, which are not usually available. -- Even if available, it takes a long time and effort to figure out how to set them up into the simulators, which would instead be preferable to spend in the inner workings of the algorithms. -- They lack flexibility: when developing an algorithm, it is necessary to test it in a vast range of scenarios, instead of just super-realistic one, to generalize the findings. -- The level of detail transportation represents an overhead: part of the com- putation time is spent in representing the detailed movement of vehicles at millisecond scale, which has no big impact on the ride sharing logic. For these reasons, transportation simulation tools are to be used a-posteriori when, for instance, a transportation authority or company wants to check what is the impact of a ride sharing strategy, already developed and thoroughly studied, on the particular scenario of interest. In this paper we present AMoDSim, a simulation framework open to re- searchers in future-generation ride-sharing systems whose design goals are: AMoDSim: An Efficient and Modular Simulation Framework 3 -- Launching massive simulation campaigns to simultaneously test the perfor- mance of the algorithms under study, under different settings, is easy and scalable. -- By means of modularity, it is easy to implement new algorithms, with min- imum modification of the other components. -- Results on the performance for both the provider and the user perspective are produced automatically and are simple to analyze. The code is available4 under the CC BY-NC-SA 4.0 license. The rest of the paper is organized as follows: In Section 2 we review the work in simulation of ride sharing systems. In Section 3 we present the model of AMoD used in AMoDSim. In Section 4 we describe its architecture and in Section 5 we showcase it in a case study in which we compare several provider and user-related metrics of two different matching algorithms. 2 Related Work In this section we describe the state of the art of the research on autonomous mobility on demand and future generation ride sharing systems, focusing on the simulation tools used. We divide this research in works that use case-specific simulators and complex transportation simulators. The limitations of both has been discussed in the previous section. 2.1 Work Based on Case-Specific Simulators We emphasize that no code has been made public with any of the studies listed in this subsection, nor the simulators have been described enough to be repro- ducible. This reinforces the utility of our effort. Santi, Frazzoli et Al. published a series of papers [22,7,23] where they proposed mathematical formulations of ride sharing problems and heuristics to solve them. Case studies are shown in New York. Similarly, Ma et Al. [19] study ride-sharing algorithms using GPS taxi trajectories collected in Bejing. Agatz et Al. [5] built a simulator for a case study in Atlanta. Within their simulator, an agent can subscribe to a provider either as a rider or a driver. The study better represents systems like BlaBla Car [1], in which a traveler can publish her future trip in a web portal and other users can hop-in. These systems are now called "carpooling" and are different from ride sharing systems like Uber and Lyft and the future AMoD, in which (i) drivers are continuously operating for hours just to serve other individuals' trips and (ii) requests for rides arrive continuously in real time and are not announced in advance. Other case-specific simulators were developed for case studies in Seul and Boston in [16] and [17], respectively. 4 https://github.com/admaria/AMoDSim 4 Di Maria et al. 2.2 Work Based on Complex Transportation Simulators Some case studies have been performed extending commercial transportation simulators, like Aimsun [20,18]. However, commercial tools are usually not avail- able to researchers and their code is closed, impeding the verification and the reproduction of results. To the best of our knowledge, three simulation tools developed by academic institutions have been extended and employed in studies related to AMoD, namely SimMobility [8] and MATSim [10,9] and SUMO[6]. The main issue with the first two is the level of complexity that the researcher is required to handle and the performance. They are agent-based, i.e., they simu- late the behavior of each single traveler through transportation-specific economic models. In order to do so, the researcher must construct first a synthetic popu- lation and describe the economic indicators of the urban network. As discussed in Section 1, this is overkill for research focused on algorithms, which is what we target here. The unsuitability of these tools is testified by the fact that: (i) they are generally used, at least as far as published research visible to us is con- cerned, only by the very same group that developed them and (ii) researchers have preferred to craft their own case-specific simulators instead of using them. SUMO is a microscopic simulator that has been employed in a recent case study on AMoD in the city of Milan[6]. However, that study does not fill the gap we aim to fill. First, SUMO is a purely microscopic simulator, i.e., it computes the detailed movement of each vehicle,5 which is an overhead that we want instead to avoid, since it has limited interest when studying the dispatching logic in an AMoD system. Second, SUMO does support natively Mobility on Demand ser- vices and the authors of [6] had to write from scratch this functionality, which, however, they do not make publicly available. Third, SUMO needs detailed in- put, that the authors needed to obtain by cross-correlating several data-sources (Google APIs, mobile phone traces, etc.), while the choice we made in AMoD is to streamline the input definition, sacrificing some realism. Finally, is it not possible in [6] to specify user-specified quality of service requirements. 2.3 Other work NOT IN THIS DRAFT 3 Model of Autonomous Mobility on Demand We now present the model of AMoD service implemented into the simulator. The model includes a fleet of vehicles, a coordinator managing it and users. Users send trip requests to the coordinator, which runs matching algorithms or simply orchestrates the distributed computation running in the vehicles, in order to decide how to match them to the available vehicles. A trip request consists of 5 A particular version of SUMO, called SUMO MESO[2], is intended to reduce the details in vehicle movement simulation. However, we are not aware of any published study on AMoD systems based on SUMO MESO. AMoDSim: An Efficient and Modular Simulation Framework 5 two stop-points, one for the pick-up and one for the drop-off. Each stop point is a tuple sp = {q, t, ∆t}, where q is the pick-up or drop-off point, t is the preferred time at which the user wishes to be picked up or dropped off, ∆t is the maximum extra-time the user tolerates to be picked up or dropped off, with respect to the preferred time. At any given time, each vehicle v has a set of planned stop-points organized in a certain sequence Sv = [sp1, sp2, . . . ], that we call schedule. Each schedule is associated with a cost c(Sv), which can be defined in different ways to take into account provider or user-related metrics. For example, this cost could be the kilometers traveled to accomplish that schedule, or some indication of the travel or waiting time of the users served by that schedule. The goal of the provider is to create and continuously update the schedule Sv of each vehicle of its fleet, in order to optimize the costs c(Sv), subject to respecting the time constraints of all the users. Observe that this model is general enough to represent different types of optimization: (i) both provider cost or user level of service can be optimized, as this boils down to the way the cost c(Sv) is defined; (ii) one can simply study the overall cost optimization, or min-max optimization, etc.; (iii) the optimization can be both centralized, in case a single coordinator decides all the schedules Sv, or distributed, in case, for instance, each vehicle v optimizes its own schedule. While the model is general, we have currently only implemented the strategies described in Section 3.2. 3.1 Time constraints We define a schedule Sv = [sp0, . . . , spn] of a vehicle v feasible, if the time constraints of all its stop-points is satisfied. Let us suppose spi = (qi, ti, ∆ti) and that bi is the time needed to complete spi, i.e., the time for the passenger to board (alight), in case of pick-up (drop-off), that the current time is tnow and the current vehicle location is qv. Let us denote with τ (q, q(cid:48)) the estimated time to go from a location q to q(cid:48). Then the estimated time at which the stop-point spi will be served is: i(cid:88) ti = tnow + τ (qv, q0) + [bj−1 + τ (qj−1, qj)] + bi j=1 The estimated delay of each stop-point di = ti − ti ≤ ∆ti, for i = 0, . . . , n. The provider must only compute feasible schedules Sv for each vehicle v in the fleet. AMoDSim is able to simulate on-line optimization algorithms, in which the schedules are continuously modified. To avoid violating some user constraints, the feasibility should be checked at any modification. For example, suppose we modify Sv by inserting a new stop-point sp = (q, t, ∆t) at position k, obtaining a new schedule S(k) v = [sp0, . . . , spk−1, sp, spk, . . . , spn]. The detour the vehicle does to serve sp determines an additional delay on all the stop-points after the k- th. If we denote with t(k) the estimated stop-point time of spi after the insertion, i 6 Di Maria et al. (cid:40) the additional delay is ∆d(k) i ≡ t(k) i − ti and it is easy to show that: ∆d(k) i = 0, τ (qk−1, q) + b + τ (q, qk) − τ (qk−1, qk), if i < k if i ≥ k where b is the time for alighting or boarding related to sp. To check whether the modified schedule is feasible, not only must we check that the time constraints of the new sp are satisfied, but also that the time constrains are satisfied for i ≤ ∆ti for all the stop-points already present in the schedules, i.e., di + ∆d(k) i = 0, . . . , n. 3.2 Examples of optimization strategies To give a more concrete idea of the model we discussed in the previous section, we now describe two heuristics we implemented in AMoDSim and some possible assumptions about the request constraints expressed by users. We adopt such heuristics and assumptions in the case study of Sec. 5. However, we emphasize that the simulator is more general and can be used in different ways. Recall a request sent by a user is composed by a stop-point sp = (q, t, ∆t) for the pick-up and another sp(cid:48) = (q(cid:48), t(cid:48), ∆t(cid:48)) for the drop-off. We assume that the user would like to be picked-up immediately, i.e., t = tnow and to be dropped- off as in the ideal case in which a vehicle is immediately at her disposal and can bring her to the destination in the shortest path, without detours, i.e., t(cid:48) = tnow + τ (q, q(cid:48)). We implement two optimization strategies, namely Radio-Taxi and Insertion Heuristic. With the former each vehicle can serve one passenger at a time, while the latter allows ride sharing, i.e., the same vehicle can serve multiple passengers at a time. as defined in Sec. 3.1, of all its stop-points, i.e., c(Sv) =(cid:80)n We first describe the Insertion Heuristic, loosely inspired by [15]. The cost function c(Sv) is chosen in order to represent the user experience. More precisely, the cost of a schedule Sv = [sp0, . . . , spn] is the sum of the estimated delays di, i=0 di. The Insertion Heuristic attempts to minimize the marginal cost when serving an additional request. Suppose a new request is sent, consisting of the stop-points sp, sp(cid:48) for the pick-up and drop-off, respectively. Assigning the new request to any vehicle, will increase the cost of its schedule, i.e., the sum of the delays suffered by its stop-points. Let us take any vehicle v and denote with S(k,k(cid:48)) the schedule obtained from Sv by inserting the pick-up sp in the k-th position and the drop- off sp(cid:48) in the k(cid:48)-th position, with k(cid:48) > k. If the modified schedule is infeasible, we set c(S(k,k(cid:48)) = ∞). We compute the best placement of drop-off and pick-up, which minimizes this increase in cost, i.e., v v (cid:16) (cid:17) (kv, k(cid:48)v) = arg min (k,k(cid:48)),k(cid:48)>k c(S(k,k(cid:48)) v ) − c(Sv) AMoDSim: An Efficient and Modular Simulation Framework 7 We repeat the same computation for all the vehicles and we choose the one whose marginal cost is minimum, i.e.: (cid:16) v∗ = arg min v (cid:17) c(S(kv,k(cid:48)v) v ) − c(Sv) Finally, we assign the request to vehicle v∗ and place the pick-up and drop-off in the kv∗-th and k(cid:48)v∗-th positions, respectively. The Radio-Taxi strategy is a constrained version of Insertion Heuristic, in that we impose that each pick-up be followed in any schedule by the correspon- dent drop-off, which ensures that at most one passenger is in the vehicle at any moment. 3.3 Vehicle Movement All vehicles travel through the links of the network, i.e., roads, at a predefined cruising speed. Each link has a length, which determines the time needed to tra- verse it. Obviously, when a vehicle alternates between a stop-point and another, its speed does not go from 0 to the cruising speed and back to 0 instantaneously. Therefore, we introduce a parameter ta (td), which represents the time lost for accelerating (decelerating). When a vehicle reaches a stop-point spi, we keep it in that node for an additional time bi + ta + td, before sending it again to the link toward the next stop-point. 4 Software Architecture AMoDSim is a simulation platform developed on top of Omnetpp[3]. It is de- signed to be configurable, modular, event-based, algorithm-oriented and exten- sible with custom optimization strategies and network topologies. The simulator models the road network as a set of nodes, i.e., geographical lo- cations that could be origins and destinations of the service requests, connected through links, i.e., road connections between different locations. A vehicle is rep- resented as a packet traveling through the links. A node is a compound-module composed of three sub-modules: queue, routing and application. A node has one queue module per each outgoing or incoming link. Each Queue module forwards (receives) packets to one of the outgoing links (from one of the incoming links). The Routing module (i) decides to which of the outgoing links a packet should be forwarded and (ii) checks, every time a vehicle passes, whether the node is one of its stop-points, in which case the vehicle is passed to the Application module. The Application module implements multiple functions: -- It generates user requests, as pairs of stop-points (one for the pick-up and one for the drop-offs). The generation obeys to a pre-determined stochastic process. So far, Poisson arrivals are implemented. -- It receives all the vehicles for which the node in question is a stop-point, checks the next stop-point, accessing a data-structure storing all the sched- ules and sends the vehicle to it. At the same time, it also notifies the coor- dinator, so that it can update the schedule in question. 8 Di Maria et al. -- It keeps the vehicles that are idling at the node with an empty schedule. In this case, it also receives a signal from the coordinator if a new schedule is assigned to the idling vehicles and sends them to their new stop-point. Fig. 1: A trip example The Coordinator manages the incoming trip requests, implements the trip allocation strategies and assigns each request to a vehicle, according to the imple- mented optimization strategy. It has been designed to be easily extensible with custom allocation strategies. We implemented a modular Coordinator within a hierarchical structure where the superclass implements the standard functions. One can extend such superclass and implement the logic of her matching algo- rithm. 4.1 AMoD Performance Metrics AMoDSim collects data during its execution and produces a set of results that enable statistical analysis related to both the point of view of the provider and of users. Regarding the provider viewpoint, AMoDSim provides the following infor- mation per-vehicle: (i) distance traveled, (ii) number of passengers on board, (iii) requests picked-up but not yet dropped-off, (iv) number of pick-ups already in the schedule but not yet completed, (v) total requests assigned, (vi) the time the vehicle has spent idle or with p passengers, where p ranges from 1 to the number of per-vehicle seats. Moreover, for each of the collected metric, AMoDSim computes aggregated fleet statistics, as sum, minimum, maximum, mean, standard deviation, median and 95th percentile. At each time frame, the following information about the users' requests re- ceived up to that time are collected: (i) length of the submitted requests, (ii) AMoDSim: An Efficient and Modular Simulation Framework 9 number of requests that the system has received and assigned to the vehicles, (iii) number of requests that the system has rejected because it could not serve them within the time constraints, (iv) number of requests that the system is processing at the snapshot time. The level of quality for the users is described by the following per-user quan- tities: (i) time that users spent in the pick-up location waiting for the vehicle, (ii) actual time that the user spent in the vehicle, (iii) Stretch, i.e., the ratio between the actual trip time and the preferred one, which is the time between the preferred pick-up and drop-off times. 5 Case Study We showcase the capabilities of AMoD in a simple case study, in which we launched a campaign of 1800 simulations. We compare the performance of the Insertion Heuristic to the Radio-Taxi. We show how AMoDSim allows to find interesting insights on the AMoD systems and answer questions like: what is the fleet size needed to sustain a certain request rate? Which kind of vehicles should be employed (of how many seats)? What is the sharing level, i.e., how effectively are we able to condensate different user rides in few vehicle schedules? By how much sharing rides allows to reduce the fleet size needed? How efficient is vehicle usage, e.g., how much time vehicles are idle? We underline that the findings we get are not necessarily general properties of every AMoD systems, but depend on the particular optimization strategy we adopt and the particular scenario. Therefore, our goal is to show how other researchers can obtain similar find- ings with AMoDSim about their strategies and their scenarios. Finally, we show the computational performance of AMoDSim. We are aware that the quality of AMoDSim cannot be validated only by the case study we present here. Part of our future work is to apply AMoDSim to different scenarios and to validate by comparing it with other simulators. This latter point requires careful thinking, since other simulators are not directly comparable, for the reasons discussed in Section 1. We also believe that the best way to make AMoDSim reach full ma- turity is its adoption by other researchers for their studies, which would help in understanding and improving its limits. 5.1 Scenario We use Manhattan Grid that covers an area of 60km2, equivalent to Manhattan, with static link travel times as in [14]. We consider different configurations of the fleet of vehicles to study the performance of multiple ride-sharing degrees and fleet size. We perform simulations starting from single-seater up to 10-seater minibus and a fleet of 500 up to 9000 vehicles. We assume a cruising speed of 35kmph and a constant acceleration and deceleration of 1.676 mpss, resulting in a ta + td = 11.5 (see Sec.3.3) as in [21]. Thus, the vehicles have a constant acceleration (deceleration) of 1.676mpss (−1.676mpss). Users submit requests with Poissonian arrivals as in [16] with rate ranging from 20 up to 640 requests 10 Di Maria et al. per hour per km2 compatible with the scenarios employed in the literature [7,15]. As for the bi of a pick-up (drop-off) stop point spi, i.e. the time need for boarding (alighting), we assume 5 seconds (10 seconds) as in [13]. All results are collected running 4h simulations. 5.2 Results In this section, we first give an example of analysis possible in AMoDSim and then discuss its computational performance. Sharing opportunities for an AMoD provider. We investigate the factors determining the sharing degree and its impact on the provider and the users. The sharing degree is the capacity of an AMoD provider to exploit the fact that a single resource (vehicle) can be used to serve multiple requests. This concept, at the core of the sharing economy, cannot be quantified in a single value, but emerges from a set of different indicators that we discuss here. Fig. 2 shows the performance of Radio-Taxi. It is clear that the system is saturated: only 35K requests are served over 65K and the number of idle vehicles goes down to zero in few minutes. Fig. 3 shows that under the same conditions, Insertion Heurisitc with a fleet of 4-seater 2K vehicles allows to meet all the requests. Observe also that the total number of kilometers traveled, a proxy for the provider cost, decreases considerably by increasing the number of seats, since the sharing opportunities increase. Fig. 2: RadioTaxi: maximum extra-time ∆t = 15min, 2K vehicles. In the left and middle figure, the rate is 320req/h/Km2 The sharing degree is well summarized by Fig.4, which shows the fraction of time vehicle spend, on average, with 0 (idle), 1, 2, ... passengers. Intuitively, if we allow users to express a tight extra-time constraint ∆t, the sharing opportunities shrink and we can just afford few passengers at a time, in order to meet the constraints of all of them. Note that, even with a long ∆t, more than 6 seats are rarely utilized. This suggests that, if we want to implement a minibus-like service, strategies different from Insertion Heuristic must be used (which is an interesting subject to investi- gate). Observe also that high capacity vehicles would be fully utilized only if ∆t is too tight. In other words the type of vehicles to be used depends on the type of service that the provider wishes to offer and the level of service users expect. AMoDSim: An Efficient and Modular Simulation Framework 11 Insertion Heuristic: ∆t = 15min. Fig. 3: 320req/h/Km2 and 2K 4-seater vehicles are used. In the left and middle plots, (a) ∆t = 5min (b) ∆t = 15min (c) ∆t = 30min Fig. 4: Vehicle occupancy with 1K 10-seater vehicles and a rate of requests 320 per hour per km2 Mean waiting time. In a RadioTaxi-based AMoD system, the only way to serve a higher service demand is to increase the fleet size. Moreover, a large fleet reduces the Waiting Time (WT), which is shown in Fig. 5a. With the Insertion Heuristic another parameter impacts the user experience, namely the vehicle seats. In Fig. 5b we use large points to indicate the first value of request rate in which we observed the system is in saturation, i.e., it is not able to serve all the requests, e.g., Fig.2. Observe that when the system is not saturated, the best WT are measured with 1 seater vehicles, since each is dedicated entirely to a single user each time and the user does not make detours due to sharing with others. However, the system saturates at only 160req/h/km2. On the contrary, larger vehicles allow to serve a more intense demand without saturation, which translates in a better WT for the users. Computation time and memory consumption. In this section we dis- cuss the single-run computation time and the peak memory consumption of AMoDSim, which we observed in our case study. Note that comparison with other simulators is not possible here for the reasons discussed in Sec.2: the case- specific simulators are not available and the transportation simulators are out of scope and would have required input data that do not exist for the scenar- 12 Di Maria et al. (a) Radio-Taxi (b) Heuristic: vehicles=2K Fig. 5: Mean waiting time with a maximum delay ∆t = 10 minutes. ios considered. Fig. 6a and 6b show how both the computation time and the memory consumption grow with the number of vehicles and the rate of requests, as expected. Fig. 6c shows how the increase in computation time is significant moving from single-seater to 2-seater vehicles and is low moving from 4-seater to 10-seater. This may be due to the fact that vehicles spend most of the time with no more than 4 passengers anyway (Fig.4). (a) rate=160, seater=4 (b) vehicles=2K, seater=4 (c) rate=160, vehicles=2K Fig. 6: Computation time and Memory consumption: ∆t = 15min 6 Conclusion NOT IN THIS DRAFT 7 Acknowledgement NOT IN THIS DRAFT. References 1. BlaBla Car. https://www.blablacar.com/. 2. MESO: Mesoscopic version of SUMO. http://sumo.dlr.de/wiki/MESO. 3. OMNeT++. https://www.omnetpp.org/. 4. AA.VV. Vehicle Routing. SIAM-MOS, 2nd edition, 2014. 5. N. A.H. Agatz, A. L. Erera, et al. Dynamic ride-sharing: A simulation study in metro Atlanta. Transport Res B-Meth, 45 '(9):1450 -- 1464, 2011. AMoDSim: An Efficient and Modular Simulation Framework 13 6. Sabina Alazzawi, Mathias Hummel, Pascal Kordt, Thorsten Sickenberger, Chris- tian Wieseotte, and Oliver Wohak. Simulating the Impact of Shared , Autonomous Vehicles on Urban Mobility - A Case Study of Milan. In SUMO User Conference, 2018. 7. J. Alonso-Mora, S. Samaranayake, et al. On-demand high-capacity ride-sharing via dynamic trip-vehicle assignment. PNAS, 114(3):462 -- 467, 2017. 8. R. Basu, A. Araldo, et al. Implementation and Policy Applications of AMoD in multi-modal activity-driven agent-based urban simulator SimMobility. Transport Res Rec, 2018. 9. J. Bischoff and M. Maciejewski. Simulation of City-wide Replacement of Private Cars with Autonomous Taxis in Berlin. In ANT. Elsevier Masson SAS, 2016. 10. P. M. Boesch, F. Ciari, et al. Autonomous Vehicle Fleet Sizes Required to Serve Different Levels of Demand. Transport Res Rec, 2542:111 -- 119, 2016. 11. S. Burgstaller, D. Flowers, et al. Rethinking Mobility: The 'pay as you go' car: Ride hailing just the start. Technical report, 2017. 12. R. R. Clewlow and G. S. Mishra. Disruptive Transportation: The Adoption, Uti- lization, and Impacts of Ride-Hailing in the United States. Technical report, UC Davis, 2017. 13. J. Elpern-Waxman. Transportation Terms: Dwell Time, 2017. 14. M. Hyland and H. Mahmassani. Dynamic Autonomous Vehicle Fleet Operations: Optimization-Based Strategies to Assign AVs to Immediate Traveler Demand Re- quests. Transport Res. C-Emer, 92:278 -- 297, 2018. 15. J. Jaeyoung, R. Jayakrishnan, et al. Design and Modeling of Real-time Shared-Taxi Dispatch Algorithms. TRB 92nd Annual Meeting, 2013. 16. J. Jung, R. Jayakrishnan, et al. Design and Modeling of Real-time Shared-Taxi Dispatch Algorithms. In TRB Annual Meeting, volume 8, 2013. 17. A. Y. S. Lam, Y. Leung, et al. Autonomous-Vehicle Public Transportation System: Scheduling and Admission Control. IEEE Trans. Intell. Transp. Syst., 17(5):1210 -- 1226, 2016. 18. M. P. Linares, L. Montero, et al. A Simulation Framework for Real-Time Assess- ment of Dynamic ride sharing demand responsive transportation models. In WSC, 2016. 19. S. Ma, Y. Zheng, et al. T-Share : A Large-Scale Dynamic Taxi Ridesharing. In ICDE, 2013. 20. L. M. Martinez, G. H. A. Correia, et al. An agent-based simulation model to assess the impacts of introducing a shared-taxi system: an application to Lisbon. JAT, 49:475 -- 495, 2015. 21. S. Robinson. Measuring bus stop dwell time and time lost serving stop with london ibus automatic vehicle location data. Transport Res Rec, 2352(1):68 -- 75, 2013. 22. P. Santi, G. Resta, et al. Quantifying the benefits of vehicle pooling with share- ability networks. Proc. Natl. Acad. Sci., 111(37):13290 -- 4, 2014. 23. M M Vazifeh, P. Santi, et al. Addressing the minimum fleet problem in on-demand urban mobility. Nature, 557(May), 2018.
1510.06587	2	1510	2016-08-30T15:15:39	Approximating Strategic Abilities under Imperfect Information: a Naive Approach	[ "cs.MA", "cs.LO" ]	Alternating-time temporal logic (ATL) allows to specify requirements on abilities that different agents should (or should not) possess in a multi-agent system. However, model checking ATL specifications in realistic systems is computationally hard. In particular, if the agents have imperfect information about the global state of the system, the complexity ranges from Delta2P to undecidable, depending on the syntactic and semantic details. The problem is also hard in practice, as evidenced by several recent attempts to tackle it. On the other hand, model checking of alternating epistemic mu-calculus can have a distinctly lower computational complexity. In this work, we look at the idea of approximating the former problem by the verification of its "naive" translations to the latter. In other words, we look at what happens when one uses the (incorrect) fixpoint algorithm to verify formulae of ATL with imperfect information.	cs.MA	cs	Approximating Strategic Abilities under Imperfect Information: a Naive Approach Wojciech Jamroga, Michal Knapik, and Damian Kurpiewski Institute of Computer Science, Polish Academy of Sciences, Poland {w.jamroga,michal.knapik,damian.kurpiewski}@ipipan.waw.pl Abstract. Alternating-time temporal logic (ATL) allows to specify re- quirements on abilities that different agents should (or should not) pos- sess in a multi-agent system. However, model checking ATL specifica- tions in realistic systems is computationally hard. In particular, if the agents have imperfect information about the global state of the system, the complexity ranges from ∆P 2 to undecidable, depending on the syn- tactic and semantic details. The problem is also hard in practice, as evi- denced by several recent attempts to tackle it. On the other hand, model checking of alternating epistemic mu-calculus can have a distinctly lower computational complexity. In this work, we look at the idea of approxi- mating the former problem by the verification of its "naive" translations to the latter. In other words, we look at what happens when one uses the (incorrect) fixpoint algorithm to verify formulae of ATL with imperfect information. 1 Introduction There is a growing number of works that study syntactic and semantic vari- ants of strategic logics, in particular the alternating-time temporal logic ATL. Conceptually, the most interesting strand builds upon reasoning about temporal patterns and outcomes of strategic play, limited by information available to the agents. The contributions are mainly theoretical, and include results concerning the conceptual soundness of a given semantics of ability [17,9,1,13], meta-logical properties [4], and the complexity of model checking [17,11,10]. However, there is very little research on the actual use of the logics, in particular on practical algorithms for reasoning and/or verification. This is somewhat easy to understand, since model checking of ATL vari- ants with imperfect information has been proved ∆P 2 - to PSPACE-complete for agents playing positional (a.k.a. memoryless) strategies [17,11] and undecidable for agents with perfect recall of the past [7]. Moreover, the imperfect informa- tion semantics of ATL does not admit fixpoint equivalences [4], which makes incremental synthesis of strategies impossible, or at least difficult to achieve. Some practical attempts at tackling the problem started to emerge only re- cently [15,5,8]. Up until now, experimental results confirm that the initial intu- ition was right: model checking strategic modalities for imperfect information is hard, and dealing with it requires innovative algorithms and verification tech- niques. One idea that has not been properly explored is that of alternating-time epistemic mu-calculus (AEµC) [3]. Since fixpoint equivalences do not hold un- der imperfect information, it follows that standard fixpoint translations of ATL modalities lead to a different interpretation of strategic ability. In fact, it can be argued that they capture existence of recomputable winning strategies. However, what especially interests us in the context of model checking is that they can make model checking computationally cheaper. Verification of AEµC is in gen- eral between NP and ∆P 2 , but the scope of backtracking is much smaller than for ATL with imperfect information, as it includes only the actions starting from a given indistinguishability class rather than all the actions in the model. Moreover, for coalitions of up to 2 agents the model checking problem is in P [3]. The question that we ask in this paper is: Is AEµC an attractive alternative for verification of strategic abilities under imperfect information? To this end, we will look at the naive AEµC approximations of formulae of ATLir (i.e., ATL with imperfect information and imperfect recall), and investigate: 1. Whether model checking of the AEµC approximations performs significantly faster than for the original ATLir formulae; 2. Whether the AEµC counterparts are indeed semantic approximations of the ATLir specifications, or they encapsulate a completely different concept of ability. As fixpoint equivalences are not valid for ATLir, we know that the naive fixpoint translation is in general incorrect. However, one can possibly ask: how often? We take on an empirical approach. More precisely, we consider two classes of benchmark models and formulae, one based on the Tian Ji scenario [14,5] and the other being the Castles benchmark from [15]. Then, for a formula ϕ, we compare the output and performance of the ATLir model checking of ϕ with the AEµC model checking of aemc(ϕ), i.e., with the straightforward (and generally incorrect) fixpoints approximation of ϕ. The work reported here is very preliminary, but it already allows to draw some conclusions, and decide on the most promising lines for future research. 2 What Agents Can Achieve under Imperfect Information In this section we provide a brief overview of the relevant variants of ATL, and the corresponding complexity results for model checking. We refer the interested reader to [2,17,3] for details. 2.1 Models The semantics for ATL is defined over a variant of transition systems where transitions are labeled with combinations of actions, one per agent. An im- perfect information concurrent game structure (ICGS) [2,17] is given by M = qoi (o ut,in) 1 (o ut,o ut) ( i n , o u t ) out1 out2 out1 in2 ( o u ( i n , i n ) 2 t , i n ) qoo (in,in) qii (out,out) qc ( i n , o u t ) 2 ( o u t , o u t ) in1 in2 in1 in2 collision ( o u t , i n ) qio (in,in) in1 out2 1 (in,o ut) Fig. 1. Autonomous vehicles at the intersection: model M1 hAgt, St, Π, π, Act, d, o, {∼a a ∈ Agt}i which includes a nonempty finite set of all agents Agt = {1, . . . , k}, a nonempty set of states St, a set of atomic propo- sitions Π and their valuation π : Prop → 2St, and a nonempty finite set of (atomic) actions Act. Function d : Agt × St → 2Act defines nonempty sets of actions available to agents at each state, and o is a (deterministic) transition function that assigns the outcome state q′ = o(q, α1, . . . , αk) to state q and a tuple of actions hα1, . . . , αki for αi ∈ d(i, q) and 1 ≤ i ≤ k, that can be executed by Agt in q. We write da(q) instead of d(a, q). Each ∼a⊆ St×St is an equivalence relation satisfying da(q) = da(q′) for q ∼a q′. Note that perfect information can be modeled by assuming each ∼a to be the minimal reflexive relation. Example 1 (Intersection with limited visibility). Consider an intersection with k autonomous vehicles around it. Each vehicle is modeled as a separate agent, whose local state is characterized by either the proposition outi (when the vehicle is outside the intersection) or ini (when the vehicle is inside it). The available actions are: in ("drive in" or "stay in", depending on the current state) and out ("drive out" or "stay out"). Transitions update the state accordingly, except for one case: when both agents are in and decide to leave at the same time, a collision occurs (collision). Furthermore, let us assume that no agent sees the location of the other vehicle. Figure 1 presents a pointed ICGS modeling the scenario for k = 2. The combinations of actions that are not displayed in the graph do not change the state of the system. The indistinguishability relations are depicted by dotted lines. ⊓⊔ A strategy of agent a is a conditional plan that specifies what a is going to do in each situation. Here, we only refer to memoryless uniform strategies (ir strategies in short), defined as functions sa : St → Act such that sa(q) ∈ da(q) for all q, and q ∼a q′ implies sa(q) = sa(q′). A collective strategy sA is a tuple of ir strategies, one per agent from A. A path λ = q0q1q2 . . . is an infinite sequence of states such that there is a transition between each qi, qi+1. We use λ[i] to denote the ith position on path λ (starting from i = 0) and λ[i, ∞] to denote the subpath of λ starting from i. Function outM (q, sA) returns the set of all paths that can result from the execution of strategy sA from state q in model M , defined formally as follows: out(q, sA) = {λ = q0, q1, q2 . . . q0 = q and for each i = 0, 1, . . . there exists a = sA[a](qi) a ∈ da(qi) for every a ∈ Agt, and αi a1 , . . . , αi hαi for every a ∈ A, and qi+1 = o(qi, αi ak i such that αi a1 , . . . , αi ak )}. Moreover, we define outir subscript M if it is clear from the context. M (q, sA) = Sa∈A Sq∼aq′ out M (q′, sA). We will omit the 2.2 Alternating Time Temporal Logic Let Agt be the set of agents and Prop be the set of atomic propositions. The language of ATL is given by the following grammar: ϕ ::= p ¬ϕ ϕ ∧ ϕ hhAiiXϕ hhAiiGϕ hhAiiϕ U ϕ where A ⊆ Agt and p ∈ Prop. Additionally, we define "sometime in the future" as Fϕ ≡ ⊤ U ϕ. The semantics of ATLir is defined by the following clauses: M, q = p iff q ∈ π(p) M, q = ¬ϕ iff M, q 6= ϕ; M, q = ϕ ∧ ψ iff M, q = ϕ and M, q = ψ; M, q = hhAiiXϕ iff there is a collective strategy sA such that, for each λ ∈ (where p ∈ Π); outir(q, sA), M, λ[1] = ϕ; M, q = hhAiiGϕ iff there is a collective strategy sA such that, for each λ ∈ outir(q, sA) and every i ≥ 0, M, λ[i] = ϕ; M, q = hhAiiϕ U ψ iff there is sA such that, for each λ ∈ outir(q, sA), there is i ≥ 0 for which M, λ[i] = ψ, and M, λ[j] = ϕ for each 0 ≤ j < i. Informally, M, q = hhAiiγ iff there exists a strategy for A such that γ holds on all the paths that the agents in A consider as possible executions of the strategy. Example 2 (Intersection with limited visibility, ctd.). Take model M1 from Ex- ample 1. Now, we have e.g. that M1, qoo = hh1iiG¬collision (it suffices that agent 1 executes action "out" regardless of anything). On the other hand, M1, qoo = ¬hh1, 2iiFcollision (the agents do not know how to make sure that a collision will happen, even if they want to). We leave it up to the interested reader to check the latter. ⊓⊔ 2.3 Verification of Strategic Abilities The model checking problem asks, given a model M , a state q in it, and a logical formula ϕ, whether ϕ holds in M, q. ATL verification is known to be tractable for perfect information models, but intractable for imperfect information. Proposition 1 ([17,12]). Model checking of ATLir is ∆P ber of states and transitions in the model, and the length of the formula. 2 -complete in the num- 2.4 Alternating Epistemic Mu-Calculus Alternating epistemic µ-calculus (AEµC) replaces the temporal-strategic oper- ators hhAiiG, hhAii U with the least fixpoint operator µ [3]: ϕ ::= p Z ¬ϕ ϕ ∧ ϕ hhAiiXϕ µZ.ϕ where Z ∈ Var is a second order variable ranging over sets of states. The great- est fixed point operator ν can be defined as dual to µ. We consider only the alternation-free fragment of AEµC, cf. [2,3] for the exact definition. A valuation of Var is a mapping V : Var → 2St. Given a set Q ⊆ St of states we define V[Z := Q] as the update of V that assigns Q to Z. The semantics of AEµC is given by the denotation function [[ϕ]]M V which defines the set of states in M that satisfy ϕ in the following way: [[p]]M V = π(p), [[Z]]M V = V(Z), [[¬ϕ]]M V = {q ∃αA ∈ dA(q)∀αAgt\A ∈ dAgt\A : o(q, (αA, αAgt\A)) ∈ [[ϕ]]M V = T{Q ⊆ St [[ϕ]]M V[Z:=Q] ⊆ Q}. Moreover, M, q = ϕ iff q ∈ [[ϕ]]M V for all valuations V. V = [[ϕ]]M V ∩ [[ψ]]M V = St \ [[ϕ]]M V , [[ϕ ∧ ψ]]M V , [[hhAiiXϕ]]M V }, and [[µZ.ϕ]]M Proposition 2 ([3]). Model checking of AEµC is P-complete if all the coali- tions in ϕ consist of at most 2 agents. For abilities of coalitions with 3 or more agents, it is between NP and ∆P 2 in the size of the largest abstraction class of relations ∼1, . . . , ∼k. Thus, verification of AEµC is potentially more attractive than ATLir. A natural idea is to use the naive translation of ATLir to AEµC, defined as follows: aemc(p) = p aemc(¬ϕ) = ¬aemc(ϕ) aemc(ϕ ∧ ψ) = aemc(ϕ) ∧ aemc(ψ) aemc(hhAiiXϕ) = hhAiiXaemc(ϕ) aemc(hhAiiGϕ) = νZ.(aemc(ϕ) ∧ hhAiiXZ) aemc(hhAiiFϕ) = µZ.(aemc(ϕ) ∨ hhAiiXZ) aemc(hhAiiϕ U ψ) = µZ.(aemc(ψ) ∨ aemc(ϕ) ∧ hhAiiXZ). We will look at how it works in the subsequent sections. Note that, in a way, aemc(hhAiiFϕ) expresses a persistent ability to achieve ϕ. Likewise, aemc(hhAiiGϕ expresses a persistent ability to maintain ϕ. This is because aemc(hhAiiFϕ) produces a strategy for A such that aemc(hhAiiFϕ) will also hold for every state reachable by the strategy. For aemc(hhAiiGϕ) the situation is analogous. 3 Fixpoint Approximation of Strategic Ability: Performance In this section, we empirically compare the performance of model checking ATLir specifications vs. their naive AEµC approximations. The experimental results on the AEµC side have been obtained by running a straightforward implementation of the fixpoint model checking algorithm, im- plemented in Python 3. The tests have been conducted on a MacBook with an Intel Core i5 CPU with dynamic clock speed of 1.4 GHz, 4 GB of RAM (one module DDR3, 1600 MHz buz clock), and OS X 10.10.5 Yosemite. The performance of ATLir model checking for the Castles model is cited after [16], and was obtained on a notebook with an Intel Core i7-3630QM CPU with dynamic clock speed of 2.4 GHz up to 3.4 GHz, and 8 GB of RAM (two modules DDR3 PC3-12800, 800 MHz bus clock, effective data rate 1600 MT/s, in dual-channel configuration). Two model checkers were used: SMC and MCMAS. The experiments with SMC were conducted on Windows 7 OS, the experiments with MCMAS on Linux Ubuntu 12.04.2. Thus, the ATLir model checking for Castles was performed on a significantly better computing equipment than the AEµC verification. We also note that SMC uses several reduction techniques to restrict the search space, and MCMAS operates on compact symbolic represen- tations of models, based on BDD's. In contrast, our model checking of AEµC was done with a straightforward implementation of the standard explicit state algorithm with no optimizations at all. The performance of ATLir model checking for the TianJi model is cited after [5], and was obtained with an experimental model checker implemented with PyNuSMV, a Python framework for prototyping and experimenting with BDD-based model-checking algorithms based on NuSMV [6].1 Besides compact symbolic representations of states and transitions, the model checker features multiple optimization techniques. The authors do not describe the computing configuration that was used for their experiments. The timeout in all cases is defined as 120 minutes. 3.1 Benchmark 1: Castles The Castles model have been proposed in [15]. The model consists of one agent called Environment that keeps track of the health points of three castles, plus 1 It should be mentioned that the results for TianJi in [5] were obtained for a slightly different semantics of ATLir, employing additional fairness constraints. AEµC (ψ ′ 1) ATLir/SMC(ψ1) ATLir/MCMAS(ψ1) Configuration 4 (1,1,1) 5 (1,1,2) 6 (2,1,2) 7 (2,2,2) 8 (3,2,2) time #sat #iter 0.011 0.024 0.386 9.231 128 256 512 1024 4352.891 5504 1 1 1 1 2 time timeout timeout timeout timeout timeout time 72 timeout timeout timeout timeout Fig. 2. Model checking performance for Castles, formula ψ ′ 1 vs. ψ1 AEµC (ψ ′ 2) ATLir/SMC(ψ2) ATLir/MCMAS(ψ2) Configuration time #sat #iter 4 (1,1,1) 5 (1,1,2) 6 (2,1,2) 7 (2,2,2) 8 (3,2,2) 8 0.004 16 0.015 32 0.050 0.225 64 1.202 128 1 1 1 1 1 time timeout timeout ? ? ? time 78 error ? ? ? Fig. 3. Model checking performance for Castles, formula ψ ′ of data about the performance of the given model checker on the given instance. 2 vs. ψ2. "?" indicates lack a number of agents called Workers each of whom works for the benefit of a castle. Health points (HP, ranging from 0 to 3) represent the current condition of the castle; 0 HP means that the castle is defeated. Workers can execute the following actions: attack a castle they do not work for, defend the castle they do work for, or do nothing. Doing nothing is the only available action to a Worker of a defeated castle. No agent can defend its castle twice in a row, it must wait one step before being able to defend again. A castle gets damaged if the number of attackers is greater than the number of defenders, and the damage is equal to the difference. In the initial state, all the castles have 3 HP and every Worker can engage in defending its castle. The indistinguishability relations for Workers are defined as follows. Every Worker knows if it can currently engage in defending its castle, and can observe for each castle if it is defeated or not. The model is parameterized by the number of agents and the allocation of Workers. For example, an instance with 1 worker assigned to the first castle, 3 workers assigned to the second and 4 to the third castle will be denoted by 9 (1,3,4). Formulae We considered the following formulas for Castles: ψ1 ≡ hhc12iiFcastle3defeated ψ2 ≡ hhw12iiFallDefeated The first formula says that the workers working for castles 1 and 2 have a collec- tive strategy to defeat castle 3, no matter what other agents do. Similarly, the second formula says that workers number 1 and 2 have a collective strategy to AEµC (φ′ 1) ATLir (φ1) Horses time #sat #iter 3 4 5 6 7 8 11 0.0002 18 0.001 153 0.014 300 0.024 0.753 2258 6.204 4900 2 2 3 3 4 4 time 2.603 8.205 30.885 99.931 586.126 ? Fig. 4. Model checking performance for TianJi, formula φ′ 1 vs. φ1 AEµC (φ′ 2) ATLir (φ2) Horses time #sat #iter 3 4 5 6 7 8 0.0002 0.001 0.003 0.004 0.016 0.075 0 0 0 0 0 0 2 2 2 2 2 2 time ≈ 2.5 ≈ 10 ≈ 650 ? ? ? Fig. 5. Model checking performance for TianJi, formula φ′ 2 vs. φ2 ensure the defeat of all castles. After the naive translation to AEµC we obtain: ψ′ ψ′ 1 ≡ µZ.(castle3defeated ∨ hhc12iiXZ) 2 ≡ µZ.(allDefeated ∨ hhw12iiXZ) Experimental Results The results for Castles are presented in Figures 2 and 3. The tables present results for a sequence of models of various sizes. The column headers are interpreted as follows: -- time: model checking time (in seconds), -- #sat: number of states in which the formula is satisfied, -- #iter: number of iterations until reaching fixpoint. Note: we only have data about #sat and #iter for the experiments with AEµC model checking. 3.2 Benchmark 2: TianJi The second series of experiments has been conducted for the TianJi variant from [5]. The model consists of two agents: Tian Ji and the king. Each agent has n horses numbered 1, . . . , n. In the game, Tian Ji and the king send their horses one by one against each other. Horse i of Tian Ji wins the race with king's horse j iff i > j. At each stage, the agents know the current score and their own remaining horses, but not those of the opponent. Moreover, the decisions at each round are made simultaneously, so one does not know which horse is currently sent by the other player. The player whose horses won most races wins the game. Configuration AEµC (ψ ′ 1) ATLir (ψ1) 4 (1,1,1) 5 (1,1,2) 6 (2,1,2) 6 (3,1,1) 7 (2,2,2) 8 (3,2,2) false false false true false true true true true true true true Fig. 6. Model checking output for Castles, formula ψ ′ 1 vs. ψ1 We considered the following formulas for TianJi: φ1 := hhTJ iiFTJWins φ2 := hhTJ iiGhhTJ iiXTJWonLess2 TJWins holds when the game is done and TianJi has won more races than the king. Similarly, TJWonLess2 is satisfied when TianJi has won at most 1 race up to the current point. After the naive translation we obtain: φ′ φ′ 1 := µZ.(TJWins ∨ hhTJ iiXZ) 2 := νZ.(hhTJ iiXTJWonLess2 ∧ hhTJ iiXZ) The results of experiments are presented in Figures 4 and 5. 3.3 Discussion The experimental results show that, for the instances of model checking that we have tested, verification of strategic abilities in alternating epistemic µ-calculus offers a dramatic speedup over model checking "standard" ATL with imperfect information. So, from the computational point of view, AEµC is definitely more attractive than ATLir. The speedup occurs despite the fact that the AEµC model checking was done by a straightforward implementation of the standard explicit state algorithm, while the results for ATLir were obtained by model checkers that use multiple optimization techniques and, in most cases, also com- pact symbolic representation of states and transitions in the model. We also note that AEµC model checking was distinctly faster than that of ATLir even when they were both (theoretically) NP-complete, i.e., for coalitions larger than 2 agents (cf. Figure 2). 4 Fixpoint Approximation of Strategic Ability: Semantics In this section, we empirically compare the output of model checking ATLir with that of model checking AEµC. That is, we compare the truth values of ATLir specifications in the benchmark models, versus the truth values of their naive fixpoint approximations. The output of AEµC model checking has been produced by our model checking algorithm. The truth values according to the ATLir semantics were determined by hand. Configuration AEµC (ψ ′ 2) ATLir (ψ2) 4 (1,1,1) 5 (1,1,2) 6 (2,1,2) 6 (3,1,1) 7 (2,2,2) 8 (3,2,2) false false false false false false false false false false false false Fig. 7. Model checking output for Castles, formula ψ ′ 2 vs. ψ2 Horses AEµC (φ′ 1) ATLir (φ1) 3 4 5 6 7 8 false false false false false false false false false false false false Fig. 8. Model checking output for TianJi, formula φ′ 1 vs. φ1 4.1 Benchmarks 1 & 2: Castles and TianJi The results for Castles are presented in Figures 6 and 7. The tables present the truth values of a given formula in the initial state of the benchmark model of a given size. Similarly, the output of model checking for TianJi is presented in Figures 8 and 9. 4.2 Benchmark 3: Modified TianJi The experiments with Castles have shown that the AEµC approximations cap- ture a much more restrictive notion of ability than the original ATLir specifica- tions. In this view, the results in Figures 8 and 9 are not very informative. Both semantics have produced the same truth values, but was it because they are indeed so close? Or rather because the ATLir semantics incidentally produced "false," i.e., the truth value that the fixpoint semantics seems to favor? To answer this question, we have prepared and executed an additional run of experiments based on a modification of the TianJi story. In "Modified TianJi," general TianJi always sees the horse selected by the king before sending his own horse to the next race. The modification significantly increases the strategic abilities of the general. The results are presented in Figure 10. 4.3 Discussion The experiments show that using AEµC to approximate model checking of ATLir in a straightforward way does not work. There is no correlation between the truth of the ATLir formulae that we have tested, and their naive AEµC Horses AEµC (φ′ 2) ATLir (φ2) 3 4 5 6 7 8 false false false false false false false false false false false false Fig. 9. Model checking output for TianJi, formula φ′ 2 vs. φ2 Horses AEµC (φ′ 1) ATLir (φ1) Horses AEµC (φ′ 2) ATLir (φ2) 3 4 5 6 7 8 false false false false false false true true true true true true 3 4 5 6 7 8 false false false false false false true true true true true true Fig. 10. Model checking output for Modified TianJi translations. In fact, the latter ones did not hold in an overwhelming majority of models that we looked at. This is due to the fact that persistent or recomputable strategic ability is a much stronger property than being able to come up with a winning strategy only in the initial state of the game. This suggests two possible ways of further study. One is to identify subclasses of concurrent game structures where recomputable strategies can be obtained. Models of agents with perfect recall seem a natural candidate in this respect. The other is to suitably weaken the AEµC translations so that they capture also existence of (some) non-recomputable strategies. We leave exploration of both paths for future research. 5 Conclusions In this paper, we have looked at verification of strategic abilities for agents with imperfect information. The aim was to investigate whether straightforward fixpoint approximations provide an interesting alternative to formulae of ATLir, for which model checking is known to be theoretically and practically hard. The answer is both yes and no. On one hand, our experimental results show that verification of "fixpoint abilities," specified in alternating epistemic µ-calculus, offers a dramatic speedup over model checking of ATLir. On the other hand, there is no correlation between satisfaction of the ATLir formulae that we have tested, and their naive AEµC translations. Thus, we conclude that AEµC is an attractive alternative to ATLir from the computational point of view, but it does not approximate model checking of ATLir in a straightforward way. Rather, it is underpinned by a distinctly different notion of ability, based on existence of persistent or recomputable strategies. In the future, we plan to identify subclasses of concurrent game structures where such recomputable strategies can be obtained. We will also investigate how to weaken the fixpoint translations so that they capture also existence of some non-recomputable strategies. References 1. T. Ågotnes. A note on syntactic characterization of incomplete information in ATEL. In Procedings of Workshop on Knowledge and Games, pages 34 -- 42, 2004. 2. R. Alur, T. A. Henzinger, and O. Kupferman. Alternating-time Temporal Logic. Journal of the ACM, 49:672 -- 713, 2002. 3. N. Bulling and W. Jamroga. Alternating epistemic mu-calculus. In Proceedings of IJCAI-11, pages 109 -- 114, 2011. 4. N. Bulling and W. Jamroga. Comparing variants of strategic ability: How uncer- tainty and memory influence general properties of games. Journal of Autonomous Agents and Multi-Agent Systems, 28(3):474 -- 518, 2014. 5. S. Busard, C. Pecheur, H. Qu, and F. Raimondi. Improving the model checking of strategies under partial observability and fairness constraints. In Formal Methods and Software Engineering, volume 8829 of Lecture Notes in Computer Science, pages 27 -- 42. Springer, 2014. 6. Simon Busard and Charles Pecheur. Pynusmv: Nusmv as a python library. In Proceedings of NASA Formal Methods, pages 453 -- 458, 2013. 7. C. Dima and F.L. Tiplea. Model-checking ATL under imperfect information and perfect recall semantics is undecidable. CoRR, abs/1102.4225, 2011. 8. X. Huang and R. van der Meyden. Symbolic model checking epistemic strategy logic. In Proceedings of AAAI, pages 1426 -- 1432, 2014. 9. W. Jamroga. Some remarks on alternating temporal epistemic logic. In B. Dunin- Keplicz and R. Verbrugge, editors, Proceedings of Formal Approaches to Multi- Agent Systems (FAMAS 2003), pages 133 -- 140, 2003. 10. W. Jamroga and T. Ågotnes. Modular interpreted systems: A preliminary report. Technical Report IfI-06-15, Clausthal University of Technology, 2006. 11. W. Jamroga and J. Dix. Model checking ATLir is indeed ∆P 2 -complete. In Proceed- ings of EUMAS'06, volume 223 of CEUR Workshop Proceedings. CEUR-WS.org, 2006. 12. W. Jamroga and J. Dix. Model checking abilities of agents: A closer look. Theory of Computing Systems, 42(3):366 -- 410, 2008. 13. W. Jamroga and W. van der Hoek. Agents that know how to play. Fundamenta Informaticae, 63(2 -- 3):185 -- 219, 2004. 14. A. Lomuscio, H. Qu, and F. Raimondi. MCMAS: An open-source model checker for the verification of multi-agent systems. International Journal on Software Tools for Technology Transfer, 2015. To appear. 15. J. Pilecki, M.A. Bednarczyk, and W. Jamroga. Synthesis and verification of uni- In Proceedings of CLIMA XV, volume form strategies for multi-agent systems. 8624 of Lecture Notes in Computer Science, pages 166 -- 182. Springer, 2014. 16. J. Pilecki, M.A. Bednarczyk, and W. Jamroga. Synthesis and verification of uni- form strategies for multi-agent systems. 2016. Journal version, under submission. 17. P. Y. Schobbens. Alternating-time logic with imperfect recall. Electronic Notes in Theoretical Computer Science, 85(2):82 -- 93, 2004.
1910.03101	1	1910	2019-10-07T21:43:22	Multi-Robot Coordinated Planning in Confined Environments under Kinematic Constraints	[ "cs.MA", "cs.RO" ]	We investigate the problem of multi-robot coordinated planning in environments where the robots may have to operate in close proximity to each other. We seek computationally efficient planners that ensure safe paths and adherence to kinematic constraints. We extend the central planner dRRT* with our variant, fast-dRRT (fdRRT), with the intention being to use in tight environments that lead to a high degree of coupling between robots. Our algorithm is empirically shown to achieve the trade-off between computational time and solution quality, especially in tight environments.	cs.MA	cs	Multi-Robot Coordinated Planning in Confined Environments under Kinematic Constraints Clayton Mangette and Pratap Tokekar 9 1 0 2 t c O 7 ] A M . s c [ 1 v 1 0 1 3 0 . 0 1 9 1 : v i X r a Abstract -- We investigate the problem of multi-robot co- ordinated planning in environments where the robots may have to operate in close proximity to each other. We seek computationally efficient planners that ensure safe paths and adherence to kinematic constraints. We extend the central planner dRRT* with our variant, fast-dRRT (fdRRT), with the intention being to use in tight environments that lead to a high degree of coupling between robots. Our algorithm is empirically shown to achieve the trade-off between computational time and solution quality, especially in tight environments. The software implementation is available online at https://github.com/ CMangette/Fast-dRRT. I. INTRODUCTION Computationally efficient multi-robot motion planning al- gorithms are highly sought after for their applications in in- dustry. In a time when automotive manufacturers are quickly approaching the advent of self-driving cars, centralized mo- tion planners in lieu of traditional traffic control structures open the possibility of increased traffic flow in busy urban environments, with studies in [9] and [22] supporting this. With an increase in automation in warehouses by companies like Amazon [1], efficient path planning of robots designed to move inventory in place of human workers has become another important use case. Beyond ground vehicles, traffic management of Unmanned Aerial Vehicles (UAVs) is iden- tified by NASA as an important area of research to ensure safe integration of aerial drones into the air space [11]. In each of the aforementioned applications, the algorithms used must be robust to planning in tight, confined envi- ronments while still ensuring that robots do not collide with one another. In the case of automated driving, urban traffic structures such as intersections and highway merging ramps constrain vehicles to a narrow set of paths. Similarly, warehouses limit robot paths due to shelving and storage units occupying the space. While not subject to high clutter, high volume air traffic can artificially restrict paths for UAVs. The planning algorithms available for such problems can be classified as centralized or decoupled. Centralized algo- rithms plan in the joint space of all robots whereas decoupled approaches only consider the space for each individual robot [12]. Decoupling interactions between robots that don't directly interact can simplify the original planning problem into a number of single-robot motion planning problems, making decoupled planners faster than centralized planners. C. Mangette is with the Department of Electrical and Computer Engi- neering, Virginia Tech, U.S.A. {mangettecj}@vt.edu of Maryland, U.S.A. {tokekar}@umd.edu P. Tokekar is with the Department of Computer Science at the University However, this can compromise completeness and allow inter- robot collisions [12]. Centralized planners, in comparison, can guarantee collision-free motions and completeness, but at the cost of solution time and scale-ability. If a decoupled planner considers a space of dimension RN for d robots, then a centralized algorithm plans over a joint space RN d. For our targeted applications, safety is of the utmost importance, so a centralized algorithm is better suited than a decoupled algorithm. Furthermore, centralized frameworks already exists in each use case. The intersection manager in [18] is a hypothetical replacement to traffic lights that can control when autonomous vehicles enter an intersection via Vehicle-to-Infrastructure (V2I) communications. A task allocation and path planning system in [8] demonstrates how to automate warehouse stock movement with kiva robots. The Unmanned Aerial System (UAS) Traffic Management (UTM) in development uses a centralized service supplier to manage requests and conflicts between UAVs operating within the same space [11]. Fig. 1: Targeted use cases: UAV coordination (left), traffic in- tersections (center), and warehouse motion planning (right). The main challenge in centralized planning is doing so in a time-efficient manner. A secondary challenge is extend- ing planning to robots with kinematic constraints, which complicates local path construction. This paper attempts solving both of these concerns by designing a framework for centralized planning in spaces with tight corridors with multiple kinematically constrained robots. State-of-the-art planners have progressed towards algo- rithms that increase efficiency while preserving complete- ness. Recognizing the shortcomings of previous algorithms that rely on explicit computation of the composite planning space, discrete RRT (dRRT) [15] and its optimal variant dRRT* [14] improve computational efficiency by offloading computations to offline tasks when possible and relying on implicit representations on the planning space. These algorithms do not encode steering constraints, but provide a general framework for fast multi-robot planning. This paper presents a variant to dRRT / dRRT* , which we call fast-dRRT (fdRRT), that returns fast but sub-optimal trajectories in tight environments requiring significant co- ordination between robots. We also extend these planners to account for robots with kinematic constraints. dRRT* and our algorithm are tested across multiple environments (Figure 1) and demonstrate fdRRT's increased computational efficiency in confined spaces. II. RELATED WORK Motion planning has been studied as one of the fun- damental problems in robotics. In the standard planning framework, a robot within a work space begins with a starting point s and goal point t, and the solution to the planning problem is to find a collision-free path connecting s and t. Grid-based methods such as Djikstra's algorithm [2] and A* [6] were developed as a means of finding shortest paths between vertices on a graph. Sampling-based motion planners became popular for their adaptability to different kinematic models and low cost by sampling points instead of searching exhaustively over the work space. A detailed review of sampled-based planning is provided in [4]. Extending motion planning to the multi-robot domain has been challenging due an increase in search space size and has led to a variety of approaches. Strategies are categorized in [21] to use cell decomposition, potential field navigation, roadmaps to plan efficient paths. Cell decomposition methods to path planning rely on discrete maps of the planning space to determine optimal paths. A sequential process in [23] splits the problem into local path planning using D* and coordination between robots to avoid entering collision re- gions simultaneously. Instead of handling spatial and velocity planning separately, Wagner and Choset developed M* , a multi-robot analogue to A* that resolves local path collisions by coupling paths only when they are found to overlap [19]. Although M* can plan paths for up to 100 robots, its performance suffers when high degrees of coupling between robots arise at choke points in the planning space. Yu and Lavalle optimize paths on a graph across various objectives and demonstrate the scaleability of their algorithm, but do not consider kinematic contraints in their models [24]. Roadmap strategies, in contrast, iteratively explore the work space instead of searching exhaustively. Van den Berg et al. provide a general framework for planning in a roadmap a sequential path planner that determines a sequential order- ing for each robot to execute its path [17]. It relies on a cou- pled motion planner for handling local connections between conflicting agents, so run time performance is dependent on the degree of coupling between robots. The coordinated path planner in [25] searches collision-free paths over an explicitly computed multi-robot work space, but is limited in scope due to the memory required to build an explicitly defined road map. Using the principle of sub-dimensional expansion, Wagner et al. designed sub-dimensional RRT (sRRT) and sub-dimensional PRM (sPRM) to plan paths for multiple robots with integrator dynamics [20], the latter using M* to query a multi-robot path. Solovey et al. also use sub-dimensional expansion in discrete RRT (dRRT) [15]. The idea of dRRT is to build road maps G = (G1, G2, ..., GN ) of collision-free motions for each robot, and then use them to build a search tree T = (V, E) implicitely embedded in G. dRRT draws samples from each road map and combines them into a composite sample, Qrand = (q1,rand, q2,rand, ...qN,rand), to which T is extended towards by selecting a composite neighboring vertex Vnew. Because G relies on pre-computed motions between configurations that have already been collision checked against environmental obstacles, dRRT can simply fetch the motions Ei ∈ Gi and check if any inter-robot collisions occur, thus relieving the algorithm of significant computational burden. Collision-free composite motions are added as vertices V to T until a goal is reached. The optimal variant of dRRT, dRRT, improves upon computation time further by carefully choosing neighbors to expand towards the goal state [14]. In addition to G , a path heuristic, H, is computed to identify configurations with short paths to Qf . This improves both solution quality and computational efficiency, making dRRT the one of the state-of-the-art algorithms in multi-robot planning. This paper presents a centralized planning strategy for kinematically constrained robots in tight environments. We demonstrate the feasibility of our kinematically constrained PRM algorithm in extending the pre-existing dRRT algo- rithm to the domain of planning under motion constraints. The central planning algorithm, which we call fast-dRRT (fdRRT), is designed to switch between randomly exploring the state space and driving greedily towards the goal state in a manner similar to dRRT. The difference in our algorithm is how expansion failures due to collisions are adjudicated. Instead of reporting an expansion failure if no collision- free connection can be established to a new node, fdRRT forces a connection by commanding some robots to stay in their previous configurations while permitting others to move forward. In practice, this makes fdRRT faster than dRRT in tight work spaces, but at the cost of solution quality. Unlike dRRT, our algorithm makes no guarantee of minimal path length, thus imposing an trade-off between solution efficiency and quality when choosing between the two algorithms. Additionally, fdRRT's incorporation of kine- matic constraints makes it a more flexible planner that can be used in different systems. III. PROBLEM FORMULATION The input to our problem is the set of start and goal positions for N robots. The two goals are to construct a local map for each robot encompassing feasible paths connecting a robot's local start and goal position, and to use these maps to construct trajectories for each robot that respect kinematic constraints and do not intersect other trajectories. Formally, given a set of initial configurations, Qinit = (q1,init, ..., qN,init) and final configurations, Qgoal = (q1,goal, ..., qN,goal), we would like to find a set of tra- jectories Π = (π1, ..., πN ), Π(0) = Qinit, Π(1) = Qgoal, such that all trajectories in Π are non-intersecting with obstacles and other robots. Time is not explicitly is part of the configuration space, but we assume that each instance of a configuration Q ∈ Π is uniformly discretized. Each robot is kinematically constrained by the motion model [ x, y, θ, κ]T = [cos(θ), sin(θ), κ, σ]T (1) For simplicity, we assume that each vehicle can only move forward. The dynamics in (1) are an extension of Dubins' steering constraints [3] that add curvature constraints. The sum of path lengths of the multi-robot trajectory is the cost metric chosen for evaluation. IV. ALGORITHM OVERVIEW Fig. 2: The central planner returns collision-free path queries by referencing pre-computed roadmaps from a local planner. Our system is illustrated in Figure 2. A local roadmap is constructed for each robot by the local planner that runs offline. The local roadmap is defined as a directed graph con- taining configurations within the robot's local configuration space and paths connecting configurations. The central planner receives path queries in the form of initial and final configurations and local roadmaps from the robots entering the planning space. To avoid re-planning due to new requests, the central planner accepts requests until a deadline Tdeadline and relegate new requests to the next planning cycle. Given the local roadmaps and initial and final configurations of each robot, the central planner returns composite path Π = (π1, π2, ..., πR) that guarantees collision free trajectories between robots. Each local trajectory is sent to its corresponding robot as a list of time-parameterized waypoints wi(t) = [xi(t), yi(t), θi(t), κi(t)]T and connect- ing paths πi(s) = [xi(s), yi(s), θi(s), κi(s)]T . The local controller on each robot determines the speed profile to follow from wi(t) and the distance travelled between consecutive waypoints. πi(s) is re-parameterized to πi(t) from the distance traveled over time, which can be tracked by a local controller using a technique such as pure- pursuit or nonlinear Model Predictive Control (MPC) [10]. A. Local Roadmaps Solovey et al. suggest using probabilistic roadmaps (PRMs) as approximations to local configuration spaces [15]. The original PRM algorithm builds a roadmap as an undirected graph G = (V, E), with each vertex v ∈ V being a unique configuration and each edge e(vi, vj) ∈ E a path in free space connecting two adjacent vertices vi and vj [7]. Configurations qrand are randomly sampled in the configuration space C and connected to any vertices in G within a connection distance r , {v ∈ V dist(qrand, v) ≤ r ∧ e(qrand, v) ∈ Cf ree)}. Construction of G continues for N iterations, after which paths between configurations are found during a query. This framework presents numerous challenges to adapting to a robot with kinematic constraints. Connections in [7] are line segments, which are sufficient under the assumption of single-integrator dynamics, but not for the dynamics in Equation (1). Numerical methods used in [5] capture both kinematic and dynamic constraints to connect two configurations in a kinematically-constrained system, but are approximate solutions. Dubins paths adhere to kinematic constraints and yield minimal path length for car-like robots [3], but require sharp changes in steering curvature that are not achievable in a real system. Scheuer and Fraichard extend Dubins paths to continuous curvature paths using clothoids to transition between changes in curvature that, while less computationally tractable than Dubins paths, are a feasible connection method [13]. Additionally, G in [7] is an undirected graph, implying that motions between connected vertices are bi-directional. Due to Dubins steering constraints and the non-holonomic constraints in (1), this is not necessarily true, and the existence of a collision-free path connecting two vertices vi to vj does not guarantee the reverse. To address this, Svestka and Overmars demonstrate that making G a directed graph is sufficient to impose this restriction [16]. Our local planner, kinematically-constrained PRM (KC- PRM), is similar to the Probabilistic Path Planner (PPP) in [16] with additional sampling and connection constraints to build a road map biased towards the optimal path that discriminates against unnecessary connections (Algorithm 1). G is initialized with an initial configuration qi (Line 1). A base path πsample is computed as the ideal path to follow from qi to qf and is used when sampling configurations (Line 2). G expands to size N by sampling random configurations qrand, attempting connections to vertices (Lines 6 -- 7), and adding connections to qrand when attempts are successful (Lines 8 -- 11). Details are provide below. RandomConfig: Random configurations are uniformly sampled along the sample path, qrand ∼ U(πsample) with additive Gaussian noise, N(0, σ) to allow for variation in qrand. The motivation behind sampling along πsample instead of the entire space is that one of the primary use cases is autonomous driving in urban environments. The space of locations that an autonomous vehicle can sample without 4 5 6 qrand ← RandomConfig(πsample); for v ∈ V do πlocal ← Steer(v, qrand); if IsReachable(v, qrand, πlocal, r) then Algorithm 1: LocalPlanner(qi, qf , N, r) 1 G ← qi; 2 πsample ← ReferencePath(qi, qf ); 3 while Size(G) < N do (G, vnew) ← Insert(qrand); end G ← Connect(v, vnew, πlocal); if qrand /∈ G then 7 8 9 10 11 12 13 14 end 15 H ← CostToGoal(G, qf ); 16 G ← PruneDeadNodes(G, H); 17 return G , H end end violating traffic norms such as staying within one's own lane is confined to the center of the lane with some allowable deviation, so qrand is restricted appropriately. Steer: Connections between adjacent vertices are at- tempted using the procedure described in [13]. Although more time consuming to compute than a Dubins' curve, this is an offline procedure, so efficiency is not a concern. IsReachable: A configuration q2 is defined to be reachable from q1 if two conditions are met: 1) the length of π12 ≤ r, the path connecting q1 to q2, is within the connection radius. 2) q2 is in front of q1. The planner in [5] imposes a similar condition by checking if q2 is in the half space of q1. We check this condition by computing the normalized distance vector between q1 and q2 , D = (x2−x2,y2−y1) , the tangent vector at q1 , T = (cos(θ1), sin(θ1)), and check if the angle between these two vectors is less than 90 degrees. dist(q1,q2) The purpose of this check is to only allow movements that would be feasible in traffic. Vehicle motions must move forward along the road in the direction of traffic, but this constraint isn't encoded into Steer. Thus, the reachability check enforces this behavior. CostToGoal:The cost to go from each vertex in G to qgoal is stored in H to be used as a heuristic in the central planner. Our implementation uses a breadth-first search. PruneDeadNodes: Due to G being a directed graph and the reachability constraints, some sampled nodes will not have a path to qf . These "dead" nodes in G are removed to avoid running into dead ends in the central planning stage. Algorithm 2: fdRRT(Qi, Qf , G, H) 1 T ← Qi; 2 Vlast ← Qi; 3 while Qf /∈ T do (T, Vlast) = Expand(T, G, H, Vlast, Qf ); if Qf ∈ T then Π ← FindPath(T, Qf ); return Π 4 5 6 7 8 9 end end B. Central Planner The algorithm structure from dRRT (Algorithm 1) is preserved with the initialization of T with Qi (Line 1). The algorithm then expands, while keeping track of the most recent expansion node Vlast to determine how it expands in the next iteration (Line 4). FindPath queries T for a path to Qf and returns a composite path Π if successful (Lines 5 -- 6). A notable difference in fdRRT is the omission of a local connector present in [15] and [14], whose purpose is to solve the multi-robot coordination problem when sufficiently close to Qf . We found this to be unnecessary in practice due to the structure of our environments. In the case of a traffic intersection, once all vehicles have passed through the physical intersection of the two roads, T tends to expand greedily towards Qf . A similar subroutine is utilized in resolving path conflict by forcing some robots to hold their positions while others move forward. Expand: Expansion of T begins with selecting a node to expand from. If a vertex Vlast was added during the previous call, then a new expansion vertex Vnew is chosen by selecting a neighbor of Vlast (Lines 2 -- 3). Otherwise, the closest neighbor Vnear of a random configuration Qrand is chosen (Lines 5 -- 6). The direction oracle subroutine selects an expansion node Vnew based on the success of the previous expansion (Line 8). If Qrand = Qf , Vnew is chosen as the tuple of individual vertices vi ∈ V that are neighbors to f ∈ Qf , and is near and have the lowest path cost to qi vi otherwise chosen as a tuple of randomly selected neighbors to vi near. We refer to [14] for a detailed explanation. All composite parents to Vnew that have already been added to T are expansion candidates to connect to Vnew (Line 9). Each candidate is evaluated base on whether the composite path between N and Vnew results in a collision- free motion and the composite path cost. Our algorithm differs from [14] when choosing the parent node to Vnew, Vbest. dRRT* chooses Vbest as the lowest cost vertex V ∈ N that is also a collision-free motion. In our algorithm, the lowest cost collision-free node, V f ree best , and the lowest cost node Vbest are selected. If no such V f ree exists, the best subroutine ForceConnect attempts forcing T to expand by creating a new hybrid node, VH, that restricts some individual nodes to hold their position at vbest, and allows Algorithm 3: Expand(T, G, H, Vlast, Qf ) 1 if Vlast = ∅ then Qrand ← RandomConfig(G); Vnear ← Nearest(T, Qrand); Qrand ← Qf ; Vnear ← Vlast; 2 3 4 else 5 6 7 end 8 Vnew ← Id(Vnear, G, H, Qf ); 9 N ← NeighborsInTree(Vnew, T); 10 (V f ree 11 if V f ree 12 best , Vbest) ← BestParent(Vnew, N ); best = ∅ then VH ← ForceConnect(Vnew, Vbest); if VH = ∅ then 13 14 15 16 17 18 19 else 20 21 22 end else return ∅ T ← Connect(Vbest, VH); return VH; end T ← Connect(V f ree return Vnew; best , Vnew); 1 ∈ V1 or moves forward towards vi others to move forward towards vnew. While forcing some vehicles to stop increases traffic delays for individual vehi- cles, ForceConnect increases computational efficiency in practice by restricting random sampling to a last resort. ForceConnect: When forcing a connection between two composite nodes V1 and V2, the ith robot either holds its 2 ∈ V2. position at vi Three sets are initialized for each robot ri ∈ R: Hi, the set of robots with higher local priority than ri, Li, the set of robots with lower priority than ri, and Ai, the set of robots that conflict with ri but have no local priority assigned. Each interaction is checked and H, L, and A are populated by LocalPriority. The local priority of ri with respect to rj is assigned according to the rules, which originate from the local connector logic in [15] and [17]: • If πi(0) blocks πj, then robot i is given priority • If πj(0) blocks the path of πi, then robot j is given priority • If πi and πj do not overlap, then there is no interaction and a priority is not assigned • Otherwise, the local priority can not be determined. This occurs when πi and πj overlap, but the starting positions of robots i and j do not block each other's paths. Either robot can be given priority, but the decision is deferred. A solution set S is then initialized to pick robots that should move forward (Line 9). Each robot is added to or rejected from S based on its own Hi, Li, andAi sets. For a robot ri, if no other robots have a higher local priority and no robots Algorithm 4: ForceConnect(V1, V2) 1 H ← ∅; 2 L ← ∅; 3 Π12 ← LocalPaths(V1, V2); 4 for πi ∈ Π12 do for πj ∈ Π12, i (cid:54)= j do (Hi, Li, Ai) ← LocalPriority(πi, πj); if Hi = ∅ & Ai = ∅ then else if Hi = ∅ & Ai (cid:54)= ∅ then 5 6 end 7 8 end 9 S ← ∅; 10 for i = 1, 2, .., N do 11 S ← S ∪ i; 12 13 if cost(i) ≤ min(cost(j ∈ Ai)) then 14 15 16 17 18 end 19 VH ← {vi 20 return VH; 2i ∈ S} ∪ {vj S ← S ∪ i; 1j /∈ S}; end end have an undetermined priority, then ri is added to S. if any vehicles have a higher priority, then ri is rejected from S. If no robots have a higher priority, but some have undetermined priorities, then the cost of adding ri is assessed. In this context, the cost refers to number of vehicles that would be excluded from S if ri was added to S. The cost of adding ri is compared to the cost of adding any of rj ∈ Ai and will be added to S if the trade-off from adding ri is lower than the trade-off from adding any other member of Ai. After all robots are either added to or rejected from S, a hybrid node VH is constructed (Line 19). V. SIMULATIONS AND RESULTS Our algorithm was implemented and tested in MATLAB. Three environments are considered for validation: a three- lane traffic intersection, a cluttered rectangular space akin to a warehouse, and a crowded space of UAVs (Figure 1). We assume each vehicle in the first environment is rectangular with length l = 3.6m and w = 1.6m, while the robots in the second and third environments disks with radius r = 0.4m and 0.2m, respectively. 1000 test cases were run for each combination of vehicles in each environment. Average search tree size, solution time, and path lengths are evaluation metrics to illustrate the trade-offs between dRRT* and fdRRT (Figures 3 -- 5). From the test results, fdRRT performs better than dRRT* in computation efficiency in the intersection and warehouse spaces. In test cases with maximum traffic, fdRRT returned solutions 57% faster in the traffic intersection and around 2000% faster in the warehouse. However, dRRT* is 12% faster in the UAV environment. This may be due to the lack of clutter within the UAV space, and thus reduced number of choke points. Under these conditions, the added (a) (a) (a) Fig. 3: Performance comparison in traffic intersection. (b) Fig. 4: Performance comparison in a warehouse space. (b) Fig. 5: Performance comparison in UAV environment. (b) (c) (c) (c) computational time in fdRRT when forcing connections may degrade performance. Solution quality metrics show the opposite trend. As more robots are added to each environment, the path quality in fdRRT degrades, with paths being 22% and 54% longer in the intersection and warehouse spaces, respectively. Paths in the UAV space are nearly identical with a 0.2% discrepancy. The trends in solution times and path lengths across the scenarios can be attributed the amount of clutter in each space. The warehouse space has more obstacles distributed across its environment, and thus more possible choke points and corridors, the traffic intersection funnels all vehicles into a single, albeit large, choke point. The UAV space, in contrast, has no obstacles and thus allows the most movement. We conclude that there's a trade-off between the two algorithms; fdRRT will generally return trajectories faster, but dRRT* will have lower cost solutions. VI. CONCLUSION We have developed a PRM planner for car-like robots to create trajectories that adhere to traffic standards, making it well suited for motion planning on roadway environments. That planning strategy was used in a central planning al- gorithm based on the previously published dRRT / dRRT* algorithm. Our implementation has demonstrated its advan- tage in computational time over dRRT* when planning in confined environments, at the cost of solution quality. The results from this study are promising, but several challenges remain. Testing the feasibility of fdRRT in a real system is one goal we would like to reach. We also plan to explore extending the planner to incorporate vehicle dy- namics in addition to vehicle kinematics. In its current form, we only consider sampling configurations q ∈ (x, y, θ, κ) and ignore constraints on vehicle speed and acceleration. Adding constraints on vehicle dynamics makes connecting [20] G. Wagner, Minsu Kang, and H. Choset, "Probabilistic path planning for multiple robots with subdimensional expansion," in 2012 IEEE International Conference on Robotics and Automation, May 2012, pp. 2886 -- 2892. [21] Z. Yan, N. Jouandeau, and A. Ali, "A survey and analysis of multi- robot coordination," International Journal of Advanced Robotic Sys- tems, vol. 10, p. 1, 12 2013. in heterogeneous [22] L. Ye and T. Yamamoto, "Modeling connected and autonomous traffic flow," Physica A: Statistical vehicles Mechanics and its Applications, vol. 490, pp. 269 -- 277, 2018. [Online]. Available: http://www.sciencedirect.com/science/article/pii/ S0378437117307392 [23] Yi Guo and L. E. Parker, "A distributed and optimal motion plan- ning approach for multiple mobile robots," in Proceedings 2002 IEEE International Conference on Robotics and Automation (Cat. No.02CH37292), vol. 3, May 2002, pp. 2612 -- 2619 vol.3. [24] J. Yu and S. M. LaValle, "Optimal multirobot path planning on graphs: Complete algorithms and effective heuristics," IEEE Transactions on Robotics, vol. 32, no. 5, pp. 1163 -- 1177, Oct 2016. [25] P. vestka and M. H. Overmars, "Coordinated path planning for multiple robots," Robotics and Autonomous Systems, vol. 23, no. 3, pp. 125 -- 152, 1998. [Online]. Available: http://www.sciencedirect. com/science/article/pii/S092188909700033X between configurations more difficult, but carries the benefit of ensuring that all paths are feasible for robots with both kinematic and dynamic constraints. REFERENCES [1] E. run able: amazon-introduces-two-new-warehouse-robots to Ackerman, this Avail- https://spectrum.ieee.org/automaton/robotics/industrial-robots/ warehouse," "Amazon [Online]. robots 2019. Jun uses 800 [2] E. W. Dijkstra, "A note on two problems in connexion with graphs," Numerische Mathematik, vol. 1, no. 1, pp. 269 -- 271, Dec 1959. [Online]. Available: https://doi.org/10.1007/BF01386390 [3] L. E. Dubins, "On curves of minimal length with a constraint on average curvature, and with prescribed initial and terminal positions and tangents," American Journal of Mathematics, vol. 79, no. 3, p. 497, jul 1957. [Online]. Available: https://doi.org/10.2307 [4] M. Elbanhawi and M. Simic, "Sampling-based robot motion planning: A review," IEEE Access, vol. 2, pp. 56 -- 77, 2014. [5] J. h. Jeon, S. Karaman, and E. Frazzoli, "Anytime computation of time-optimal off-road vehicle maneuvers using the rrt," in 2011 50th IEEE Conference on Decision and Control and European Control Conference, Dec 2011, pp. 3276 -- 3282. [6] P. E. Hart, N. J. Nilsson, and B. Raphael, "A formal basis for the heuristic determination of minimum cost paths," IEEE Transactions on Systems Science and Cybernetics, vol. 4, no. 2, pp. 100 -- 107, July 1968. [7] L. E. Kavraki, P. Svestka, J. . Latombe, and M. H. Overmars, "Proba- bilistic roadmaps for path planning in high-dimensional configuration spaces," IEEE Transactions on Robotics and Automation, vol. 12, no. 4, pp. 566 -- 580, Aug 1996. [8] J.-T. Li and H.-J. Liu, "Design optimization of amazon robotics," 2016. [9] B. Liu and A. El Kamel, "V2x-based decentralized cooperative adap- tive cruise control in the vicinity of intersections," IEEE Transactions on Intelligent Transportation Systems, vol. 17, no. 3, pp. 644 -- 658, March 2016. [10] B. Paden, M. C´ap, S. Z. Yong, D. S. Yershov, and E. Frazzoli, "A survey of motion planning and control techniques for self- driving urban vehicles," CoRR, vol. abs/1604.07446, 2016. [Online]. Available: http://arxiv.org/abs/1604.07446 [11] J. L. Rios, L. Martin, and J. Mercer, "Use of UAS Reports (UREPs) during TCL3 Field Testing," National Aeronautics and Space Admin- istration,, Tech. Rep., 07 2017. [12] G. Sanchez and J. . Latombe, "Using a prm planner to compare central- ized and decoupled planning for multi-robot systems," in Proceedings 2002 IEEE International Conference on Robotics and Automation (Cat. No.02CH37292), vol. 2, May 2002, pp. 2112 -- 2119 vol.2. [13] A. Scheuer and T. Fraichard, "Continuous-curvature path planning for car-like vehicles," 10 1997, pp. 997 -- 1003 vol.2. [14] R. Shome, K. Solovey, A. Dobson, D. Halperin, and K. E. Bekris, "drrt: Scalable and informed asymptotically-optimal multi-robot motion planning," CoRR, vol. abs/1903.00994, 2019. [Online]. Available: http://arxiv.org/abs/1903.00994 [15] K. Solovey, O. Salzman, and D. Halperin, "Finding a needle in an exponential haystack: Discrete RRT for exploration of implicit roadmaps in multi-robot motion planning," CoRR, vol. abs/1305.2889, 2013. [Online]. Available: http://arxiv.org/abs/1305.2889 [16] P. Svestka and M. H. Overmars, "Motion planning for carlike robots using a probabilistic learning approach," The International Journal of Robotics Research, vol. 16, no. 2, pp. 119 -- 143, 1997. [Online]. Available: https://doi.org/10.1177/027836499701600201 [17] J. van den Berg, J. Snoeyink, M. Lin, and D. Manocha, "Centralized path planning for multiple robots: Optimal decoupling into sequential plans," 06 2009. [18] J. J. B. Vial, W. E. Devanny, D. Eppstein, and M. T. Goodrich, "Scheduling autonomous vehicle platoons through an unregulated intersection," CoRR, vol. abs/1609.04512, 2016. [Online]. Available: http://arxiv.org/abs/1609.04512 [19] G. Wagner and H. Choset, "M*: A complete multirobot path planning algorithm with performance bounds," in 2011 IEEE/RSJ International Conference on Intelligent Robots and Systems, Sep. 2011, pp. 3260 -- 3267.
cs/0509017	1	0509	2005-09-06T17:02:52	Traders imprint themselves by adaptively updating their own avatar	[ "cs.MA", "cs.CE" ]	Simulations of artificial stock markets were considered as early as 1964 and multi-agent ones were introduced as early as 1989. Starting the early 90's, collaborations of economists and physicists produced increasingly realistic simulation platforms. Currently, the market stylized facts are easily reproduced and one has now to address the realistic details of the Market Microstructure and of the Traders Behaviour. This calls for new methods and tools capable of bridging smoothly between simulations and experiments in economics. We propose here the following Avatar-Based Method (ABM). The subjects implement and maintain their Avatars (programs encoding their personal decision making procedures) on NatLab, a market simulation platform. Once these procedures are fed in a computer edible format, they can be operationally used as such without the need for belabouring, interpreting or conceptualising them. Thus ABM short-circuits the usual behavioural economics experiments that search for the psychological mechanisms underlying the subjects behaviour. Finally, ABM maintains a level of objectivity close to the classical behaviourism while extending its scope to subjects' decision making mechanisms. We report on experiments where Avatars designed and maintained by humans from different backgrounds (including real traders) compete in a continuous double-auction market. We hope this unbiased way of capturing the adaptive evolution of real subjects behaviour may lead to a new kind of behavioural economics experiments with a high degree of reliability, analysability and reproducibility.	cs.MA	cs	Traders imprint themselves by adaptively updating their own avatar Gilles Daniel1, Lev Muchnik2, and Sorin Solomon3 1 School of Computer Science, University of Manchester, UK [email protected] 2 Department of Physics, Bar Ilan University, Ramat Gan, Israel [email protected] 3 Racah Institute of Physics, Hebrew University of Jerusalem and Lagrange Laboratory for Excellence in Complexity, ISI Foundation, Torino [email protected] Simulations of artificial stock markets were considered as early as 1964 [Sti64] and multi-agent ones were introduced as early as 1989 [KM89]. Starting the early 90's [PAH+94, LLS00, Ter00], collaborations of economists and physicists produced in- creasingly realistic simulation platforms. Currently, the market stylized facts are easily reproduced and one has now to address the realistic details of the Market Microstructure and of the Traders Behaviour. This calls for new methods and tools capable of bridging smoothly between simulations and experiments in economics. We propose here the following Avatar-Based Method (ABM). The subjects im- plement and maintain their Avatars (programs encoding their personal decision mak- ing procedures) on NatLab, a market simulation platform. Once these procedures are fed in a computer edible format, they can be operationally used as such without the need for belabouring, interpreting or conceptualising them. Thus ABM short- circuits the usual behavioural economics experiments that search for the psychologi- cal mechanisms underlying the subjects behaviour. Finally, ABM maintains a level of objectivity close to the classical behaviourism while extending its scope to subjects' decision making mechanisms. We report on experiments where Avatars designed and maintained by humans from different backgrounds (including real traders) compete in a continuous double- auction market. Instead of viewing this as a collectively authored computer simula- tion, we consider it rather as a new type of computer aided experiment. Indeed we consider the Avatars as a medium on which the subjects can imprint and refine inter- actively representations of their internal decision making processes. Avatars can be objectively validated (as carriers of a faithful replica of the subject decision making process) by comparing their actions with the ones that the subjects would take in sim- ilar situations. We hope this unbiased way of capturing the adaptive evolution of real subjects behaviour may lead to a new kind of behavioural economics experiments with a high degree of reliability, analysability and reproducibility. 2 Gilles Daniel, Lev Muchnik, and Sorin Solomon 1 Introduction In the last decade, generic stylized facts were reproduced with very simple agents by a wide range of models [BPS97, LGC+99, LM99, CB00, MSS03, GB03]. By the very nature of their generic properties, those models teach us little on real particular effects taking place as result of real particular conditions within the market. In order to understand such specific market phenomena, one may need to go beyond "simple- stupid" traders behaviour [Axe97]. Thus the task of the present generation of models is to describe and explain the observed collective market phenomena in terms of the actual behaviour of the individuals. For a long while, classical economics assumed individuals were homogeneous and behaved rationally. Thus it was not necessary to study real people behaviour since (presumably) there is only one way to be rational. Even after the conditions of rationality and homogeneity were relaxed, many models did it by postulating arbi- trary departures not necessarily based on actual experiments. When the connection to the real subjects behaviour was considered [KT79], an entire host of puzzles and paradoxes appeared even in the simplest artificial (laboratory) conditions. Thus the inclusion of real trader behaviour in the next generation of models and simulations is hampered by the inexistence of comprehensive, systematic, reliable data. Given the present state of the art in psychological experiments, where even the behaviour of single subjects is difficult to assess, we are lead to look for alternative ways to elicit the necessary input for agent-based market modelling. In this paper we propose a way out of this impasse. Rather than considering the computer as a passive receiver of the behavioural information elicited by psycholog- ical experiments, we use the computer itself as an instrument to extract some of the missing information. More precisely, we ask the subjects to write and update adap- tively, between simulation runs (or virtual trading sessions) their own avatars. By gradual corrections, those avatars converge to satisfactory representations of the sub- jects' behaviour, in situations created by their own collective co-evolution. The fact that the co-evolution takes place through the intermediary of the avatars interaction provides an objective detailed documentation of the process. More important, the dialogue with the avatars, their actions and their collective consequences assist the subjects in expressing in a more and more precise way their take on the evolving situation and validate the avatar as an expression of the subject internal decision mechanisms. Ultimately, the avatar becomes the objective reposi- tory of the subject decision making process. Thus we extend, with the help of com- puters, the behaviorist realm of objectivity to a new area of decision making dynam- ics. The classical behaviourism limits legitimate research access to external overt behaviour, restraining its scope to the external effects produced by a putative men- tal dynamics. The method above enables us to study the subjects decision making dynamics without relying on ambiguous records of overt subjects behaviour nor on subjective introspective records of their mental state and motivations. Far from invalidating the psychological experimental framework, the present method offers psychological experiments a wide new source of information in prob- ing humans mind. The competitive ego-engaging character of the realistic NatLab Traders imprint themselves by adaptively updating their own avatar 3 market platform [Muc] puts humans in very interesting, authentic and revealing situ- ations in a well controlled and documented environment. Thus standard psycholog- ical techniques can exploit it e.g. by interviewing the subjects before and after their updated strategies are applied and succeed (or fail!). 2 Avatars The program sketched in the previous section suggests a Behavioural Finance view- point, in a realistic simulation framework. More precisely, the avatars acting in such an environment are able to elicit from the subjects operationally precise and arbi- trarily refined descriptions of their decision processes. In particular, by analysing the successive avatar versions that passed the validation of its owner, one can learn how the owner behaves in this market environment, how (s)he designs his/her strategies, how (s)he decides to depart from them, how (s)he updates them iteratively, etc. Thus the new environment acquires a mixed computational and experimental laboratory character. In this respect, the present study owes to previous research that involved simulations / experiments combining human beings and artificial agents, in real-time [Cap05] or off-line [MS03, BPBK04] – see [Duf04] for a review of computational vs experimental laboratories. The heart of the new simulation-experimentation platform is the co-evolving set of Avatars. They constitute both the interacting actors and the medium for recording the chronicles of the emergent collective dynamics of the subjects. As a medium for capturing cognitive behaviour, the avatars help extend the behaviorist objectivity cri- teria to processes that until now would be considered as off-limits. We are achieving it by trying to elicit from humans operational instructions for reaching decisions that they want implemented by their market representatives - the avatars. There is an im- portant twist in this procedure: we are not trying to obtain from the subjects reports of their internal state of mind and its evolution; we are just eliciting instructions for objective actions in specific circumstances. They are however formulated in terms of conditional clauses that capture users intentionality, evaluations, preferences and internal logics. 2.1 Principle At the beginning of a run, every participant designs his own avatar which is used as a basis to generate an entire family of artificial agents whose individuality is expressed by various (may be stochastically generated) values of their parameters. The resulting set of artificial agents compete against each other in our market environment; see Fig. 1. We use many instances, rather than a single instance of the avatar for each subject, for the following reasons: • • having a realistic number of traders that carry a certain strategy, trading policy or behaviour profile having enough statistics on the performance of each avatar and information on the actual distribution of this performance 4 Gilles Daniel, Lev Muchnik, and Sorin Solomon Fig. 1. The Avatar-Based Method: Human subjects design their avatar in the NatLab environ- ment. From each avatar, a family of artificial agents is generated and included in the market. Once the population of agents is generated, a first simulation run is performed. A typical run lasts about 10 minutes of CPU , which may represent years of trad- ing on the artificial time scale. At the end of each run, the results are processed and presented to the participants. In our experiments until now, both private and public information were made available. In particular, the price (and volume) trajectory, the (relative) individuals wealth in terms of cash holdings, stock holdings, and their evolution were publicly displayed. The avatar codes were also disclosed and the par- ticipants were asked to describe publicly their strategy and the design of their avatar. After being presented with the results (whether full or only public information) of the previous run, the participants are allowed to modify their own avatar and submit an upgraded version for the next run, as described in Fig. 2. The goal of this iterative process, co-evolving subjects thinking with computer simulations, is to converge in two respects; the subject understands better and better: • • the consequences of his/her own strategy how to get the avatars to execute it faithfully 2.2 Comparison between approaches In this section, we discuss the relevance of our method in the context of other works in economics. The economics field spans a wide range of fields and approaches. In the table displayed in Fig. 3, the four rows classify the activities in terms of their con- text and environment, starting with the DESK at the bottom of the table, extending it to the use of computers, then to the laboratory and ultimately to the real unstructured world. Traders imprint themselves by adaptively updating their own avatar 5 Fig. 2. Iterative process: participants design and improve their avatar in between every simu- lation run CLASSICAL METHODS AVATAR-BASED METHOD Wide Wild World LAB Field Studies Econometrics Behavioral Economics WWW based large scale experiments Interactive subjects experiments Single Subject cognitive experiments COMPUTER Agent-Based and Social Networks Simulations Computational finance Numerical Model Solving DESK Game Theory and other Analytical work Economic Analysis Remote extraction of avatars from special distant subjects: Important bank executives, nationals of various cultures etc. Intimate dialogue between Subjects and Computer via Avatar Update alternated with NatLab Market Runs Validation Relaxing in a controlled way the assumptions of the models. Expressing operationally qualitative descriptions. Fig. 3. Positioning the Avatar-Based Method 6 Gilles Daniel, Lev Muchnik, and Sorin Solomon The two columns of the table refere to the usual methods and the "Avatar-Based Method" (ABM from now on). One sees that in our view, the ABM constitutes a rather uniform way to treat economic behaviour and is capable of bridging appli- cation areas that were until now disjoint. This is clearly the case for the LAB and COMPUTER rows, where we even erased the separation line, but it has implications for the other rows too. For instance the Avatars, and especially within the NatLab environment, have been used already to extend to more realistic conditions some theoretical models (row 4 of the table) and results [MSS03]. At the other extreme (rows 1-2 in the table), the Avatar-Based Method can help correct a perennial problem of economic studies: the biased and sometimes unrep- resentative profile of the involved subjects. Indeed, it is very difficult to involve in those studies decision making officials from financial institutions or traders. Substi- tuting them by BA Undergrads is hardly a step towards realistic emulation of the real world. It is much more likely that these important players, rather than coming to a lab, will agree to provide the elements for creating their Avatars. Similar prob- lems can be solved by including in the ABM experiments subjects from far away cultures or environments, without the necessity for distant travels and without sepa- rating them from their usual motivations and environment. Moreover, the information provided once by such subjects that are difficult to access can be used now repeatedly by playing their Avatars. Thus ABM has a good chance to bridge the gap between field studies and lab experiments too (rows 1-2 in the table). In fact as opposed to experiments that do not involve a market mechanism with capital gain and loss, in NatLab, incompetent non-representative subjects will naturally be eliminated since their Avatars loose very quickly their capital. Another point on which the ABM procedures are offering new hope is the well known problem of subjects motivation. Within the usual experimental frameworks, it is very difficult to motivate subjects, especially competent important people. From our experience, the NatLab realistic framework and the direct identification of the subjects with their Avatars successes and failures, lead to a very intensive and en- thusiastic participation of the subjects even for experiments that last for a few days. In fact, beyond the question of "prestige", even seasoned professionals reported to have gained new insights in their own thinking during the sessions. Another promise that ABM is yet to deliver is that by isolating and documenting the Avatar update at discrete times, one will be able to contribute to neighbouring cognitive fields such as learning. 3 Method Validation A piece of software is not an experimental set-up. With all its power, the value of the platform and of the "Avatar-Based Experiments" method has to be realized in real life and an elaborate technical and procedural set-up has to be created. The ba- sic condition for the very applicability of our method is the humans capability to faithfully, precisely and consistently express their decision making in terms of com- puter feedable procedures. Thus we concentrated our first validation efforts in this Traders imprint themselves by adaptively updating their own avatar 7 direction, adapting platform and procedural features to accommodate humans. Many other experimental aspects have to be standardised and calibrated, but in those ex- periments we concentrated on this crucial sine qua non issue. We can conclude at this stage that while there are humans (even economist professionals and successful traders) that could not express their "system" in a computer feedable format ("buy low , sell high"), by-and-large the participants in our experiments were able to con- firm at some point that their avatar behaved in agreement with their own behaviour. This happened even with subjects with no particular computer (or market) skills. 3.1 Experimental setup Our experiment features a continuous double-auction implemented on the NatLab simulation platform. Every participant received extensive support from a computer scientist to implement his/her avatar in C++ on the platform. NatLab platform The NatLab has the capability to simulate in great detail the continuum time asyn- chronous real world [MS03]. Bilateral and multilateral communication between agents outside and in parallel with the market is made possible by NatLab. How- ever, given that this experiment focuses mainly on the participants behaviour, we kept the market mechanism (the rules of the game) as simple as possible, while re- taining the concept of continuous double-auction, essential to understand the price formation dynamics. NatLab was initially engineered as a simulation platform but its use is now in three distinct directions: 1. the platform provides a realistic framework for the individuals to act within. Pro- viding this "reality" is independent of whether one is interested in its characteris- tics; it just allows an interactive continuous extraction of information from each of the participants and thereby refining our understanding on their approach, reactions and decision mechanisms; 2. the platform is part of a recent wide effort to understand the emergence of col- lective complex dynamics out of interacting agents with well defined, relatively simple individual behaviour; and 3. the platform, due to its realistic features and its asynchronous continuous time microstructure, is a reliable way to reproduce and maybe in the future predict real market behaviour. Market microstructure Our market implements a continuous double-auction mechanism, where agents can submit, asynchronously and at any time, limit or market orders to a single public book. Orders are sorted by price and then by time, as on the NYSE for instance. Every agent acts as a simple trader, and we do not include brokers or market makers at this stage. In this simple setup, agents balance their portfolio between a risky asset (a stock distributing or not a dividend) and a riskless one (a bond yielding or not an interest rate). Agents can communicate with each other through pairwise messages, and react to external news according to an idiosyncratic sensibility. 8 Gilles Daniel, Lev Muchnik, and Sorin Solomon Avatars We organise our experiment as a competition between participants through the inter- mediary of their avatars. Avatars generate, by assigning values to their parameters, families of agents that act as independent (but possibly interacting) individuals in the market. The subjects' aim in each run is to generate a family of artificial agents that perform well against other families throughout the simulation run. A typical simula- tion run is exhibited in Fig. 4. Families were compared by their average wealth, but an average utility (given some utility function) or a certain bonus for minimising risk could be used in the future. We give our participants total liberty while implementing their avatar. They can define their own time horizon and design trading strategies as simple or complex as needed, but in the future we may tax agents with heavy data processing by imposing a fine or a specific time lag in the order execution. 3.2 Preliminary results We have run two sets of experiments so far, with different participants including prac- titioners (real traders) and academics, either economists, physicists, psychologists or computer scientists. Each experiment included seven participants. The first experi- ment took place on July 19-31 2004, in Lyon, during the SCSHS Summer School on Models for Complex Systems in Human and Social Sciences organised by the Ecole Normale Suprieure de Lyon. The second was organised on January 12-16, in Turin, during the Winter Seminar on Agent-based Models of Financial Markets organised by the ISI Foundation. A typical run, with a preliminary analysis of the price time series and relative evolution of populations, is presented on Fig. 4. We report here on some of the non trivial aspects of the participants behaviour during the experiments, while creating and updating their avatars. Imprinting oneself We noticed, specially at the beginning of the process, that some of our participants encountered some difficulties to express themselves in terms of computer feedable strategies. However, this improved dramatically during the iterative process itself. This is clearly linked to the learning process that one has to face while performing any experiment, especially computerised ones. Conscious / Unconscious decisions The very nature of our method barely allows such things as intuition, improvisation or unconscious decisions to be operationally expressed in the avatar. In fact, after a few runs, avatars capture exclusively the conscious part of our subjects decision making process. Since we we do not know to what extent markets dynamics are driven by unconscious choices, it would be interesting to design a double experiment, comparing subjects and their own avatar in the same market microstructure. 0.16 a 0.12 0.08 F C A 0.04 0.00 c 14 e c i r P 12 10 120 e s k c o t S 100 80 Time (au) Time (au) y t i l i b a b o r P 0 -0.04 -0.03 -0.02 -0.01 0.00 0.01 0.02 0.03 0.04 h t l a e W e v i t l a e R Log Return Time (au) d 0.172 0.170 0.168 0.166 0.164 0.162 f 1200 Traders imprint themselves by adaptively updating their own avatar 9 Returns Volatility 1 b Data Gaussian 1000 h s a C 800 600 Time (au) Time (au) Fig. 4. Typical run with 7 avatars, 1000 agents each, for above 350 000 transaction ticks. (a) Autocorrelation functions – absence of raw returns autocorrelation and long-term autocorre- lation of volatility, as defined as absolute returns, as observed in empirical data [LGC+99];(b) Normality Plot – fat tailed distribution of returns; (c) Price trajectory; (d) Relative wealth of agents populations – measure the relative success of competing avatars; (e) Stock holdings – some strategies are clearly buy-and-hold, others interact with each other; and (f) Cash holdings Convergence There are two different but related convergence processes that took place during the successive iterations: the first was the convergence of the avatar's behaviour to its creator's intended strategy, while the second involved the evolution of subjects strat- egy itself to beat other participants. While it appeared relatively easy after a couple of runs to get an avatar successfully reproducing their initial intended behaviour, subjects, driven by competition, kept refining and complexifying their strategy. 10 Gilles Daniel, Lev Muchnik, and Sorin Solomon Strategies An interesting panel of strategies was proposed and grown by the participants, that could loosely be termed random trader, momentum trader, oscillatory trader, diver- sified Bollinger Bands trader, volume seeker, Neural Network based trader and evo- lutionary trader. Practitioners clearly outmarked themselves by their ability to think out of the box, the creativity of their strategies, their high analysis power and ability to quickly understand what was going on and spot opportunities to arbitrage other participants' strategies. We also observed the emergence of cooperation between par- ticipants to hunt for the leader, trying to bring down the winning strategy by copying and modifying it or even custom-designing new strategies for this specific purpose. Fundamental Value In the two experiments we ran, our computer simulation figured a closed artificial market, with no creation of stocks, no distribution of dividends and no interest rate associated with the riskless asset, cash. In those conditions, we observed that after a transition period, characterised by high volumes, during which assets were heavily reallocated between agents, the price kept fluctuating around a steady state equi- librium price. This price, emerging from the interactions between heterogeneous, relative risk aversion agents, was generally different from the fundamental value we could have expected from rational agents with homogeneous preferences. 4 Conclusion The rapidly growing field of Agent-based Computational Finance comes naturally as a complementary approach to the other Finance subfields: Behavioural Finance, Laboratory experiments, Econometrics, Game Theory, etc. The field is definitely out of his infancy and a rather wide range of choices is available to academics and prac- titioners that wish to define and test concrete real and realistic systems or new mod- els of individual and market behaviour. The next step is to set common standards for the platforms that propose to represent and simulate artificial financial markets [JLM04, Cap03, SMLS00, BPBK04]. One possible goal is to transform them in vir- tual or even real laboratories capable to implement and test in realistic conditions arbitrarily sophisticated experiments. One way to solve the problems of realistic trader behaviour is the Avatar-Based Method introduced in the present paper. Even though there are many obstacles not even yet uncovered in realizing its ambitions, the method is already providing new insight and definitely even if its main ambitions are going to remain unfulfilled, it is guaranteed to provide fresh unexpected and valuable material to the existing methods. Among the fundamental issues which the ABM can address is the mystery of price formation by providing in great detail, reliability and reproducibility, the traders decision making mechanisms. Occasionally the Avatars are going to be caught un- prepared and inadequate to deal with some instances that were not previewed by their Traders imprint themselves by adaptively updating their own avatar 11 owners. By the virtue of this very instance, they will become effective labels for the emergence of novelty in the market. Thus in such instances, even in its failure, the ABM will provide precious behavioural and conceptual information. ABM can serve as a design tool for practitioners in the development of new trading strategies and the design of trading automata. Moreover, we hope that this approach will provide new ways to address some of the fundamental problems un- derlying the economics field: • • • how people depart from rationality how out-of-equilibrium markets achieve or not efficiency how extreme events due to a shifting composition of markets participants could be anticipated The experiments we ran, beyond eliciting information, provided a very special and novel framework of interaction between practitioners and academics. Thus Nat- Lab and ABM might have an impact on the community by providing a common language and vocabulary to bring together academics and much needed practition- ers. As a consequence, it appears necessary to gather interdisciplinary projects that would house within the same team the psychologists that run experiments on peo- ple's behaviour, computer scientists that canonise this behaviour into artificial agents, practitioners that relate those experiments to real markets and economists that assess the consequences in terms of policy making. Acknowledgements We would like to thank the participants of our experiments for their time and com- mitment, together with the participants of the Seminar on (Un)Realistic Simulations of Financial Markets at ISI Foundation, Turin, Italy, on April 1-5 2005, for their enlightening comments, from which this paper largely benefitted. Finally, we are largely indebted to Alessandro Cappellini and Pietro Terna for their views and ex- perience on online laboratory experiments of stock markets, as well as Martin Hos- nisch, Diana Mangalagiu and Tom Erez. The research of SS was supported in part by a grant from the Israeli Academy of Science, and the research of LM was sup- ported by a grant from the Centre for Complexity Science. All errors are our own responsibilities. References [Axe97] R. Axelrod. The Complexity of Cooperation: AgentBased Models of Competition and Collaboration. Princeton University Press, 1997. [BPBK04] K. Boer, M. Polman, A. Bruin, and U. Kaymak. An agent-based framework for artificial stock markets. In 16th Belgian-Dutch Conference on Artificial Intelligence (BNAIC), 2004. [BPS97] P. Bak, M. Paczuski, and M. Shubik. Price variations in a stock market with many agents. Physica A, 246:430–453, 1997. 12 Gilles Daniel, Lev Muchnik, and Sorin Solomon [Cap03] A. N. Cappellini. Esperimenti su mercati finanziari con agenti naturali ed artifi- ciali. Master's thesis, Dipartimento di Scienze Economiche e Finanziarie, Facolta di Economia, Universita di Torino, Italy, 2003. [Cap05] A. Cappellini. Avatar e simulazioni. Sistemi intelligenti, 1:45–58, 2005. [CB00] R. Cont and J.-P. Bouchaud. Herd behaviour and aggregate fluctuations in financial markets. Macroeconomic Dynamics, 4:170–196, 2000. J. Duffy. Agent-based models and human subject experiments. Computational Economics 0412001, Economics Working Paper Archive at WUSTL, December 2004. available at http://ideas.repec.org/p/wpa/wuwpco/0412001.html. I. Giardina and J.-P. Bouchaud. Bubbles, crashes and intermittency in agent based market models. The European Physical Journal B, 31:421–537, 2003. [Duf04] [GB03] [JLM04] B. I. Jacobs, K. N. Levy, and H. Markowitz. Financial market simulations. Journal of Portfolio Management, 30th Anniversary, 2004. [KM89] G. Kim and H. Markowitz. Investment rules, margin, and market volatility. Journal of Portfolio Management, 16(1):45–52, 1989. [KT79] D. Kahneman and A. Tversky. Prospect theory: An analysis of decision under risk. Econometrica, 47(2):263–292, 1979. [LGC+99] Y Liu, P. Gopikrishnan, P. Cizeau, M. Meyer, C. Peng, and H. E. Stanley. Statis- tical properties of the volatility of price fluctuations. Physical Review E, 60:1390– 1400, 1999. [LM99] [LLS00] H. Levy, M. Levy, and S. Solomon. Microscopic Simulation of Financial Markets: From Investor Behavior to Market Phenomena. Berkeley, CA: Academic Press, 2000. T. Lux and M. Marchesi. Scaling and criticality in a stochastic multi-agent model of a financial market. Nature, 397:498–500, 1999. L. Muchnik and S. Solomon. Statistical mechanics of conventional traders may lead to non-conventional market behavior. Physica Scripta, T106:41–47, 2003. [MS03] [MSS03] L. Muchnik, F. Slanina, and S. Solomon. The interacting gaps model: reconciling theoretical and numerical approaches to limit-order models. Physica A, 330:232– 239, 2003. Lev Muchnik. Simulating emergence of complex collective dynamics in the stock markets. http://shum.huji.ac.il/sorin/ccs/Lev-Thesis.pdf. [Muc] [PAH+94] R. G. Palmer, W. B. Arthur, J. H. Holland, B. LeBaron, and P. Tayler. Artificial economic life: a simple model of a stock market. Physica D, 75:264–274, 1994. [Sti64] [SMLS00] M. Shatner, L. Muchnik, M. Leshno, and S. Solomon. A continuous time asyn- chronous model of the stock market; beyond the lls model. In Economic Dynamics from the Physics Point of View. Physikzentrum Bad Honnef, Germany, 2000. G. J. Stigler. Public regulation of the securities market. Journal of Business, 37(2):117–142, 1964. P. Terna. Sum: A surprising (un)realistic market - building a simple stock mar- ket structure with swarm. In Computing in Economics and Finance. Society for Computational Economics, 2000. [Ter00]
1909.06168	1	1909	2019-09-13T12:27:00	A Particle Swarm Based Algorithm for Functional Distributed Constraint Optimization Problems	[ "cs.MA" ]	Distributed Constraint Optimization Problems (DCOPs) are a widely studied constraint handling framework. The objective of a DCOP algorithm is to optimize a global objective function that can be described as the aggregation of a number of distributed constraint cost functions. In a DCOP, each of these functions is defined by a set of discrete variables. However, in many applications, such as target tracking or sleep scheduling in sensor networks, continuous valued variables are more suited than the discrete ones. Considering this, Functional DCOPs (F-DCOPs) have been proposed that is able to explicitly model a problem containing continuous variables. Nevertheless, the state-of-the-art F-DCOPs approaches experience onerous memory or computation overhead. To address this issue, we propose a new F-DCOP algorithm, namely Particle Swarm Based F-DCOP (PFD), which is inspired by a meta-heuristic, Particle Swarm Optimization (PSO). Although it has been successfully applied to many continuous optimization problems, the potential of PSO has not been utilized in F-DCOPs. To be exact, PFD devises a distributed method of solution construction while significantly reducing the computation and memory requirements. Moreover, we theoretically prove that PFD is an anytime algorithm. Finally, our empirical results indicate that PFD outperforms the state-of-the-art approaches in terms of solution quality and computation overhead.	cs.MA	cs	A Particle Swarm Based Algorithm for Functional Distributed Constraint Optimization Problems Moumita Choudhury, Saaduddin Mahmud and Md. Mosaddek Khan Department of Computer Science and Engineering, University of Dhaka {moumitach22, saadmahmud14}@gmail.com, [email protected] 9 1 0 2 p e S 3 1 ] A M . s c [ 1 v 8 6 1 6 0 . 9 0 9 1 : v i X r a Abstract Distributed Constraint Optimization Problems (DCOPs) are a widely studied constraint handling framework. The objective of a DCOP algorithm is to optimize a global objective func- tion that can be described as the aggregation of a number of distributed constraint cost functions. In a DCOP, each of these functions is defined by a set of discrete variables. However, in many applications, such as target tracking or sleep schedul- ing in sensor networks, continuous valued variables are more suited than the discrete ones. Considering this, Functional DCOPs (F-DCOPs) have been proposed that is able to explic- itly model a problem containing continuous variables. Never- theless, the state-of-the-art F-DCOPs approaches experience onerous memory or computation overhead. To address this issue, we propose a new F-DCOP algorithm, namely Particle Swarm Based F-DCOP (PFD), which is inspired by a meta- heuristic, Particle Swarm Optimization (PSO). Although it has been successfully applied to many continuous optimiza- tion problems, the potential of PSO has not been utilized in F-DCOPs. To be exact, PFD devises a distributed method of solution construction while significantly reducing the compu- tation and memory requirements. Moreover, we theoretically prove that PFD is an anytime algorithm. Finally, our empiri- cal results indicate that PFD outperforms the state-of-the-art approaches in terms of solution quality and computation over- head. Introduction Distributed Constraint Optimization Problems (DCOPs) are an important constraint handling framework of multi-agent systems in which multiple agents communicate with each other in order to optimize a global objective. The global ob- jective is defined as the aggregation of cost functions (i.e. constraints) among the agents. The cost functions can be de- fined by a set of variables controlled by the corresponding agents. DCOPs have been widely applied to solve a number of multi-agent coordination problems including, multi-agent task scheduling (Sultanik, Modi, and Regli 2007), sensor networks (Farinelli, Rogers, and Jennings 2014), multirobot coordination (Yedidsion and Zivan 2016). Over the years, a number of algorithms have been pro- posed to solve DCOPs which includes both exact and non- exact algorithms. Exact algorithms, such as ADOPT (Modi et al. 2005), DPOP (Petcu and Faltings 2005) and PT-FB (Litov and Meisels 2017), are designed in such a way that provide the global optimal solutions of a given DCOP. How- ever, exact solutions experience either, or both, of the ex- ponential memory requirements and computational cost as the system grows. On the contrary, non-exact algorithms such as, DSA (Zhang et al. 2005), MGM & MGM2 (Ma- heswaran, Pearce, and Tambe 2004), Max-Sum (Farinelli et al. 2008), CoCoA (van Leeuwen and Pawelczak 2017), and ACO DCOP (Chen et al. 2018) compromise some solution quality for scalability. In general, DCOPs assume that participating agents' vari- ables are discrete. Nevertheless, many real world applica- tions (e.g. target tracking sensor orientation (Fitzpatrick and Meetrens 2003), sleep scheduling of wireless sensors (Hsin and Liu 2004)) can be best modelled with continuous vari- ables. Therefore, for discrete DCOPs to be able to apply in such problems, we need to discretize the continuous do- mains of the variables. However, the discretization process needs to be coarse for a problem to be tractable and must be sufficiently fine to find high quality solutions of the prob- lem(Stranders et al. 2009). To overcome this issue, Stranders et al. has proposed a continuous version of DCOP which is later referred as Functional DCOP (F-DCOP) (Hoang et al. 2019). There are two main differences between F-DCOP and DCOP. Firstly, instead of having discrete decision variables, F-DCOP has continuous variables that can take any value between a range. Secondly, the constraint functions are rep- resented in functional forms in F-DCOP rather than in the tabular forms in DCOP. To cope with the modification of the DCOP formulation, Continuous Max-Sum (CMS) has been proposed which is an extension of the discrete Max-Sum (Stranders et al. 2009). However, this paper approximates the constraint utility func- tions as piece-wise linear functions which is often not ap- plicable in practice since a handful of real life applications deals with only peice-wise linear functions. To address this limiting assumption of CMS, Hybrid Max-Sum (HCMS) has been proposed in which continuous non-linear optimization methods are combined with the discrete Max Sum algorithm (Voice et al. 2010). However, continuous optimization meth- ods such as, gradient based optimization require derivative calculations, and thus they are not suitable for non differ- entiable optimization problems. The latest contribution in this field has been done by Hoang et al., 2019. In this pa- per, authors propose one exact, Exact Functional DPOP (EF- DPOP) and two approximate versions, Approximate Func- tional DPOP (AF-DPOP), and Clustered AF-DPOP (CAF- DPOP) of DPOP for solving F-DCOP (Hoang et al. 2019). The key limitation of these algorithms is that both AF-DPOP and CAF-DPOP incur exponential memory and computation overhead even though the latter cuts the communication cost by providing a bound on message size. Against this background, we propose a Particle Swarm Optimization based F-DCOP algorithm(PFD). Particle Swarm Optimization (PSO) (Eberhart and Kennedy 1995) is a stochastic optimization technique inspired by the so- cial metaphor of bird flocking that has been successfully ap- plied to many optimization problems such as Function Mini- mization (Shi and Eberhart 1999), Neural Network Training (Zhang et al. 2007) and Power-System Stabilizers Design Problems (Abido 2002). However, to the best of our knowl- edge no previous work has been done to incorporate PSO in DCOP or F-DCOP. In PFD, agents cooperatively keep a set of particles where each particle represents a candidate so- lution and iteratively improve the solutions over time. Since PSO requires only primitive mathematical operators such as, addition and multiplication, it is computationally inexpen- sive (both in memory and speed) than the gradient based op- timization methods. Specifically, We empirically show that PFD can not only find better solution quality by exploring a large search space but it is also computationally inexpen- sive both in terms of memory and computation cost than the existing FDCOP solvers. Background and Problem Formulation In this section, we formulate the problem and discuss the background necessary to understand our proposed method. We first describe the general DCOP framework and then move on the F-DCOP framework which is the main prob- lem that we want to solve. We then discuss the central- ized PSO algorithm and the challenges remain to incorporate PSO with F-DCOP framework. Distributed Constraint Optimization Problem A DCOP can be defined as a tuple (cid:104)A, X, D, F, α(cid:105) (Modi et al. 2005) where, • A is a set of agents {a1, a2, ..., an}. • X is a set of discrete variables {x1, x2, ..., xm}, where each variable xi is controlled by one of the agents ai ∈ A. • D is a set of discrete domains {D1, D2, ..., Dm}, where each Di corresponds to the domain of variable xi. • F is a set of cost functions {f1, f2, ..., fl}, where each fi ∈ F is defined over a subset xi = {xi1, xi2, ..., xik} of variables X and the cost for the function fi is defined for every possible value assignment of xi, that is, fi: Di1 × Di2 ×...× Dik → R. The cost functions can be of Di = [−10, 10] f (x1, x2) = x2 1 − x2 2 x1 x3 x4 x2 f (x1, x3) = x2 f (x1, x4) = 2x2 1 + 2x1x3 1 − 2x2 4 f (x3, x4) = x2 3 + 3x2 4 (a) Constraint Graph (b) Cost Functions Figure 1: Example of an F-DCOP any arity but for simplicity we assume binary constraints throughout the paper. • α : X → A is a variable to agent mapping function which assigns the control of each variable xi ∈ X to an agent ai ∈ A. Each agent can hold several variables. However, for the ease of understanding, in this paper we assume each agent controls only one variable. The solution of a DCOP is an assignment X∗ that minimizes the sum of cost functions as shown in Equation 1. l(cid:88) X∗ = argmin X i=1 fi(xi) (1) are controlled by agents A. Functional Distributed Constraint Optimization Problem Similar to the DCOP formulation, F-DCOP can be defined as a tuple (cid:104)A, X, D, F, α(cid:105). In F-DCOP, A, F and α are the same as defined in DCOP. Nonetheless, the set of variables, X and the set of Domains, D are defined as follows - • X is the set of continuous variables {x1, x2, ..., xm} that • D is a set of continuous domains {D1, D2, ..., Dm}, where each variable xi can take any value between a range, Di = [LBi, U Bi]. As discussed in the previous section, a notable difference between F-DCOP and DCOP is found in the representation of cost function. In DCOP, the cost functions are conven- tionally represented in tabular form, while in F-DCOP each constraint is represented in the form a function (Hoang et al. 2019). However, the goal remains the same as depicted in Equation 1. Figure 1 presents the example of an F-DCOP where Figure 1a represents the constraint graph with four variables where each of the variable xi is controlled by an agent ai. Each edge in Figure 1a stands for a constraint function and the definition of each function is shown in Fig- ure 1b. Each variable xi can take values from the range [-10, 10] in this particular example. Algorithm 1: Particle Swarm Optimization 1 Generate an n-dimensional population, P 2 Initialize positions and velocities of each particle 3 while Termination condition not met do 4 5 6 7 calculate current velocity and position if current position < personal best then for each particle Pi ∈ P do update personal best 8 9 if current position < global best then update global best Particle Swarm Optimization PSO is a population based optimization technique inspired by the movement of a bird flock or a fish school. In PSO, each individual of the population is called a particle. PSO solves the problem by moving the particles in a multi- dimensional search space by adjusting the particle's posi- tion and velocity. As shown in Algorithm 1, initially each particle is assigned a random position and velocity. A fitness function is defined which is used to evaluate the position of each particle. For simplicity, we are going to consider the optimization and minimization interchangeably throughout the paper. In each iteration, the movement of a particle is guided by its personal best position found so far in the search space, as well as the global best position found by the en- tire swarm (Algorithm 1: Lines 4 − 5). The combination of the personal best and the global best position ensures that when a better position is found through the search process, the particles will move closer to that position and explore the surrounding search space more thoroughly considering it as a potential solution. The personal best position of each particle and the global best position of the entire population is updated if necessary (Algorithm 1: Lines 6 − 9). Over the last couple of decades, several versions of PSO have been developed. The standard PSO has a tendency to converge to a suboptimal solution since the velocity component of the global best particle tends to zero after some iterations. Consequently, the global best position stops moving and the swarm behavior of all other particles leads them to follow the global best particle. To cope with the premature conver- gence property of standard PSO, Guaranteed Convergence PSO (GCPSO) has been proposed that provides convergence guarantees to local optima (van den Bergh and Engelbrecht 2002). To adapt similar convergence behavior to F-DCOP, we choose to adapt GCPSO in our proposed method. Challenges The following challenges must be addressed to develop an anytime F-DCOP algorithm that adapts the guaranteed con- vergence PSO: • Particles and Fitness Representation: We need to define a representation for the particles where each particle rep- resents a solution of the F-DCOP. Moreover, a distributed x1 x3 x2 x4 x2 x1 x3 x4 (a) BFS psuedo tree (b) Ordered arrangement Figure 2: Pseudo tree construction and ordered arrangement method for calculating the fitness for each of the particle need to be devised. • Creating Population: In centralized optimization prob- lems, creating the initial population is a trivial task. But in case of F-DCOP, different agents control different vari- ables. Hence, a method need to be devised to coopera- tively generate initial population. • Evaluation: Centralized PSO deals with an n- dimensional optimization task. In F-DCOP, each agent holds k variables (k ≤ n) and each agent is responsible for solving k-dimensional optimization task where the global objective is still an n-dimensional optimization process. • Maintaining Anytime Property: To maintain the any- time property in a F-DCOP model we need to identify the global best particle and the personal best position for each particle. A distribution method needs to be devised to notify all the agents when a new global best particle or personal best position is found. Finally, a coordination method is needed among the agents to update the position and velocity considering the current best position. In the following section we devise a novel method to apply PSO in F-DCOP while maintaining the balance between lo- cal benefit and global benefit. Proposed Method PFD is a PSO based iterative algorithm consisting of three phases: Initialization, Evaluation and Update. In the initial- ization phase, a pseudo-tree is constructed, initial popula- tion is created and parameters are initialized. In the evalu- ation phase, agents distributedly calculate the fitness func- tion for each particle. The update phase keeps track of the best solution found so far and propagates the information to the agents and updates the assignments accordingly. The detailed algorithm can be found in Algorithm 2. Initialization starts with ordering the agents in a Breadth First Search(BFS) pseudo-tree (Chen, He, and He 2017). The pseudo-tree serves the purpose of defining a message passing order which is used in the evaluation and update phase. In the ordered arrangement the agents with lower depths have higher priorities over the agents with higher depths and ties are broken in the alphabetical order. Figure 2(a) and (b) illustrates the BFS pseudo tree and the ordered arrangement. In Figure 2(b), x1 is the root and the arrows represent the message passing direction. From this point for an agent ai, we refer Ni as the set of neighbors, Hi ⊆ Ni and Li ⊆ Ni as the set of higher priority and lower pri- ority neighbors respectively. In Figure 2(b) for agent x3, Ni = {x1, x4}, Hi = {x1} and Li = {x4}. At the beginning of algorithm, PFD takes input from the users to initialize all the parameters where K is the num- ber of particles. The parameters depend on the experiments; the recommended settings for our experiments are discussed later in the text. We define P as the set of K particles which is maintained by the agents where each agent holds com- ponent(s) of the particles. Each particle Pk ∈ P has a ve- locity and a position attribute. The velocity attribute defines the movement directions and position attribute defines the variable values associated with the variables that the agent holds. Then each agent ai executes Init(Algorithm 2: Lines 3 to 8) and initializes the the velocity component, vi to 0 and position component, xi to a random value from its domain Di for each particle Pk. For the example of Figure 2(b), let us assume number of particles, K = 2, and the set of par- ticles, P = {P1, P2} . Here, P1.V = P2.V = {0, 0, 0, 0} shows the complete assignment for velocity attribute of two particles and the complete assignment for position attribute can be shown as, P1.X = {x1 = −1, x2 = 0, x3 = 2, x4 = 9.5}, P2.X = {x1 = 3.5, x2 = 4.9, x3 = 1, x4 = 0}. We define Pk.xi and Pk.vi as the position and velocity component of particle Pk set by agent ai. In this example P1.x3 = 2 which is the value of variable x3 of particle P1 set by agent a3. After selecting the value of its variable each agent shares the particle set, P.xi to its lower priority neigh- bors, Li. For this example, agent a3 sends P.x3 = {2, 1} to its lower priority neighbor a4. Evaluation phase of PFD calculates the fitness of each particle, Pk using a fitness function shown in Equation 2 where Pk.X represents the complete assignment of vari- ables in X. Pk.f itness = fi(Pk.xi) (2) (cid:88) fi∈F This phase starts after the agents receive value assignments from all the higher priority neighbors. Each agent ai is re- sponsible for calculating the constraint cost associated with each of its higher priority neighbors from Hi (Algorithm 2: Lines 13-16). We define Pk.f itness as the local fitness of each particle Pk of the particle set P . When an agent ai re- ceives value assignments P.xi, from a higher priority neigh- bor Hij ∈ Hi, it calculates the constraint cost between them and sends it to Hij . Additionally, each agent except the leaf agents need to pass the constraints cost upward the pseudo- tree calculated by lower priority neighbors, Li (Algorithm 2: Lines 18-19) For the example shown in Figure 1, agent a4 sends the fitness {P1 = 274.75, P2 = 1} to a3 and fitness {P1 = −178.5, P2 = 24.5} to a1. Agent a2 calculates the fitness {P1 = −3, P2 = 19.25} and sends it to a1. Furthermore, a3 receives the fitness from a4 and passes it to a1. Similarly, a2 sends the fitness {P1 = 1, P2 = −11.76} to a1. Update phase consists of two parts: pbest, gbest update Algorithm 2: Particle Swarm F-DCOP 1 Construct BFS pseudo-tree 2 Initialize parameters: K, w, c1, c2, maxsc , maxfc 3 P ← set of K particles 4 Function Init(): 5 Pk.vi ← 0 6 Pk.xi ← a random value from Di 7 for each particle Pk ∈ P do Init() Sends P.xi to agents in Li 8 9 for each agent ai do 10 11 while Termination condition not met each agent ai do 12 13 14 for P.xi received from Hij ∈ Hi do Pk.f itness ← Costi,j(Pk.xi, Pk.xj) for each particle Pk ∈ P do 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 Sends P.f itness to agents in Hij Wait until P.f itness received from all agent in Li if Li (cid:54)= 0 and P.f itness received from all agent in Li then Pk.f itness ←(cid:80) for each particle Pk ∈ P do j∈Li if ai (cid:54)= root then Sends P.f itness to an Hij ∈ Hi P.f itness if ai = root then Update(P.f itness) Wait until P.pbest and P.gbest receives from Hi if P.pbest and P.gbest receives from Hi then Calculate sc and fc according to equation 7, 8 for each particle Pk ∈ P do if P.gbest = Pk then Calculate Pk.vi and Pk.xi according to equation 3, 5 else Calculate Pk.vi and Pk.xi according to equation 4, 5 if Li (cid:54)= 0 then Sends P.xi to agents in Li Sends P.pbest and P.gbest to agents in Li 35 Function Update(P.f itness): P.pbest ← {} 36 for each particle Pk ∈ P do 37 38 39 40 Pk.pbest ← Pk P.pbest ← Pk.pbest ∪ P.pbest if Pk.f itness < Pk.pbest.f itness then 41 42 43 if Pk.f itness < P.gbest.f itness then P.gbest ← Pk Sends P.pbest and P.gbest to agents in Li and variable update. We define pbest to be the personal best position achieved so far by each particle and gbest to be the global best position among all the particles. Since each agent calculates and passes the cost of the constraints to the agents in Hi, the fitness of all the particles propagate to the root. The root agent then sums the fitness values received from the agents in Li for each of the particles, Pk. Then the root agent checks and updates the pbest for Pk ∈ P and gbest for P and sends the new values to the agents in Li (Algorithm 2: Lines 38-44). When an agent ai receives pbest and gbest of the previous iteration, it updates the the velocity component Pk.vi and position component Pk.xi for Pk ∈ P . To adapt the guaranteed convergence method to PFD, two types of update equations for velocity component are defined. If the particle is the current global best particle, the update equa- tion is defined as follows: Pk.vi(t) = −Pk.xi(t − 1) + P.gbest(t − 1)+ w ∗ Pk.vi(t − 1) + ρ ∗ (1 − 2r2) For all other particles, the velocity update equation is defined as follows: Pk.vi(t) = w ∗ Pk.vi(t − 1) + r1 ∗ c1∗ (Pk.pbest(t − 1) − Pk.xi(t − 1)) + r2 ∗ c2∗ (P.gbest(t − 1) − Pk.xi(t − 1)) (3) (4) The position component update equation is same for all the particles which is defined in the following equation: Pk.xi(t) = Pk.xi(t − 1) + Pk.vi(t) (5) In equations 3 4, and 5, Pk.vi(t) and Pk.xi(t) refers to the velocity and position components controlled by agent ai for particle Pk in tth iteration. Here, an iteration refers to a com- plete round of the Evaluation and Update phase (Algorithm 2: Line 12). w is the inertia weight which defines the influ- ence of current velocity on the updated velocity, r1 and r2 are two random values between [0, 1] and c1, c2 are two constants. Combinations of c1 and c2 define the magnitude of influence personal best and global best have on the up- dated particle position. In equation 3, ρ is used to explore a random area near the position of the global best particle. To be precise, ρ defines the diameter of this area that the par- ticles can explore. The value of ρ is adjusted according to equation 6.  ρ(t) = 1 2 ∗ ρ(t − 1) 0.5 ∗ ρ(t − 1) ρ(t − 1) t = 0 sc > maxsc fc > maxfc otherwise (6) In equation 6, sc and fc are the count of consecutive suc- cess and failures respectively. A success is defined when the global best particle updates its personal best position. Sim- ilarly, a failure is defined when the position of the global best particle remains unchanged. The parameters maxsc and maxfc are the upper bound of sc and fc. The following equations define sc and fc. sc(t) = sc(t − 1) + 1 0 if PG.pbest(t) < P.gbest(t − 1) otherwise (7) (cid:40) (cid:40) fc(t) = fc(t − 1) + 1 0 if P.gbest(t) = P.gbest(t − 1) otherwise (8) In equation 7, PG defines the global best particle of iteration t − 1. Each agent calculates sc and fc according to equa- tions 7 and 8 after receiving PG.pbest and P.gbest from their higher priority neighbors, Hi (Algorithm 2: Line 27). Consider agent a1 in Figure 2. When a1 receives fitness value from all of its lower priority neighbors, it is ready to calculate the P.pbest and P.gbest. The final updated fitness value, P.f itness = {94.25, 33}. Based on the updated val- ues ai constructs P.pbest = {94.25, 33} and P.gbest = 33 and notifies the agents in Li. Then each agent calculates sc and fc and updates the values based on equation 3, 4, and 5. Theoretical Analysis In this section, we first prove PFD is an anytime algorithm that is, solution quality improves and never degrades over time. Later, we provide the theoretical complexity analysis in terms of communication, computation and memory. Lemma 1: At iteration 1 t+d , root is aware of the P.pbest and P.gbest up to iteration t, where d is the longest path be- tween root and any node in the pseudo-tree. To prove this lemma it is sufficient to show that, at iter- ation t + d, root agent has enough information to calculate P.pbest and P.gbest up to iteration t, that is, root agent knows the fitness of each particle. To calculate the fitness of each particle Pk.f itness according to equation 2, the root agent needs cost messages from the agents in Lai. The cost mes- sages from agents at d distance from root will need d − 1 iteration to reach agents in Lai. By induction, it will take t + d iterations to reach the cost messages calculated at iter- ation t from the agents with distance d to root. Lemma 2: At iteration t + 2d, each agent is aware of the P.pbest and P.gbest up to iteration t In PFD, for any agent ai, the value message passing length and cost message passing length from the root are same. So, it takes d iterations to reach the P.pbest and P.gbest to the agents at distance d from the root. Using lemma 1 and the above claim, it takes t+d+d = t+2d iter- ations to reach P.pbest and P.gbest to the agent at d distance from the root. Proposition 1: PFD is an anytime algorithm. By lemma 2, at iteration t + 2d and t + 2d + δ (δ ≥ 0) each agent is aware of the P.pbest and P.gbest up to iteration t and t+δ respectively. Let us assume, P.pbest and P.gbest at iteration t+2d+δ ≥ t+2d. But for any δ ≥ 0, t+2d+δ >= t + 2d and P.pbest and P.gbest only gets updated if a better solution is found. Therefore, using proof by contradiction, P.pbest and P.gbest at iteration t + 2d + δ ≤ t + 2d that is, solution quality improves monotonically as the number 1For the theoretical analysis section, iteration refers to the com- munication steps required. In one communication step agents only directly communicate with the neighbors. of iterations increases. Thus we prove, PFD is an anytime algorithm. Complexity Analysis We define, the total number of agents a = n and the total number of neighbors of an agent ai ∈ a, Ni = Li +Hi. In PFD, during the Initialization and Update phase an agent sends Li messages. Additionally, during the Evaluation phase an agent sends Hi + 1 messages. After one round of completion of Initialization, Evaluation and Update phases, an agent ai sends 2 ∗ Li + Hi = Li + Ni messages. In the worst case, the graph is complete where Ni = n. In a complete graph if Li = n, then Hi = 0. There- fore, the total number of messages sent by an agent ai is O(2 ∗ Li + Hi) = O(2n) in the worst case. The size of each message can be calculated as the size of each particle multiplied by the number of particles. If the total number of particle is K, at each iteration the total mes- sage size for an agent ai is O(K ∗ n ∗ 2n) = O(n2) in the worst case. During an iteration, an agent only needs to calculate Pk.vi and Pk.xi for each of the particle Pk. Hence, the to- tal computation complexity per agent during an iteration is O(2 ∗ K) = O(K) where K is the number of particles. Experimental Results In this section, we empirically evaluate the quality of solu- tions produced by PFD with HCMS and AF-DPOP on two types of graphs: Random Graphs and Random Trees. How- ever, CMS is not used in comparison because it only works with peicewise linear functions which is not applicable in most of the real world applications. Although Hoang et al., proposed three versions Functional DPOP, we only compare with AF-DPOP here. The reason is AF-DPOP is reported to provide the best solution among the approximate algorithms proposed in their work. For the experimental performance evaluation, binary quadratic functions are used which are of form ax2 + bxy + cy2. Note that, although we choose binary quadratic functions for evaluation, PFD is broadly applica- ble to other class of problems. The experiments are carried out on a laptop with an Intel Core i5-6200U CPU, 2.3 GHz processor and 8 GB RAM. The detailed experimental set- tings are described below. Random Graphs: For random graphs we use three set- tings - sparse, dense and scale-free. Figure 3 shows the comparison of average costs on Erdos-R´enyi topology with sparse settings (edge probability 0.2) varying the number of agents. We choose coefficients of the cost functions (a, b, c) randomly between [−5, 5] and set the domains of each agents to [−50, 50]. For our proposed algorithm PFD, we set the parameters, K = 2000, w = 0.9, c1 = 0.9, c2 = 0.1, maxfc = 5, and maxsc = 15. For both HCMS and AF-DPOP we choose the number of discrete points to be 3. The discrete points are chosen randomly between the domain range. The averages are taken over 50 randomly gen- erated problems. Figure 3 shows that PFD performs bet- Figure 3: Solution Cost Comparison of PFD and the com- peting algorithms varying number of agents (sparse graphs) Figure 4: Solution Cost Comparison of PFD and the com- peting algorithms with iterations (sparse graphs) ter than both HCMS and AF-DPOP on average. Notably, the performance of HCMS varies significantly which re- sults in a high standard deviation. The reason behind the high standard deviation is that, the performance of HCMS on cyclic graph varies on the initial discretization of do- mains of the agents. For no. of agents ≥ 20, AF-DPOP ran out of memory. Thus, we omit the result of AF-DPOP for no. of agents ≥ 20. Figure 4 shows the comparison between PFD and HCMS on sparse graph settings with increasing number of itera- tions. We set the number of agents to 50 and other settings are same as the above experiment. Moreover, we stop both algorithms after 500 iterations. HCMS initially performs slightly better than PFD till 50 iterations since the particles of PFD initially start from random positions and require few iterations to move the particles towards the best position. However, PFD outperforms HCMS later and the improve- ment rate of PFD is steadier than HCMS. Note that, for 50 agents, AF-DPOP run out of memory in our settings. Thus, we omit the result of AF-DPOP here. To compare with the performance of AF-DPOP on larger graphs we use scale-free graphs. Figure 5 shows the aver- age cost comparison between the three algorithms with in- creasing number of agents. PFD shows a comparable perfor- Figure 5: Solution Cost Comparison of PFD and the compet- ing algorithms varying number of agents (scale-free graphs) Figure 7: Solution Cost Comparison of PFD and the com- peting algorithms varying number of agents (random trees) ables. F-DCOP framework is a variant of DCOP frame- work that can model such problems effectively. To solve F- DCOPs, we propose an anytime algorithm called PFD that is inspired by Particle Swarm Optimization (PSO) technique. To be precise, PFD devises a new method to calculate and propagate the best particle information across all the agents which influence the swarm to move towards a better solu- tion. We also theoretically prove that our proposed algorithm PFD is anytime. Moreover, the guaranteed convergence ver- sion of PSO is tailored in PFD which ensures its conver- gence to a local optima. We empirically evaluate our algo- rithm in a number of settings, and compare the results with the state-of-the-art algorithms, HCMS and AF-DPOP. In all of the settings, PFD markedly outperforms its counterparts in terms of solution quality. In the future, we would like to further investigate the potential of PFD on various F-DCOP applications. We also want to explore whether PFD can be extended for multi-objective F-DCOP settings. Figure 6: Solution Cost Comparison of PFD and the com- peting algorithms varying number of agents (dense graphs) mance with HCMS upto 30 agents and outperforms HCMS afterwards. Both PFD and HCMS outperforms AF-DPOP. The huge standard deviation of HCMS results into the com- parable performance with PFD for smaller agents. We choose dense graphs as our final random graph set- tings. Figure 6 shows comparison between the PFD and HCMS on Erdos-R´enyi topology with dense settings (edge probability 0.6). PFD shows comparatively better perfor- mance than HCMS. Note than, AF-DPOP is not used in dense graph due to the huge computation overhead. Random Trees: We use the random tree configuration in our last experimental settings since the memory require- ment of AF-DPOP is less on trees. The experimental config- urations are similar to the random graph settings. Figure 7 shows the comparison graph between PFD and the compet- ing algorithms on random trees. The closest competitor of PFD in this setting is HCMS. On an average, PFD outper- forms HCMS which in turn outperforms AF-DPOP. When the number of agent is 50, PFD shows better performance than AF-DPOP at a significant level. Conclusions In order to model many real world problems, continuous val- ued variables are more suitable than discrete valued vari- A scalable method for multiagent constraint optimization. In IJCAI. [Shi and Eberhart 1999] Shi, Y., and Eberhart, R. C. 1999. Empirical study of particle swarm optimization. In Proceed- ings of the 1999 Congress on Evolutionary Computation- CEC99 (Cat. No. 99TH8406), volume 3, 1945 -- 1950. IEEE. [Stranders et al. 2009] Stranders, R.; Farinelli, A.; Rogers, A.; and Jennings, N. R. 2009. Decentralised coordina- tion of continuously valued control parameters using the In Proceedings of The 8th Interna- max-sum algorithm. tional Conference on Autonomous Agents and Multiagent Systems-Volume 1, 601 -- 608. International Foundation for Autonomous Agents and Multiagent Systems. [Sultanik, Modi, and Regli 2007] Sultanik, E.; Modi, P. J.; and Regli, W. C. 2007. On modeling multiagent task scheduling as a distributed constraint optimization problem. In IJCAI, 1531 -- 1536. [van den Bergh and Engelbrecht 2002] van den Bergh, F., and Engelbrecht, A. P. 2002. A new locally convergent particle swarm optimiser. In IEEE International conference on systems, man and cybernetics, volume 3, 6 -- pp. IEEE. [van Leeuwen and Pawelczak 2017] van Leeuwen, C. J., and Pawelczak, P. 2017. Cocoa: A non-iterative approach to a local search (a)dcop solver. In AAAI. [Voice et al. 2010] Voice, T.; Stranders, R.; Rogers, A.; and Jennings, N. R. 2010. A hybrid continuous max-sum algo- rithm for decentralised coordination. In ECAI, 61 -- 66. [Yedidsion and Zivan 2016] Yedidsion, H., and Zivan, R. 2016. Applying dcop mst to a team of mobile robots with directional sensing abilities: (extended abstract). In AAMAS. [Zhang et al. 2005] Zhang, W.; Wang, G.; Xing, Z.; and Wit- tenburg, L. 2005. Distributed stochastic search and dis- tributed breakout: properties, comparison and applications to constraint optimization problems in sensor networks. Ar- tificial Intelligence 161(1-2):55 -- 87. [Zhang et al. 2007] Zhang, J.-R.; Zhang, J.; Lok, T.-M.; and Lyu, M. R. 2007. A hybrid particle swarm optimization -- back-propagation algorithm for feedforward neural net- Applied mathematics and computation work training. 185(2):1026 -- 1037. References [Abido 2002] Abido, M. 2002. Optimal design of power- system stabilizers using particle swarm optimization. IEEE transactions on energy conversion 17(3):406 -- 413. [Chen et al. 2018] Chen, Z.; Wu, T.; Deng, Y.; and Zhang, C. 2018. An ant-based algorithm to solve distributed constraint optimization problems. In Thirty-Second AAAI Conference on Artificial Intelligence. [Chen, He, and He 2017] Chen, Z.; He, Z.; and He, C. 2017. An improved dpop algorithm based on breadth first search pseudo-tree for distributed constraint optimization. Applied Intelligence 47(3):607 -- 623. [Eberhart and Kennedy 1995] Eberhart, R., and Kennedy, J. In Proceedings of 1995. Particle swarm optimization. the IEEE international conference on neural networks, vol- ume 4, 1942 -- 1948. Citeseer. [Farinelli et al. 2008] Farinelli, A.; Rogers, A.; Petcu, A.; and Jennings, N. R. 2008. Decentralised coordination of low-power embedded devices using the max-sum algorithm. In Proceedings of the 7th international joint conference on Autonomous agents and multiagent systems-Volume 2, 639 -- 646. International Foundation for Autonomous Agents and Multiagent Systems. [Farinelli, Rogers, and Jennings 2014] Farinelli, A.; Rogers, A.; and Jennings, N. R. 2014. Agent-based decen- tralised coordination for sensor networks using the max- sum algorithm. Autonomous agents and multi-agent systems 28(3):337 -- 380. [Fitzpatrick and Meetrens 2003] Fitzpatrick, S., and Mee- trens, L. 2003. Distributed sensor networks a multiagent perspective, chapter distributed coordination through anar- chic optimization. [Hoang et al. 2019] Hoang, K. D.; Yeoh, W.; Yokoo, M.; and Rabinovich, Z. 2019. New algorithms for functional dis- tributed constraint optimization problems. arXiv preprint arXiv:1905.13275. [Hsin and Liu 2004] Hsin, C.-f., and Liu, M. 2004. Network coverage using low duty-cycled sensors: random & coordi- nated sleep algorithms. In Proceedings of the 3rd interna- tional symposium on Information processing in sensor net- works, 433 -- 442. ACM. [Litov and Meisels 2017] Litov, O., and Meisels, A. 2017. Forward bounding on pseudo-trees for dcops and adcops. Artificial Intelligence 252:83 -- 99. [Maheswaran, Pearce, and Tambe 2004] Maheswaran, R. T.; Pearce, J. P.; and Tambe, M. 2004. Distributed algorithms for dcop: A graphical-game-based approach. In ISCA PDCS, 432 -- 439. [Modi et al. 2005] Modi, P. J.; Shen, W.-M.; Tambe, M.; and Yokoo, M. 2005. Adopt: Asynchronous distributed con- straint optimization with quality guarantees. Artificial Intel- ligence 161(1-2):149 -- 180. [Petcu and Faltings 2005] Petcu, A., and Faltings, B. 2005.
1910.00767	1	1910	2019-10-02T03:54:11	Cognitive Agent Based Simulation Model For Improving Disaster Response Procedures	[ "cs.MA", "cs.AI" ]	In the event of a disaster, saving human lives is of utmost importance. For developing proper evacuation procedures and guidance systems, behavioural data on how people respond during panic and stress is crucial. In the absence of real human data on building evacuation, there is a need for a crowd simulator to model egress and decision-making under uncertainty. In this paper, we propose an agent-based simulation tool, which is grounded in human cognition and decision-making, for evaluating and improving the effectiveness of building evacuation procedures and guidance systems during a disaster. Specifically, we propose a predictive agent-wayfinding framework based on information theory that is applied at intersections with variable route choices where it fuses N dynamic information sources. The proposed framework can be used to visualize trajectories and prediction results (i.e., total evacuation time, number of people evacuated) for different combinations of reinforcing or contradicting information sources (i.e., signage, crowd flow, familiarity, and spatial layout). This tool can enable designers to recreate various disaster scenarios and generate simulation data for improving the evacuation procedures and existing guidance systems.	cs.MA	cs	Cognitive Agent Based Simulation Model For Improving Disaster Response Procedures Rohit K. Dubey D-INFK, ETH-Zurich Samuel S. Sohn Computer Science, Rutgers University Future Cities Laboratory, Singapore-ETH Centre New Jersey, USA Singapore [email protected] [email protected] Christoph Hoelscher D-GESS, ETH-Zurich Mubbasir Kapadia Computer Science, Rutgers University Future Cities Laboratory, Singapore-ETH Centre New Jersey, USA Singapore [email protected] [email protected] Abstract In the event of disaster, saving human lives is of utmost importance. For developing proper evacuation procedures and guidance systems, behavioural data on how people respond during panic and stress is crucial. In the absence of real human data on building evacuation, there is need for a crowd simulator to model egress and decision-making under uncertainty. In this paper, we propose an agent-based simulation tool, which is grounded in human cognition and decision-making, for evaluating and improving the effectiveness of building evacuation procedures and guidance systems during a disaster. Specifically, we propose a predictive agent- wayfinding framework based on information theory that is applied at intersections with variable route choices where it fuses N dynamic information sources. The proposed framework can be used to visualize trajectories and prediction results (i.e., total evacuation time, number of people evacuated) for different combinations of reinforcing or contradicting information sources (i.e., signage, crowd flow, familiar- ity, and spatial layout). This tool can enable designers to recreate various disaster scenarios and generate simulation data for improving the evacuation procedures and existing guidance systems. 1 Introduction The ability to evacuate people from densely populated, large, and complex buildings during a natural or human-made disaster is an essential design issue. To this end, engineers and designers rely on the conventional evacuation design codes and standards (e.g., door dimensions and the minimum number of exits) [1]. The analysis of recent disastrous incidents in buildings indicates that the conventional design codes and standards are not sufficient by themselves and highlights the need to study evacuation guidelines and the occupants' interactions with the building [2]. Cognitive agent-based simulation tools may aid the designers to (a) evaluate the evacuation procedures in both existing and future buildings and (b) highlight some of the problem areas (e.g., choke points during chaotic crowd motion and the lack of exit signs). Moreover, by simulating various unforeseen circumstances, such tools may aid in the training and decision-making strategies of first responders and building security personnel. Wayfinding in a complex indoor environment is a dynamic process which is mentally demanding. Occupants have to continuously pick up relevant wayfinding cues from the environment, interpret them, and make route decisions accordingly. This process is dependent on physical (e.g., height, Preprint. Under review. visual acuity) and psychological (e.g., attention, stress, anxiety, panic) factors. Stress and panic can influence navigational behaviour, which may differ vastly between an egress scenario and general circulation. Thus, there is a need to generate simulation data on how occupants might behave in the event of an emergency under panic, stress, chaotic crowd movement, and uncertainty. In the proposed cognitive agent-based framework, we decompose the directional decision into high and low levels, where a high-level decision (i.e., a macro-decision) chooses a global route to travel along, and a low-level decision (i.e., a micro-decision) chooses a local direction to move in. The wayfinding process becomes more pronounced when the occupant has to decide at an intersection. In this paper, we propose a simulation framework to model the macro- and micro-decision-making of agents at intersections based on previous research findings. We specifically study the effects of four dynamic information sources under the influence of stress due to panic, but our framework generalizes to a variable number of sources: Signage - Occupants rely on signage in the absence of other wayfinding cues [3]. Crowd flow - Theoretical work has suggested that during egress and under stress, occupants may develop a tendency to follow others, a phenomenon called the "herding effect" [4, 5, 6]. Spatial layout - A corridor with longer radial line of sight [7, 8] and higher occlusivity [9] tends to bias human path choices. Memory - Familiarity to the environment along with reinforced information from other directional information such as signage or crowd flow influences human decision-making at a decision point. [10]. Moreover, when making a decision, an occupant will occasionally be confronted with conflicting information from different sources [11] (e.g., the person may receive a conflicting direction from a security guard compared to the direction provided by an "EXIT" sign during an evacuation). Therefore, it is essential to study the impact of these conflicts on a simulated agent's decision-making. To this end, we evaluate our proposed framework both under the dynamic change of an individual information source and in various combinations of either reinforcing (e.g., signage (S) + spatial layout (P) , spatial layout (P) + crowd flow (C), etc.) or contradicting (e.g., S - P, P - C, S + C - P, etc.) information sources. The proposed framework enables the facility manager/designer of a building to systematically evaluate the influence of environmental and psychological factors on egress performance in large and complex buildings. Moreover, the cognitively inspired decision-making model based on human uncertainty could help to improve the research in the field of disaster responses. 2 Related Work Experiments and simulations which study crowd evacuation from a building during emergencies already exist. However, due to ethical and safety-related reasons, it is not possible to conduct an experiment with real participants. Therefore, in order to understand human behaviours, researchers typically focus on analysing previous events or develop computational models using human-like virtual agents. Here, we briefly mention some of the work done in agent-based egress models. EvacSim models the egress of tall buildings with a large number of agents [12]. In this simulation, designers have the flexibility of selecting building behaviours by choice or by assigning a probability. Agents can interact with each other and with the environment. MASSEgress is an example of a pattern-based model wherein, agent behaviours are dependant on the surrounding environment, past experiences, and social or rational inferences [13]. One of the most successful pattern-based evacuation tools is buildingEXODUS [14]. Occupant, movement, behaviour, toxicity, and hazard are the five interacting elements which govern the simulation in buildingEXODUS. One of the drawbacks in pattern-based models is the predefinition of agent interactions which is computationally expensive and prohibits the modelling of unforeseen situations. In the last two decades, many force-based egress models have been studied. Social force model [15] was the first one to study evacuee motion using a mixture of real (physical) and virtual (social) forces. Later on, many revisions of the social force model were implemented to improve its functionality [16, 17]. In the more recent past, force-based models have been criticised by a few researchers Limitations. Some limitations of the works mentioned above are that they model the agent's decision- making either by predetermining their behaviour or by considering environment factors (e.g., signage, spatial layout, and crowd flow) in isolation. These models fail to both account for the absence of this isolation in varying degrees and generalize to new information. Bode et al. [10] show that while one-directional information cues (e.g., crowd flow and memory) may not affect occupants' decisions during an evacuation in isolation, they can have an influence when combined with other information sources. 2 3 Agent Wayfinding Prediction Model In this section, we propose an agent wayfinding prediction model that fuses multiple information sources. 3.1 General Formalism The preliminaries are defined as follows. X is the set of all M macro-decisions. Vector ot consists of the observations made of the N information sources at time t from location l. Γ is the set of neighboring positions for location l. The functions fi are constituent macro-decision-making models based on N 1 physical information sources and function fmem is a model based on memory as its information source. Matrix FM N consists of the constituent models' probability distributions. Function G fuses F into a single probability distribution over X. − × Function ∆ outputs the macro-decisions made by the information-theoretic framework, which thresholds the maximal macro-decision to determine whether to output it. Function δ evaluates the neighboring positions in Γ and outputs the one (i.e., a micro-decision) where the information sources at that position maximize function C. At every time step t, ot is updated and function δ is evaluated. Every W time steps, function ∆ is evaluated after having accumulated observations in memory. 3.2 Quantifying Directional Information from Multiple Sources During the wayfinding decision process, an agent employs its perception model to quantify the in- formation from N sources (e.g., signage, spatial layout, crowd flow, and memory) that guides its macro-decisions in a virtual environment. The nav- igable areas are divided into an array of rectangular grid cells, which serve as reference points for an agent's location. The size of a grid cell is set to 0.5 meters by 0.5 meters because it approximates the average step length and size of an adult. Grid cell locations are pre-computed in the virtual environ- ment and used as nodes in an 8-connected graph Figure 1: Three hypothetical probability distribu- for agent navigation. The connectivity of this graph tions (Sign, Crowd, and Space) are employed to compute the confidence probability in choosing a determines the function Γ(l), which computes the corridor. neighboring cells at a given cell l (Section 3.1). In order to realistically model the interaction between agents and the environment, a human-like visual perception model is used, which collects information at time t into ot. The effective horizontal field of view (FOV) is 120 degrees in order to account for human neck rotation. Although the proposed framework supports N information sources, we specifically use four information sources in the context of this paper to create decision-making models fi. ∈ Signage. We use an entropy-based information-theoretic principle to quantify the information provided by a sign in a virtual environment. In particular, a function is used to compute the visibility information from a sign s ot at location l. The visibility-based confidence in a macro-decision is proportional to the relative angle and the distance between sign s and location l [3]. An entropy-based decision confidence distribution (Figure 1) is then employed to compute the confidence probability fsign(ot) of perceiving the directional information for each possible macro-decision x ∈ X. Spatial Layout. To quantify space, we rely on four isovist measures (i.e., max radial line, isovist area, isovist perimeter, and isovist occlusivity) [18]. Instead of using one single isovist per intersection based on the agent's FOV, we divide it into M -many (i.e., the number of possible macro-decisions X) smaller partial isovists. The aforementioned isovist measures are computed for each partial isovist based on [18] and then aggregated. To compute the confidence probability fspace(O) afforded to each macro-decision, we employ a hypothetical distribution (Figure 1) based on the ratio of the partial isovist measures [8, 7, 19]. 3 Crowd Flow. To compute an agent's perception of crowd flow for each macro-decision, we have a function fcrowd compute the number of visible agents inside the isovist polygon for that route. Then, a hypothetical distribution (Figure 1) is used to generate the confidence probability fcrowd(ot) for each macro-decision. Memory. The memory information source applies the function fmem to fuse the probability dis- tributions of the N 1 other information sources over W = 3 time steps using observations ot W +1, , ot. The probability distributions are first converted to beliefs and then combined using a temporal weighted combination rule [20], which effectively weights time steps to value newer information over older information. · · ·− − 3.3 Information-Theoretical Framework The framework described in this section takes as input N probability distributions in the form of F and outputs either one of M macro-decisions or no decision. Based on [21], a multi-source information fusion method is proposed that considers Jensen-Shannon divergence (JSD) and Shannon entropy (H) in order to determine the confidence in each of the macro-decisions. JSD is employed to measure uncertainty between information sources and entropy is used to measure uncertainty within information sources. The steps involved in this information-theoretical approach are described below. Step 1: We compute the JSD between each pair of sources, where A = 2·Fi(x) Fi(x)+Fj (x) and B = 2·Fj (x) . Fi(x)+Fj (x) Step 2: The average JSD (JSDµ) of information source i i can be calculated by Equation 2. Step 3: The support degree Supi of information source i is defined in Equation 3, where E = 10−5 is used in practice. Step 4: The credibility degree Crdi of information source i is defined in Equation 4, where the range of Crdi is [0, 1]. Step 5: We then measure the normalized Shannon entropy of each information source i, which is the entropy of source i divided by the maximum possible entropy for X. Step 6: Based on the normalized entropy H i, the credibil- ity degree Crdi is adjusted, giving the confidence in the information provided by source i. On account of the confidence in each information source i, the confidence distribution over X will be obtained by Equation 6. Steps 1 through 6 correspond to G(F) (Section 3.1), which transforms the input F into the output confidence distribution over X. The rule that this framework uses to output a macro-decision is based on the highest confidence in the output distribution. If this value exceeds θ, the framework is sufficiently confident in the corresponding macro-decision, which it makes. Otherwise, the framework will not make a decision. This decision rule corresponds to function ∆(G(F)). 3.4 Agent Decision Model According to [22], agents should take three steps between making macro-decisions in order to simulate realistic wayfinding. The memory information source is the only source that spans multiple time steps, and it accommodates this type of decision-making by ensuring that no time step has its observations ignored. However, without changing the location of the agent, the memory will not be fusing different probability distributions per each information sources. Therefore, we must have the agent move while it is deliberating on its macro-decision by making one micro-decision δ per time √ 2 meter step in the direction γ that step (Section 3.1). This micro-decision is either a 1 meter or 2 maximizes function C, which takes the maximum probability of a macro-decision given by either signage fsign(Ot(γ)) or spatial layout fspace(Ot(γ)). This micro- and macro-decision-making cycle repeats until a physical user-defined threshold is reached or the agent within a certain proximity to the intersection. In either case, the final macro-decision made by function ∆ is chosen as the agent's goal direction, while all prior macro-decisions are predictions of the agent's goal direction at that point in time. 2 4 4 Experiments & Results This section describes the simulation performed to verify the proposed dynamic, uncertain information fusion framework. The general test-case we present is a wayfinding decision-making problem at an intersection/decision point with two and four route choices under the influence of multiple information sources. 4.1 Effects of Reinforced and Contradictory Combinations of Multiple Information Sources P S C P S C Left (L ) Right ( R) Test Cases 100 agents were spawned randomly and assigned a wayfinding task of finding a target location (e.g., Find Gate A2) for each test-case. In Table 1, we presents the prediction results for eight test-cases. Sign+Crowd+Space S+, C+, P+ : Test-case 1 represents the reinforced combination. All three information sources are strongly affording the directional information to take "Left". We found that almost 90% of the agents predicted to go left. It is in agreement with previous work stating that the presence of sign has a strong effect on human wayfinding decision making. Sign - (Crowd+Space) S+, C−, P− : Test- cases 2 and 3 represent the conflicting informa- tion from the sign and reinforced information Prediction Results Information Theory 1 (S+, C+, P+) Yes High High No Med Med L (89%) R (11%) 2 (S+, C−, P−) Yes Low Low No High High L (4%) R (96%) 3 (S+, C−, P−) Yes Med Low No High High L (5%) R (95%) No Low Med No Med Med L (5%) R (95%) No Low High No Med Low L (6%) R (94%) No Low High No Low Low L (27%) R (73%) No High Med No Low High L (44%) R (56%) No High Low No Med High L (11%) R (89%) Table 1: Prediction results from the proposed method are shown in various combinations of information sources (Signage S, Crowd Flow C, and Spatial Layout P). 4 (C−, P+) 5 (C+, P−) 6 (C−, P+) 7 (C−, P+) 8 (C+, P−) ( ) ( ) ) ( from the combination of crowd and space. In both test-cases, almost all agents' decisions were influenced by the strong presence of the crowd and space. Crowd - Space S+, S− : Test- cases 4 to 8 represent the effect of conflicting information from crowd and space in the absence of a sign. By decreasing the confi- dence from the crowd's presence (for right corridor) in test-case 6, we observe a considerable reduc- tion in the number of agents select- ing that corridor in comparison to test-case 5. Figure 2: We highlight two sections in the agent decision-making process separated over time. In the beginning phase, agents predict to take the left route due to its position and orientation. Agents perceive higher confidence from the crowd on the left corridor, and the difference in the confidence between the left and right corridors afforded by space is small. In the later phase (middle image), agent perception of both crowd and space changes and the confidence afforded by space and crowd gradually increases towards the right corridor. We replicate the same set-up and increase the crowd flow at the far end of the left corridor (rightmost image). The agent is exposed to this new crowd only at the later phase of decision-making, resulting in the change of its prediction. The results observed from the pro- posed framework highlights its non-deterministic nature. We be- lieve this is due to two main rea- sons. Firstly, the direction of ap- proach of an agent at an inter- section influences the confidence afforded by various information sources and secondly, the continuous change in the information over a temporal axis. 5 Conclusions In the absence of real crowd behaviour data in disasters and emergencies, a simulation tool which mimics human decision making capability and models the crowd emergence behaviour due to micro- level occupant behaviour is a viable alternative. In this paper, we have proposed an information- theoretic agent wayfinding prediction framework which predicts an agent's navigational decision at an exit/intersection with M route choices under the influence of N information sources. The simulation results highlight the non-deterministic nature of our framework and produces realistic results that are consistent with previous works. We demonstrate that our information-theoretic method is also able to fuse the uncertain and dynamically changing information over time. The proposed work can be used to model the individual-level interactions and decision-making of an agent and can be used to study the evacuation behaviour or general circulation of a crowd in an indoor environment. One limitation of our proposed method is its dependence on probability distributions assigned to 5 individual information sources. A small error in the distribution can result in faulty decisions. Also, the mutual information between N information source is not considered. We aim to extend our model by utilizing context as an information source of information as proposed in [23]. References [1] Richard W Bukowski and Jeffrey S Tubbs. Egress concepts and design approaches. In SFPE Handbook of Fire Protection Engineering, pages 2012 -- 2046. Springer, 2016. [2] Suvek Salankar, SM Tauseef, and RK Sharma. Need for better high-rise building evacuation practices. In Advances in Fire and Process Safety, pages 191 -- 205. Springer, 2018. [3] Lazaros Filippidis, Edwin R Galea, Steve Gwynne, and Peter J Lawrence. Representing the influence of signage on evacuation behavior within an evacuation model. Journal of Fire Protection Engineering, 16(1):37 -- 73, 2006. [4] Dirk Helbing, Illés Farkas, and Tamas Vicsek. Simulating dynamical features of escape panic. Nature, 407(6803):487, 2000. [5] Hassan A Karimi. Indoor wayfinding and navigation. CRC Press, 2015. [6] Mei Ling Chu and Kincho H. Law. Incorporating individual behavior, knowledge, and roles in simulating evacuation. Fire Technology, Jul 2018. [7] B Hillier. Moving diagonally: Some results and some conjectures. London: University College, 1997. [8] Jan M Wiener, Christoph Hölscher, Simon Büchner, and Lars Konieczny. Gaze behaviour during space perception and spatial decision making. Psychological research, 76(6):713 -- 729, 2012. [9] Stephen Kaplan. Environmental preference in a knowledge-seeking, knowledge-using organism. 1992. [10] Nikolai WF Bode, Armel U Kemloh Wagoum, and Edward A Codling. Human responses to multiple sources of directional information in virtual crowd evacuations. Journal of The Royal Society Interface, 11(91):20130904, 2014. [11] CW Johnson. Lessons from the evacuation of the world trade centre, 9/11 2001 for the development of computer-based simulations. Cognition, Technology & Work, 7(4):214 -- 240, 2005. [12] Leong S Poon. Evacsim: A simulation model of occupants with behavioural attributes in emergency evacuation of high-rise building fires. Fire Safety Science, 4:681 -- 692, 1994. [13] Xiaoshan Pan. Computational modeling of human and social behaviors for emergency egress analysis. PhD thesis, Stanford University Stanford, CA, USA, 2006. [14] S Gwynne, ER Galea, Peter J Lawrence, and L Filippidis. Modelling occupant interaction with fire conditions using the buildingexodus evacuation model. Fire Safety Journal, 36(4):327 -- 357, 2001. [15] Dirk Helbing and Peter Molnar. Social force model for pedestrian dynamics. Physical review E, 51(5):4282, 1995. [16] Paul A Langston, Robert Masling, and Basel N Asmar. Crowd dynamics discrete element multi-circle model. Safety Science, 44(5):395 -- 417, 2006. [17] Simo Heliövaara, Timo Korhonen, Simo Hostikka, and Harri Ehtamo. Counterflow model for agent-based simulation of crowd dynamics. Building and Environment, 48:89 -- 100, 2012. [18] Michael L Benedikt. To take hold of space: isovists and isovist fields. Environment and Planning B: Planning and design, 6(1):47 -- 65, 1979. [19] Ruth Dalton. The secret is to follow your nose: Route path selection and angularity. 2001. [20] Xiaoyun Wang and Tingdi Zhao. Operation condition monitoring using temporal weighted dempster-shafer theory. Technical report, BeiHang University Bejiing China, 2014. [21] Fuyuan Xiao. Multi-sensor data fusion based on the belief divergence measure of evidences and the belief entropy. Information Fusion, 46:23 -- 32, 2019. [22] Alasdair Turner and Alan Penn. Encoding natural movement as an agent-based system: An investigation into human pedestrian behaviour in the built environment. 2002. [23] Galina L Rogova and Lauro Snidaro. Considerations of context and quality in information fusion. In 2018 21st International Conference on Information Fusion (FUSION), pages 1925 -- 1932. IEEE, 2018. 6 6 Appendix 6.1 Influence of Memory In the proposed temporal weighted evidence com- bination model we encode the memory of an agent as a fourth information source. We model memory as the continuous process of informa- tion retention over time. The results generated in this paper (see Table 1) is by considering mem- ory for W = 3 time steps (i.e., information from past three time steps is fused according to their temporal weight during decision-making). Our hypothesis is that the uncertainty in an agent's prediction will reduce with the increase in mem- ory time steps. In Figure 3 we plot the impact of the increase in memory duration W on the agent's prediction entropy. We notice a gradual reduction in prediction entropy with the increase in memory. How long the memory should be retained in the context of wayfinding decision-making needs a thorough examination which is beyond the scope of this paper. The result demonstrates that our framework is capable of modelling working memory in an agent wayfinding prediction model. Figure 3: Influence of memory duration W on agent's prediction uncertainty. 6.2 Generalization of The Proposed Model In Figure 4 we expand the number of possible route choices from two to four to showcase the generalization capability of the proposed model to M number of possible route choices. We present four different examples under various combina- tion of information sources. In the first example (Fig. 4(a)), we only keep the impact of spatial layout. The highest confidence of spatial layout evidence is highlighted in light orange (Route 2). Unsurprisingly, the agent decides to take Route 2. In Figure 4(b), we remove the influence of signage and our model adapts to N = 2 informa- tion sources (i.e., crowd flow and spatial layout). Different prediction results at different stages of agent wayfinding are color-coded. We believe, the change in prediction at different time intervals models the fluctuation in the human decisions during a wayfinding task due to the increase or decrease in evidences' confidences. The model successfully captures the temporal property of in- formation sources. In Figure 4(c), we introduce a directional sign which directs to Route 1. In Figure 4(d), we remove the sign and allow the agent to get influenced by the fused information sources (i.e., C and P). Similarly to Figure 4(b), the prediction fluctuates over time and finally, the agent decides to choose Route 3. 7 Figure 4: Visualization of an agent prediction model at an intersection with four route choices. Representation of the proportion of crowd in each corridor is shown in black dots. Isovist polygon for each corridor is shown in a light blue semi-transparent polygon. Agent's tra- jectory is colour-coded to represent its prediction state in that time.
1912.04442	1	1912	2019-12-10T01:26:03	A Study on Accelerating Average Consensus Algorithms Using Delayed Feedback	[ "cs.MA", "math.OC" ]	In this paper, we study accelerating a Laplacian-based dynamic average consensus algorithm by splitting the conventional delay-free disagreement feedback into weighted summation of a current and an outdated term. We determine for what weighted sum there exists a range of time delay that results in the higher rate of convergence for the algorithm. For such weights, using the Lambert W function, we obtain the rate increasing range of the time delay, the maximum reachable rate and comment on the value of the corresponding maximizer delay. We also study the effect of use of outdated feedback on the control effort of the agents and show that only for some specific affine combination of the immediate and outdated feedback the control effort of the agents does not go beyond that of the delay-free algorithm. Additionally, we demonstrate that using outdated feedback does not increase the steady state tracking error of the average consensus algorithm. Lastly, we determine the optimum combination of the current and the outdated feedback weights to achieve the maximum increase in the rate of convergence without increasing the control effort of the agents. We demonstrate our results through a numerical example.	cs.MA	cs	A Study on Accelerating Average Consensus Algorithms Using Delayed Feedback Hossein Moradian, Student Member, IEEE and Solmaz S. Kia, Senior Member, IEEE i 9 1 0 2 c e D 0 1 ] A M . s c [ 1 v 2 4 4 4 0 . 2 1 9 1 : v i X r a Abstract -- In this paper, we study accelerating a Laplacian- based dynamic average consensus algorithm by splitting the conventional delay-free disagreement feedback into weighted summation of a current and an outdated term. We determine for what weighted sum there exists a range of time delay that results in the higher rate of convergence for the algorithm. For such weights, using the Lambert W function, we obtain the rate increasing range of the time delay, the maximum reachable rate and comment on the value of the corresponding maximizer delay. We also study the effect of use of outdated feedback on the control effort of the agents and show that only for some specific affine combination of the immediate and outdated feedback the control effort of the agents does not go beyond that of the delay- free algorithm. Additionally, we demonstrate that using outdated feedback does not increase the steady state tracking error of the average consensus algorithm. Lastly, we determine the optimum combination of the current and the outdated feedback weights to achieve maximum increase in the rate of convergence without increasing the control effort of the agents. We demonstrate our results through a numerical example. I. INTRODUCTION The average consensus problem for a group of networked agents each endowed with a reference input signal (dynamic or static) is defined as designing a distributed interaction policy for each agent such that a local agreement state converges asymptotically to the average of the reference signals across the network. For this problem, in continuous time domain, when the reference signals of all the agents are static, the well-known distributed solution is the Laplacian consensus al- gorithm [1] -- [4]. In the Laplacian consensus, each agent initial- izes its first order integrator dynamics with its local reference value and uses the weighted sum of the difference between its local state and those of its neighbors (disagreement feedback) to drive its local dynamics to the average of the reference signals across the network. When the reference signals are dynamics, agents use a combination of the Laplacian input and their local reference signal and/or its derivative to drive their local integrator dynamics; see [5] for examples of dynamic average consensus algorithms. Average consensus algorithms are of interest in various multi-agent applications such as sensor fusion [6] -- [9], robot coordination [10], [11], formation control [12], distributed optimal resource allocation [13], [14], distributed estimation [15] and distributed tracking [16]. For these cooperative tasks, it is highly desired that the consensus among the agents is obtained fast, i.e., the consensus algorithm converges fast. For a connected network with undirected communication, it is well understood that the convergence rate of the average consensus algorithms is associated with The authors are with the Department of Mechanical and Aerospace Irvine, CA 92697, Engineering, University of California {hmoradia,solmaz}@uci.edu Irvine, the connectivity of the graph [17], specified by the smallest non-zero eigenvalue of the Laplacian matrix [1], [5]. Given this connection, various efforts such as optimal adja- cency weight selection for a given topology by maximizing the smallest non-zero eigenvalue of the Laplacian matrix [3], [18] or rewiring the graph to create topologies such as small-world network [19], [20] with high connectivity have been proposed in the literature. In this paper, we study use of outdated disagreement feedback to increase the convergence rate of a dynamic average consensus algorithm. Our method can be applied in conjunction with the aforementioned topology designs to maximize the acceleration effect. In performance analysis of dynamical systems intuition inad- vertently ties time delay to sluggishness and adverse effects on system response. However, some work such as [21] -- [28] point to the positive effect of time delay on increasing stability margin and rate of convergence of time-delayed systems. Specifically, the positive effect of time-delayed feedback in accelerating the convergence of the static average consensus Laplacian algorithm is reported in [26] -- [28]. The study in [27] and [28] consider delaying the immediate Laplacian disagree- ment feedback, and show that when the network topology is connected there always exists a range of delay (0, τ ) such that the rate of convergence of the modified algorithm is faster. The technical results also include specifying τ and also showing that the maximum attainable convergence rate due to employing delayed feedback is the Euler number times the rate without delay. [27] also specifies the delay for which maximum convergence rate is attained. However, the effect of use of outdated feedback on the control effort of the agents is left unexplored in [27] and [28]. This study is of importance because one can always argue that the convergence rate of the original Laplacian algorithm can be increased by multiplying the Laplacian input with a gain greater than one. But, this choice leads to increase in the control effort of the agents. On the other hand, [26] studies a modified Laplacian algorithm where the immediate Laplacian disagreement feedback input is broken in half and one half is replaced by outdated delayed feedback. For this modified algorithm, [26] shows that it is possible to increase the convergence rate for some values of delay. They also show that this increase in rate is without increasing the control effort of the agents. However, [26] falls short of specifying the exact range of delay for which the convergence rate can be increased by employing the outdated feedback and also quantifying the maximum rate and its corresponding maximizer delay. In this paper, we study the use of outdated disagreement feedback to increase the convergence rate of the dynamic average consensus algorithm of [29]. This algorithm, when the reference signal of the agents are all static, simplifies to the static Laplacian average consensus algorithm [4]. In our study, we split the disagreement Laplacian feedback into two components of immediate and outdated feedback. However in- stead of equal contribution, we consider the affine combination of the current and outdated feedback to investigate the effect of the relative size of the outdated and immediate feedback terms on the induced acceleration. Our comprehensive study includes [26], [28] and [27] as special cases. We note here that the analysis methods used in [26], [28] and [27] do not generalize to study the case of affine combination of the immediate and outdated feedback. This is due to the technical challenges involved with study of the variation of the infinite number of the roots of the characteristic equation of the linear time delayed systems with delay, which often are resolved via methods that conform closely to the specific algebraic structure of the system under study. We recall here that the exact value of the worst convergence rate of a linear time- invariant system, with or without delay, is determined by the magnitude of the real part of the right most root of its characteristic equation [30], [31]. We start our study by characterizing the admissible range of delay for which the average consensus tracking is maintained. Then, we show that for the delays in the admissible range, the ultimate tracking error of our modified average consensus al- gorithm of interest is not affected by use of outdated feedback regardless of the affine combination's split factor. However, we show that the control effort of the agents does not increase only for a specific range of the split factor of the affine combination of the outdated and immediate feedback. Our results also specify (a) for what values of the system parameters the rate of convergence in the presence of delay can increase, (b) the exact values of delay for which the rate of convergence increases, and (c) the optimum value of τ corresponding to the maximum rate of convergence in the presence of delay. In light of all the aforementioned study, we summarize our results in the remark that discuses the trade off between the performance (maximum convergence rate) and the robustness of the algorithm to the delay as well as the level of control effort. Our study relies on use of the Lambert W function [32], [33] to obtain the exact value of the characteristic roots of the internal dynamics of our dynamic consensus algorithm. Via careful study of variation of the right most root in the complex plan with respect to delay we then proceed to conduct our study to establish our results. Organization: Notations and preliminaries including a brief review of the relevant properties of the Lambert W function and the graph theoretic definitions are given in Section II. Problem definitions and the objective statements are given in Section III, while the main results are given in Sec- tion IV.Numerical simulations to illustrate our results are given in Section V. Section VI summarizes our concluding remarks. Finally, the appendices contain the auxiliary lemmas that we use to develop our main results. II. NOTATIONS AND PRELIMINARIES In this section, we review our notations, definitions and auxiliary results that we use in our developments. ii 1 + ··· + x2 We let R, R>0, R≥0, Z, and C denote the set of real, positive real, non-negative real, integer, and complex num- bers, respectively. Given i, j ∈ Z with i < j, we define i = {i, i + 1,··· , j}. For s ∈ C, Re(s) and Im(s) represent, Zj respectively, the real and imaginary parts of s. Moreover, s =(cid:112)Re(s)2 + Im(s)2 and arg(s) = atan2(Im(s), Re(s)). For any vector x ∈ Rn, we let (cid:107)x(cid:107) = (cid:112)x2 n and (cid:107)x(cid:107)∞ = max{xi}n i=1. For a measurable locally essentially bounded function u : R≥0 → Rm, we define u∞ = ess sup{(cid:107)u(t)(cid:107)∞, t ≥ 0}. For a matrix A, its ith row is denoted by [A]i. For a linear time-delayed system, admissible delay range is the range of time delay for which the internal dynamics of the system is stable. We recall that for linear time-delayed systems with exponentially stable dynamics when delay is set to zero, by virtue of the continuity stability property theorem [34, Proposition 3.1], the admissible delay range is a connected range (0, ¯τ ) ⊂ R>0 where ¯τ ∈ R>0 is the critical delay bound beyond which the system is always unstable. Lemma II.1 (Admissible delay bound for a scalar time-de- layed system [34, Proposition 3.15]). Consider x(t) = a x(t − τ ) + b x(t), x(η) ∈ R, t ∈ R≥0, η ∈ [−τ, 0], (1) where a ∈ R\{0} and a+b < 0. Then, the following assertions hold (a) For b ≤ −a, system (1) is exponentially stable indepen- (b) For a < −b, system (1) is exponentially stable if and dent of the value of τ ∈ R≥0, i.e., ¯τ = ∞. only if τ ∈ [0, ¯τ ) where ¯τ = arccos(− b √ a ) a2 − b2 . (2) (cid:3) Lambert W function specifies the solutions of s es = z for a given z ∈ C, i.e., s = W (z). It is a multivalued function with infinite number of solutions denoted by Wk(z), k ∈ Z, where Wk is called the kth branch of W function. For any z ∈ C, Wk(z) can readily be evaluated in Matlab or Mathematica. Below are some of the intrinsic properties of the Lambert W function, which we use (see [32], [33]), (3a) for z (cid:54)= 1/ e, Wk(z)/z = 1, lim z→0 d Wk(z)/d z = 1/(z + eWk(z)), (3b) for k ∈ Z. For any z∈R, the value of all the branches of the Lambert W function except for the branch 0 and the branch −1 are complex (non-zero imaginary part). Moreover, the zero branch satisfies W0(−1/ e) =−1, W0(0) = 0 and z ∈ [−1/ e,∞), z ∈ C\[−1/ e,∞), z ∈ R\{−1/ e}, (4a) (4b) (4c) Lemma II.2 (Maximum real part of Lambert W function [32]). For any z ∈ C, the following holds Im(W0(z)) ∈ (−π, π)\{0}, Re(W0(z)) > −1, Re(W0(z)) ≥ max(cid:8) Re(Wk(z)) k ∈ Z\{0}(cid:9). W0(z) ∈ R, (5) 1 d x = x+eW0(x) > 0. e ) where we have Re(W0(z)) = Re(W−1(z)). The equality holds between branch 0 and −1 over z ∈ (−∞,− 1 (cid:3) Lemma II.3 (W0(x) is an increasing function of x ∈ R>0). For any x, y ∈ R>0 if x < y, then W0(x) < W0(y). Proof. The proof follows from the fact that for x ∈ R>0, W0(x) ∈ R>0. Therefore, d W0(x) We follow [35] to define our graph related terminologies and notations. In a network of N agents, we model the inter-agent interaction topology by the undirected connected graph G(V,E, A) where V is the node set, E ⊂ V × V is the edge set and A = [aij] is the adjacency matrix of the graph. Recall that aii = 0, aij ∈ R>0 if j ∈ V can send information to agent i ∈ V, and zero otherwise. Moreover, a graph is undirected if the connection between the nodes is bidirectional and aij = aji if (i, j) ∈ E. Finally, an undirected graph is connected if there is a path from every agent to every other agent in the network (see e.g. Fig. 1). Here, L = Diag(A1N ) − A is the Laplacian matrix of the graph G. The Laplacian matrix of a connected undirected graph is a symmetric positive semi-definite matrix that has a simple λ1 = 0 eigenvalue, and the rest of its eigenvalues satisfy λ1 = 0 < λ2 ≤ ··· ≤ λN . Moreover, L1N = 0. Since L of a connected undirected graph is a symmetric and real 1N , v2,··· , vN matrix, its normalized eigenvectors v1 = 1√ are mutually orthogonal. Moreover for N (cid:105) , R =(cid:2)v2 ··· (cid:3) N T = 1N R (6) we have T(cid:62)LT = Λ = Diag(0, λ2,··· , λN ). We note that for any q ∈ RN , we have (cid:107)R(cid:62)q(cid:107) = (cid:107)(IN − 1 N 1N 1(cid:62) N ) q(cid:107). vN (cid:104) 1√ III. PROBLEM DEFINITION We consider a group of N agents each endowed with a one-sided time-varying measurable locally essentially bounded signal ri : R≥0 → R, interacting over a connected undirected graph G(V,E, A). To obtain the average of their reference inputs, ravg(t) = 1 i=1 ri(t), these agents implement the N distributed algorithm aij(xj(t)−xi(t)) + ri, i ∈ V, (7) (cid:80)N (cid:88)N j=1 xi(t) = −α xi(0) = ri(0), (cid:80)N where α ∈ R>0. When the reference inputs of the agents are all static, i.e., ri = 0 for all i ∈ V, (7) becomes the well-known Laplacian static average consensus algorithm that converges exponentially to xavg(0) = ravg = 1 j=1 rj, with N the rate of convergence ρ0 = αλ2 (for details see [4]). When one or more of the input signals are time-varying, (7) is the dynamic average consensus algorithm of [29].The convergence guarantee of (7) is as follows. Theorem III.1 (Convergence of (7) over an undirected con- nected graph [5]). Let G be a connected undirected graph. Let (cid:107)(IN − 1 N )r(cid:107)∞ = γ < ∞. Then, for any α ∈ R>0, the trajectories of algorithm (7) are bounded and satisfy N 1N 1(cid:62) (cid:12)(cid:12)(cid:12)xi(t) − ravg(t) (cid:12)(cid:12)(cid:12) ≤ 0, lim t→∞ i ∈ V, (8) For further discussion about the convergence rate of linear time delayed systems see [31, Corollary 1]. iii where 0 = γ and ρ0 = αλ2. Moreover, The rate of ρ0 convergence to this error neighborhood is no worse than ρ0. In this paper, with the intention of using outdated information to accelerate the convergence, we alter the average consensus algorithm (7) to (compact representation) x(t) = −α (1 − k) L x(t) − α k L x(t − τ ) + r, xi(0) = ri(0), xi(η) = 0 for η ∈ [−τ, 0), (9a) (9b) for t ∈ R≥0, where k ∈ R and τ ∈ R≥0. For k = 0, (9) recovers the original algorithm (7). We refer to k as split factor. To simply analyzing the convergence properties of (9), we implement the change of variable i ∈ V, z(t) = T(cid:62)(x(t) − ravg(t)1N ) (recall (6)) to write (9) in equivalent form z1(t) = 0, z1(0) = 0, z2:N (t) = −α(1 − k) ¯Λ z2:N (t) − αk ¯Λz2:N (t − τ ) + R(cid:62) r(t), z2:N (0) = R(cid:62)r(0), z2:N (η) = 0 for η ∈ [−τ, 0), (11b) (11c) where ¯Λ = Diag(λ2,··· , λN ). Under the given initial condi- tion, the tracking error then is x(t) − ravg(t)1N = R z2:N (t), t ∈ R≥0. (12) Using the method that specifies the solution of linear time- delayed systems [36], the trajectory of (11b) under initial condition (11c) is (10) (11a) (cid:90) t (cid:88) 0 j∈Z eSj t Cj z2:N (0)+ (cid:48) eSj (t−ζ) C jR(cid:62) r(ζ)dζ, (13) (cid:88) j∈Z z2:N (t) = where j , . . . , CN−1 1 , j Sj = Diag(S1 j , . . . , SN−1 j ), (14a) Wj(−αkλi+1τ eα(1−k)λi+1τ )−α(1 − k)λi+1 (14b) (14c) ), Si j = 1 τ Cj = Diag(C1 Ci (14d) j τ j = 1 − αkλi+1τ e−Si and C(cid:48) j = Cj because of the given initial conditions.When the reference input signals satisfy the condition given in Theorem III.1, it follows from (13) that z2:N in the ad- missible delay range should converge exponentially to some neighborhood of zero, whose size is proportional to γ. More- over, the rate of convergence of algorithm (9) is ρτ (k) = min{{− Re(Si j=−∞. By invoking Lemma II.2, ρτ (k) simplifies to ρτ (k) = min{− Re(Si i=1 , which reads as i=1 }∞ 0)}N−1 j)}N−1 (cid:110) − Re( ρτ (k) = min 1 τ W0(−αkλiτ eα(1−k)λiτ )) + α(1 − k)λi (cid:111)N . (15) i=2 Our objective in this paper is to show that by splitting the disagreement feedback into a current −α (1 − k) L x(t) and an outdated −α (1−k) L x(t−τ ) components, it is possible to increase the rate of convergence of algorithm (9). Specifically, we determine for what values of k, there exists ranges of time delay that the rate of convergence of (9) increases (ranges of delay for which decay rate of the transient response of (9) increases). We also specify the maximum reachable rate due to delay and its corresponding maximizer delay. One may argue that the rate of convergence of (7) can be increased by 'cranking up' the gain α. However, this choice leads to increase in the control effort of the agents. In our study, then, we set to identify values of split factor k for which for a fixed α the increase in the convergence rate of (9) due to delay in comparison to (7) is without increasing the control effort. Finally, we prove that for delays in the admissible delay bound, the ultimate tracking error of (9) is the same as (8). This assertion, increases the appeal of the modified average consensus algorithm (9) as an effective algorithm that yields faster convergence than the original algorithm (7). We close this section by noting that following the change of variable method proposed in [5], algorithm (9a) can be implemented in the alternative way (recall that ri is a one-sided signal) yi(t) = −α (1 − k) L x(t) − α k L x(t − τ ), xi(t) = yi(t) + ri(t), yi(0) = 0, xi(η) = 0 for η ∈ [−τ, 0), i ∈ V, which does not require knowledge of derivative of the refer- ence input of the agents . IV. ACCELERATING AVERAGE CONSENSUS USING OUTDATED FEEDBACK In this section, we study the effect of the outdated feedback on the convergence rate and the ultimate tracking response of the modified average consensus algorithm (9). To start our study, we identify the admissible delay range (0, ¯τ ) for algorithm (9) for different values of split factor k. Given the tracking error (12), the admissible delay bound is determined by the ranges of delay for which the zero input dynamics of (11b) preserves its exponential stability. Lemma IV.1 (Admissible rage of delay for internal stability of algorithm (9)). The following assertions hold for the modified average consensus algorithm (9) over an undirected connected graph (recall (14b)). (a) For k ≤ 0.5, the modified average consensus algo- rithm (9) is internally stable for any τ ∈ R≥0, i.e., ¯τ = ∞. the modified average consensus algo- rithm (9) is internally stable if and only if τ ∈ [0, ¯τ ), where (b) For k > 0.5, ¯τ = arccos(1 − 1/k)/(αλN (16) j t = 0, i ∈ Also, for any τ ∈ [0, ¯τ ), we have limt→∞ eSi and j ∈ Z. Moreover, under the initial condition (9b), ZN−1 the trajectories of xi, i ∈ V of the zero-input dynamics of algorithm (9) converges exponentially fast to xavg(0). 2k − 1). 1 √ iv Proof. Consider the zero-input dynamics of (11), the equiva- lent representation of zero dynamics of algorithm (9). It is evident that the delay tolerance of (11) is defined by the dynamics of states z2:N . Note that (11b) because of definition of ¯Λ reads also as zi(t) =−α(1 − k)λi zi(t) − αk λizi(t − τ ), 2 . (17) When k ≤ 0.5 we have −αλi(1 − k) ≤ αλik, while when k > 0.5 we have −αλik < −αλi(1 − k). Therefore, the admissible delay ranges stated in the statement (a) and the statement (b) follow, respectively, from the statements (a) and (b) of Lemma II.1. To establish (16), we used ¯τ = min{¯τi}N where according to (2) we have √ i ∈ ZN i=2 (18) . ¯τi = arccos(1 − 1 k ) 2k − 1 αλi In admissible delay bound, the time-delayed systems (17) for i ∈ {2,··· , N} are exponentially stable, i.e., zi → 0 as t → ∞, i ∈ {2,··· , N}. As a result, limt→∞ eSi j t = 0, i ∈ ZN−1 and j ∈ Z can be certified from (13) when the second term in the right-hand side is removed (zero-input response). Moreover since z(t) = T(cid:62)x(t) (in zero-input dynamics), we then obtain that in the stated admissible delay ranges in the statements (a) and (b), x(t) converges exponentially fast to 1√ z1(0)1N = 1√ j=1 xi(0))1N = xavg(0). This completes the proof. N (cid:80)N ( 1√ N N 1 The results of Lemma IV.1 includes the result in [37], which specifies the admissible range of delay for when k = 1, as special case. Next, we study the ultimate tracking bound of the modified average consensus algorithm (9). We show that for delays in the admissible delay bound the ultimate tracking error is still 0 as defined in Lemma IV.1. Theorem IV.1 (Convergence of (9) over connected graphs when τ ∈ [0, ¯τ ) ). Let G be a connected undirected graph with communication delay in τ ∈ [0, ¯τ ) where ¯τ is specified in Lemma IV.1. Let (cid:107)(IN − 1 N )r(cid:107)∞ = γ < ∞. Then, for any α ∈ R>0, the trajectories of algorithm (9) for any k ∈ R are bounded and satisfy (8). Moreover, The rate of convergence to this error neighborhood is no worse than ρτ (k) defined in (15). N 1N 1(cid:62) Proof. To establish our proof we consider (11), the equivalent representation of algorithm (9). Recall (11a) which along with the given initial condition gives z1(t) = 0 for t ∈ R≥0. Also, given (13), the trajectories of t (cid:55)→ z2:N for t ∈ R≥0 satisfy (cid:107)z2:N (t)(cid:107)≤ Diag(C1 j eS1 j t,··· , CN−1 eSN−1 j t) j (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:107)z2:N (0)(cid:107) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) = t)) j∈Z (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) (cid:110)(cid:12)(cid:12)(cid:88) (cid:110)(cid:12)(cid:12)(cid:88) j∈Z j∈Z Ci Diag( j∈Z Ci j Si j j eSi + γ max γ max (1 − eS1 j t),··· , (1 − eSN−1 j CN−1 SN−1 j j C1 j S1 j j t(cid:12)(cid:12)(cid:111)N−1 i=1 (1 − eSi (cid:107)z2:N (0)(cid:107) + j t)(cid:12)(cid:12)(cid:111)N−1 i=1 . (19) j∈Z 1 it j∈Z j∈Z Ci j Si j 1 αλi+1 Ci j Si j j∈ZeSi j t Ci Ci j Si j using (17) for any i ∈ ZN−1 }N−1 for any i ∈ ZN−1 Here, we used (cid:107)R(cid:62) r(cid:107) ≤ γ. Furthermore, using (10) we obtain xi(t) − ravg(t) ≤ (cid:107)x(t) − ravg(t)1N(cid:107) = (cid:107)z(t)(cid:107) = follows from (19) that limt→∞ xi(t)−ravg(t) ≤ limt→∞ (cid:107)z2:N (t)(cid:107) = = . To this end, note that from zero- j) zi(0) 0 zi(t)dt. On the other hand (cid:112)z1(t)2 + (cid:107)z2:N (t)(cid:107)2 = (cid:107)z2:N (t)(cid:107). Then, i=1 . Next, we show that (cid:80) γ max{(cid:80) input response of (13) we have zi(t) = ((cid:80) )zi(0) =(cid:82) ∞ which gives ((cid:80) (cid:90) ∞ (cid:90) ∞ Recalling (11c), we get(cid:82) 0 plies(cid:82) ∞ dition zi+1(0) ∈ R. Therefore, we get(cid:80) −τ zi+1(t)dt = 0, which along with the fact that under admissible range limt→∞ zi+1(t) = 0, im- which holds for any initial con- and consequently limt→∞ xi(t) − ravg(t) ≤ ≤ γ . ρ0 Moreover, the maximum rate of convergence corresponds to the worst rate of the exponential terms in (19), or equivalently min{{− Re(Si we have zi+1(t)dt − αλi+1k zi+1(t)dt =−αλi+1 j)}N−1 i=1 }j∈Z given in (15). = 1 γ j∈Z α min{λi}N 0 zi+1(t)dt = zi+1(0) (cid:90) 0 zi+1(t)dt. Ci j Si j αλi+1 i=2 1 0 −τ αλi 0 So far we have shown that splitting the immediate disagree- ment feedback of (7) into current and outdated components as in (9) does not have adverse effect on the tracking per- formance. Next, we show that this action interestingly can lead to increase in the rate of the converge at some specific values of k and τ. As we noted earlier, the rate of convergence of (9) is determined by behavior of its transient response that is governed by its zero-input dynamics. Consequently, we study the stability of the zero-input dynamics of the modified average consensus algorithm (9) and examine how its exponential rate of convergence to the average of its initial condition at time t = 0 changes due to delay at various values of k ∈ R/{0}. For any given value of k and τ, in what follows, we let ρτ (k) be the rate of convergence of (9) and uτ,k(t) = −α (1 − k) L x(t) − α k L x(t − τ ) be the control effort to steer the zero-input dynamics of (9). Specifically, we show that for all k ∈ R>0, there always exists a range of delay (0, τk) such that ρτ (k) > ρ0(0) = ρ0 = αλ2 for any τ ∈ (0, τk). We show however that only for k ∈ (0, 1] we can guarantee uτ,k∞ ≤ u0,0∞, for τ ∈ (0, τk). In what follows, we also investigate what the maximum value of ρτ (k) and the k ∈ (0, τk) are for a given k ∈ R>0. corresponding maximizer τ (cid:63) We start our analysis, by defining the delay gain function (cid:40) 1 g(γ, x) = x Re(W0(x eγ x)), 1, x ∈ R\{0}, x = 0, with x, γ ∈ R, to write ρτ (k) in (15) as ρτ (k) = min{ρτ,i(k)}N ρτ,i(k) = (kg(1 − 1 k i=2, ,−kλiατ )+(1 − k)) αλi. (20) (21) It follows from (3a) that limx→0 g(γ, x) = 1. Therefore, as expected, limτ→0 ρτ (k) = ρ0 = αλ2. We note that in fact ρτ,i, i ∈ {2,··· , N}, defines the rate of convergence of zi v in (17). In what follows, when emphasis on k is not necessary, to simplify the notation we write ρτ (k) as ρτ . In Appendix A, we study the variation of delay gain function versus x ∈ R≥0 for given values of γ. We show that for some specific values of γ there always exists a subset of the admissible delay range that the delay gain g(γ, ατ ) is greater than 1. In what follows, we use these results to determine ranges of delay and k we have ρτ (k) > αλ2. We also identify the optimum value of the delay τ (cid:63) for which ρτ has its maximum value, i.e., we identify the solution for min{ρτ,i}N (22) i=2. τ (cid:63) = argmax τ∈(0,¯τ ) ρτ = argmax τ∈(0,¯τ ) As we showed in Appendix A, the variation of the delay gain function g with x ∈ R≥0 for given values of γ is not monotone. Therefore, the solution to (22) is not trivial. Our careful characterization of variation of g vs. x ∈ R≥0 in Appendix A however, let us achieve our goal. In what follows, we set ρτ,i, τ (cid:63) i = argmax τ∈(0,¯τi) τi = {τ ∈ (0, ¯τi) g(1 − 1 k is given in (18). With the notation defined, ,−kαλiτ )) = 1}. where ¯τi the next theorem examines the effect of outdated feedback on the rate of convergence of modified consensus algorithm (9) for different value of k ∈ R/{0}. Theorem IV.2 (Effect of outdated feedback on the rate of con- vergence of average consensus algorithm (9)). The following assertions hold for the modified average consensus dynam- ics (9) over a connected graph whose rate of convergence is specified in (21): (a) For k < 0 the rate of convergence of the consensus algorithm (9) decreases by increasing τ ∈ R≥0. (b) For k > 0, ρτ > ρ0 if and only if τ ∈ [0, τ ) ⊂ [0, ¯τ ) where τ = min{τi}N i=2 with τi = {τ ∈ R>0ρτ,i = ρ0} and satisfies τN ≤ τ ≤ min{τ2, ¯τ}. Moreover, the opti- mum time delay τ (cid:63) corresponding to the maximum rate of convergence of the consensus algorithm (9) satisfies N , min{τ (cid:63) τ (cid:63) ∈ [τ (cid:63) W0( 1−k k e ) N = k e ), and is given by τ (cid:63) = {τ ∈ and τ (cid:63) 2 = α(1−k)λ2 2 , τ}] ρτ,2 = min{ρτ,i}N N , min{τ (cid:63) [τ (cid:63) 2 , τ}], where τ (cid:63) W0( 1−k α(1−k)λN i=3}. 1 1 Proof. Recall that the rate of convergence of algorithm (9) is specified by (21) (equivalent representation of (15)), which is the minimum of the rate of convergence of zi, i ∈ {2,··· , N} dynamics given in (17). Then, the proof of part (a) follows directly from statement (a) of Theorem A.1, which states that the rate of convergence of each zi, i ∈ {2,··· , N} dynamics decreases by increasing delay τ ∈ R>0 (note that in Theorem A.1 each zi dynamics reads as a = −αkλi > 0 and b = −α(1 − k)λi < 0). To prove statement (b) we proceed as follows. For k > 0, because of the statement (b) of Theorem A.1 for each zi, i ∈ {2,··· , N}, dynamics (a =−αkλi < 0) we have the guarantees that ρτ,i = (kg(1 − 1 k ,−kλiατ )+(1 − k)) αλi > ρ0,i ≥ ρ0, for τ ∈ (0, τi). Since α > 0, λN ≥ λN−1 ≥ ··· ≥ λ2 > 0 and ρ0,N ≥ ρ0,N−1 ≥ ··· ≥ ρ0,2, we have i ∈ {3,··· , N}, N−1 ≤ ··· ≤ τ (cid:63) 2 . i < τi ≤ τi < ¯τi, τ (cid:63) τN ≤ τN−1 ≤ ··· ≤ τ2, ¯τ = ¯τN ≤ ¯τN−1 ≤ ··· ≤ ¯τ2, N ≤ τ (cid:63) τ (cid:63) (23a) (23b) (23c) (23d) and τ2 = τ2. Since g(1− 1 k ,−kλiατ ) is a decreasing function of τ for any τ ∈ (τi, ¯τi) ⊂ (τ (cid:63) i , ¯τi) (Recall Lemma A.1), it follows that for any τ ∈ [0, τj) we have ρτ,j > ρ0 and for any τ ∈ [τj, ¯τ ) we have ρτ,j < ρ0. Because ρτ = min{ρτ,j}N j=2, we have ρτ > ρ0, if and only if τ ∈ (0, τ ) where τ = min{τj}N j=2. From (23a) and (23b), it follows that τN ≤ τ. Moreover, since ρτ,2 > ρ0 for τ ∈ (0, τ2), we obtain τ ≤ min{τ2, ¯τ}. This concludes the proof of the first part of statement (b). To obtain τ (cid:63) ∈ (0, τ ) which gives the maximum attain- able ρ(cid:63) τ we proceed as follows. First, note that statement (b) of Theorem A.1 indicates that ρτ,i, i ∈ {2,··· , N} is a monotonically increasing (resp. decreasing) function of τ ∈ (0, τ (cid:63) i , ¯τi)). Then because of (23d), we have the guarantees that ρτ is a monotonically increasing function of τ ∈ (0, τ (cid:63) N ), and decreasing function of τ for 2 . Therefore, the maximum value of ρτ should any τ > τ (cid:63) N , min{τ (cid:63) 2 , τ}] 2 ] ∩ (0, τ )) ⊆ [τ (cid:63) be attained at τ (cid:63) ∈ ([τ (cid:63) N , τ (cid:63) W0( 1−k with τ (cid:63) k e ). Now let j = min{i ∈ {2,··· , N}τ (cid:63) i ≤ τ (cid:63)}. Then, given (23d), for any i ∈ {2,··· , N} such that i < j (resp. i ≥ j) by virtue of state- ment (e) of Lemma A.1 we know dg(1− 1 k ,−kλiατ )/dτ > 0 (resp. < 0) and consequently dρτ,i/dτ > 0 (resp. < 0) at τ = τ (cid:63). Since ρτ = min{ρτ,i}N i=2, the maximum value of ρτ is attained at τ = τ (cid:63) at which i ) (resp. τ ∈ (τ (cid:63) k e ) and τ (cid:63) N = 1 λN W0( 1−k 2 = 1 λ2 min{ρτ,i}N Since λ2τ (cid:63) ≤ ··· ≤ λj−1τ (cid:63) and dg(1 − 1 for i ∈ {2,··· , j− 1}, we have g(1− 1 k ,−kλj−2ατ (cid:63)) ≥ ··· ≥ g(1 − 1 follows from (21) that at τ = τ (cid:63) we have min{ρτ,i}j−1 ρτ,2, which given (24) completes our proof. i=j = min{ρτ,i}j−1 (24) i=2. k ,−kλiατ (cid:63))/dτ > 0 k ,−kλj−1ατ (cid:63)) ≥ g(1− k ,−kλ2ατ (cid:63)). As a result, it i=2 = 1 Theorem IV.2 indicates that for any k > 0 there always exists a range of delay in (0, ¯τ ] for which faster response can be achieved for the modified average consensus algorithm (9) relative to the original one (7). Next, our goal is to identify values of k ∈ R>0 for which the maximum driving effort uτ,k(t) does not exceed the one for the original algorithm (7) (for zero-input dynamics). However, before that we make the following statement about the maximum attainable rate by using outdated feedback. Lemma IV.2 (Ultimate bound on the maximum attainable increase in the rate of convergence of (9)). For any k ∈ R≥0, the ultimate bound on the maximum attainable rate of convergence for (9) by using outdated feedback is equal to (1 − k)(1 + Proof. It follows from part (f) of Lemma A.1 that g(1 − for any i ∈ {2,··· , N}. Then, k ,−kλiατ (cid:63) W0( 1−k k e ) i ) = )ρ0. 1 1 1−k kW0( 1−k k e ) given (21) we have ρτ ≤ ρτ,2 ≤ ρτ (cid:63),2 = (cid:0)kg(1 − 2 ) + (1− k)(cid:1)αλ2 = (1− k)(1 + )ρ0, which 1 vi 1 k ,−kλ2ατ (cid:63) concludes our proof. W0( 1−k k e ) Next, we study how the maximum control effort of the agents while implementing for the modified algorithm (9) compares to that of the original average consensus algorithm (7) any k ∈ R>0. The theorem below indicates that for any k ∈ (0, 1] using the outdated feedback does not increase the maximum control effort while for k > 1 the maximum control effort is greater than the one of the original algorithm (7). Theorem IV.3 (The maximum control effort for steering the zero-input dynamics of the algorithm (9)). For a given α ∈ R>0, let u0,0, and uτ,k(t) be respectively the network aggregated control input of the zero-input dynamics of (7), and (9) for any k ∈ R>0 and τ ∈ R>0. Then, for any τ ∈ [0, ¯τ ], where admissible delay bound ¯τ is given in Lemma IV.1, the following assertions hold for t ∈ R≥0: (a) For k ∈ (0, 1] we have uτ,k(t)∞ ≤ u0,0(t)∞. (b) For k > 1 we have uτ,k(t)∞ ≥ e(k−1)αλ2τ u0,0(t)∞. Proof. Consider the zero-input dynamics of (11), the equiva- lent representation of algorithm (9). For the maximum control effort of algorithm (9) we have uτ,k(t)∞ = − α (1 − k) Λ z(t) − α k Λ z(t − τ )∞ = α max{(1−k) λi zi(t) + k λi zi(t − τ )∞}N i=2. (25) Here we used the fact that z1(t) = 0. Also, recalling (17), for τ = 0 and any i ∈ {2,··· , N} we have zi(t) = e−λit zi(0), which gives u0,0(t)∞ = u0,0(0)∞ = α max{λizi(0)}N i=2. Next, we show that for any τ ∈ (0, ¯τ ) and k ∈ (0, 1] uτ,k(t)∞ ≤ α{λizi(0)}N i=2. Notice that from (25) we have uτ,k(t)∞ ≤ α(1 − k) max{λizi(t)∞} + αk max{λiz(t − τ )∞}. Also, recall that for t ∈ [0, τ ) we have zi(t − τ ) = 0. Thus, to validate the statement (a) it suffices to show that zi(t)∞ = zi(0). To this aim, consider the trajectories t → z2:N of (17). Since set of dynamics (17) are exponentially stable with −α(1 − k)λi ≤ 0 and −αkλi ≤ 0, recalling Lemma A.2 for any delay in the admissible range we have zi(t)∞ = maxs∈[−τ,2τ ] zi(s) for any i ∈ {2,··· , N}. Also, note that from (17) we get zi(t) = 0, zi(t) = e−α(1−k)λit zi(0), zi(t) = e−α(1−k)λitzi(0)(1 + t ∈ [−τ, 0), t ∈ [0, τ ), k (1 − k) (26a) (26b) (e−α(1−k)λi(t−τ )−1)) t ∈ [τ, 2τ ], (26c) which results in maxs∈[−τ,2τ ] zi(s) = zi(0), and conse- quently zi(t)∞ = zi(0), which concludes statement (a). To validate part (b) we proceed as follows. Recalling (25) for k > 1 we have uτ,k(2τ )∞ = α max{kλizi(τ ) − (k − 1)λizi(2τ )}N i=2 ≥ α max{kλizi(τ )− (k− 1)λizi(2τ )}N i=2. Also, from (26c) for t ∈ [τ, 2τ ) we have zi(2τ ) ≤ zi(τ ), which gives uτ,k(2τ )∞ = α max{kλizi(τ ) − (k − 1)λizi(τ )}N i=2. Moreover, (26b) implies that zi(τ ) = eα(k−1)λiτ zi(0), which deduces i=2 = α(2k − 1) max{λizi(τ )}N 1 1 2 1 5 1 3 1 1 1 4 1 1  L =  3 −1 0 −1 4 −1 −1 −1 −1 −1 0 −1 −1 −1 0 3 −1 −1 −1 −1 3 3 −1 0 −1 Fig. 1: A connected graph of 5 nodes. i=2 = (2k− uτ,k(2τ ) ≥ α(2k−1) eα(k−1)λiτ max{λizi(0)}N 1) eα(k−1)λ2τ u0,0∞. Knowing 2k − 1 ≥ 1 and uτ,k(t)∞ ≥ uτ,k(2τ )∞ we can conclude the proof. We close this section by a remark on how the split factor can be chosen based on the expectations on the convergence rate, robustness to delay and managing the control effort. Remark IV.1 (Selecting k in the algorithm (9)). Lemma IV.1, Theorem IV.2 and Theorem IV.3 give insights on how we can choose the slit factor k ∈ R given expectations on the algorithm's acceleration, robustness to delay and control effort. Theorem IV.3 certifies that for any k ∈ (0, 1], the rate of convergence we observe for any τ ∈ [0, ¯τ ] is attained without imposing any extra control effort on the agents. Therefore, assuming that the acceleration is expected without increasing the control effort, the split factor should be selected to satisfy k ∈ (0, 1]. According to Lemma IV.2 the maximum attainable rate of convergence is an increasing function of k ∈ (0, 1]. Moreover, as k → 1 the ultimate bound on the rate of convergence converges to e ρ0, which recovers the same bound established in [27, Theorem 4.4]. On the other hand, as ex- pected, as k → 0 the ultimate bound on the rate of convergence converges to ρ0. Finally, we observe from Lemma IV.1 that for k > 0.5 the admissible delay bound is finite, and thus the robustness of the algorithm to delay is not strong. In Contrary, the algorithm is robust with respect to any perturbation in delay for k ∈ (0, 0.5], because the admissible delay range for such split factors is R≥0. Taking these observations into account, there exists a trade-off between robustness to delay and achieving higher acceleration when comes to choosing the split factor; k = 1 gives the maximum rate of convergence with the corresponding optimum delay while k = 0.5 results in robustness as well as higher rate of convergence relative to the original system (7). V. NUMERICAL EXAMPLE We consider the modified average consensus algorithm (9) over the graph depicted in Fig. 1. The reference input of each agent i ∈ {1, . . . , 5} is chosen according to the first numerical example in [5] to be the zero-order hold sampled points from the signal ri(t) = ai(2 + sin(ω(t)t + φ(t)) + bi. The idea discussed in [5] is that the sensor agents sample the signal and should obtain the average of these sampled points before the next sampling arrives. The parameters ai (the multiplica- tive sampling error) and bi (additive bias), i ∈ {1, . . . , 5}, are chosen as the ith element of [1.1, 1, 0.9, 1.05, 0.96] and [−0.55, 1, 0.6,−0.9,−0.6], respectively. At each sampling time ω and φ are chosen randomly according to N (0, 0.25) vii and N (0, (π/2)2), where N (µ, σ) indicates the Gaussian distribution with mean µ and variance σ. We set the sampling rate at 2Hz. This numerical example can be viewed as a simple abstraction for decentralized operations such as distributed sensor fusion where a dynamic or static average consensus algorithm is used to create the additive fusion terms in a distributed manner, e.g., [6], [8]. Since the convergence of the average consensus algorithm is asymptotic, there is always an error when the algorithm is terminated in the finite inter- sampling time. Faster convergence is desired to reduce the residual error. For this example, in what follows, we study the response of the modified average consensus algorithm (9) for k ∈ {−0.5, 0, 0.5, 1, 1.5}. We note that the case of k = 0 gives the original (delay free) dynamic average consensus algorithm (7) and thus is the baseline case that the rest of the cases should be compared to. For k ∈ {−0.5, 0, 0.5, 1, 1.5}, the critical delay value ¯τ of the admissible delay range (0, ¯τ ) of (9), respectively, is {∞,∞,∞, 0.32, 0.18} seconds. Figure 2 il- lustrates how ρτ changes with τ. First, we note that for k = −0.5 the rate of convergence decreases with delay. However, for positive values of k there is a range (0, τ ) for which ρτ > ρ0. For positive values of k we also observe monotonic increase until reaching τ (cid:63) and then the monotonic decrease afterwards. The trend observed is in accordance with the results of Theorem IV.2.We also can observe that as the k increases the maximum achievable rate of convergence increases also. Figure 3 shows the tracking response of agent 2 for k ∈ {−0.5, 0, 0.5, 1, 1.5} when the delay is τ = 0.1 (similar trend is observed for the other agents). As seen, the convergence rate of (9) is different for each value of k. The fastest response is observed for k = 1.5 while k = −0.5 shows the lowest one. The decrease of rate of convergence for k = −0.5 and its increase for the positive values of k is in accordance with the trend certified by Theorem IV.2 (note that as seen in Figure 2, τ = 0.1 is in the rate increasing delay range of (0, τ of the cases corresponding to k ∈ {0.5, 1, 1.5}). The desired effect of fast convergence shows itself in the smaller tracking error that is observed at the end of each sampling time, e.g., the tracking error in the first epoch for k ∈ {0.5, 1, 1.5} is, respectfully, %13, %9, and %0.5 that is an improvement over %15 that corresponds to k = 0 (case of original algorithm). We note here that as can be seen in Fig. 2, τ = 0.1 is close to τ (cid:63) of the case corresponding to k = 1.5. The same level of fast convergence can be achieved for the cases of k = 1 and k = 0.5 if one uses τ (cid:63) corresponding to these split factors. Figure 4 shows the maximum control effort of zero-input dynamics of the algorithm (9) over time corresponding to τ = 0.1 and different values of k ∈ {0, 0.5, 1, 1.5}. For k = 1.5 the maximum control effort exceeds the value for the original consensus algorithm (case of k = 0). But, for k = 1 and k = 0.5 the maximum control effort is equal or less than the case k = 0. The trend observed above is in accordance with Theorem IV.3. viii Fig. 2: The rate of convergence ρτ of the modified average consensus algorithm (9) over the graph in Fig. 1 for different values of feedback gain k ∈ {−0.5, 0, 0.5, 1, 1.5}. For the example case of k = 1.5, note that τ = 0.14 and the maximum rate of convergence that can be achieved is ρ(cid:63) τ ≈ 2ρ0 at τ (cid:63) = 0.11. Fig. 4: The maximum control effort executed by the algorithm (9) over the graph in Fig. 1 for τ = 0.1 and different values of feedback gain k ∈ {0, 0.5, 1, 1.5}. [2] W. Reb and R. W. Beard, "Consensus seeking in multi-agent systems under dynamically changing interaction topologies," IEEE Transactions on Automatic Control, vol. 50, no. 5, pp. 655 -- 661, 2005. [3] L. Xiao and S. Boyd, "Fast linear iterations for distributed averaging," Systems and Control Letters, vol. 53, pp. 65 -- 78, 2004. [4] R. Olfati-Saber, J. A. Fax, and R. M. Murray, "Consensus and coop- eration in networked multi-agent systems," Proceedings of the IEEE, vol. 95, no. 1, pp. 215 -- 233, 2007. [5] S. S. Kia, B. V. Scoy, J. Cort´es, R. A. Freeman, K. M. Lynch, and S. Mart´ınez, "Tutorial on dynamic average consensus: The problem, its applications, and the algorithms," IEEE Control Systems Magazine, vol. 39, no. 3, pp. 40 -- 72, 2019. [6] R. Olfati-Saber and J. S. Shamma, "Consensus filters for sensor networks and distributed sensor fusion," in IEEE Int. Conf. on Decision and Control and European Control Conference, (Seville, Spain), pp. 6698 -- 6703, December 2005. [7] R. Olfati-Saber, "Distributed kalman filtering for sensor networks," in IEEE Int. Conf. on Decision and Control, (New Orleans, USA), pp. 5492 -- 5498, December 2007. [8] T. A. Kamal, J. A. Farrell, and A. K. Roy-Chowdhury, "Information weighted consensus filters and their application in distributed camera networks," IEEE Transactions on Automatic Control, vol. 58, no. 12, pp. 3112 -- 3125, 2013. [9] W. Ren and U. M. Al-Saggaf, "Distributed Kalman-Bucy filter with embedded dynamic averaging algorithm," IEEE Systems Journal, no. 99, pp. 1 -- 9, 2017. [10] P. Yang, R. A. Freeman, and K. M. Lynch, "Multi-agent coordination by decentralized estimation and control," IEEE Transactions on Automatic Control, vol. 53, no. 11, pp. 2480 -- 2496, 2008. [11] Y. Chung and S. S. Kia, "Distributed dynamic containment control over a strongly connected and weight-balanced digraph," in IFAC Workshop on Distributed Estimation and Control in Networked Systems, (Chicago, IL), 2019. [12] J. Fax and R. Murray, "Information flow and cooperative control of vehicle formations," IEEE Transactions on Automatic Control, vol. 42, no. 2, pp. 465 -- 1476, 2004. [13] A. Cherukuri and J. Cort´es, "Initialization-free distributed coordination for economic dispatch under varying loads and generator commitment," Automatica, vol. 74, no. 12, pp. 183 -- 193, 2016. [14] S. S. Kia, "Distributed optimal in-network resource allocation algorithm design via a control theoretic approach," Systems and Control Letters, vol. 107, pp. 49 -- 57, 2017. [15] S. Meyn, Control Techniques for Complex Networks. Cambridge University Press, 2007. [16] P. Yang, R. A. Freeman, and K. M. Lynch, "Distributed cooperative active sensing using consensus filters," in IEEE Int. Conf. on Robotics and Automation, (Roma, Italy), pp. 405 -- 410, April 2007. [17] M. Fiedler, "Algebraic connectivity of graphs," Czechoslovak Mathemat- ical Journal, vol. 23, no. 2, pp. 298 -- 305, 1973. [18] S. Boyd, A. Ghosh, B. Prabhakar, and D. Shah, "Randomized gossip algorithms," IEEE Information Theory Society, vol. 52, no. 6, pp. 2508 -- 2530, 2006. [19] S. Kar and J. M. F. Moura, "Topology for global average consensus," in Fortieth Asilomar Conference on Signals, Systems and Computers, (Pacific Grove, CA, USA), 2006. [20] P. Hovareshti, J. S. Baras, and V. Gupta, "Average consensus over small world networks: A probabilistic framework," in IEEE Int. Conf. on Decision and Control, (Cancun, Mexico,), 2008. Fig. 3: The trajectory of local state of agent 2 executed by the algorithm (9) over the graph in Fig. 1 for τ = 0.1 and different different values of feedback gain k ∈ {−0.5, 0, 0.5, 1, 1.5}. VI. CONCLUSION We analyzed the effect of using an affine combination of immediate and outdated disagreement feedbacks in increasing the rate of convergence of a dynamic average consensus algorithm. The modified algorithm has the same ultimate tracking accuracy but with the right choices of the delay and the affine combination factor, can have faster convergence. Our study produced a set of closed-form expressions to specify the admissible delay range, the delay range for which the system experiences increase in its rate of convergence and a range that the optimum time delay corresponding to the maximum rate of convergence lies. We also examined the range of affine combination factor for which the outdated feedback can be used to improve the convergence of the algorithm without increasing the control effort. To develop our results we used the Lambert W function to obtain the rate of convergence of our algorithm under study in the presence of the delay. Our future work includes extending our results for dynamic consensus algorithms over directed graphs and also investigating the use of outdated feedback in increasing the rate of convergence of other distributed algorithms for networked systems such as leader-follower algorithms. REFERENCES [1] 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. 00.20.40.60.8102468k=-0.5k=0k=0.5k=1k=1.501Time23x2(t)k=-0.5k=0k=0.5k=1k=1.500.511.52Time051015 u(t) k=0k=0.5k=1k=1.5[21] B. Ghosh, S. Muthukrishnan, and M. Schultz, "First and second-order diffusive methods for rapid, coarse, distributed load balancing," Theory of Computing Systems, vol. 31, pp. 331 -- 354, 1998. [22] Y. Ghaedsharaf, M. Siami, C. Somarakis, and N. Motee, "Interplay be- tween performance and communication delay in noisy linear consensus networks," in European Control Conference, (Aalborg, Denmark), 2017. [23] M. Cao, D. Spielman, and E. Yeh, "Accelerated gossip algorithms for distributed computation," In Proceedings of the 44th Annual Allerton Conference, pp. 952 -- 959, 2006. [24] Z. Meng, Y. Cao, and W. Ren, "Stability and convergence analysis of multi-agent consensus with information reuse," International Journal of Control, vol. 83, no. 5, pp. 1081 -- 1092, 2010. [25] A. G. Ulsoy, "Improving stability margins via time-delayed vibration control," in Time Delay Systems: Theory, Numerics, Applications, and Experiments (T. Insperger, T. Ersal, and G. Orosz, eds.), pp. 235 -- 247, Springer, 2017. [26] Y. Cao and W. Ren, "Multi-agent consensus using both current and out- dated states with fixed and undirected interaction," Journal of Intelligent and Robotic Systems, vol. 58, no. 1, pp. 95 -- 106, 2010. [27] H. Moradian and S. Kia, "A study on rate of convergence increase due to time delay for a class of linear systems," in IEEE Int. Conf. on Decision and Control, (Miami, US), 2018. [28] W. Qiao and R. Sipahi, "A linear time-invariant consensus dynamics with homogeneous delays: analytical study and synthesis of rightmost eigenvalues," SIAM Journal on Control and Optimization, vol. 51, no. 5, p. 39713991, 2013. [29] D. P. Spanos, R. Olfati-Saber, and R. M. Murray, "Dynamic consensus on mobile networks," in IFAC World Congress, (Prague, Czech Repub- lic), July 2005. [30] T. Hu, Z. Lin, and Y. Shamash, "On maximizing the convergence rate for linear systems with input saturation," IEEE Transactions on Automatic Control, vol. 48, no. 6, pp. 1249 -- 1253, 2003. [31] S. Duan, J. Ni, and A. G. Ulsoy, "Decay function estimation for linear time delay systems via the Lambert W function," Journal of Vibration and Control, vol. 18, no. 10, pp. 1462 -- 1473, 2011. [32] H. Shinozaki and T. Mori, "Robust stability analysis of linear time- delay systems by Lambert W function: Some extreme point results," Automatica, vol. 42, no. 10, pp. 1791 -- 1799, 2006. [33] R. M. Corless, G. Gonnet, D. E. G. Hare, D. J. Jeffrey, and D. E. Knuth, "On the Lambert W function," Advances in Computational Mathematics, vol. 5, pp. 329 -- 359, 1996. [34] S. Niculescu, Delay effects on stability: A robust control approach. New York: Springer, 2001. [35] F. Bullo, J. Cort´es, and S. Mart´ınez, Distributed Control of Robotic Networks. Applied Mathematics Series, Princeton University Press, 2009. [36] S. Yi, P. W. Nelson, and A. G. Ulsoy, Time-Delay Systems: Analysis and Control Using the Lambert W Function. World Scientific Publishing Company, 2010. [37] H. Moradian and S. Kia, "On robustness analysis of a dynamic average consensus algorithm to communication delay," IEEE Transactions on Control of Network Systems, vol. 6, no. 2, pp. 633 -- 641, 2018. [38] A. Ivanov, E. Liz, and S. Trofimchuk, "Halanay inequality, Yorke 3/2 stability criterion, and differential equations with maxima," Tohoku Mathematical Journal, vol. 54, no. 2, pp. 277 -- 295, 2002. APPENDIX A DELAY GAIN FUNCTION The lemma below highlights some of the properties of the delay gain function g(γ, x). Figure 5 gives some graphical representation for the properties discussed in this lemma. Lemma A.1 (Properties of g(γ, x)). The following assertions hold for the delay gain function (20) with γ, x ∈ R : (a) For any γ ∈ R we have limx→0 g(γ, x) = 1. (b) For any γ > 1 and x ∈ R>0 we have g(γ, x) < γ. (c) For any γ > 1 and x ∈ R>0, g(γ, x) is a strictly (d) Let x ∈ (¯x, 0), where ¯x = arccos(γ)/(cid:112)1 − γ2. Then, for increasing function of x. any γ < 1 (respectively γ > 1) we have g(γ, x) > γ (respectively g(γ, x) < γ). ix (a) γ > 1 (b) 0 < γ < 1 (c) −1 < γ < 0 (d) γ < −1 Fig. 5: The delay gain function for different values of x, γ. γ W0(− γ e ) when γ (cid:54)= 0 and x(cid:63) = − 1 (e) For any γ < 1 and x ∈ R<0, g(γ, x) is a strictly decreasing function of x for any x ∈ [x(cid:63), 0) ⊂ (¯x, 0), and a strictly increasing function of x for any x < x(cid:63), where x(cid:63) = 1 e when γ = 0. (f) For any γ < 1 and x ∈ R<0, the maximum value of g(γ, x) occurs at x(cid:63) = 1 e ) where g(γ, x(cid:63)) = −γ e where g(γ, x(cid:63)) = e W0(− γ when γ = 0. (g) For any γ < 1 and x ∈ R<0, g(γ, x) > 1 if and only if x ∈ (x, 0) where x is the unique solution of g(γ, x) = 1 in (¯x, 0). γ W0(− γ e ) when γ (cid:54)= 0, and at x(cid:63) = − 1 The proof of this lemma invokes various properties of the Lambert W function listed in Section II and is given in Appendix B. The next theorem, whose proof relies on the results of Lemma A.1, and is also given in Appendix B, characterizes the effect of delay on the rate of convergence of scalar time-delayed system (1). The tightest estimate of the rate of convergence of (1) is characterized by the magnitude of the real part of the rightmost root of its characteristic equation τ W0(α τ e−τ b) + b (recall Lemma II.2 and (4a)). That s = 1 is (see [31, Corollary 1]) ρτ = − 1 τ Re(W0(a τ e−τ b)) − b. (27) -505x0510g(,x)=2=5=10-505x-2024g(,x)=0.02=0.1=0.8-505x-2024g(,x)=-0.02=-0.1=-0.8-505x-10010g(,x)=-2=-5=-10Recalling (20), we write (27) as ρτ = −(g(γ, x) a + b) = −(g(γ, x) − γ) a a . It follows from (3a) that where x = aτ and γ =− b lim τ→0 g(γ, aτ ) = 1. (28) (29) Therefore, as expected, limτ→0 ρτ = ρ0, where ρ0 = −(a + b) = −(1 − γ)a. (30) system (1) in terms of different values of a, b ∈ R, a (cid:54)= 0 satisfying a + b < 0. Theorem A.1 (Effect of delay on the rate of convergence of delayed system (1)). Consider system (1) with a ∈ R\{0} and b ∈ R such that a + b < 0, whose rate of convergence ρτ is specified by (28). Consider also the delay gain function (20) with γ = − b (a) for a > 0 and b < 0 the system (1) is exponentially stable for any τ ∈ R≥0. Moreover, the rate of convergence decreases by increasing τ ∈ R≥0. (b) for a < 0 and b ∈ R, ρτ > ρ0 if and only if τ ∈ [0, τ ) ⊂ [0, ¯τ ) where τ is the unique solution of g(γ, aτ ) = 1 in (0, ¯τ ) and ¯τ is specified by a and x = aτ. Then, ¯τ = arccos(−b/a)/ a2 − b2. (31) (cid:112) Moreover, ρτ is monotonically increasing (resp. decreas- ing) with τ for any τ ∈ [0, τ (cid:63)) ⊂ [0, ¯τ ) (resp. τ ∈ a e ) when b (cid:54)= 0 (τ (cid:63), ¯τ ) ⊂ [0, ¯τ )), where τ (cid:63) = − 1 and τ (cid:63) = − 1 a e when b = 0. Finally, the maximum rate )b when b (cid:54)= 0 and τ = −(1 + 1 of convergence of ρ(cid:63) τ = −a e when b = 0 is obtained at τ = τ (cid:63). ρ(cid:63) b W0( b W0( b a e ) In developing our results we also invoke the following result. Lemma A.2. (maximum value of the trajectory of (1) [38, Theorem 2.10]) For the time delay system (1) and any τ ∈ (0, ¯τ ] with a, b ∈ R<0 the following holds x(t)∞ = maxs∈[−τ,2τ ]x(s). (32) APPENDIX B PROOFS OF LEMMA A.1 AND THEOREM A.1 Proof of Lemma A.1. Part (a) can be readily deduced by in- voking (3b) since W0(x eγ x) → x eγ x as x → 0. To prove statement (b) we proceed as follows. Let q = x eγ x. Since x ∈ R>0, then q ∈ R>0. As a result, given the properties of Lambert W function reviewed in Section II, we can write γ W0(γq) and Re(W0(q)) = W0(q), which allows us to x = 1 represent g(γ, x) as g(γ, x) = W0(q) W0(γ q) γ, for x ∈ R>0. (33) Since for γ > 1 we have q < γ q, by invoking Lemma II.3 W0(γ q) < 1, which together with W0(q) ∈ R>0 we obtain W0(q) and W0(γ q) ∈ R>0 validates statement (b) from (33). x = Re( (34) )− Next, we validate statement (c). The derivative of g(γ, x) with respect to x ∈ R is d g(γ, x) (1 + γ x) eγ x 1 d x x2 Re( (1 + γ x) W0(x eγ x) + 1 x eγ x + eW0(x eγ x) W0(x eγ x) )− (γ x−W0(x eγ x))W0(x eγ x) x 1 x2 Re(W0(x eγ x)) = 1 x2 Re(W0(x eγ x)) = for x eγ x (cid:54)= − 1 e . Recall (4c) that Re(W0(z)) + 1 > 0 for any z ∈ R\{− 1 e} and Re(W0(z)) = W0(z) > 0 for any z ∈ R>0. Note also that we have already shown that for any γ > 1 and x > 0 we have g(γ, x) < γ which gives γ x − W0(x eγ x) > 0. Therefore, for γ > 1 and x ∈ R>0 from (34) we obtain (W0(x eγ x) + 1) 1 x2 Re( ), d g(γ,x) d x > 0, which validates statement (c). e γ W0(− γ γ W0(− γ e ), 0) we have x eγx ∈ (− 1 e ), 0) for γ (cid:54)= 0 and x ∈ [− 1 To validate statement (d), consider x ∈ (¯x, 0]. For x → 0− we have g(γ, x) → 1. So, for γ < 1 (respectively γ > 1) we get g(γ, x) > γ (respectively g(γ, x) < γ) as x → 0−. Moreover, we know that the admissible bound, x = ¯x is the first point that g(γ, x) = γ holds. So, since g(γ, x) is a continuous function, for any x ∈ (¯x, 0] we have g(γ, x) > γ for γ < 1, and g(γ, x) < γ for γ > 1. For proof of statement (e) we proceed as follows. Recall the properties of Lambert W0 function in (4). Note that for 0 < γ < 1, we have −1 < W0(− γ e ) < 0 and for γ < 0, we have W0(− γ e ) > 0. Also recall that W0(0) = 0. Therefore, for γ < 1 and γ (cid:54)= 0, we have 1 e ) ∈ R<0. Now for γ < 1 γ W0(− γ consider x ∈ [ 1 e , 0) for γ = 0. For such x, we have x eγ x ∈ R<0. For f (x) = x eγx, with x, γ ∈ R we know d f d x = (1 + γ x) eγx > 0 for any x ∈ (− 1 e , 0] and γ < 1, i.e., f (x) is a strictly increasing continuous function. Because the solutions of z eγz = − 1 γ Wl(− γ e ), l = {−1, 0} for γ (cid:54)= 0 and z = − 1 are z = 1 e for γ = 0, for x ∈ [ 1 e , 0] and then W0(x eγ x) ∈ R<0 (recall (4a)). Next, note that by statement (d) we have γ x − W0(x eγ x) = x(γ − g(x, γ)) > 0 for x ∈ (¯x, 0]. Therefore d g(x,γ) d x < 0 can be inferred from (34). Next, for x < 1 e ), let W0(x eγ x) = w + i u. Then, (34) can be written as d g(γ,x) ) = d x x2((w+1)2+u2) ((γx − w)(w2 + u2 + w) + u2). In addition, we have w = −u cot u since Im(x eγ x) = 0, which gives x2u2((cot u+1)2+1) ((γx + u cot u)(u2 cot2 u + u2 − u cot u) + u2) > 0. Here, we used u2 cot2 u + u2 − u cot u = sin u − cos u) > 0, and γx + u cot u = γx − w = sin u ( u γx − Re(W0(x eγ x)) = x(γ − g(γ, x)) > 0, which holds for any x ∈ [¯x, 0)(recall statement (d)), which finalize our proof for statement (e). For proof of statement (f), notice that statement (e) explicitly implies that max(g(γ, x)) = g(γ, x(cid:63)) for any x ∈ R<0 where x(cid:63) eγ x(cid:63) e ) for γ (cid:54)= 0, and x(cid:63) = − 1 Proof of statement (g) is as follows. In statement (a) we showed that g(γ, x) → 1 as x → 0−. Moreover, g(γ, x) is a continuous ascending function in x ∈ (−∞, 1 e )], and descending function in x ∈ [ 1 e ), 0). So, continuity e , which is equivalent to x(cid:63) = 1 x2 Re( (γ x−(w+i u))(w+i u) γ W0(− γ γ W0(− γ γ W0(− γ γ W0(− γ e for γ = 0. = − 1 d x = ((w+i u)+1) = 1 d g(γ,x) u 1 1 xi implies that there exists a x ∈ (¯x, x(cid:63)) such that g(γ, x) = 1, or equivalently Re(W0(x eγ x)) = x, and also g(γ, x) > 1 holds for any x ∈ (x, 0). (cid:3) Proof of Theorem A.1. Because by assumption we have α + b < 0, a > 0 implies that b < −a < 0, resulting in γ > 1 and x = aτ > 0 for τ ∈ R>0. Therefore, invoking Lemma A.1 statement (b) we get g(γ, x) < γ. Thus, (28) implies that system (1) is exponentially stable regardless of value of τ ∈ R≥0. Moreover, by taking derivative of ρτ with respect to τ, we obtain d ρτ d τ = ( d g(γ, x) d x )( d x d τ ) = −a d g(γ, x) d x . (35) Lemma A.1 part (c) states that d g(γ,x) d x > 0 for any γ > 1 and x > 0. Hence, for a > 0 we have d ρτ d τ < 0 which concludes our proof of part (a). For a < 0 and b ∈ R, from (28) it follows that ρτ > ρ0 if and only if g(γ, aτ ) > 1. In this case, because of a + b < 0, we have γ < 1 and x = aτ < 0 for τ ∈ R>0. Therefore, by virtue of statement (g) of Lemma A.1 we have ρτ > ρ0 if and only if τ ∈ [0, τ ) ⊂ [0, ¯τ ) where τ is the unique solution of g(γ, aτ ) = 1 in (0, ¯τ ). Additionally, by virtue of part (e) of Lemma A.1, ρτ , whose rate of change with respect to τ is specified by (35), is monotonically increasing (resp. decreasing) with τ for any τ ∈ [0, τ (cid:63)) ⊂ [0, ¯τ ) (resp. τ ∈ (τ (cid:63), ¯τ ) ⊂ [0, ¯τ )) where τ (cid:63) = x(cid:63) b W0( b a e ) for b (cid:54)= 0 and τ (cid:63) = x(cid:63) a = − 1 a e for b = 0. Moreover, by virtue of part (f) of Lemma A.1 we conclude that the maximum value of g(γ, x) occurs at x(cid:63) = aτ (cid:63) where g(γ, x(cid:63)) = e ) for b (cid:54)= 0, which gives ρ(cid:63) )b. For b = 0, we have τ = −a e. ρ(cid:63) τ = −(1 + 1 W0( b (cid:3) −γ W0(− γ a e ) a = 1 γa W0(− γ e ) = − 1
1905.04077	1	1905	2019-05-10T11:30:20	Emergent Escape-based Flocking Behavior using Multi-Agent Reinforcement Learning	[ "cs.MA", "cs.AI" ]	In nature, flocking or swarm behavior is observed in many species as it has beneficial properties like reducing the probability of being caught by a predator. In this paper, we propose SELFish (Swarm Emergent Learning Fish), an approach with multiple autonomous agents which can freely move in a continuous space with the objective to avoid being caught by a present predator. The predator has the property that it might get distracted by multiple possible preys in its vicinity. We show that this property in interaction with self-interested agents which are trained with reinforcement learning to solely survive as long as possible leads to flocking behavior similar to Boids, a common simulation for flocking behavior. Furthermore we present interesting insights in the swarming behavior and in the process of agents being caught in our modeled environment.	cs.MA	cs	Emergent Escape-based Flocking Behavior using Multi-Agent Reinforcement Learning Carsten Hahn1, Thomy Phan1, Thomas Gabor1, Lenz Belzner2 and Claudia Linnhoff-Popien1 1Mobile and Distributed Systems Group, LMU Munich, Munich, Germany 2MaibornWolff, Munich, Germany [email protected] 9 1 0 2 y a M 0 1 ] A M . s c [ 1 v 7 7 0 4 0 . 5 0 9 1 : v i X r a Abstract In nature, flocking or swarm behavior is observed in many species as it has beneficial properties like reducing the prob- ability of being caught by a predator. In this paper, we pro- pose SELFish (Swarm Emergent Learning Fish), an approach with multiple autonomous agents which can freely move in a continuous space with the objective to avoid being caught by a present predator. The predator has the property that it might get distracted by multiple possible preys in its vicinity. We show that this property in interaction with self-interested agents which are trained with reinforcement learning to solely survive as long as possible leads to flocking behavior similar to Boids, a common simulation for flocking behavior. Fur- thermore we present interesting insights in the swarming be- havior and in the process of agents being caught in our mod- eled environment. Introduction Flocking or swarm behavior is observed in many species in nature. A prominent example is fish schooling, where mul- tiple fishes do not only stay close to each other for social reasons but coordinate their actions collectively. That means that an individual fish aligns its direction in regard to fishes that are close to it, while maintaining a certain cohesion of the group and still avoiding collisions with other individuals. However, flocking behavior does not only exist as an end in itself. In nature, a schooling fish benefits from schooling in multiple ways: The swarm increases one's hydrodynamic efficiency or mating chances. Also, flocking enhances for- aging success as collaborative observation is superior to a single individual's. The same is true for predator detection. Even further, the probability of being caught decreases for an individual with regard to certain predator behaviors. Reynolds (1987) showed that algorithmically implement- ing the three rules of alignment, cohesion and separation leads to flocking behavior while an individual only needs local knowledge about its surrounding neighbors (called Boids). In order to overcome these static flocking rules Morihiro et al. (2008) used reinforcement learning to train an individual to justify the rules stated above in order to form a swarm. This was done by shaping the reward sig- nal according to distances between the individuals and lim- iting their actions to be attracted to another fish, be repulsed from another fish and move parallel in the same or opposite direction of another fish, respectively. With SELFish we investigate the case that an individual tries to optimize its behavior with respect to the objective of surviving as long as possible in the presence of a predator (which might get distracted by multiple preys). We show that this simple objective leads to emergent flocking behav- ior (similar to Boids) in a multi-agent reinforcement learning setting, without the need to explicitly enforce it. Reinforcement Learning Reinforcement Learning denotes a machine learning paradigm in which an agent interacts with its environment and receives a certain reward for its action accompanied with an observation of the new state of the environment. Such scenarios are usually modeled as Markov Decision Processes (MDPs), where S denotes the set of states of the environment, A denotes the set of actions an agent can take and r(st, at) is the intermediate reward received after action at was taken in state st at time step t. Also, the process moves to a new state st+1 influenced by the action at, with the Markov property being that the new probability of tran- sitioning into state st+1 only depends on state st and the chosen action at: P(st+1st, at). The goal is to find a pol- icy π : S → A which maximizes the accumulated reward i=t γi−tr(si, ai) from time step t to the simulation horizon T with a discounting factor γ ∈ [0, 1]. Rt =(cid:80)T In SELFish the state is partially observable, which means that instead of using the full state description st to determine the action at = π(st), the agent only uses an observation ot ∈ O (where O is the space of all possible observations) as input to a policy function π : O → A to compute the action at = π(ot). Furthermore the observation may be dif- ferent for every agent. However, we focus on a deterministic domain, so P(st+1st, at) ∈ {0, 1}. Deep Learning In Reinforcement Learning the policy or intermediate func- tions, which help to derive it, are usually expressed as deep artificial neural networks. Neural networks can viewed as a directed graph of nodes, called neurons, which are intercon- nected by weighted edges. A neuron receives inputs over its ingoing edges, usually computes the weighted sum of the inputs, applies a non-linear function to this weighted sum and forwards its output to subsequent neurons via its out- going edges. The neurons are usually arranged in layers, where layers between the input layer and the output layer of the network are referred to as hidden layers. Networks with multiple hidden layers are called deep neural networks. Artificial neural networks serve as biologically inspired function approximators which can be trained by example to approximate a function f mapping an input vector x ∈ IRn to an output vector y ∈ IRm depending on the weights of the edges θ. The goal in training a neural network is to minimize the error between the networks' output y(cid:48) = f (x; θ) and the known desired (example) output y by adjusting the weights θ accordingly. This can be done with the Backpropagation method combined with a gradient descent strategy. Deep Q-Learning (DQN) Q-Learning is a value-based approach named after the action-value function Qπ : S×A → IR, which describes the expected accumulated reward Qπ(st, at) after taking action at in state st and following the policy π in all subsequent states. The goal is to find an optimal action-value function Q∗, which yields the highest accumulated reward. Q∗ can be approximated through Bellman's principle based on the intuition that for an optimal policy, independently of the ini- tial state and initial decision, all remaining decisions must constitute an optimal policy with regard to the state result- ing from the first decision (Bellman (1957)). Starting from an initial guess for Q, it can be iteratively updated via Q(st, at) ← Q(st, at)+α[rt+γ max Q(st+1, a)−Q(st, at)] where the learning rate α ∈ (0, 1) is a parameter to be specified. The learned action-value function Q converges to Q∗, from which an optimal policy can be derived via π∗(st) = arg maxa Q(st, a). a In Deep Q-Learning (DQN) (Mnih et al. (2013)) an ar- tificial neural network is used to represent the action-value function Q. Also, to minimize correlations between sam- ples and to alleviate non-stationary distributions an experi- ence replay mechanism is used (Mnih et al. (2013)) which randomly samples previous state action transitions to train the neural network. Deep Deterministic Policy Gradient (DDPG) To overcome the limitation of Q-Learning, which cannot di- rectly be applied to continuous action spaces, efforts were ] made to learn the policy µ(sθµ) directly with a parame- terized objective function J(θ) (Silver et al. (2014); Lilli- crap et al. (2015)). In addition it was proposed to split the learning process in two components to reduce the gradient variance, called actor-critic approach. The critic learns the action-value function Q(s, a) using the Bellman equation as in Q-learning. The actor then updates the policy parameters θµ in the direction suggested by the critic: ∇θµJ = Est[∇aQ(s, aθQ)s=st,a=µ(st)∇θµµ(sθµ)s=st Multi-Agent Case Many approaches have been suggested for the case that there are multiple agents present which are either self-interested or have to work together to achieve a cooperative goal. A straightforward idea in the case that there are multiple agents that act in their self-interest, which means that they only maximize their own accumulated reward, is deploying a standard reinforcement learning algorithm (as in the single- agent case) in each individual agent in the multi-agent set- ting and let all agents learn simultaneously. This straightfor- ward approach bears the problem of non-stationarity in the state transitions. As one agent tries to adapt its actions in certain states, other agents, which are considered as part of the environment for the first agent, do so as well. This makes it difficult to learn a policy depending on the observed state, which no longer satisfies the Markov property. Egorov (2016) approaches a pursuit-evasion game with reinforcement learning. There are multiple pursuers and multiple evaders. Only one agent of each kind is trained through Q-Learning at a time while the policies of the other agents are fixed. After a number of iterations the policy of the learning agent is distributed to all other agents of the same type. Through this process the policy of one set of agents is improved incrementally over time. This mitigates the problem of non-stationarity. Further- more it seems reasonable to copy the policy of one agent throughout multiple homogenous agents as all are alike and pursue the same self-interested goal. This observation is also relevant for flocking or swarm behavior of multiple agents as we will demonstrate below. Swarm Behavior In 1987, Craig Reynolds (Reynolds (1987)) described three basic rules through which flocking behavior can be mod- eled. For these rules an individual only needs local knowl- edge about its neighbors within a certain distance. These rules are: • Alignment: Steer towards the average heading direction of local flockmates • Cohesion: Steer towards the average position (center of mass) of local flockmates • Separation: Steer to avoid crowding local flockmates If each individual (called Boids by Reynolds as he thought of bird-like creatures) follows these rules, a swarm forma- tion emerges. In an implementation, they can be expressed as physical forces which act upon an individual. Supple- mentary forces can be introduced, which repel an individual from an enemy or from obstacles, for example. To overcome these static rules definitions, Morihiro et al. (2008) used Reinforcement Learning, particularly Q- Learning, to train agents to follow these rules. In their model the agents iteratively learn while at every time step an agent i only considers one other agent j. Agent i receives the eu- clidean distance to j as observation and can choose among four actions to execute. These actions are to move towards agent j, away from agent j or parallel to agent j either in the same or opposite direction. The reward agent i receives for an action depends on the previously mentioned distance to agent j and is shaped in a way that it intuitively represents the cohesion and separation rule. In this regard agent i re- ceives a positive reward if it steers so to keep its distance to j within predefined boundaries. While the previously mentioned approaches lead to flock- ing behavior, they neglect the beneficial properties flocking behavior might have for the individuals. One of those ben- efits could be the increased likehood to survive in the pres- ence of predators, as they might get distracted by the sheer amount of possible targets. The question arises whether flocking behavior occurs in a scenario with such properties where agents solely try to maximize their survival time. In contrast to Morihiro et al. (2008), we pursue a scenario in which agents are trained with reinforcement learning solely on the objective to survive, without explicitly enforcing swarm behavior. Additionally, we demonstrate that SELFish also works for a continuous action space of the agents. Emergent Swarm Behavior In order to investigate whether the objective to survive in the presence of a predator would lead to flocking behavior in a multi-agent setting, we created a model that facilitates such a behavior. In the following the properties of the environment will be explained. This is followed by a description of the action and observation space as well as the reward structure which was used to train the agents. Environment The agents, which are the prey in this scenario, can freely move in a continuous two-dimensional space, visualized as a square with predefined edge lengths (see Figure 1). An agent itself is represented as a circle with a surface substan- tially smaller than the space it is moving in. There are nei- ther obstacles nor walls in the environment. Furthermore agents do not collide with each other. To ease free roaming of the agents, the space has the special characteristic that it wraps around at the edges forming a torus. That means that if an agents leaves the square visualization to the right, it Figure 1: Example of the space with 60 agents (green) and one predator (orange). will immediately enter it again from the left (same for the other direction or around top and bottom). Together with the agents there also exists a predator in the environment. The predator is also represented as a circle. The goal of the predator is to catch the agents by moving to their position. As soon as the predator collides with an agent, the execution of the concerning agent will end and a new agent is spawned immediately at a random position to keep the number of agents in the system constant. If there are multiple agents within a certain distance around the predator, it will choose one for a target at random (oth- erwise it will move to the closest agent's direction). This means that the predator can be distracted by multiple agents in its proximity. Thus it might be beneficial for an agent to move towards other agents as the predator might get dis- tracted, which is essential for flocking behavior. However, to prevent the predator from constantly changing targets it will follow a chosen target for a certain time before a new target will be chosen. By default, the agents and the predator move at the same speed. This would allow an agent to turn in the opposite direction of the predator and move away without the predator having a chance to catch up. That is why the predator will accelerate occasionally for a short amount of time, which simulates a leap forward to catch the prey it is following. The policy of the predator is static and does not change over time. Objective of an Agent The goal of the agents is not to collide with the predator. For this they receive a reward of +1 for each step/frame they live and -1000 for the collision with the predator which ends their life. With this reward structure the objective of the agents can be viewed as "surviving as long as possi- ble". As there are no obstacles in the environment and the agents do not collide with each other, there are no other re- wards/penalties. Action Space The action space of the agents only comprises of the angle they want to turn each time step. The movement speed of the agents is constant and cannot be altered by them for now. The action a, which represents the turning angle that can be chosen from discrete steps or out of a continuous interval by the agent, depends on the reinforcement learning strategy which is used later on. In the case that DQN is used, the actions an agent can chose from comprise five discrete de- gree values {−90◦,−45◦, 0◦, +45◦, +90◦}. The agent can choose any real-valued degree as turning angle in the case of DDPG. As a side node, the predator can only take limited real- valued turns {x ∈ IR − 45◦ ≤ x ≤ 45◦} at every step with the goal to give the agents a higher maneuverability than the predator. Observation Space In order to facilitate the scalability to many autonomous agents, one agent cannot observe the full state of the environ- ment; instead its observation is limited to itself, the predator and the n nearest neighboring agents. This approach can be explained biologically, where, for example, a fish in a swarm cannot observe the whole swarm but only its local neighbors. But it is also in line with related work, for ex- ample Boids, where also only local neighborhoods between agents are regarded. Furthermore it eases computation and has the nice property that the observation vector, which is forwarded through the reinforcement learning algorithm in order to obtain an action, has a constant length (cf. the fol- lowing section). For every observable entity e, the agent receives a 3-tuple which contains the euclidean distance between the entity and the agent, the angle the agent would have to turn to face to- wards the observed entity and the absolute orientation of the entity in the environment: (diste, directione, orientatione). As the environment is a torus, the distances are also cal- culated around the edges of the visualized square, with the shorter distance being taken (with the directione correspond- ing to this). The absolute orientation of an entity is measured in degrees [0◦, 360◦), where facing east corresponds to 0◦, measuring the angle counter-clockwise. The angle an agent would have to turn to face towards another entity is mea- sured in degrees in the range of (−180◦, 180◦]. Accordingly, an agent receives the following observation for the predator, itself and the n nearest neighboring agents, in which the n neighbors are ordered by their distance.  distpredator directionpredator 0 distneighbor1 distneighbor2 distneighborn 0 ... directionneighbor1 directionneighbor2 directionneighborn orientationpredator orientationself orientationneighbor1 orientationneighbor2 orientationneighborn  Hyperparameter Training Steps Hidden Layer Neurons in Layers Hidden Layer Activation Last Layer Activation γ Optimizer Learning Rate Replay Buffer Size Batch Size Exploration DQN 500,000 10 16 relu linear 0.999999 Adam 0.001 50,000 64 -Greedy = 0.1 Observable neighboring agents 5 DDPG 500,000 5 Actor: 16 Critic: 32 relu linear 0.999999 Adam 0.001 100,000 512 Ornstein Uhlenbeck θ = 0.15, µ = 0.0, σ = 0.3 1 Table 1: Hyperparameters for Reinforcement Learning Training As mentioned before, a valid way for training multiple ho- mogeneous agents through reinforcement learning is to train only one instance and then to copy the learned policy to all instances of the homogeneous group (Egorov (2016)). This also resembles nature, where for example multiple schooling fish follow the same behavioral policy. For this purpose, the DQN and DDPG implementations of Keras-RL (cf. Plappert (2016)) were used. Keras-RL is originally developed for OpenAI Gym Environments (Plap- pert (2016)), in which only single agents interact with these environments through a step(action)-method, which is given an action and returns an observation, a reward and a done flag, indicating whether the current episode is finished. This interface was also used in the proposed swarm environment to train a single agent to avoid the present predator with the previously mentioned rewards, action and observation spaces. During the training of one agent, the other agents are present as well, onto which the policy (i.e. the neural network) of the learning agent is copied after each episode. An episode ends if the learning agent is caught by the preda- tor or 10, 000 steps (frames) were executed. During training, the edge lengths of the space were 40×40 pixels, although it wraps around at the edges. Please note that the agents and the predator could be positioned at any real value in the interval [0, 40]. However, the values in the 3-tuples of the observation were normalized to [0, 1] anyway. The agents and the predator were represented by circles of radius 1, with an agent being caught if the distance of its position and the position of the predator is below 2. Also, during training only 10 agents were present. In order to find a good configuration for the parameters Figure 2: Swarm formation in the first 40 frames of an episode of SELFishDQN. Agents (white) and predator (red) were randomly initialized. Kernel Density Estimation Phillips et al. (2006) was used to highlight the dense regions of the multi-agent swarm. Note that the space wraps around the edges. of the reinforcement learning algorithms, many runs were executed. The quality of the parameter configuration of the training run was assessed during a test phase based on the cumulative reward the learning agent could acquire, which essentially equals the number of time steps it could survive. The number of neighboring agents that could be observed was also varied as parameter. See Table 1 for the best pa- rameters found. Even for the small number of agents which were present during the training, a swarming behavior could be observed when the learned behavior of one agent was transferred to the others. Since the observation of an agent is partial and thus limited to the 3-tuple (distneighbori , directionneighbori, orientationneighbori) for the n nearest neighbors, the number of agents as well as the size of the space can be increased without breaking the learned policy. With this even better swarming behavior can be ob- served, which shall be further evaluated in the next section. Simulations and Results First we want to give an impression of the swarms that are forming from reinforcement learning. See Figure 2 for the formation of a swarm in the first 40 frames of a test episode of SELFishDQN. With a continuous action space, SELFishDDPG, exhibits similar behavior although the swarm tends to be more dense. The swarm presumably forms be- cause one agent learns that the predator might get distracted from it if it stays close to other agents which prolongs its life and thereby its accumulated reward. Boids enforces the alignment, cohesion and separation of neighboring agents. This can be expressed by vector cal- culations together with weights which set these three rules in context. To make the scenario more similar to the rein- forcement learning setting, another force which pushes the Boids away from the predator was added (altogether with a weight for this behavior which sets it in context to the other rules). To find a good configuration for the alignment, co- hesion, separation and predator avoidance weight, multiple runs with different parameter setting were executed. Again, the quality of a setting was evaluated based on the number of time steps a certain boid could survive. If it is only about the survival of an agent, a simple strat- egy one could think of is to simply turn in the opposite direc- tion of the predator and to move away from it regardless of the surrounding agents. This policy will be called TurnAway in the following and will be given for comparison1. Alignment and Cohesion As Boids enforces the alignment and the cohesion of the agents, we want to compare the swarms resulting from predator avoidance through reinforcement learning to Boids by these means. As the orientation of an agent is measured as angle in [0◦, 360◦) (facing east corresponds to 0◦), the alignment of the agents can be measured as deviation from a mean angle of a group (see Figure 3). The absolute deviation of each agent from this mean angle was summed and aver- aged over the number of agents. To measure the cohesion of the swarm, the average distance between the agents was calculated. For this the distance between all agents i and j was summed and averaged by the number of pairs of agents. Considering that the agents flee from a predator and the space wraps around at the edges, multiple flocks with dif- ferent orientations, depending on their position in regard to the predator, might form, as it is already evident from the Figures 1 and 2. That is why it did not seem sensible to cal- culate alignment and cohesion over all agents in the space. To counter this, the density-based clustering method DB- SCAN (Ester et al. (1996)) and particularly its scikit-learn implementation (Pedregosa et al. (2011)) was used before- hand and the average deviation from the mean angle and the average distance between two agents was only calculated for agents in a specific cluster (see Figure 4 for an example). The measurements over all agents are given for comparison. 1For a short video showing all implemented policies please re- fer to https://youtu.be/SY59CYaqWpE 4035302520151050Figure 3: Considering the orientation of five agents in space, a mean angle (black) and the deviation from this in (−180◦, 180◦] can be computed (Watson (1983)). Figure 4: Example Clustering for SELFishDQN with 40 agents (predator as red dot). Figure 5 shows the number of agents in a specific cluster, when 40 agents were present in a space of 40 × 40 pixels. It is visible that the TurnAway strategy produces many noise points on average. The clusters that are found for TurnAway are mostly due to the agents moving in the same direction to avoid the predator and also overlapping when wrapping around the edges of the space. Boids and the two rein- forcement learning approaches used in SELFish, DQN and DDPG, produce rather similar cluster numbers and sizes on average, with DDPG having a tendency to form one large cluster. By looking at the average deviation from the mean ori- entation angle of the agents inside clusters (see Figure 6) Figure 6: Average deviation from the mean orientation angle of the agents over clusters. one can see that Boids produces the most aligned groups of agents which generally move in the same direction. SELFishDQN and SELFishDDPG are deviating more, presum- ably because agents following these policies tend to kind of quiver. Also these agents show the behavior of creat- ing a line at the point at which they would again move to- wards the predator because of the torus environment. At these lines the agents circulate until the predator moves into their direction. For TurnAway only groups of agents mov- ing in the same direction are detected anyway, with the aver- age angle deviation being distorted by agents coming from the other side of the space and moving in the opposite di- rection. One might question whether the swarms (respec- tively clusters) found for SELFishDQN or SELFishDDPG also solely result from the fact that the agents learned to turn away from the predator and thereby move in the same di- rection. This can be countered by the observation that if the predator is pinned down at a fixed position (it cannot be re- moved completely as it is part of the agents' observation), the learning agents still form a swarm at the greatest possi- ble distance from the predator where they circulate around each other. Figure 7 shows the average pairwise distance between agents either inside clusters, between noise points or between all agents, which is homogeneous over all four agent policies, with only SELFishDDPG tending to produce somewhat denser agent groups. The homogeneity between the behavioral strategies with regard to the average pairwise distance also results from the DBSCAN clustering. Figure 5: Average number of agents in a respective cluster (cluster ID given) with noise points being agents that could not be assigned to a specific cluster. Figure 7: Average pairwise distance between agents ei- ther inside clusters, between noise points or over all. Edge lengths of the space normalized to 1 for distance calculation. avgAngle: 111.06diff. to avgAngle: -98.94diff. to avgAngle: 24.06diff. to avgAngle: 95.06diff. to avgAngle: 13.06diff. to avgAngle: -39.940.00.20.40.60.81.00.00.20.40.60.81.0BoidsSELFishDDPGSELFishDQNTurnAway0102030Avg.no.ofagentsperclusterNoisePoints0123BoidsSELFishDDPGSELFishDQNTurnAway01020304050Avg.deviationfromthemeanangle[degree]OverallNoisePoints012BoidsSELFishDDPGSELFishDQNTurnAway0.00.10.20.3Avg.distbetweenagentsOverallNoisePoints012Figure 8: Average episode length for each of the behavior strategies with varying number of agents in the environment. Figure 10: Density of an agent in accordance to the Ker- nel Density Estimation in the last 100 time steps before it is caught (mean for multiple agents). the reinforcement learning could potentially be explained considering the Prisoner's Dilemma (Poundstone (1992)). In this game-theoretical example, prisoners A and B are kept in arrest without means to communicate. Simultaneously, both are given the opportunity either to betray the other by testifying that the other committed the crime, or to cooperate with the other by remaining silent with the respective pay- offs shown in Table 2. The only Nash equilibrium (Nash (1951)) is that both prisoners defect as this yields less charge for each of them than if one stays silent while the other pris- oner keeps its strategy unchanged and testifies that the other committed the crime (betrays). The dilemma is that mutual cooperation yields a better outcome although it is not ratio- nal from a self-interested perspective. For our reinforcement learning setting it could be the case that the TurnAway strat- egy was not found because the learning process got stuck in the Nash equilibrium of staying with the swarm (analo- gous to the mutual defection in the Prisoner's Dilemma). If all agents keep their policy of staying close to each other, the one agent deviating has a higher chance of being chosen as prey. Our learning procedure is in conformity with this as one learning agent adjusts its policy in such a way that it obtains the highest reward while the policies of the other agents stay unchanged (during an episode). This assumption is also supported by looking at the procedure how agents are caught (see Figure 11): When the predator moves in the di- rection of the swarm, it collaboratively moves away, with a few agents being left behind. The community of the agents gets smaller and smaller as some sheer off until one is sepa- B B stays silent (cooperates) -1 A A stays silent (cooperates) A betrays (defects) -1 -3 0 B betrays (defects) 0 -2 -3 -2 Table 2: Prisoner's dilemma payoff matrix Figure 9: Number of caught agents divided by the time it took with varying number of agents in the environment. Agent Survival For the reinforcement learning algorithms the reward was defined such that the single learning agent received +1 for every step and −1000 for being caught. The maximization of the accumulated reward should encourage it to stay alive as long as possible. After the end of an episode, which ended when the learning agent was caught or 10, 000 steps passed, the learned policy was copied to all other agents. Figure 8 shows the mean episode length for the different policies, which essentially corresponds to the mean accumulated re- ward of the learning agents. For the static policies, Boids and TurnAway, it corresponds to the time it took until a cer- tain agent was caught. Note that although the number of agents in the environment is varied, the parameter for Boids or the policies for SELFishDQN/DDPG are still those that were determined in smaller settings with only 10 agents. It turns out that when evaluating the actual survival rate of every single agent, the best strategy to survive is to simply turn away from the predator. This is also true considering the whole swarm, i.e. all agents. In Figure 9, the absolute number of caught agents in an episode was divided by the length of the episode (reduced by a transient phase of 100 frames for swarm formation). These measurements were then again averaged over multiple episodes and runs (with different seeds). This raises the question why this behavior was not found by the reinforcement learning algorithms. The outcome of 20406080100Numberofagents10001500200025003000350040004500Avg.episodelenghtBoidsSELFishDDPGSELFishDQNTurnAway20406080100Numberofagents0.020.030.040.050.060.07Avg.caughtagentsperframeBoidsSELFishDDPGSELFishDQNTurnAway-100-80-60-40-200Timestepsbeforebeingcaught0.20.30.40.50.6DensityBoidsSELFishDDPGSELFishDQNTurnAwayFigure 11: Separation of agents from the swarm before being caught. rated and picked as prey. This is also evident considering the density measurements of agents in the time steps before it is being caught. Figure 10 shows the density around an agent in accordance to the Kernel Density Estimation (cf. Figure 2 and 11) in the last 100 time steps of its life. Conclusion and Future Work With SELFish we showed that flocking behavior can emerge solely from the fact that agents trained by multi-agent re- inforcement learning try to avoid being caught by a preda- tor, given the circumstance that flocking yields a benefit like distracting the predator. Only one agent was trained at a time with a reward structure that encourages to avoid being caught as long as possible. After each episode the learn- ing policy was copied onto all other agents. The results for SELFishDQN and SELFishDDPG concerning the alignment and cohesion but also with regard to the survival chances of the agents were compared with Boids, a common approach for algorithmic flocking simulations. Out results show, that the measurements for the swarm are quite similar to Boids. Considering the survival of an agent, surprisingly, the re- inforcement learning algorithms did not find the policy of simply turning away from the predator (without caring about flocking) although it yields higher accumulated rewards w.r.t our reward structure. We propose that staying in the swarm is a Nash equilibrium (comparable to defecting in the Pris- oner's dilemma) and want to further investigate this assump- tion. Also, we would like to examine if other beneficial properties of a swarm, like increased hydrodynamic effi- ciency or easier search for food, which were not modeled by us, also lead to flocking behavior in a reinforcement sce- nario. This would probably also facilitate the steering of the swarm. Co-evolution of the behavior of the predator and its prey through reinforcement learning could be further investi- gated in our continuous environment. In our setting, agents could freely roam in a torus-like environment without ob- stacles or collisions. Naturally, there are enhancements to this like adding walls, obstacles and collisions between the agents. References Bellman, R. (1957). Dynamic programming. Princeton Press. Egorov, M. (2016). Multi-agent deep reinforcement learning. CS231n: Convolutional Neural Networks for Visual Recog- nition. Ester, M., Kriegel, H.-P., Sander, J., Xu, X., et al. (1996). A density-based algorithm for discovering clusters in large spa- tial databases with noise. In Kdd, volume 96. Lillicrap, T. P., Hunt, J. J., Pritzel, A., Heess, N., Erez, T., Tassa, Y., Silver, D., and Wierstra, D. (2015). Continuous arXiv preprint control with deep reinforcement learning. arXiv:1509.02971. Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, Playing arXiv preprint I., Wierstra, D., and Riedmiller, M. (2013). atari with deep reinforcement arXiv:1312.5602. learning. Morihiro, K., Nishimura, H., Isokawa, T., and Matsui, N. (2008). Learning grouping and anti-predator behaviors for multi- In Int'l Conf. on Knowledge-Based and In- agent systems. telligent Information and Engineering Systems. Springer. Nash, J. (1951). Non-cooperative games. Annals of Mathematics. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., and Duchesnay, E. (2011). Scikit- Journal of Machine learn: Machine learning in Python. Learning Research, 12. Phillips, S. J., Anderson, R. P., and Schapire, R. E. (2006). Max- imum entropy modeling of species geographic distributions. Ecological modelling, 190(3-4). Plappert, M. (2016). keras-rl. https://github.com/keras-rl/keras-rl. Poundstone, W. (1992). Prisoner's Dilemma. Doubleday. Reynolds, C. W. (1987). Flocks, herds and schools: A distributed behavioral model. In ACM SIGGRAPH computer graphics, volume 21. ACM. Silver, D., Lever, G., Heess, N., Degris, T., Wierstra, D., and Ried- miller, M. (2014). Deterministic policy gradient algorithms. In ICML. Watson, G. (1983). Statistics on spheres. University of Arkansas lecture notes in the mathematical sciences. Wiley. 185180175170165160155150145
1803.07401	2	1803	2018-07-25T08:49:00	Asynchronous opinion dynamics on the $k$-nearest-neighbors graph	[ "cs.MA", "eess.SY", "math.DS" ]	This paper is about a new model of opinion dynamics with opinion-dependent connectivity. We assume that agents update their opinions asynchronously and that each agent's new opinion depends on the opinions of the $k$ agents that are closest to it. We show that the resulting dynamics is substantially different from comparable models in the literature, such as bounded-confidence models. We study the equilibria of the dynamics, observing that they are robust to perturbations caused by the introduction of new agents. We also prove that if the number of agents $n$ is smaller than $2k$, the dynamics converge to consensus. This condition is only sufficient.	cs.MA	cs	Asynchronous opinion dynamics on the k-nearest-neighbors graph Wilbert Samuel Rossi and Paolo Frasca∗† July 26, 2018 Abstract This paper is about a new model of opinion dynamics with opinion- dependent connectivity. We assume that agents update their opinions asynchronously and that each agent's new opinion depends on the opinions of the k agents that are closest to it. We show that the resulting dynamics is substantially different from comparable models in the literature, such as bounded-confidence models. We study the equilibria of the dynamics, observing that they are robust to perturbations caused by the introduction of new agents. We also prove that if the number of agents n is smaller than 2k, the dynamics converge to consensus. This condition is only sufficient. 1 Introduction Driven by the evolution of digital communication, there is an increasing interest for mathematical models of opinion dynamics in social networks. A few such models have become popular in the control community, see the surveys [1, 2]. In the perspective of the control community, opinion dynamics distinguish them- selves from consensus dynamics because consensus is prevented by some other dynamical feature. In many popular models, this feature is an opinion depen- dent limitation of the connectivity. This is the case of bounded confidence (BC) models [3, 4], where social agents influence each other iff their opinions are closer than a threshold. This way of defining connectivity, however, seems at odds with several social situations, since it may require an agent to be in- fluenced by an unbounded number of fellow agents. Instead, the number of possible interactions is capped in practice by the limited capability of attention by the individuals. For instance, online social network services are based on ∗W.S. Rossi and P. Frasca are with Department of Applied Mathematics, University of Twente, 7500 AE Enschede, The Netherlands [email protected]. P. Frasca is with Univ. Grenoble Alpes, CNRS, Inria, Grenoble INP, GIPSA-lab, F-38000 Grenoble, France [email protected] †This work has been partly supported by IDEX Universit´e Grenoble Alpes under C2S2 "Strategic Research Initiative" grant. The authors also acknowledge the inspiring conversa- tions with J.M. Hendrickx and S. Martin. recommender systems that select a certain number of news items, those which are closer to the user's presumed tastes. However, to the best of our knowledge, this important observation has not been incorporated in any suitable model of opinion dynamics, with the partial exception of [5]. The latter paper compares different models of interaction, including one in which each agent is influenced by a fixed number of neighbors. In a striking contrast, this observation has been made in the field of biology by a number of quantitative studies about flocking in animal groups (these include both theoretical and experimental works) [6, 7, 8, 9]. The importance of this way of defining connectivity has been also captured by graph theorists, who have studied a the properties of what they call k-nearest-neighbors graph. For instance, it is known that k must be logarithmic in n to ensure connectivity [10] and flocking behavior [11]. In this paper, we provide the first analysis of the k-nearest-neighbor opinion dynamics. In this analysis, our contribution is threefold: (1) We describe the equilibria of the dynamics, distinguishing a special type of clustered equilibria that are constituted of separate clusters; (2) We discuss the robustness of clus- tered equilibria to perturbations consisting in the addition of new agents; (3) We provide a proof of convergence for small groups, that is, groups such that n < 2k. Our work differs from [5] in several aspects. As per the model, the dynamical model in [5] is synchronous and continuous-time, whereas ours is asynchronous and discrete-time. As per the analysis, [5] focuses on the equilibria and their properties (for instance, the distribution of their clusters' sizes) are studied by extensive simulations, whereas we study the dynamical properties (robustness to perturbations, convergence) by a mix of simulations and analytical results. Our robustness analysis is based on the approach taken by Blondel, Hendrickx and Tsitsiklis for BC models [12]. Our convergence result is inspired by classical proofs of convergence for randomized consensus dynamics [13, Chapter 3], but its interest and difficulty originate from the lack of reciprocity in the interactions: this feature clearly distinguishes our model from bounded confidence models, where interactions are reciprocal as long as the interaction thresholds are equal for all agents [3, 5, 14, 15, 16, 17]. 2 The dynamical model Let n and k be two integers with 1 ≤ k ≤ n, and let V = {1, . . . , n} be the set of agents. Each agent is endowed with a scalar opinion xi ∈ R, to be updated asynchronously. The update law x+ = f (x, i) (1) goes as follows. An agent i is selected from V ; the elements of V are ordered by increasing values of xj −xi; then, the first k elements of the list (i.e. those with smallest distance from i) form the set Ni of current neighbors of i. Should a tie between two or more agents arise, priority is given to agents with lower index. Agent i may but not necessarily does belong to Ni. Once Ni is determined, agent i updates his opinion xi to (cid:88) j∈Ni x+ i = 1 k xj , while the remaining agents do not change their opinions x+ j = xj for every j (cid:54)= i . We show a couple of simulations to illustrate the possible behaviours of the model, see Figure 1 and 2. For these simulations we set n = 20, k = 5 and choose the initial opinion of every agent uniformly at random in [0, 1]. At every step, we choose from V the node that updates opinion, independently and uniformly at random. The simulation of Figure 1 shows a typical outcome: the agents form two distinct groups (of 10 agents each) with homogeneous opinions; for every agent, his neighbors at time t = 1000 have almost the same opinion. This last observation does not hold in the simulation of Figure 2: the two pairs of agents that at time t = 1000 have opinion about 0.6 and about 0.7, respectively, have neighbors with different opinions. These distinct behaviors lead us to distinguish different kinds of equilibria: this will be the topic of the next section. 3 Equilibria In this section we discuss some properties of the equilibria of system (1). Mo- tivated by the simulations, we introduce the following terminology. Given a configuration x ∈ Rn, the directed graph that represents the possible interac- tions (i.e. the opinion dependancies for any possible selection of the node to be updated) is G(x) = (V, E(x)) with E(x) = {(i, j), j ∈ Ni} , (cid:91) i∈V where Ni is the set of neighbors of i, should i be selected to update his opinion. Clearly, if k = n the graph G(x) = (V, V × V ) is complete. A configuration x ∈ Rn is an equilibrium for the asynchronous dynamics if x = f (x, i) for every i . If k = 1, then G(x) contains only links between nodes with the same opinion: in this trivial case, every configuration is an equilibrium because agents cannot change opinion. A configuration x is called clustered if xNi = xi1Ni for every i , Figure 1: Simulation of the model (1) with n = 20, k = 50, initial opinions chosen uniformly at random in [0, 1] and update sequence chosen uniformly at random. The plot contains a typical trajectory that converges to a clustered equilibrium. 10010110210300.20.40.60.81Figure 2: Simulation of the model (1) with n = 20, k = 50, initial opinions chosen uniformly at random in [0, 1] and update sequence chosen uniformly at random. The plot contains a less common trajectory that converges to a non-clustered equilibrium. 10010110210300.20.40.60.81that is, if for every node all of his neighbors have the same opinion. Furthermore, a clustered configuration x = c1 for some c ∈ R is called consensus. It is immediate to see that clustered configurations are equilibria. However, there exist equilibria that are not clustered. It is possible to obtain a simple counterexample with n = 7 and k = 3 and exploiting the tie break rule. Consider any configuration x ∈ R7 of the form x{1,3,5} = α 1{1,3,5} , x{2,4,6} = β 1{2,4,6} , x7 = α+β 2 , where α, β ∈ R and α < β. The above is an equilibrium even if xN7 = x{1,2,7} (cid:54)= 1 2 (α + β)1{1,2,7}. The tie breaking rule is not central for the existence of non-clustered equi- libria, as one can see in the following example inspired by Figure 2. Example 1. Consider x ∈ R20 with x{1,2,...,11} = α 1{1,2,...,11} , x12 = x13 = 3α+2β x14 = x15 = 2α+3β x{16,17,...,20} = β 1{16,17,...,20} , 5 5 , , where α, β ∈ R and α < β. For instance, the neighbors of agent 12 are N12 = {1, 12, 13, 14, 15} because x12 − x12 = x13 − x12 = 0 , x12 − x14 = x12 − x15 = 1 x12 − x1 = 2 5 (β − α) , 5 (β − α) , while the remaining agents are at distance 2 is an equilibrium with xN12 (cid:54)= x12 1N12 . 5 (β−α) or larger. Such configuration A simple analysis shows that clustered configuration are those in which the agents form clusters of at least k participants with the same opinion. To make this claim formal, let Vi = {j : xj = xi} be the set of nodes that share the same opinion of i. Lemma 1. A configuration is clustered if and only if Vi ≥ k for every i. Proof. By definition, in a clustered configuration Ni ⊆ Vi for every i. Assume Vi ≥ k for every i. For any i there are at least k nodes j (including i) with xj = xi: such nodes have zero distance from i and hence Ni ⊆ Vi. This holds for every i so the configuration is clustered. On the other hand, assume that exists i with Vi ≤ k − 1. The set Ni must contain a node j with xj (cid:54)= xi so not in Vi, violating the definition of clustered configuration. From this result, it follows that a clustered configuration allows up to (cid:106) n (cid:107) k distinct sets Vi (and this bound is tight). For the special case of consensus, this claim becomes the following corollary. Corollary 2. Consensus is the only possible clustered configuration if and only if n < 2k . 4 ROBUSTNESS OF THE EQUILIBRIA The clustered equilibria of the dynamics described above have interesting ro- bustness properties regarding the addition of new nodes or the removal of nodes. The model shows different behavior with respect to a standard Asynchronous Bounded Confidence (ABC) model. In this section, we briefly introduce for comparison the ABC model; then we provide a few simulations to motivate the following discussion of the robustness properties. 4.1 ABC model Given a fixed range of confidence d > 0, we introduce the Asynchronous Bounded Confidence (ABC) update law x+ = fABC(x, i) . (2) where i is the agent that updates his opinion. The neighborhood of i is N ABC {j : xj − xi ≤ d} and always contains i itself. The new opinion of agent i is i = (cid:88) x+ i = 1 N ABC i xj , j∈N ABC i while the remaining agents do do not change opinion for every j (cid:54)= i . x+ j = xj 4.2 Simulations We present a simulation to show the difference between model (1) and model (2) when a few agents are added to a consensus configuration (which is an equi- librium for both models). We set k = 5 for model (1) and d = 0.25 for model (2). We start with 10 agents sharing opinion 0.4; at steps t = 2, 3, 4, 5 we add a new agent, with opinion chosen uniformly at random in [0, 1]. We select the agent that updates his opinion among those present at that time, independently and uniformly at random: the same selection is used in both models. Figure 3 contains the plots of the simulation. The upper plot regards the dynamics of model (1): the four new agents converge to the consensus opinion, which does not change; they are too few to form a new cluster. The lower plot contains the dynamics of the model (2): the consensus configuration is not preserved and Figure 3: The addition of four new nodes to a consensus configuration with ten nodes. Upper plot: the trajectory of the model (1) with k = 5. Lower plot: the trajectory of the model (2) with d = 0.25. The same initial conditions and update order are used. 10010110210300.20.40.60.8110010110210300.20.40.60.81the agent added at step t = 3 remains isolated during the dynamics and keeps his opinion. The other three new agents join the original ten; this group of 13 agents converge to the same opinion which however is different from the original consensus value. 4.3 Robustness of the equilibria We now provide a general discussion that explains the observations from Fig- ure 3. Let n, k with 1 ≤ k ≤ n be given and consider a clustered equilibria x ∈ Rn of the model (1). We first discuss the addition of a new agent with opin- ion xn+1 = α to the configuration x, that becomes [x; α] ∈ Rn(cid:48) with n(cid:48) = n + 1. Before the addition of the new node, clusters have to contain at least k agents. This fact remains true after the addition and we have that f ([x; α], i) = [x; α(cid:48)] for every i, meaning that the original (clustered) portion of the configuration [x; α] remains unperturbed. For a generic value of α the limit of the dynamics has the same cluster locations of x, with one of the clusters getting a new member. For some specific values, it may happen that the configuration [x; α] is a non-clustered equilibrium. In any case, none of the original agents changes opinion. Instead, in the metric ABC model (2) with uniform visibility radius d, either the new agent is further apart from the original agents and nothing happens or he falls within the visibility radius of a cluster of agents. In the latter case both the new agents and the agents in the cluster change opinions, converging to an intermediate value. Assuming n sufficiently large, the removal of an agent from a clustered equi- librium presents interesting differences too. In the metric ABC model (2) the removal of an agent does not trigger any dynamics in the remaining agents. In model (1), if the agent is removed from a cluster with k + 1 agents or more, nothing happens. But if the agent is removed from a cluster with k agents, the new configuration is not an equilibrium anymore and the remaining nodes from that group will evolve towards some new equilibrium. 5 Convergence to consensus In this section we show that process (1) converges to a consensus, provided n < 2k and the choice of the agent that updates his opinion at time t is an i.i.d. uniform random variable over V . We recall from Section 3 that the consensus is the unique clustered equilibrium for n < 2k. For t ≥ 0, let x(t) ∈ Rn be the sequence of opinion vectors and I(t) ∈ V a sequence of agents. Given an initial configuration x(0) = x0, we consider the dynamics for every t ≥ 0 , where I(t) is the agent that updates his opinion at time t. x(t + 1) = f (x(t), I(t)) (3) We introduce two functions µ, M : Rn → V that, given an opinion vector x, return respectively the index of the smallest and largest components, with ties sorted µ(x) = min(arg min xi) , i M (x) = min(arg max xi) . i The outer min sorts possible ties; note that M (x) = µ(−x). In the following two lemmas we prove the properties of the dynamics in which the agent with smallest opinion is the one that updates his opinion. Lemma 3. Given n, k with 1 ≤ k ≤ n and an initial configuration x0 ∈ Rn consider dynamics (3) with I(t) = µ(x(t)) and the scalar sequence y(t) := maxi∈Nµ(x(t)) xi(t). Then: • the set sequence Nµ(x(t)) and the scalar sequence y(t) are constant; • for every i ∈ Nµ(x(0)) the sequences xi(t) are non-decreasing and satisfy xi(t) ≤ y(0); • for every i /∈ Nµ(x(0)) the sequences xi(t) are constant. Proof. The proof goes by induction. First, consider the trivial case with xµ(x(t))(t) = y(t). This condition means xi(t) = y(t) for every i ∈ Nµ(x(t)) and thus xµ(x(t))(t+ 1) = xµ(x(t))(t) so everything remains unchanged. Next, consider the case with xµ(x(t))(t) < y(t). We have (cid:88) xj(t) ∈(cid:0)xµ(x(t))(t), y(t)(cid:1) . xµ(x(t))(t + 1) = 1 k j∈Nµ(x(t)) Therefore, and {i : xi(t) < y(t)} = {i : xi(t + 1) < y(t)} {i : xi(t) = y(t)} = {i : xi(t + 1) = y(t)} . Moreover, the cardinality of the set {i : xi(t) < y(t)} is strictly smaller than k. This implies that Nµ(x(t+1)) = Nµ(x(t)) and also y(t + 1) = y(t). The claims follow by induction and by observing that only the agents i ∈ Nµ(x(0)) can update their opinions at some time t ≥ 0 and the updated value xi(t + 1) belongs to [xi(t), y(t)]. Lemma 4. Given n, k with 1 ≤ k ≤ n and an initial configuration x0 ∈ Rn consider the dynamics (3) with I(t) = µ(x(t)) and the scalar sequence y(t) = maxi∈Nµ(x(t)) xi(t). Then Proof. First, compute xµ(x(t))(t + 1) for a generic t ≥ 0. We have (cid:1)(cid:0)y(0) − min xi(0)(cid:1) i y(k−1) − min i k xi(k−1) ≤(cid:0)1 − 1 (cid:80) (cid:80) (cid:80) xµ(x(t))(t + 1) = 1 k = 1 k ≥ 1 k j∈Nµ(x(t)) j∈Nµ(x(0)) j∈Nµ(x(0)) xj(t) xj(t) xj(0) thanks to Lemma 3. Then, xµ(x(t))(t + 1) ≥ k−1 k xµ(x(0))(0) + 1 = xµ(x(0))(0) + 1 k S(t) =(cid:8)i : xi(t) < xµ(x(0))(0) + 1 k k y(0) (cid:0)y(0) − xµ(x(0))(0)(cid:1) . (cid:0)y(0) − xµ(x(0))(0)(cid:1)(cid:9) , Next, consider the set and observe that either S(t) = ∅ or S(t + 1) = S(t) − 1 because µ(x(t)) /∈ S(t + 1). Since the set S(0) contains at most k − 1 elements, the set S(k−1) is empty. Hence, xi(k − 1) ≥ xµ(x(0))(0) + 1 k for every i, a fact that implies xµ(x(k−1))(k − 1) ≥ xµ(x(0))(0) + 1 (cid:0)y(0) − xµ(x(0))(0)(cid:1) (cid:0)y(0) − xµ(x(0))(0)(cid:1) . k Using Lemma 3 we know that Nµ(x(t)) = Nµ(x(0)) for every t ≥ 0 and that for every i therein, xi(t) ≤ y(t) = y(0). Therefore y(k−1) − xµ(x(k−1))(k−1) ≤ y(0) − xµ(x(0))(0) (cid:0)y(0) − xµ(x(0))(0)(cid:1) − 1 k and the thesis follows because xµ(x(t)) = mini xi(t). The following lemma follows from Lemma 3 and 4 using the property M (x) = µ(−x). Lemma 5. Given n, k with 1 ≤ k ≤ n and an initial configuration x0 ∈ Rn consider the dynamics (3) with I(t) = M (x(t)) and the scalar sequence z(t) := mini∈NM (x(t)) xi(t). Then: • the set sequence NM (x(t)) and the scalar sequence z(t) are constant; • for every i ∈ NM (x(0)) the sequences xi(t) are non-increasing and satisfy xi(t) ≥ z(0); • for every i /∈ NM (x(0)) the sequences xi(t) are constant. Moreover, max i x(k−1) − z(k−1) ≤(cid:0)1 − 1 (cid:1)(cid:0) max xi(0) − z(0)(cid:1) . k i The next equivalence will be crucial in the following. Lemma 6. Given n, k with 1 ≤ k ≤ n, consider x ∈ Rn and define the quanti- ties y := max i∈Nµ(x) xi and z := min i∈NM (x) xi . Then, z ≤ y for every x ∈ Rn if and only if n < 2k. Proof. We prove the equivalent claim that x ∈ Rn with z > y exists if and only if n ≥ 2k. Indeed, if n ≥ 2k consider the vector x ∈ Rn such that x1 ≤ x2 ≤ . . . ≤ xk < xk+1 ≤ . . . ≤ xn−k+1 ≤ . . . ≤ xn where n − k + 1 > k. The set Nµ(x) contains the k smallest elements of x so y = xk, while the set NM (x) contains the k largest elements of x, so z = xn−k+1 > xk = y. For the converse, assume that x with z > y exists, meaning (cid:0)maxi∈Nµ(x) xi (cid:1) <(cid:0)mini∈NM (x) xi (cid:1) . Both sets Nµ(x) and NM (x) contain k elements, so the sets {j : xj ≤ maxi∈Nµ(x) xi} and {j : xj ≥ mini∈NM (x) xi} contain at least k elements each. These two sets are disjoint, thus the vector x ∈ Rn has at least n ≥ 2k components. The next lemma describes a "shrinking sequence". Lemma 7. Given n, k with 1 ≤ k ≤ n and an initial configuration x0 ∈ Rn consider the dynamics (3) with (cid:40) µ(x(t)) M (x(t)) I(t) = for t ∈ {0, . . . , k − 2} for t ∈ {k − 1, . . . , 2k − 3} xi(T ) ≤(cid:0)1− 1 k (cid:1)(cid:0) max i xi(0) − min i xi(0)(cid:1) If n < 2k then xi(T ) − min i max i where T = 2k−2. Proof. For the sake of compactness, we set α(t) := min i xi(t) , β(t) := max i xi(t) , introduce the two sequences γ :=(cid:0)1 − 1 k (cid:1) , y(t) := max i∈Nµ(x(t)) xi(t) and z(t) := min i∈NM (x(t)) xi(t) , and set R = k−1. We have using Lemma 5 with initial configuration x(R). Then β(T ) − α(T ) = β(T ) − z(T ) + z(T ) − α(T ) ≤ γ(cid:0)β(R)−z(R)(cid:1) + z(R)−α(R) = γ(cid:0)β(R)−y(R)(cid:1) + γ(cid:0)y(R)−z(R)(cid:1) + z(R)−α(R) ≤ γ(cid:0)β(R)−y(R)(cid:1) +(cid:0)y(R)−z(R)(cid:1) + z(R)−α(R) = γ(cid:0)β(R)−y(R)(cid:1) + y(R)−α(R) ≤ γ(cid:0)β(0) − y(0)(cid:1) + γ(cid:0)y(0) − α(0)(cid:1) = γ(cid:0)β(0) − α(0)(cid:1) β(T ) − α(T ) ≤ γ(cid:0)β(0) − α(0)(cid:1). since γ < 1 and since y(R) − z(R) ≥ 0 if n < 2k by Lemma 6. Then using Lemma 3 and 4 with initial configuration x(0). We have finally obtained If n < 2k and the agent I(t) that updates his opinion at time t is chosen independently and uniformly at random over V , then process (3) converges almost surely to a consensus, from any initial configuration. The almost sure convergence is guaranteed because the finite sequence of updates introduced in the Lemma 7 appears infinitely often with probability one. This fact is proved in the following theorem, which provides the desired converge result. Theorem 8. Let n, k with 1 ≤ k ≤ n be given. Let {I(t), t ≥ 0} be a sequence of independent and uniformly distributed random variables over {1, . . . , n} and consider dynamics (3). If n < 2k, then lim t→∞ x(t) = 1c almost surely for any x0 ∈ Rn, with c ∈ [mini(x0 Proof. Let δ(t) = maxi xi(t) − mini xi(t) and observe that, for any x(0) = x0 and {I(t), t ≥ 0}, i ), maxi(x0 i )]. δ(0) ≥ 0 and 0 ≤ δ(t + 1) ≤ δ(t) for every t ≥ 0 , because the updates in the dynamics (3), based on model (1), involve convex combinations: the element with highest opinion cannot increase it and the el- ement with lowest opinion cannot decrease it. We introduce the sequence of events {At, t ≥ 2k−3} with At =(cid:8)I(s) = µ(x(s)) for s ∈ {t−2k+3, . . . , t−k+1} and I(s) = M (x(s)) for s ∈ {t−k+2, . . . , t}(cid:9) , i.e. the event At is the occurrence of the finite sequence introduced in Lemma 7 in the time window {t− (2k− 3), . . . , t}. In the same lemma we proved that, k ) δ(t−2k+3). Observe that given the occurrence of At, we have δ(t+1) ≤ (1 − 1 0 ≤ lim t→∞ δ(t) ≤ lim t→∞ (cid:1)nt δ(0) (cid:0)1 − 1 k where nt is the number of times At occurred up to time t. If P(At infinitely often) = 1 then nt → ∞ for t → ∞ and the rightmost limit above is zero almost surely. Hence, limt→∞ δ(t) almost surely, which implies the convergence to consensus. Moreover, c ∈ [mini(x0 i )] because every update in (3) is a convex combination of a subset of the current opinions. i ), maxi(x0 It remains to prove P(At infinitely often) = 1. The events of the sequence {At, t ≥ 2k−3} are not independent but the events in the subsequence {Ath , h ≥ 1} where th = h(2k − 2) − 1 are. Each of these events has probability thus(cid:80)∞ P(Ath ) = ∞. Hence, {At i.o.} ⊃ {Athi.o.}. From the second Borel- Cantelli lemma [18, Ch. 2, Thm 18.2] P(At infinitely often) ≥ P(Ath infinitely often) = 1 . h=1 (cid:18) 1 (cid:19)2k−2 n P(Ath ) = , The result continues to hold for dynamics where I(t) is not uniformly dis- tributed over {1, . . . , n}, as long as the probability to sample each agent is constant and positive. The proof has been based on exhibiting one suitable "shrinking sequence": however, it is clear that plenty of other sequences could do the job and actually play a role in inducing convergence of the dynamics. Therefore, the proof does not imply any good estimate of the convergence time. 6 Conclusion In this paper we have introduced a new model of opinion dynamics with opinion- dependent connectivity following the k-nearest-neighbors graph. The model is motivated by the rise of online social network services, where recommender systems select a certain number of news items to present to users, reducing the number of possible interactions to those which are closer to the user's presumed tastes. The resulting dynamics is substantially different from comparable models in the literature, such as bounded-confidence models. One key difference is the inherent lack of reciprocity of the interactions, which makes all convergence analysis challenging. Another key difference is the robustness of the formed clusters, whose opinions are hard to sway by external leader nodes. This feature makes control approaches based on leadership, like [19], unsuitable to k-nearest- neighbors dynamics. References [1] A. V. Proskurnikov and R. Tempo, "A tutorial on modeling and analysis of dynamic social networks. Part I," Annual Reviews in Control, vol. 43, pp. 65 -- 79, Mar. 2017. [2] A. Proskurnikov and R. Tempo, "A tutorial on modeling and analysis of dynamic social networks. Part II," Annual Reviews in Control, vol. 45, pp. 166 -- 190, 2018. [3] U. Krause, "A discrete nonlinear and non-autonomous model of consensus formation," Communications in Difference Equations, pp. 227 -- 236, 2000. [4] G. Deffuant, D. Neau, F. Amblard, and G. Weisbuch, "Mixing beliefs among interacting agents," Advances in Complex Systems, vol. 03, no. 01n04, pp. 87 -- 98, 2000. [5] A. Aydogdu, M. Caponigro, S. McQuade, B. Piccoli, N. Pouradier Duteil, F. Rossi, and E. Tr´elat, "Interaction network, state space, and control in social dynamics," in Active Particles, Volume 1: Advances in Theory, Models, and Applications, N. Bellomo, P. Degond, and E. Tadmor, Eds. Springer, 2017, pp. 99 -- 140. [6] M. Ballerini, N. Cabibbo, R. Candelier, A. Cavagna, E. Cisbani, I. Gi- ardina, V. Lecomte, A. Orlandi, G. Parisi, A. Procaccini, M. Viale, and V. Zdravkovic, "Interaction ruling animal collective behavior depends on topological rather than metric distance: Evidence from a field study," Pro- ceedings of the National Academy of Sciences, vol. 105, no. 4, pp. 1232 -- 1237, 2008. [7] I. Giardina, "Collective behavior in animal groups: Theoretical models and empirical studies," HFSP Journal, vol. 2, no. 4, pp. 205 -- 219, 2008, pMID: 19404431. [8] E. Cristiani, P. Frasca, and B. Piccoli, "Effects of anisotropic interactions on the structure of animal groups," Journal of Mathematical Biology, vol. 62, no. 4, pp. 569 -- 588, 2011. [9] A. Aydogdu, P. Frasca, C. D'Apice, R. Manzo, J. Thornton, B. Gachomo, T. Wilson, B. Cheung, U. Tariq, W. Saidel, and B. Piccoli, "Modeling birds on wires," Journal of Theoretical Biology, vol. 415, pp. 102 -- 112, 2017. [10] P. Balister, B. Bollob´as, A. Sarkar, and M. Walters, "Connectivity of ran- dom k-nearest-neighbour graphs," Advances in Applied Probability, vol. 37, no. 1, pp. 1 -- 24, 2005. [11] C. Chen, G. Chen, and L. Guo, "On the minimum number of neighbors needed for consensus of flocks," Control Theory and Technology, vol. 15, no. 4, pp. 327 -- 339, 2017. [12] V. D. Blondel, J. M. Hendrickx, and J. N. Tsitsiklis, "On Krause's multi- agent consensus model with state-dependent connectivity," IEEE Transac- tions on Automatic Control, vol. 54, no. 11, pp. 2586 -- 2597, 2009. [13] F. Fagnani and P. Frasca, Introduction to Averaging Dynamics over Net- works, ser. Lecture Notes in Control and Information Sciences. Springer Nature, 2017. [14] V. D. Blondel, J. M. Hendrickx, and J. N. Tsitsiklis, "Continuous-time average-preserving opinion dynamics with opinion-dependent communica- tions," SIAM Journal on Control and Optimization, vol. 48, no. 8, pp. 5214 -- 5240, 2010. [15] A. Mirtabatabaei and F. Bullo, "Opinion dynamics in heterogeneous net- works: Convergence conjectures and theorems," SIAM Journal on Control and Optimization, vol. 50, no. 5, pp. 2763 -- 2785, 2012. [16] C. Canuto, F. Fagnani, and P. Tilli, "An Eulerian approach to the analysis of Krause's consensus models," SIAM Journal on Control and Optimiza- tion, vol. 50, no. 1, pp. 243 -- 265, 2012. [17] F. Ceragioli and P. Frasca, "Continuous and discontinuous opinion dynam- ics with bounded confidence," NonLinear Analysis and its Applications B, vol. 13, no. 3, pp. 1239 -- 1251, 2012. [18] A. Gut, Probability: A Graduate Course, 2nd ed. Springer, 2013. [19] F. Dietrich, S. Martin, and M. Jungers, "Control via leadership of opinion dynamics with state and time-dependent interactions," IEEE Transactions on Automatic Control, vol. 63, no. 4, pp. 1200 -- 1207, 2018.
1807.05826	1	1807	2018-07-16T12:57:00	Multi-agents features on Android platforms	[ "cs.MA", "cs.CY" ]	The current paper shows the multi-agents capabilities to make a valid and flexible application when using a framework. Agent-based functions were used within JADE framework to make an Android messenger application with all requirements included. In the paper are described the architecture, the main functions and the databases integration of the user friendly agent-based application. There are included existing and possible multi-agents characteristics to provide integration with mobile platforms and storage challenges to improve the user experience through data mining.	cs.MA	cs	Multi-agents features on Android platforms Camelia-M. Pinteaa, Andreea Camelia Triponb, Anca Avrama, Gloria-Cerasela Crisanc aTechnical University Cluj Napoca, North University Center Baia Mare, Romania bRodeApps, Cluj-Napoca, Romania c"Vasile Alecsandri" University, Bacau, Romania 8 1 0 2 l u J 6 1 ] A M . s c [ 1 v 6 2 8 5 0 . 7 0 8 1 : v i X r a Abstract The current paper shows the multi-agents capabilities to make a valid and flexible application when using a framework. Agent-based functions were used within JADE framework to make an Android messenger application with all require- ments included. In the paper are described the architecture, the main functions and the databases integration of the user friendly agent-based application. There are included existing and possible multi-agents characteristics to provide integration with mobile platforms and storage challenges to improve the user experience through data mining. 1. Introduction Multi-agent systems are involved today for solving dif- ferent type of problems. They could be used in real-time applications and for solving complex problems in different domains as bio-informatics, ambient intelligence, semantic web [10, 16]. The main properties of agents in general, including the JADE [2] platform used here, are the auton- omy, reactivity, pro-activeness, cooperation, mobility and not at last learning capabilities. JADE (Java Agent Development Framework) is a Java- based framework used to specify the multi-agent systems. A JADE-based system can be distributed over many sys- tems and its configuration could be controlled for example by a user-friendly graphical application. The communication architecture of JADE offers an ef- ficient and flexible transmission of messages based on a pri- vate queue of messages under Agent Communication Lan- guage (ACL) format for each agent. The agents can easily identify the ACL messages received from other agents and have access to their queue of messages. The Android applications are nowadays some of the most used application, especially due to the free operating system features. Known to be frequently used, the messen- ger applications are real-time communication means be- tween people. The messenger could be installed and used through any existing Android devices as smartphones or tablets. The structure of the paper includes multi-agents with specific features for Android applications in the second section. In Section 3 the messenger Application based on JADE is introduced with the description of the structure Email addresses: [email protected] (Camelia-M. Pintea), [email protected] (Andreea Camelia Tripon), [email protected] (Anca Avram), [email protected] (Gloria-Cerasela Crisan ) and the main functionalities of the messenger application. Section 4 includes new approaches on agents features for Android applications including GPS/GIS characteristics. Section 5 illustrates the storage challenges and improving the user experience through data mining. The conclusion of the paper presents the new JADE-based messenger and possible multi-agent future improvements. 2. Multi-Agent Systems enhance the Android ap- plications features In several disciplines as Artificial Intelligence, in human- computer interface design and in object-based systems [10] the intelligent agents have been used. The agents are in- spired from real living "agents" as humans, insects or ani- mals capable to react to their own environment, with own objectives to reach and the autonomous capability to make the proper activities to achieve its goals [5, 6]. In general, it is assumed that an agent has the following properties [8, 16]: • The autonomy shows the ability of the agent to op- erate by itself without any other intervention. • The reactivity shows the ability of the agent to know the environment and react to the changes from its environment. • The pro-activeness shows the ability of the agent to have initiative and pursue its own goals. • The cooperation shows the capability of the agent to interact with others, agents or humans, through a specific communication language. • The learning ability of the agent is activated while the agent interacts with its environment. • The mobility shows the ability of an agent to move in a self-directed way around a network. Based on the particularities of a problem to be solved, the agents could be endowed with other features as for example rationality or sensitivity. The agents from the Multi-Agent System (MAS) are autonomous and hetero- geneous agents capable of interaction. In MAS the com- putation is asynchronous and has no global control [10]. In [10] is specified that negotiation, as a coordination process, is essential in MAS to solve conflicts. The com- munication [10] in MAS is a must due to exchanging infor- mation or other inter-operation tasks between agents. The communication process between agents requires the Agent Communication Language (ACL) and understanding the concepts exchanged by agents. The specific literature [3, 19, 18] presents many con- nections between the evolution and the organization of a MAS and the way an Android applications works and is coded. These strong connections generated in fact the JADE project, its worldwide success both in academia and in industry. Basically, the event-driven paradigm materialized by a generic Android application can be seen as the theoretical specifi- cation of the behaviour of MAS. In the following it is pre- sented such an application that couples the MAS paradigm with Android. The minimal installation and running requirements are: minimum JDK 5, Android Studio and Android SDK. Once the Main Container is started, the server starts too. The other agents will run from the same or different system and will connect with the Main Container. After that, the application is ready to function as a real-time messenger application. When an agent is connected to the Main Container, a new container will be created and will be visible; to run other agents we can install the messenger application on an Android phone, run the application and then connect to the Main Container. The Main Container is the container that starts at first and the other containers will login to it; the Main Con- tainer includes two special agents: • The Agent Management System (AMS) is the au- thority of the platform and it is the only agent that could manage the platform (starting or stopping agents or stopping the entire platform). • The Direction Facilities (DF) is a service through which the agents could publish the tasks offered and give the possibility to find other agents based on the task. 3. A new messenger application based on JADE An application based on JADE is made from a set of components called agents. The agents are uniquely iden- tified by their names. The agents execute tasks and in- teract by changing messages. The agents are kept "alive" through a platform offering services. A platform has one or many containers that could be run on systems with different hosts. Each container could have zero or many agents. Each platform has a special container called the Main Container with different agents (Figure 1). Figure 1: The main architecture of JADE framework [2]. 2 The communication between agents is made no matter if the agents are or not in the same container or if they belong or not to the same platform. The messages for- mat ACL is defined by Foundation for Intelligent Physical Agents (FIPA). An ACL message includes the sender, the receiver, the communication task and the content of the message. There are twenty two communication FIPA sets, each one well defined and having a well defined semantic. For example INFORM message: "INFORM 7 May 2016 rains", or REQUEST message requesting the receiver to make a task. 3.1. The messenger application Multi-agent system is used for the application, a JADE- based messenger on the Android operating system. A de- vice (phone/tablet) with Android operating system con- nects with the server address; after the connection is es- tablished, the agent could send and receive messages. The application considers the client-server principle: through the already existing connections, the clients ask the server to do some tasks like sending messages. When the server receives the message, the server sends it fur- ther to all the connected clients. The application has two modules: 1. The Android Messenger Client permits the messen- ger connection from an Android device; it includes a graphical interface and an agent managing the in- teractions with the other components. 2. The Messenger Server is the platform called Main Container including an agent called ChatManagerA- gent keeping the evidence of all agents connected to the messenger. The steps to connect to the Main Container are the fol- lowing: • The installation of application on the Android phone or on a simulator. • The IP address of the server is set from the menu and should be the same with the address of the server. • User must choose a name, which should be different from the name of the other messenger-users. • If there are no errors, the messenger-user could send messages to the others messenger-users. 3.2. The structure of the messenger The application has two parts. The first part makes the connection with the JADE agent and the second one is the graphical interface. They are included in the pack- ages agent and gui. Based on the Android architecture, the JADE agent is integrated in the project through jade- Android.jar library; it includes the following services to connect to the server: RuntimeService and MicroRuntime- Service. In this particular application it is used the MicroRun- timeService. The connectivity with the service is made by the class MainActivity through the following code: serviceConnection = new ServiceConnection(){ public void onServiceConnected(ComponentName name, IBinder service){ microRuntimeServiceBinder = (MicroRuntimeServiceBinder) }; public void onServiceDisconnected(ComponentName name){ microRuntimeServiceBinder = null; service; } }; bindService(new Intent(getApplicationContext(), MicroRuntimeService.class),serviceConnection, Context.BIND_AUTO_CREATE); Once connected to the service, we can go to the next step, which is to create the container and to start the agent. If it is a success, then the application can send and receive mes- sages. As a JADE-messenger novelty, there are included many actions further described including for example the groups of users, the messages with a single user, the mes- sages within a group, blocking/unblocking users, etc. void getAllUsers(); void getUsersNotInGroup(Group group); void getBlockedUsers(); void getGroupsConversations(); void getUsersFromGroup(Group group); void getMessagesFromUser(User user); void getMessagesFromGroup(Group group); void getUserConversations(); 3 void createGroup(Group group); void leaveGroup(Group group); void addToGroup(User user, Group group); void blockUser(User user); void unblockUser(User user); void register(User user); void login(User user); void sendMessageToGroup(Message message); void sendMessageToUser(Message message); void deleteMessage(Message message); ArrayList<User> getParticipantNames(); The first method is called when a message is sent from the phone and the second when the user wants to see the other messenger-users connected to the same server. These two actions are on the ChatClientAgent class, im- plementing the ChatClientInterface. This class deals with sending and receiving the messages further to/from the ChatActivity class in order to visualize the messages. In the ChatClientAgent class it is implemented a listener; when the server send a message, the message is shown through a broadcast receiver. 3.3. The main functions of the application The introduced application is a stand-alone messenger application for Android with the main functions of any Google Play Store application with a user-friendly design. The application starts from the main menu, after the user logged in or registered in the application. If some of the settings are not correct, the user could change the port or the host address. To set the connection the address where the JADE agent would be connected, the host and the port should be specified. Implicitly are used the last ones used for a connection. Logging into the system is beneficent to keep all the user information saved and the user could use different mobile phone to send/receive the messages, or see the older messages. The button used to show all the users will open a page with the status of all registered users. The current user could add any user of an existing group or block another user. If a user is blocked, he does not have the possibility to send messages to the one who blocked him. In the same context, you cannot see the status of an blocked user and cannot add him to a group. These could be possible after the user is unblocked (Figure 2). A new group is created easily using a particular but- ton and by specifying its name. The name is verified to be unique in the database and an appropriate message is shown. The new group is automatically added to the other existing groups. Nevertheless, the only ones who can add other members to a group are the members of the groups. When selecting a conversation with a friend all the sent and received messages from that user will be shown. On the toolbar there are the user-name, a button to add the friend to the group, a button to add the user in the block- ing list and a button to delete all the conversation. In the left side of the window there are the received messages • Flexibility: Android is a multi-channel, multi-carrier, freely distributed OS, based on Java, and has a mar- ket share of 82.5% in 2015 [17]. The implicit intent object is an example of how the Android operating system adapts to the environ- ment [18]. Likewise, current complex social problems are modeled by MAS's that need to perform well in open environments. For example, the MIT Robust Open Multi-Agent Systems (ROMA) Research Group is dedicated to "learning how we can develop multi-agent systems for open contexts where the constituent agents can come from anywhere, may be buggy or even mali- cious, and must run in the dynamic and potentially failure-prone environments at hand" [19]. The agents involved in the JADE messenger applica- tion have limited features. Here we introduce several new features to be included in the JADE framework. The Multi-agent system of JADE could include agents with different levels of sensitivity related to their environ- ment. Based on the agents sensitivity, the messenger ap- plication could include other facilities as: • the messenger could start when the user is closer to a already set point; for example one could set the messenger application to start when approaching to a building, let's say a museum, when the application starts automatically based on the GPS/GIS; so, the agent is sensitive to each "museum" encountered to exchange impressions with the other users. • the messenger could include a blocking option based on the agent sensitivity: an user is automatically blocked or unblocked based on its behavior: – if the words used by another user are "bad", so they are in a database, locally or cloud-based, the user is automatically blocked and an appro- priate message is sent to the other user; – an user could be unblocked if it is identified a good behavior on its personal social platform; for example on its social website there are in- cluded voluntary actions, promoting "good words" and facts. This feature could be added or not by the user. • another improvement could be considered the auto- matically connection to a close or extended group when an earthquake is ongoing based on a specific sensor of the device or by an earthquake alert. Sim- ilar features could be included for floods or other related crisis. The GPS/GIS features could be used as in [11, 12, 13, 14]. In [12] is shown an implementation of an intelligent, multi-modal, multi-user geographic in- formation environment (GCCM Connect) used on 4 Figure 2: Examples of the messenger-JADE features: ing/unblocking users (left) and managing groups of users (right). block- and in the right side the sent messages. All the messages include the time and date when they were sent. At the bot- tom of the window are the buttons for adding and sending new messages. To delete a message, one can use a long touch on that message and the message could be deleted after a confirmation. For the conversations in a group there is a similar func- tionality; the differences are on the toolbar: adding a new member to the group, visualization of all the group mem- bers and their status, the button to leave the group. All the messages received from any member of the group will be placed in the left side of the window and the sent ones in the right side of the window. 4. Theoretic approach on new agents features for Android applications Several common features for Multi Agent Systems and an Android application are: • Concurrency: the Android components are activated by intents (decentralized events trigger an agent be- havior or the Yellow Pages service provided by the DF special JADE agent [2]). The agents from MAS act simultaneously, in a com- mon environment, pursuing their individual goals, but possibly ready to draw coalitions or to co-operate. • Loosed coupling: like agents in MAS, the Android application target components are activated by ex- plicit or implicit intents [3]. • Asynchronous communication: both for the agents in MAs and for the Android application components, the self-decided conversation initiation is a manifes- tation of the individual, autonomous behavior. a spatial decision-making contexts. Collaboration and share knowledge is a must especially when is in- volved a critical problem as for example earthquakes or floods. Crisis management demands information technology and individuals /organizations to share information and expertise on decision-making. • An idea of improvement in terms of functionality would be to have the existing list of named servers available and user to choose from the server list in- stead of having to configure the server himself the host and port. When blocking another user, one could invoke a reason for doing that if he/she wants. There could be a list of predefined reasons or user can invoke a new one. Later on, analysis of data could show the potential users that have malicious intents. Auto-complete features could be provided based on analyzing the most used phrases and offer edit support to the users using the application. All the new and the common characteristics generate tight connections, allowing the researchers to design and to pro- vide useful integration of the mobile platform with Multi Agent System features. 5. Storage challenges and improving the user ex- perience through data mining For the messenger application presented, there was the need of having information stored in a database. The infor- mation that is stored consists of details related to users, lo- gin information, friends of a user, blocked/unblocked users; messenger groups; messages exchanged between users; agent behaviors. As soon as the application is used by more and more users, the volume of the data increases and a proper database maintenance plan needs to be in place. That means there will be a need for performing actions like: • Cleaning unnecessary data - logs of the status of ac- tions performed by the agents for example, are only needed for a short amount of time (in case of an error for example, these logs could provide a meaningful support for finding the cause of the problem). • Archiving the old data. A proper system will make sure that data that is not in use anymore is handled (for example messages exchanged in groups that no longer contain active users). Archiving could mean storing the plain data to a different server (less per- forming, less expensive). Another approach would be obtaining a more compact version of that data and storing it in a different format. • Having an indexing strategy that is reviewed period- ically. Indexes are data structures that improve the speed of data retrieval operations on a table. Having proper indexes will mean a faster time to obtain the 5 results, but reviewing the indexes periodically is a must, because the increase of data could mean that an index that was performing at some point might no longer be efficient. The database could also provide the means for obtain- ing relevant information related to the behavior of the ap- plication and agents by performing data mining. Currently data mining is widely recognized as the process of discov- ering relevant patterns in large sets of data, patterns that can be later used. To have these patterns discovered, intel- ligent methods are applied. Data mining can be applied to any kind of data, as long as it delivers meaningful results to a target application [20]. In the context of the JADE messenger application, data mining could be applied on the stored data related to the users to obtain relevant information on the user behavior. For example, discovering the use patterns for a specific functionality could give an idea on what type of users are using that functionality, what are the topics that are the most tackled, what is the geographical distribution of the users, whether the application is more widely used in the urban or rural area. This type of knowledge could lead fur- ther on to developing new functions that are targeted to improving user experience; remove/make less visible func- tions that do not come with an adequate return of invest- ment. Furthermore, in [21] the concept of agent mining refers to the application of autonomous intelligent agents in the field of data mining to support and improve the knowledge discovery and decision-making process. Because agents have autonomy, flexibility, mobility, adaptability and have a rational nature, they prove to be a perfect choice for parallel, multisource, distributed mining. In the context of the JADE framework, that could mean integrating agents responsible with mining the data. 6. Conclusions and future work The paper introduces a multi-agent-based messenger application using JADE framework. JADE has several components successfully used for the current messenger application as the Agent Communication Language (ACL) to send/receive the messages. The user-friendly applica- tion includes several features as using groups of users or blocking/unblocking users on any Android devices. The application could be improved with for example audio/video messages and especially with other multi-agents features, as sensitivity based on GPS/GIS location. Acknowledgments. The study was conducted under the auspices of the IEEE-CIS Interdisciplinary Emergent Technologies TF. We thank Alexandru Pintea for carrying out Latex editing tasks. References [1] Hitchcock, F. L., The distribution of a product from several sources to numerous localities. J Math Physic, 20:224–230, 1941 [2] JADE Framework http://jade.tilab.com [3] Poslad, S.: Specifying protocols for multi-agent systems interac- tion. ACM Trans. Auton. Adapt. Syst. 2(4), Article 15Novem- ber 2007. [4] Camazine, S. et al.: Self organization in biological systems. Princeton: Princeton University Press, 2001. [5] Chira, C., Pintea, C.M., Crisan, G.C., Dumitrescu, D.: Solving the linear ordering problem using ant models. GECCO 2009, 1803–1804, 2009. [6] Pintea, C.M., Crisan, G.C., Chira, C: Hybrid ant models with a transition policy for solving a complex problem. Logic Journal of IGPL, 20 (3):560–569, 2012. [7] Finin, T., Labrou, Y., Mayfield, J.: KQML as an Agent Com- munication Language, Software Agents, B.M. Jeffrey (ed.), 1997. [8] Franklin, S., Graesser, A.: Is It an Agent, or Just a Program?: A Taxonomy for Autonomous Agents. Proceedings of the Third International Workshop on Agent Theories, Architectures and Languages, 1996. [9] Michener, C.D.: The social behavior of bees: A comparative study. Cambdridge. Harvard University Press, 1974. [10] Jennings, N.R., Sycara, K.P., Wooldridge, M.: A Roadmap of Agent Research and Development. J. Autonomous Agents and Multi-Agent Systems. 1(1):7–36, 1998. [11] Moratis, P., Petraki, E., Spanoudakis, N.I.: Providing ad- vanced, personalised infomobility services using agent technol- ogy. Applications and innovations in intelligent systems XI, 35– 48, 2004. [13] Lei, Ye, Hui, The Web Integration of [12] Cai, Guoray: Extending distributed GIS to support geo- collaborative crisis management. Geographic information sci- ences, 2005, 11(1):4–14. Lin: the GPS+GPRS+GIS Tracking System and Real-Time Moni- toring System Based on MAS, Web and Wireless Geographical Information Systems: W2GIS 2006, Hong Kong 2006. Proceed- ings, 2006, 54–65. [14] Crisan, G.C., Pintea, C-M., Palade, V.: Emergency Man- agement Using Geographic Information Systems. Application to the first Romanian Traveling Salesman Problem Instance. Knowledge and Information Systems. 1–21, 2016. [15] Warneke, B. Last, M. Liebowitz, B. Pister, K.S.J. : Smart Dust: communicating with a cubic-millimeter computer, Computer (2001), 34, 44–51. [16] Wooldridge, M.: Intelligent Agents, An Introduction to Multi- agent Systems. In G. Weiss (ed.) MIT Press (1999). [17] Smartphone market share, http://www.idc.com/prodserv/ smartphone-os-market-share.jsp. [18] Android Developer, https://developer.android.com/guide/ components/intents-filters.html. [19] MIT Robust Open Multi-Agent Systems, http://ccs.mit.edu/ roma/. [20] Han, J., Kamber, M., Pei, J.: Data Mining: Concepts and Techniques, The Morgan Kaufmann Series, 2011. [21] Cao, C., Gorodetsky, V., Mitkas, P.M.: Agent Mining: The Synergy of Agents and Data Mining, 2009. 6
1707.09183	2	1707	2019-03-11T20:17:29	A Survey of Learning in Multiagent Environments: Dealing with Non-Stationarity	[ "cs.MA", "cs.LG" ]	The key challenge in multiagent learning is learning a best response to the behaviour of other agents, which may be non-stationary: if the other agents adapt their strategy as well, the learning target moves. Disparate streams of research have approached non-stationarity from several angles, which make a variety of implicit assumptions that make it hard to keep an overview of the state of the art and to validate the innovation and significance of new works. This survey presents a coherent overview of work that addresses opponent-induced non-stationarity with tools from game theory, reinforcement learning and multi-armed bandits. Further, we reflect on the principle approaches how algorithms model and cope with this non-stationarity, arriving at a new framework and five categories (in increasing order of sophistication): ignore, forget, respond to target models, learn models, and theory of mind. A wide range of state-of-the-art algorithms is classified into a taxonomy, using these categories and key characteristics of the environment (e.g., observability) and adaptation behaviour of the opponents (e.g., smooth, abrupt). To clarify even further we present illustrative variations of one domain, contrasting the strengths and limitations of each category. Finally, we discuss in which environments the different approaches yield most merit, and point to promising avenues of future research.	cs.MA	cs	A Survey of Learning in Multiagent Environments A Survey of Learning in Multiagent Environments: Dealing with Non-Stationarity Pablo Hernandez-Leal Michael Kaisers Tim Baarslag Intelligent and Autonomous Systems Group Centrum Wiskunde & Informatica Amsterdam, The Netherlands [email protected] [email protected] [email protected] Enrique Munoz de Cote Instituto Nacional de Astrof´ısica, ´Optica y Electr´onica, Puebla, M´exico PROWLER.io Ltd., Cambridge, United Kingdom [email protected] Editor: Abstract The key challenge in multiagent learning is learning a best response to the behaviour of other agents, which may be non-stationary: if the other agents adapt their strategy as well, the learning target moves. Disparate streams of research have approached non- stationarity from several angles, which make a variety of implicit assumptions that make it hard to keep an overview of the state of the art and to validate the innovation and significance of new works. This survey presents a coherent overview of work that addresses opponent-induced non-stationarity with tools from game theory, reinforcement learning and multi-armed bandits. Further, we reflect on the principle approaches how algorithms model and cope with this non-stationarity, arriving at a new framework and five categories (in increasing order of sophistication): ignore, forget, respond to target models, learn models, and theory of mind. A wide range of state-of-the-art algorithms is classified into a taxonomy, using these categories and key characteristics of the environment (e.g., observability) and adaptation behaviour of the opponents (e.g., smooth, abrupt). To clarify even further we present illustrative variations of one domain, contrasting the strengths and limitations of each category. Finally, we discuss in which environments the different approaches yield most merit, and point to promising avenues of future research. Keywords: Multiagent learning, reinforcement learning, multi-armed bandits, game theory 1. Introduction There are many successful applications of multiagent systems (MAS) in the real world. Examples are ubiquitous in energy applications, for example, to implement a network to distribute electricity (Pipattanasomporn et al., 2009) or to coordinate the charging of elec- tric vehicles (Valogianni et al., 2015), in security, to patrol the Los Angeles airport (Pita et al., 2009) and in disaster management to assign a set of resources to tasks (Ramchurn et al., 2010). Multiagent systems include a set of autonomous entities (agents) that share 1 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote a common environment and where each agent can independently perceive the environment, act according to its individual objectives and as a consequence, modify the environment. How the environment changes as a consequence of an agent exerting an action is known as the environment dynamics. In order to act optimally with respect to its objectives these dynamics need to either be known (a priori) by the agent or otherwise be learned by experience, i.e., by interacting many times with the environment. Once the environment dynamics have been learned, the agent can then adapt its behaviour and act according to its target objective. We know a lot about the single agent case, where only one agent is learning and adapting its behaviour, but most of the results break apart when two or more agents share an environment and they all learn and adapt their behaviour concurrently. The problem with this concurrency is that the action executed by one agent affects the goals and objectives of the rest, and vice-versa. To tackle this, each agent will need to account for how the other agents are behaving and adapt according to the joint behaviour. Needless to say, this joint behaviour needs to be learned by each agent, and due to the fact that all agents are performing the same operations of learning and adapting concurrently, the joint behaviour -- and therefore the environment -- is perceived by each agent as non- stationary. This non-stationarity (sometimes referred to as the moving target problem, see Tuyls and Weiss, 2012) sets multiagent learning apart from single-agent learning, for which it suffices to converge to a fixed optimal strategy. Most learning algorithms to date are not well suited to deal with non-stationary en- vironments,1 and usually, such non-stationarity is caused by changes in the behaviour of the participating agents. For example, a charging vehicle in the smart grid might change its behavioural pattern (Marinescu et al., 2015); robot soccer teams may change between pre-defined behaviours depending on the situation (MacAlpine et al., 2012); and attack- ers change their behaviours to keep security guards guessing in domains involving frequent adversary interactions, such as wildlife and fishery protection (Fang et al., 2015). Previous works in reinforcement learning (RL), MAS and multi-armed bandits (to name a few) have all acknowledged the fact that specialized targeted work is needed that explic- itly addresses non-stationary environments (Sutton et al., 2007; Panait and Luke, 2005; Garivier and Moulines, 2011; Matignon et al., 2012; Lakkaraju et al., 2017). Against this background, this survey fills this gap with an extensive analysis of the state of the art. Previous surveys have proposed different ways to categorise MAS algorithms (Panait and Luke, 2005; Shoham et al., 2007; Busoniu et al., 2010), others have divided them by the type of learning (Tuyls and Weiss, 2012; Bloembergen et al., 2015) and another group have proposed properties that MAS algorithms should have (Bowling and Veloso, 2002; Powers et al., 2007; Crandall and Goodrich, 2011). In contrast, we propose another view, which has been mostly neglected, focused on how algorithms deal with non-stationarity, providing an illustrative categorization with increasing order of sophistication where each algorithm is analysed along with related characteristics (observability and opponent adaptation). The questions addressed by the surveyed algorithms are illustrated by the following simple scenario comprising two agents: Predator. The agent under our control. 1. Environments, in which all counterpart agents are perceived as part of the environment. 2 A Survey of Learning in Multiagent Environments Prey. The opponent agent.2 Both agents engage in repeated rounds of interactions (possibly infinite), there is no com- munication and the rewards received depend on the joint action. The prey has several (possibly infinite) strategies at its disposal (ways to select its actions) and it can change from one to another during the interaction. In this context we can raise several questions: • Should the predator assume the prey will behave in a certain way (e.g., minimizing the predator's reward, enacting their part of the Nash Equilibrium or playing as a teammate)? • Should the predator learn to optimise against a single opponent strategy or should it generalise by learning a more robust strategy against a class of possible strategies? • Should the predator assume the prey is modelling the predator's strategy? • Should the predator assume the prey will use a stationary strategy? If not, will the prey change its behaviour slowly or drastically? Different research communities make different assumptions that give rise to distinct answers to these questions. While there is some awareness within each community of the work outside that community, it remains a challenge to keep up to date with the recent literature due to this fragmentation, which impedes AI research in its entirety (Eaton et al., 2016). For example, many game theory algorithms focus on finding equilibria in self-play. Multi- armed bandits either assume a stochastic or adversarial setting and try to optimize against that behaviour. Some basic approaches of reinforcement learning ignore other agents and optimise a policy assuming a stationary environment, essentially treating non-stationary aspects like stochastic fluctuations. Other approaches learn a model of the other agents to predict their actions to remove the non-stationary behaviour. Finally, algorithms from behavioural game theory and planning have proposed recursive modelling approaches that assume opponents are capable of performing strategic reasoning and modelling of the rest of the agents. In this context, the main contributions of this survey are the following: • Provide a coherent view of how state-of-the-art algorithms in reinforcement learning, multi-armed bandits and game theory tackle the planning problem of long-term sum of expected rewards in non-stationary environments. • Propose a new framework for multiagent systems (see Section 3.2). This framework allows to describe a categorisation with increasing order of sophistication with respect to how non-stationarity is handled, arriving at five categories: ignore, forget, respond to target opponents, learn opponent models and theory of mind (see Section 3.3). • Describe the fundamental algorithms of each category using an illustrative example highlighting their strengths and limitations (see Section 4). 2. In this work we use the word "opponent" when referring to another agent in the environment irrespective of the domain, and irrespective of its adversarial or cooperative nature. 3 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote • Categorise most significant learning algorithms while also describing their main char- acteristics with respect to the environment and opponent assumptions (see Section 5). • Provide a structured set of open questions with promising avenues of future research in multiagent learning (see Section 6.5). With its tailored scope, this survey aims to establish a structure to think clearly about all the assumptions, characteristics and concepts related to the challenge of addressing non-stationarity in multiagent learning. 1.1 Related work and demarcation Multiagent learning has received a lot of attention in the past years and some previous sur- veys have emerged with different motivations and outcomes. Shoham et al. (2007) presented a general survey of multiagent learning providing some interesting foundational questions and identifying five different agendas in this research community. Tuyls and Weiss (2012) presented a bird's eye view about the AI problem of multiagent learning, identifying the milestones achieved by the community and mentioning the open challenges at the time.3 Panait and Luke (2005) presented an extensive analysis of cooperative multiagent learn- ing algorithms, dividing them into two categories: single learner in a multiagent problem (team learning) and multiple learners (concurrent learning). Matignon et al. (2012) fo- cused on the evaluation of independent RL algorithms on cooperative stochastic games. Busoniu et al. (2010) presented a thorough survey on multiagent RL where they identi- fied a taxonomy and several properties for algorithms in multiagent reinforcement learning (MARL). Crandall and Goodrich (2011) assessed the state of the art in two-player repeated games with respect to three properties: security, cooperation and compromise, which they propose as important to act in a variety of different games. Muller and Fischer (2014) presented an application-oriented survey, highlighting applications that use or are based on MAS. Weiss (2013) edited a book about multiagent systems; in particular there is a chapter dedicated to multiagent learning where they present state-of-the-art algorithms dividing them into joint action, gradient, Nash and other learners (see Weiss, 2013, chap. 10). A recent survey analysed methods from evolutionary game theory and its relation with mul- tiagent learning (Bloembergen et al., 2015). Finally, the recent area of multiagent deep reinforcement learning gained a lot of interest with two recent surveys (Nguyen et al., 2018; Hernandez-Leal et al., 2018). None of these survey articles provide an explicit treatment of the non-stationarity approaches taken in various algorithms. Our survey exceeds previous work in scope of different domains and coverage measured by number of algorithms, and fills the gap of reflecting on non-stationarity. In contrast to previous works, we provide a detailed analysis of algorithms from multi-armed bandits (for stochastic and adversarial environments), single agent RL (model-based and model-free approaches), multiagent RL and game theory (mainly for repeated and stochastic games) in both competitive and cooperative scenarios. We provide a full taxonomy of how algorithms cope with non-stationarity, and describe opponent and environment characteristics. 3. We reflect on the relation between those challenges and the promising avenues of future research in Section 6.5. 4 A Survey of Learning in Multiagent Environments This survey does not cover work related to learning in dynamic environments that do not have other active autonomous and automated agents, such as recommender systems of news articles (Liu et al., 2010) or online supervised learning scenarios (see Section 6.4). 1.2 How to read this survey Since different audiences are expected to read this survey, this section provides forward references to key insights and sections for different target groups: • For researchers seeking an introduction to multiagent learning we propose to follow the current structure of the paper sequentially, progressing through each section in order. • For experienced researchers we recommend starting with the new framework pro- posed in Section 3.2, followed by the high level vision of the categorisation of algo- rithms depicted in Figure 4 in Section 4; to be followed by the extensive categorization in Section 5; in particular given in Table 2 and Figure 5. • We encourage researchers seeking guidance on promising research directions to consult the discussion in Section 6, in particular to find common types of results in Section 6.3 and interesting open problems in Section 6.5. Finally, we encourage all readers to position their future work in this framework, as delin- eated in Section 3, for ease of reference and navigation of related (future) work. 1.3 Paper overview This paper aims to provide a general overview of how different algorithms cope with the problem of learning in multiagent systems where it is necessary to deal with non-stationary behaviour. In Section 2, we review formal models used in this context ; in particular we review multi-armed bandits, reinforcement learning and game theory. Section 3 describes the main challenge of non-stationarity in multiagent systems together with a new framework that naturally models its key elements, and lastly presents the proposed categorization of how algorithms deal with non-stationarity. Section 4 illustrate the categories using a simple scenario. Section 5 presents an extensive list of works of multi-armed bandits, RL and game theory categorised by the taxonomy proposed in this survey. Section 6 provides a discussion about the strengths and limitations of each category, describes the common experimental settings, presents a summary of the theoretical results and pinpoints interesting open problems highlighting promising lines of future research. Finally, Section 7 summarizes the conclusions and contributions of this survey. 2. Formal approaches from different domains that model non-stationarity This section describes the formal models used in multi-armed bandits, reinforcement learn- ing, and game theory, and contrasts how they capture non-stationarity. Each domain makes different assumptions about a priori information about the interaction, as well as about online observability of the environment and opponents during the interaction. This dis- crimination forms the basis of the environment characteristics in the next section. In line 5 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote with available information, the solution concept for finding good behaviour may be charac- terized correspondingly by more a priori or more online reasoning, exhibiting characteristic behaviour and the ability to cope with certain types of non-stationarity in opponent be- haviour. In order to present our behavioural categorization, we here present a synopsis of the different approaches from the literature. Note that different areas provide different terminology. Therefore, we will use the terms player and agent interchangeably; similarly for reward and payoff; and for rounds and steps. Finally, we will refer to other agents in the environment as opponents irrespective of the domain's or agent's cooperative or adversarial nature. 2.1 Multi-armed bandits The simplest possible reinforcement-learning problem is known as the multi-armed bandit problem (Robbins, 1985): the agent is in a room with multiple gambling machines (called "one-armed bandits"). At each time-step the agent pulls the arm of one of the machines and receives a reward. The agent is permitted a fixed number of pulls. The agent's purpose is to maximise its total reward over a sequence of trials. Usually each arm is assumed to have a different distribution of rewards, therefore, the goal is to find the arm with the best expected return as early as possible, and then to keep gambling using that arm. A K−armed (stochastic) bandit can be formalised as a set of real distributions B = {R1, R2, . . . , Rk}, with the set of arms I = {1, . . . , K}, such that each arm yields a stochastic reward ri following the distribution Ri. Let µ1, µ2, . . . , µk be the mean values associated with these reward distributions. A policy, or allocation strategy, is an algorithm that chooses the next machine to play based on the sequence of past plays and obtained rewards. The policy selects one arm at each round and observes the reward, this process is repeated for T rounds. This problem illustrates the fundamental trade-off between exploration and exploitation: should the agent choose the arm with the highest average reward observed so far (exploit), or should it choose another one for which it has less information, so as to find out if it in fact exceeds the first one (explore)? The regret ∆R is a common measure used to evaluate different algorithms in multi- armed bandits. The regret is the difference (necessarily a loss) between the chosen policy π and the optimal policy π∗. In the multi-armed bandit setting the optimal policy would choose the arm i∗ with the highest expected reward at all times, i.e., ∀t : π∗ t = i∗, while πt = i(t) may vary over time. For an episode of T steps, the stochastic regret yields T(cid:88) ri∗ − T(cid:88) ∆R = ri(t). t=1 t=1 With this concept in mind, some approaches guarantee low regret under certain conditions. These policies work by associating a quantity called upper confidence index to each arm. This index relies on the sequence of rewards obtained so far from a given arm and is used by the policy as an estimate for the corresponding reward expectation (Auer et al., 2002a). The UCB1 (Upper Confidence Bounds) algorithm achieves logarithmic regret assuming bounded rewards, without further constraints on the reward distributions (Auer et al., 2002a). UCB uses the principle of optimism in the face of uncertainty to select its actions, i.e., the algorithm selects arms by an optimistic estimate on the expected rewards of certain 6 A Survey of Learning in Multiagent Environments arms to balance exploration and exploitation. The UCB1 algorithm is extremely simple, it initially plays each arm once, and subsequently selects the arm i(t): (cid:32) (cid:115) (cid:33) i(t) = arg max j ¯rj + 2 ln t nj where ¯rj is the average reward obtained from arm j, nj is the number of times arm j has been played, and t is the total number of rounds so far. Algorithm 1: Multi-armed bandit: stochastic (s) or adversarial (a) (Bubeck and Slivkins, 2012) Input: K arms, T rounds (T ≥ K ≥ 2). for t=1, . . . , T do Algorithm chooses one arm i(t) ∈ {1, . . . , K} Adversary selects rewards gt = (g1,t, . . . , gK,t) ∈ [0, 1]K Stochastic environment produces reward gi,t ∼ Ri (drawn independently) 1 2-a 2-s 3 The goal is to minimise the regret, defined by : Receive reward gi(t),t (does not observe the other arms) In the adversarial model, ∆r = maxj∈{1,...,K}(cid:80)T In the stochastic model, ∆r =(cid:80)T t=1(maxj∈{1,...,K}µj − µi(t)) t=1 gj,t −(cid:80)T t=1 gi(t),t The stochastic bandit scenario is useful to model decision-making in stationary but stochastic settings. A direct extension of this setting is the adversarial model, which as- sumes that rewards of each arm are controlled by an adversary, i.e., the reward distribution associated with each arm at every round is fixed in advance by an adversary before the game starts (Auer et al., 2002b); see Algorithm 1 that juxtaposes both scenarios. However, when relaxing the assumptions made in the problem definition even more by assuming online adaptive adversaries, the standard definition of regret is no longer adequate (due to adap- tivity of the adversary, the optimal action might change at different steps, see Arora et al., 2012). Because of that, different variations of the regret measure have been proposed (Arora et al., 2012; Crandall, 2014). It is worth mentioning that there are further extensions to the bandit scenario (Pandey et al., 2007; Beygelzimer et al., 2011; Tran-Thanh et al., 2012), which are beyond the scope of this survey. However, we refer the interested reader to the discussion in related work (Bubeck and Cesa-Bianchi, 2012). 2.2 Reinforcement Learning Reinforcement learning (RL) is one important area of machine learning that formalises the interaction of an agent with its environment (Puterman, 1994). A Markov Decision Process (MDP) can be seen as a model of an agent interacting with the world (see Figure 1), where the agent takes the state s of the world as input and generates an action a as output that affects the world. There is a transition function T that describes how an action affects the environment in a given state. The component Z represents the agent's perception function, which is used to obtain an observation z from the state s. In an MDP it is assumed there is no uncertainty in where the agent is. This implies that the agent has full and perfect 7 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote Figure 1: The agent interacts with the environment, performing an action a that affects the state environment according to a function T , producing the state s. The agent perceives an observation z about the environment (given by a function Z) and obtains a reward r (given by a function R). Figure 2: A Markov decision process (MDP) with four states S0, S1, S2, S3 and two actions a1, a2. The arrows denote the tuple: action, transition probability and reward. perception capabilities and knows the true state of the environment (what it perceives is the actual state, z = s). The component R is the reward function, the rewards give an indication of the quality of which actions the agent needs to choose. However, the reward function is not always simple to define (for example, it may be stochastic or delayed). Formally, Definition 1 (Markov decision process) An MDP is defined by the tuple (cid:104)S, A, R, T(cid:105) where S represent the world divided up into a finite set of possible states. A represents a finite set of available actions. The transition function T : S × A → ∆(S) maps each state-action pair to a probability distribution over the possible successor states, where ∆(S) denotes the set of all probability distributions over S. Thus, for each s, s(cid:48) ∈ S and a ∈ A, the function T determines the probability of a transition from state s to state s(cid:48) after executing action a. The reward function R : S × A × S → R defines the immediate and possibly stochastic reward that an agent would receive for being in state s, executing action a and transitioning to state s(cid:48). An example of an MDP with 4 states and 2 actions is depicted in Figure 2, where ovals represent states of the environment. Each arrow has a triplet an, p, r representing the action, the transition probability and the reward, respectively. 8 TZRazrsAgentEnvironmentS2S0a1,1,10S3S1a2,1,5a2,0.2,1a1,1,-100a2,1,5a2,0.8,1a1,1,5A Survey of Learning in Multiagent Environments The key assumption in defining MDPs is that they are stationary, i.e., particularly the transition probabilities and the reward distributions do not change in time.4 MDPs are adequate models to obtain optimal decisions in environments with a single agent. Solving an MDP will yield a policy π : S → A, which is a mapping from states to actions. An optimal policy π∗ is the one that maximises the expected reward. There are different techniques for solving MDPs assuming a complete description of all its elements. One of the most common techniques is the value iteration algorithm (Bellman, 1957) which is based on the Bellman equation: (cid:88) a∈A (cid:88) s(cid:48)∈S V π(s) = π(s, a) T (s, a, s(cid:48))[R(s, a, s(cid:48)) + γV π(s(cid:48))], (1) with γ ∈ [0, 1]. This equation expresses the value of a state which can be used to obtain the optimal policy π∗ = arg maxπ V π(s), i.e., the one that maximises that value function, and the optimal value function V ∗(s). V ∗(s) = max π V π(s) ∀s ∈ S. Finding the optimal policy for an MDP using value iteration requires the MDP to be fully known, including a complete and accurate representation of states, actions, rewards and transitions. However, this may be difficult if not impossible to obtain in many domains. For this reason, RL algorithms have been devised that learn the optimal policy from experience and without having a complete description of the MDP a priori. An RL agent interacts with the environment in discrete time-steps. At each time, the agent chooses an action from the set of actions available, which is subsequently executed in the environment. The environment moves to a new state and the reward associated with the transition is emitted (see Figure 1). The goal of a RL agent is to maximise the expected reward. In this type of learning the learner is not told which actions to take, but instead must discover which actions yield the best reward by trial and error. Q-learning (Watkins, 1989) is one well known value-based algorithm for RL. It has been devised for stationary, single-agent, fully observable environments with discrete actions. In its general form, a Q-learning agent can be in any state s ∈ S and can choose an action a ∈ A. It keeps a data structure Q(s, a) that represents the estimate of its expected payoff starting in state s, taking action a. Each entry Q(s, a) is an estimate of the corresponding optimal Q∗ function that maps state-action pairs to the discounted sum of future rewards when starting with the given action and following the optimal policy thereafter. Each time the agent makes a transition from a state s to a state s(cid:48) via action a receiving payoff r, the Q table is updated as follows: Q(s, a) = Q(s, a) + α[(r + γ max (2) with the learning rate α and the discount factor γ ∈ [0, 1] being parameters of the algo- rithm, with α typically decreasing over the course of many iterations. Q-learning is proved b Q(s(cid:48), b)) − Q(s, a)] 4. Formally, an MDP assumes S, A, R and T to be stationary. These sets and function must be unchanged over time, albeit that is compatible with making the action set A stochastic or dependent on the state. 9 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote to converge towards Q∗ if each state-action pair is visited infinitely often under specific pa- rameters (Watkins, 1989). Q-learning is said to be an off-policy method since it estimates the sum of discounted rewards of the optimal policy (aka. target policy) while actually exe- cuting an exploration policy (aka. behavior policy) distinct from it.5 In contrast, on-policy methods refer to algorithms that estimation the value of the executed (exploration) pol- icy. Since the exploration policy is commonly non-stationary, primarily due to the decrease of exploration parameters over time, the target value Q∗ to approximate changes with it, making it more intricate to provide convergence results.6 One classic on-policy algorithm is SARSA (statet, actiont, rewardt, statet+1, actiont+1) (Sutton and Barto, 1998) which uses a variation of Equation (2): Q(s, a) = Q(s, a) + α[(r + γ Q(s(cid:48), a(cid:48))) − Q(s, a)]. (3) By using Q-learning it is possible to learn an optimal policy without knowing T or R beforehand, and even without learning these functions (Littman, 1996). For this reason, this type of learning is known as model free RL. In contrast, model-based RL aims to learn a model of its environment, specifically approximating T and R. Such models are then used by the agent to predict the consequences of actions before they are taken, facilitating planning ahead of time. One example of this type of algorithms is Dyna-Q (Sutton and Barto, 1998). Exploration vs exploitation Similar to multi-armed bandits, in RL one main concern is to develop algorithms that balance exploration and exploitation well. However, in con- trast to bandits where algorithms are evaluated in terms of regret, the RL community has proposed different measures to determine efficient exploration. An important concept is the sample complexity (Vapnik, 1998), which was first defined in the context of supervised learning. Loosely speaking, sample complexity is the number of examples needed to bring the estimate of a target function within a given error range. Kakade (2003) studied sample complexity in a RL context. Consider an agent interacting in an environment. The steps of the agent can be roughly classified into two categories: steps in which the agent acts near-optimally as "exploitation" and steps in which the agent is not acting near optimally as "exploration". Subsequently, it is possible to see the number of times in which the agent is not acting near-optimally as the sample complexity of exploration (Kakade, 2003) for which some algorithms have guarantees (Brafman and Tennenholtz, 2003). Before formalizing the problem of learning in multiagent environments (Section 3) we will discuss game theory in the next section, as it is a classical area that addresses the interaction, reasoning and decision-making of multiple agents in strategic conflicts of interest. 2.3 Game theory Game theory studies decision problems when several agents interact (Fudenberg and Tirole, 1991). The terminology in this area is different, agents are usually called players, a single 5. The reason that the exploration policy is not the optimal policy is that 1. the optimal policy is not know yet to the agent, and 2. that the action that is most informative is not necessarily the one leading to the highest expected reward. 6. Convergence proofs for on-policy methods usually require more details to be specified than for off-policy algorithms (Singh et al., 2000; Van Seijen et al., 2009). 10 A Survey of Learning in Multiagent Environments Table 1: The normal-form representation of the prisoner's dilemma game. Each cell represents the utilities given to the players (left value for A and right one for O), rpd, tpd, spd, ppd ∈ R where the following conditions must hold tpd > rpd > ppd > spd and 2rpd > ppd + spd. Player O cooperate rpd, rpd tpd, spd defect spd, tpd ppd, ppd Player A cooperate defect interaction between players is represented as a game, and rewards obtained by the players are called payoffs. The most common way of presenting a game is by using a matrix that denote the utilities obtained by each agent, this is the normal-form game. Definition 2 (Normal-form game) A (finite, I-person) normal-form game Γ, is a tuple (cid:104)N , A, u(cid:105), where: N is a finite set of I players, indexed by i; A = A1 × ··· × AI, where Ai is a finite set of actions available to player i. Each vector a = (a1, . . . , aI) ∈ A is called an action profile; u = (u1, . . . , uI) where ui : A (cid:55)→ R is a real-valued utility or payoff function for player i. For example, Table 1 shows a two-action two-player game, known as the Prisoner's Dilemma (PD). Each row corresponds to a possible action for player A and each column corresponds to a possible action for player O. Player's payoffs are provided in the corre- sponding cells of the joint action, with player A's utility listed first. In the example, each player has two actions {cooperate, defect}. A strategy specifies a method for choosing an action. One kind of strategy is to select a single action and play it, this is a pure strategy. In general, a mixed strategy specifies a probability distribution over actions. Definition 3 (Mixed strategy) Let (I, A, u) be a normal-form game, and for any set X, let ∆(X) be the set of all probability distributions over X, then the set of mixed strategies for player i is Si = ∆(Ai) In this context, it is important to define what is a good strategy, i.e., the best response. Definition 4 (Best response) Player i's best response to the strategy profile s−i is a mixed strategy s∗ i , s−i) ≥ ui(si, s−i) for all strategies si ∈ Si. i ∈ Si such that ui(s∗ where s−i = s1, . . . , si−1, si+1, . . . , sn represents the strategies of all players except i. Thus, a best response for an agent is the strategy (or strategies) that produce the most favourable outcome for a player, taking other players' strategies as given. Another common strategy is the minimax strategy that ensures a security level for the player. 11 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote Definition 5 (Minimax Strategy.) Strategy that maximizes its payoff assuming the op- ponent will make this value as small as possible. Definition 6 (Security level) The security level is the expected payoff a player can guar- antee itself using a minimax strategy. In single-agent decision theory, the notion of optimal strategy refers to the one that maximises the agent's expected payoff for a given environment. In multiagent settings the situation is more complex, and the optimal strategy for a given agent may now vary, since the best response strategy depends on the choices of others. In order to draw conclusions on the joint behavior in games, game theory has identified certain subsets of outcomes, called solution concepts, such as the Nash equilibrium (NE). Suppose that all players have a fixed strategy profile in a given game, if no player can increase its utility by unilaterally changing its strategy, then the strategies are in Nash equilibrium. Formally it is defined by: Definition 7 (Nash equilibrium; Nash, 1950b) A set of strategies s = (s1, . . . , sn) is a Nash equilibrium if, for all agents i, si is a best response to s−i. Even when it is proved that in every game exists a Nash equilibrium, this solution concept has limitations. One problem is that there may be multiple equilibria in a game, and it is not an easy task to select one (Harsanyi and Selten, 1988). Also several experiments involving humans have shown that that people usually do not follow the actions prescribed by the theory (Kahneman and Tversky, 1979; Risse, 2000; Goeree and Holt, 2001; Camerer, 2003). Extensive-form games Another common representation for games is the extensive-form in which it is easier to describe the sequential structure of the decisions (for example, this is useful to represent poker games). Commonly, the game is described as a tree where nodes represent actions taken by the players. Extensive-form games can be finite or infinite-horizon (regarding the length of the longest possible path), with observable or non-observable ac- tions and with complete or incomplete information (observability of the opponent pay- offs) (Fudenberg and Tirole, 1991). Most of the games represented in extensive form can be converted into a normal-form representation, however, this generally results in a matrix which is exponential in the size of the original game. For this reason, it is common to find a solution in the original game tree. Repeated and stochastic games Previous concepts (e.g., best response, Nash equilib- rium) were defined for one-shot games (one single interaction), however, it could be the case that more than one decision has to be made. For example, repeating the same game, or having a set of possible games. Definition 8 (Stochastic game) A stochastic game (also known as a Markov game) is a tuple (S,N , A, T, R), where: S is a finite set of states, N is a finite set of I players, A = A1 ×···× AI where Ai is finite set of actions available to player i, T : S × A× S → R is the transition probability function; T (s, a, s) is the probability of transitioning from state s to state s after action profile a, and R = r1, . . . , rI where ri : S × A → R is a real valued payoff function for player i. 12 A Survey of Learning in Multiagent Environments (a) (b) Figure 3: (a) The automata that describes the TFT strategy, depending of the opponent action (c or d) it transitions between the two states C and D. (b) The automata describing Pavlov strategy, it consists of four states formed by the last action of both agents (CC, CD, DC, DD). In a stochastic game, agents repeatedly play games (states) from a collection. The particular game played at any given iteration depends probabilistically on the previous played game (state), and on the actions taken by all agents in that game (Shoham and Leyton-Brown, 2008). Definition 9 (Repeated game) A repeated game is a stochastic game in which there is only one game (called stage game). To exemplify a repeated game, recall the prisoner's dilemma presented in Table 1. Re- peating the game for a number of rounds results in the iterated prisoner's dilemma (iPD), which has been the subject of different experiments and for which there are diverse well- known strategies. A successful strategy which won Axelrod's tournament7 is called Tit- for-Tat (TFT) (Axelrod and Hamilton, 1981); it starts by cooperating, and does whatever the opponent did in the previous round: it will cooperate if the opponent cooperated, and will defect if the opponent defected. Another important strategy is called Pavlov, which cooperates if both players performed the same action and defect whenever they used dif- ferent actions in the past round. The finite-state machines describing TFT and Pavlov are depicted in Figure 3. It should be noticed that these strategies can be described in terms of the past actions and therefore do not depend on the time index; they are stationary strategies. Having presented the formal models of multi-armed bandits, reinforcement learning and game theory, the next Section highlights the challenge of non-stationarity in multiagent systems, followed by a new framework for this setting. Moreover, we present our proposed categorisation on how algorithms cope with non-stationary behaviour. 3. Learning in multiagent environments The following subsection pinpoints where and how the main challenge of non-stationarity arises in multiagent environments. This provides a crisp basis problem definition against which the approaches of algorithms can be positioned. Next, we present a new abstract 7. Robert Axelrod held a tournament of various strategies for the iterated prisoner's dilemma. Strategies were run by computers. In the tournament, programs played games against each other and themselves repeatedly. 13 CDddccCCDDCDDCcdddcdccHernandez-Leal, Kaisers, Baarslag and Munoz de Cote framework for multiagent learning algorithms that naturally models (and emphasizes) three components of accounting for other reasoning agents. Finally, we present a taxonomy of multiagent learning algorithms, aligned with assumptions they make in light of this framework. 3.1 The problem Learning in a multiagent environments is inherently more complex than in the single-agent case, as agents interact at the same time with environment and potentially with each other (Busoniu et al., 2008). Transferring single-agent algorithms to the multiagent set- ting is a natural heuristic approach (see Section 3.3) -- even if assumptions under which these algorithms were derived are violated. In particular the Markov property, denoting a stationary environment, does not hold, thus invalidating guarantees derived for the single- agent case (Tuyls and Weiss, 2012). Since this approach of applying single-agent algorithms ignores the multiagent nature of the setting entirely, it can fail when an opponent may adapt its choice of actions based on the past history of the game (Shoham et al., 2007). In order to expose why multiagent domains are non-stationary from agents' local per- spectives, consider a stochastic game (S,N , A, T, R). Given a learning agent i and using the common shorthand notation −i = N \ {i} for the set of opponents, the value function j πj(s, aj): now depends on the joint action a = (ai, a−i), and the joint policy π(s, a) =(cid:81) V π i (s) = π(s, a) T (s, ai, a−i, s(cid:48))[R(s, ai, a−i, s(cid:48)) + γVi(s(cid:48))]. (4) (cid:88) a∈A (cid:88) s(cid:48)∈S Consequently, the optimal policy is a best response dependent on the other agents' policies, π∗ i (s, ai, π−i) = BRi(π−i) = arg max (πi,π−i) i (s) πi πi(s, ai)π−i(s, a−i) T (s, ai, a−i, s(cid:48))[R(s, ai, a−i, s(cid:48)) + γV (πi,π−i) i (s(cid:48))]. (cid:88) a∈A = arg max πi V (cid:88) s(cid:48)∈S Specifically the opponents' joint policy π−i(s, a−i) could be non-stationary, (for example when opponents' are learning) thus becoming the parameter of the best response function. If the opponents' are not learning, e.g., they are using a stochastic policy, then the environment is Markovian and single-agent learning algorithms suffice. Next, we propose a general framework for multiagent learning algorithms, separating three steps of modelling opponents' behaviour to tackle the problem of non-stationary op- ponent policies. 3.2 A new framework for multiagent learning algorithms Before going into the formal definitions of the abstract concepts, consider an intuitive de- scription of the three components of our proposed framework: • Policy generating functions τ ∈ T , describe how an opponent j obtains its policy πj. • Belief βj, i.e., a probability distribution over τ , measures an agent's belief about each opponent's reasoning. 14 A Survey of Learning in Multiagent Environments • Influence function θ partitions beliefs according to equivalent best responses. Some definitions are required to describe each component in detail: let ht = (z0, z1, . . . , zt) denote the observation history of t observations, and let H t be the set of all possible histo- ries of this length. Note that observations in the stochastic game are given by state, action sequences, but this more general representation also subsumes models of partial observabil- ity, such as POMDPs. While rewards are commonly treated separately in the literature, they may simply be added as part of the observation history in our model. Some work may presume or learn a model of reward functions, as in POMDPs for the agent (Kaelbling et al., 1998) or the frame of I-POMDPs for opponents (Gmytrasiewicz and Doshi, 2005), which is equally compatible with our model. Definition 10 (Policy generating functions) A policy generation function (PGF) τ maps the history of observations into a policy π, τ : H t → Π On the one hand, this definition could be extended to the stochastic case τ : H t → ∆(Π), where ∆(·) indicates the simplex function, i.e., here denoting a probability distribution or probability mass function over policies. However, this complexity appears unnecessary for the exposition we aim for in this section. On the other hand, the composition of τ and π could be chosen as an alternative definition, thus mapping histories directly to actions. In contrast to alternative definitions, deterministic policy generating functions are a par- ticularly relevant category since they capture memory-bounded models with hidden states, while maintaining the structure of policies. This enables additional assumptions over the rate of change in policies, or the set of policies that are (re-) visited by the algorithm. Such models subsume learning algorithms, e.g., Q-learning (Watkins, 1989), weight matri- ces describing neural networks (Bengio, 2009), and MDPs, e.g., mapping a sliding window of the histories to an action or policy, as finite state automata over a predefined set of policies (Banerjee and Peng, 2005; Chakraborty and Stone, 2013). The PGFs capture the adaptation dynamics of agents, and research articles derive in- sights related to learning algorithms within a scope delimited by an implicitly or explicitly defined set of PGFs for any opponents. One of the more general assumptions is given by the frame defined in I-POMDPs (Gmytrasiewicz and Doshi, 2005), which assumes further structure on the PGFs, such as ascribing rewards and optimality criteria to opponents. Our taxonomy below employs PGF assumptions as a main criterion for classifying algorithms. The next step is to define how the agent uses those PGFs. Note that observations are local to each agent, i.e., an agent i can only infer another agent's local perceived observation history hj by a probability distribution p(hjhi) using its own observations hi together with any available a priori knowledge, e.g., about the structure of the game. In stochastic games, state, action sequences are joint observations,8 thus hj = hi if rewards are treated separately. Definition 11 (Belief ) A belief β ∈ B indicates for each opponent j the likelihood βj(τhj) for each policy generating function τ given opponent experience hj. 8. Stochastic games usually assume that agents have complete information about the state of the game, a more general model are partially observable stochastic games (POSGs) (Bernstein et al., 2004). 15 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote Since hj is local information, it must be inferred from agent i's observations, i.e., hj ∼ p(hjhi). If the presumed set T gives rise to distinguishable policies, then the belief may identify each opponent's PGF in a crisp belief (assigning probability one to a specific τ for each opponent). Even if unique identification is not possible, this poses a classification task, and a unique assignment may be used as an approximation of the belief. On the other hand, full belief representations over multiple τ for each opponent are common in Bayesian reasoning (Ghavamzadeh et al., 2015). The last step defines how the belief could be filtered (e.g., to reduce complexity) by means of an influence function. Definition 12 (Influence function for multiagent learning) The co-domain of the in- fluence function θ over the belief is a k-dimensional influence space Θ: θ : B → Θ. Assumptions about the influence function may significantly alter the complexity of the algo- rithm, and the validity of such assumptions differentiates whether resulting model insights hold or reduce to heuristic approaches; below we provide some examples. • In single-agent learning, the assumption is that θ maps onto a singleton set. • On the opposite side of the spectrum, taking the identity function as θ is equivalent to not modelling θ at all, thus also not limiting the validity at this step. • However, imposing -- or learning -- a structure of θ would cluster equivalent best re- sponses (Bard et al., 2015) and may lead to more sample efficient learning of best response approximations. One example instantiation of the influence function may encode abductive reasoning by mapping mixed beliefs to crisp classifications, as men- tioned in the above discussion of beliefs. • Furthermore, in symmetric games with distinguishable policies, θ may encode the strategy histogram (counting players for each τ ), as by definition the payoffs of a player only depend on the strategies employed, and not on who is playing them. This structure is used in heuristic payoff tables to compress utility representations and corresponding best response mappings (Walsh et al., 2002). Overall, an influence function typically reduces the complexity, either as a lossless compres- sion or as a heuristic to reduce the set of best responses and the computational complexity of deriving them. Definition 13 (Best response in multiagent learning) A multiagent learning algorithm computes the best response to the influence state of its belief, given an a priori assumed T : BRi(θ) = π∗ i (s, a, θ) = BRi (π−iπj ∼ βj(τhj), hj ∼ p(hjhi)) , for any β that satisfies θ(β) = θ. This new framework maintains agent independence by mutual non-observability of in- dividual policies, and inherently models agent autonomy by the independent choice of best response policies. 16 A Survey of Learning in Multiagent Environments 3.3 Taxonomic Categories: Environment, Opponent and Agent Now, we present a taxonomy in terms of the environment (observability) and the oppo- nent characteristics (learning capabilities). Then, we provide an overview of the proposed categories of how algorithms deal with non-stationarity. 3.3.1 Environment: observability One crucial aspect that provides information on how to tackle a learning problem is observ- ability of actions and rewards, for both the learning agent and the opponent. Depending on the restriction of the domain, there are four categories in increasing order of observability. Local reward. The most basic information that an algorithm commonly observes are its own immediate rewards. Opponent actions. Most algorithms also assume that is possible to observe the opponent actions (but not the their rewards). Opponent actions and payoffs. Some algorithms assume to observe the action and also the actual payoffs of the opponents (which may hold more naturally in cooperative scenarios). Complete a priori knowledge. Similar to the previous category algorithms observe re- wards and actions, however, in this category the algorithms know from start the complete reward function. 3.3.2 Opponent: adaptation capabilities The capability of the opponent to adapt and change its behaviour provides another source of important information to be used while learning. Roughly, we distinguish three categories: No adaptation. ∀ht : τ (ht) = π These are opponents that follow a stationary strategy during the complete period of interaction. Slow adaptation. ∃ << 1,∀t : d(cid:0)τ (ht+1), τ (ht)(cid:1) < These opponents show non-stationary (cid:112)JSD (τs(ht+1)τs(ht)). behaviour. However, it is a limited adaptation, for example providing bounds to the possible change in the current strategy between rounds. Candidate metrics are Man- hattan distance d1 or the average Jensen-Shannon distance over all states, which with base 2 logarithm is bound to [0, 1]: d(cid:0)τ (ht+1), τ (ht)(cid:1) = 1S (cid:80) s Drastic or abrupt adaptation. If the above assumptions are not in place, non-stationary opponents may show abrupt changes in their behaviour, for example changing to a different strategy (no limits) from one step to the next. 3.3.3 Agent: dealing with non-stationarity Previous surveys have proposed different ways to categorise algorithms in multiagent sys- tems such as: team vs concurrent learning (Panait and Luke, 2005); temporal difference, game theory and direct policy search (Busoniu et al., 2010); model-based, model-free and re- gret minimization (Shoham et al., 2007); and joint action, gradient, Nash and other learners 17 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote (see Weiss, 2013, chap. 10). There are also some previous categories for the type of learn- ing used: multiplied, divided and interactive (Tuyls and Weiss, 2012); and independent, joint-action and gradient ascent (Bloembergen et al., 2015). Another group of works have proposed properties that MAS algorithms should have: Bowl- ing and Veloso (2002) propose rationality and convergence. The former needs the learning algorithm to converge to a stationary policy that is a best-response to the other players policies if the other players policies converge to stationary policies; the latter refers to the need of the agent to necessarily converge to a stationary policy. Powers and Shoham (2004) proposed: targeted optimality, compatibility and safety. The first one needs the agent to achieve within of the expected value of the best response to the actual opponent. Compat- ibility needs the algorithm to achieve at least within of the payoff of some Nash equilibrium that is not Pareto dominated by another NE (during self-play), and safety needs the agent to receive at least within of the security value for the game. Crandall and Goodrich (2011) proposed: security, coordination and cooperation. Security refers to long-term average pay- offs meet a minimum threshold, coordination refers to the ability to coordinate behaviour when associates share common interests, and cooperation is the ability to make compro- mises that approach or exceed the value of the Nash bargaining solution (Nash, 1950a) in games of conflicting interest. In contrast with previous works, we propose another view focused on how algorithms deal with non-stationary behaviour. We propose five categories in increasing order of so- phistication which we summarize as follows: 1. Ignore. The most basic approach which assumes a stationary environment. 2. Forget. These algorithms adapt to the changing environment by forgetting information and at the same time updating with recent observations, usually they are model-free approaches. 3. Respond to target opponents. Algorithms in this group have a clear and defined target opponent in mind and optimize against that opponent strategy. 4. Learn opponent models. These are model-based approaches that learn how the op- ponent is behaving and use that model to derive an acting policy. When the opponent changes they need to update its model and policy. 5. Theory of mind. These algorithms model the opponent assuming the opponent is modelling them, creating a recursive reasoning. Note that this order is according to the sophistication in terms of complexity in assump- tions and approach -- it is throughout possible that the elegance of solutions does not follow this ordering. Moreover, we acknowledge that some algorithms could fit in more than one category. To better understand the high level behaviour of these categories, the next section presents an illustrative example using a simple domain. 18 A Survey of Learning in Multiagent Environments (a) Ignore: A assumes an opponent which is station- ary (S) for the complete interaction period. Examples are Q-learning (Watkins and Dayan, 1992) and fictitious play (Brown, 1951). (b) Forget: A learns an initial strategy (S1) which is continually updated (S1(cid:48), S1(cid:48)(cid:48), . . . ) with recent observa- tions, one example is WoLF-PHC (Bowling and Veloso, 2002). (c) Respond to target opponents: One example is Minimax-Q (Littman, 1994) where the learning agent assumes the opponent tries to minimise the rewards. (d) Learn: A learns a model of the opponent strategy (S?) and derives an acting policy; opponent changes are infrequent, e.g., RL-CD (Da Silva et al., 2006). (e) Theory of mind: O reasons about how A might act and obtains a best response against that behaviour, BR(A). A repeats that process with the model of O, BR(BR(A)) (Gmytrasiewicz and Durfee, 2000). Figure 4: A learning agent A (outside the cloud) and how it models one opponent O (inside the cloud) exemplifying the 5 categories of how to handle non-stationary behaviour. 19 TimeSS…Learning agentOpponent/environmentTimeS1…S1'…S1''…S1'''S1''''…Learning agentOpponent/environmentTime…Min( )Max( )Learning agentOpponent/environmentTime…S?S?…S?Learning agentOpponent/environmentTime…BR( )BR( )S1S3S2Learning agentOpponent/environmentHernandez-Leal, Kaisers, Baarslag and Munoz de Cote 4. Illustrative Example - Iterated Prisoner's Dilemma In this section we exemplify each category and contrast them by using the same domain (see Figure 4). Our example is presented in the context of the iterated prisoner's dilemma (see Section 2.3) where two agents A and O play the infinite-horizon version of this game. Definition 14 Prisoner's dilemma (PD) is a normal-form game (cid:104)N , A, u(cid:105), where: The set of players N = {A,O} are the two agents. The set of actions is the same for both agents, they have two possible actions A = {C, D}. u = (uA, uO) where ui : A (cid:55)→ R the payoff function for player i as shown in Table 1, satisfying: tpd > rpd > ppd > spd and 2rpd > ppd + spd. In the PD game when both players cooperate they both obtain the reward rpd. If both defect, they get a punishment reward ppd. If a player chooses to cooperate with someone who defects receives the sucker's payoff spd, whereas the defecting player gains the temptation to defect, tpd. We now present slight variations of the above scenario exemplifying the assumptions made by algorithms in each category, pointing out where they are most useful and where their main assumptions do not hold. 4.1 Ignore In this category algorithms can be useful with simple opponents or by making probably unrealistic assumptions, ignoring the non-stationary behaviour. For example, assume the opponent uses a mixed (stationary) strategy, πm = (0.25, 0.75) with higher probability of se- lecting defect. If the assumption is correct, the learning agent can use fictitious play (Brown, 1951) to learn an optimal policy against O. However, consider the case that after A has learned the optimal policy, O decides to change to a Tit-for-Tat strategy πT F T , thus A's learned policy will no longer be optimal. 4.2 Forget Now, consider a different set of assumptions where A is interested in converging to a station- ary policy and O has the same interest. Thus, both agents need to adapt to the changing (non-stationary) behaviour of the other (see Figure 4(b)). One algorithm that is especially useful in this scenario is WoLF-PHC (Bowling and Veloso, 2002), the algorithm generalizes Q-learning, but it was proposed to converge to a stationary policy in self-play. We can view WOLF-PHC as continuously learning (and forgetting), adjusting its learning rate to cope with the changing behaviour of the opponent. Note that, if we remove the assump- tion of self-play (and we assume O uses a different behaviour), then WOLF-PHC loses its convergence guarantees. 4.3 Respond to target opponents If the learning agent knows (or assumes) the opponent will behave in specific ways, that information can be used to target classes of opponents. For example, assume A knows that the opponent will use the set of strategies {Tit-for-Tat, Pavlov, Bully} and change 20 A Survey of Learning in Multiagent Environments among them stochastically. In this case HM-MDPs (Choi et al., 1999) can target that type of opponent since they assume the environment can be represented in different stationary modes (MDPs) with stochastic transitions among themselves. Note that there are different algorithms that target a variety of classes (see Section 5.3). However, if the assumptions about the opponent do not hold (in this case, adding a new strategy to the initial set) these algorithms provide restricted adaptability and therefore the policy will be suboptimal after most opponent changes. 4.4 Learn opponent models In this category, agent A learns a model of the opponent which is used to derive an optimal acting policy. In this case, the learning agent starts without predefined opponent strategies or policies (Da Silva et al., 2006; Hernandez-Leal et al., 2014a). Instead, A assumes the opponent will use several stationary strategies with infrequent changes among them. For example, in the iPD the opponent could start with Pavlov and later change to Tit-for-Tat. Moreover, if the opponent returns to a previous learned strategy, A should be able to detect and change its policy without relearning the model (Hernandez-Leal et al., 2016a). However, one limitation of these algorithms is that they do not consider the strategic behaviour of O (an opponent that reasons about the agent A). 4.5 Theory of mind In the last category, the learning agent assumes an opponent that is performing strategic reasoning. This is, in the lowest level O reasons about A, in an upper level A reasons about O reasoning about A. Best responding to a reasoning level is the way to obtain an acting policy. For example, assume the opponent thinks A uses a set of strategies to act {Bully, random, Pavlov}, a distribution of those strategies represent the zero level or reasoning, L0. With the previous information O can compute a best response (BR) against L0, called level 1 strategy, L1 = BR(L0). Moreover, A can compute a best response against a distribution of the previous two levels, to obtain an acting policy (level 2) L2 = BR({L1, L0}). Note that this recursive reasoning could continue upwards and is the base of many approaches (Camerer et al., 2004; Gmytrasiewicz and Doshi, 2005; Wunder et al., 2009, 2012). A limitation is that the basic strategies need to be specified a priori and computing optimal policies can be computationally expensive (Gmytrasiewicz and Doshi, 2005). In the next section we present an extensive list of state-of-the-art algorithms in from game theory, multi-armed bandits and RL and where they fall into each category of sophis- tication along with their environment and opponent characteristics. 5. Algorithms In this section we present an extensive list of algorithms categorised with respect to how they deal with non-stationarity. Table 2 summarises this section by providing for each algorithm its category and some related characteristics such as observability, opponent adaptation and the environment it was designed for. Similarly, Figure 5 depicts a diagram highlighting 21 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote Table 2: A categorisation of different algorithms in terms of how they handle non-stationarity and with respect to related characteristics such as observability, opponent adaptation (de- scribed in Section 3.3) and the domain they were designed for: one-shot games (OSG), repeated games (RG), stochastic games (SG), extensive-form games (EG), sequential de- cision tasks (SDT) and multi-armed bandit scenarios (MAB). Algorithm Observability Opp. adaptation Designed for No No No No No No Slow Slow Slow Slow Slow Drastic Drastic Drastic Slow Slow Both Drastic Drastic Slow Slow Drastic Slow Drastic Slow Slow Drastic Slow Slow Slow Slow Drastic Drastic Slow Slow Drastic Slow Slow Drastic Drastic Slow Drastic Drastic Slow Slow Drastic Drastic Drastic Drastic No Slow No No No No Slow RG SDT RG MAB MAB SG SG SG RG RG RG MAB MAB MAB RG RG MAB SG RG RG MAB SG SG SDT SG SG MAB SG SG SG RG RG RG RG EG SDT RG SG SDT SDT RG SG SDT SDT RG RG MAB RG RG SDT RG OSG OSG SDT RG RG/SG Category Ignore Forget Fictitious play (Brown, 1951) Q-learning (Watkins, 1989) JAL (Claus and Boutilier, 1998) UCB (Auer et al., 2002a) Exp3 and Exp4 (Auer et al., 2002b) R-max (Brafman and Tennenholtz, 2003) O. actions Local reward O. actions Local reward Local reward O. actions and payoffs WOLF-IGA (Bowling and Veloso, 2002) WOLF-PHC (Bowling and Veloso, 2002) GIGA-WOLF (Bowling, 2004) COLF (Munoz de Cote et al., 2006) WPL (Abdallah and Lesser, 2008) WMD-UCB (Yu and Mannor, 2009a) D-UCB (Garivier and Moulines, 2011) SW-UCB (Garivier and Moulines, 2011) FAQL / IQ (Kaisers and Tuyls, 2010) LFAQ (Bloembergen et al., 2010) Rexp3 (Besbes et al., 2014) FAL-SG (Elidrisi et al., 2014) R-max# (Hernandez-Leal et al., 2017b) RUQ (Abdallah and Kaisers, 2013, 2016) UUB (Lakkaraju et al., 2017) Target Minimax-Q (Littman, 1994) Nash-Q (Hu and Wellman, 1998) HM-MDPs (Choi et al., 1999) FF-Q Littman (2001) EXORL (Suematsu and Hayashi, 2002) Exp3.S (Auer, 2002) Hyper-Q (Tesauro, 2003) Correlated-Q (Greenwald and Hall, 2003) NSCP (Weinberg and Rosenschein, 2004) ReDVaLeR (Banerjee and Peng, 2004) MetaStrategy (Powers and Shoham, 2004) Manipulator (Powers and Shoham, 2005; Powers O. actions and rewards Local rewards Local rewards O. actions Local reward Local rewards Local rewards Local rewards O. actions O. actions Local rewards O. actions O. actions O. actions Local rewards O. actions O. actions and payoffs O. actions O. actions and payoffs O. actions and payoffs Local rewards O. actions O. actions and payoffs O. actions O. actions O. actions and payoffs O. actions and payoffs et al., 2007) AWESOME (Conitzer and Sandholm, 2006) RNR and DBR (Johanson et al., 2007; Johanson and Bowling, 2009) Local rewards O. actions ORDP (Yu and Mannor, 2009b) M-Qubed (Crandall and Goodrich, 2011) Pepper (Crandall, 2012) MDP-A and BPR (Mahmud and Ramamoorthy, O. actions O. actions O. actions O. actions 2013; Rosman et al., 2016) HS3MDPs (Hadoux et al., 2014b) RSRS (Damer and Gini, 2017) OLSI (Hernandez-Leal and Kaisers, 2017a) O. actions O. actions O. actions Learn RL-CD (Da Silva et al., 2006; Hadoux et al., 2014a) ζ−R-MAX (Lopes et al., 2012) CMLeS (Chakraborty and Stone, 2013) MDP-CL (Hernandez-Leal et al., 2014a) Restless Markov bandits (Ortner et al., 2014) DriftER (Hernandez-Leal et al., 2017a) BPR+ (Hernandez-Leal et al., 2016a,b) Theory of mind O. actions O. actions O. actions O. actions Local rewards O. actions O. actions RMM (Gmytrasiewicz and Durfee, 2000) s-EWA (Camerer et al., 2002) Level-K (Costa Gomes et al., 2001) Cognitive Hierarchy (Camerer et al., 2004) I-POMDP (Gmytrasiewicz and Doshi, 2005) PI-POMDP (Wunder et al., 2011, 2012) ToM/MToM (de Weerd et al., 2013; Van der Osten O. actions and payoffs O. actions and payoffs O. actions O. actions O. actions O. actions O. actions et al., 2017) 22 A Survey of Learning in Multiagent Environments Figure 5: Diagram of the algorithms (coloured boxes; each colour represent one experimental do- main) analysed in this survey divided in 5 categories (dashed lines) on how they handle non-stationarity. We present how they are connected to each other (arrows) and highlight those algorithms that are representative of each category (double box). the connections among the algorithms and showing the most representative ones of each category. 5.1 Ignore Game theory is the study of strategic interactions among several agents, with the central concept of equilibrium among players denoting a mutual best response. While such rea- soning does account for opponent strategies, classic algorithms typically do not account for changes in opponent strategies. One early work for learning in repeated games is fictitious play (Brown, 1951). The model maintains a count of the plays by the opponent in the past. The opponent is assumed to be playing a stationary mixed strategy and the observed frequencies are taken to represent the opponent's mixed strategy. However, if the opponent does not follow a stationary strategy the method will not compute a best response. 23 UCBFPQ-learningR-maxWOLF-IGACOLFRUQFAQL/IQR-MAX#Correlated-QNash-QEXORLAWESOMEMetaStrategyManipulatorMinimax-QMDP-ADBRORDPFAL-SG-R-MAXCMLeSMDP-CLDriftERBPR+RL-CDPI-POMPLevel-kCHI-POMDPRMMWOLF-PHCExp3IgnoreForgetLearnTheory of mindTargetJALs-EWASW-UCBRNRBPRExp4Exp3.SFF-QGIGA-WoLFReDVaLeRWMD-UCBWPLPCM(A)Extensive-form gamesStochastic gamesRepeated gamesOne-shot gamesSequential decision tasksMulti-armed banditsRestless Markov banditsUUBHS3MDPsRexp3RSRSMToMToMPepperOLSID-UCBA priori MDP-CLHM-MDPsM-QubedNSCPHyper-QLFAQHernandez-Leal, Kaisers, Baarslag and Munoz de Cote Techniques from multi-armed bandits have been used to deal with the exploration- exploitation trade-off. In the classical bandit setting some assumptions are made regarding the rewards, e.g., the rewards are drawn independently from some fixed (stationary), but unknown distributions. In this context, the UCB (upper confidence bounds) algo- rithm (Auer et al., 2002a) guarantees low regret under certain conditions. UCB uses the principle of optimism in the face of uncertainty to select its actions (assumes an optimistic guess on the expected rewards). A different setting is the adversarial bandit setting where no statistical assumptions are made about the generation of rewards (Auer et al., 2002b). Instead, the reward associated with each arm at every round are fixed in advance by an ad- versary before the game starts. Even in this complicated scenario, the exponential-weight algorithm for exploration and exploitation (Exp3) provides theoretical bounds for expected rewards (Auer et al., 2002b). Finally, a different algorithm based on Exp3 is the exponential-weight algorithm for exploration and exploitation using expert advice (Exp4) (Auer et al., 2002b). The scenario is different since now the algorithm assumes a set of "experts" that provide a mechanism to select an action. Exp4 provides bounds of expected utility to perform nearly as well as the best expert in hindsight. Note that these bandit algorithms assume the setting is fixed in advance and does not account for changes during the interaction (i.e., ignore). In the context of RL, a model-based algorithm for acting optimally in adversarial en- vironments is R-max (Brafman and Tennenholtz, 2003). The algorithm uses an MDP to model the environment which is initialised optimistically assuming all actions return the maximum possible reward, (r-max). After several experiences with the environment R-max updates and fixes a part of the model (i.e., state-action pairs). The policy efficiently leads the agent to less known state-action pairs or exploits known ones with high utility. R-max promotes an efficient sample complexity of exploration (Kakade, 2003), this means that R-max has theoretical guarantees for obtaining near-optimal expected rewards. However, R-max alone will not work when the environment presents non-stationary behaviour (Lopes et al., 2012) since it fails to adjust its model if the environment changes. The classic model-free RL algorithm of Q-learning assumes a stationary environment. However, it has been applied with some success in different multiagent scenarios (Tan, 1993; Sen et al., 1994; Crites and Barto, 1998). The simple approach of using plain Q-learning, i.e., ignoring other agents in the environment, is known as independent learners. In contrast, joint-action learners (JALs) model the strategies of the opponents explicitly by taking into account the joint-action of all the agents in the Q-learning update, which implies the agent can observe the actions of others (Claus and Boutilier, 1998). However, even when they have more information, convergence is not dramatically enhanced. Moreover, in JALs it is required a considerable amount of contrary experience to be overcome some changing behaviour (Claus and Boutilier, 1998). 5.2 Forget Failing to update with current information is the main limitation of the algorithms in the previous category. A solution is to forget old information and update with recent one, which has been experimentally noted to improve learning algorithms in repeated games (Bouzy and 24 A Survey of Learning in Multiagent Environments M´etivier, 2010). While this category could be more aptly described as adaptive discounting of experiences, we chose the label forget as an intuitive concept and mnemonic. Even though UCB has been empirically shown to not work well on non-stationary envi- ronments (Hartland et al., 2006), it has inspired two algorithms that can adapt to sudden changes, i.e., when the distributions of rewards changes abruptly (not depending on the policy of the player or on the sequence of rewards). Garivier and Moulines (2011) proposed two methods: the discounted UCB (D-UCB) whose policy averages past rewards with a discount factor giving more weight to recent observations; and the sliding-window ap- proach (SW-UCB) which relies on a local empirical information of the last τ plays (τ being a parameter of the algorithm). Lakkaraju et al. (2017) proposed a variant of D-UCB to solve an exploration problem where the expected utility of each arm is non-stationary. However, instead of assuming arbitrary changes in the utility distribution (as D-UCB), their setting has certain structure which is encoded in their proposed bandit for unknown unknowns (UUB) algorithm. Yu and Mannor (2009a) tackle a specific non-stationary bandit problem with two main characteristics: (i) the rewards are piecewise-stationary, i.e., the reward distribution changes arbitrarily and at arbitrary time instants, but it remains stationary on intervals; (ii) the agent can observe some of the past outcomes of arms that have not been picked. Yu and Mannor (2009a) propose the windowed mean-shift de- tection (WMD)-UCB to cope with these scenarios. The algorithm works by detecting changes in the environment using a statistical test on the most recent τ time-steps (i.e., sliding window), when this happens the algorithm resets. Note that discounting or using a sliding window approach have the same effect, give more weight to recent observations and forgetting the old one. A different way to model non-stationarity in multi-armed bandit scenarios is to assume the total variation in expected reward is bounded by a (known) variation budget. This allows to model diverse reward changes, e.g., both slow and continuous or drastic jumps. Besbes et al. (2014) proposed the Rexp3 algorithm (based on Exp3) for this setting and their results highlight a trade-off that exists between retaining and forgetting information, i.e., the fewer past observations to recall, the larger the associated error; the more past observations, the higher the chances of these being biased towards outdated information. In the context of efficiently exploring adversarial environments one example of the for- getting behaviour is the R-max# algorithm (Hernandez-Leal et al., 2017b). R-max# proposes a drift exploration to detect changes that happened in the opponent, but that may have not been noticed, which results in suboptimal behaviour. This effect is known as shadowing (Fulda and Ventura, 2007) or observationally equivalent models (Doshi and Gmytrasiewicz, 2006). To avoid this effect, the solution is to continually revisit states that have not been visited recently (which is determined by a parameter). Therefore R-max# proposes to reset (to r-max ) those state-action pairs and then update the model and policy which will implicitly re-explore those parts of the environment. R-max# provides theoret- ical results showing that under some assumptions it is guaranteed to learn a new model within finite sample complexity. Note that, in contrast to the classic R-max which fixes one part of its model and later is never allowed to update that same part; R-max# is con- tinually updating its model (and policy) to keep up with the non-stationary environment. However, the approach may not be easily scalable to scenarios with many agents. 25 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote In the model-free context of RL there are two variants of Q-learning that achieve con- vergence of self-play in specific games by updating action-value estimators equally fast, even when one action is more frequently selected than another: The first has been studied under the name frequency-adjusted Q-learning (FAQL) (Kaisers and Tuyls, 2010, 2011) and individual Q-learning (IQ) (Leslie and Collins, 2005), and the second one is repeated up- date Q-learning (RUQ) (Abdallah and Kaisers, 2013, 2016). Intuitively, the action that receives fewer updates needs to make larger adjustments to keep up, which is implemented with a learning rate modulation in FAQL/IQ (inversely proportional to the probability of the action's selection probability), and by repeated updates in RUQ. As a result, all actions receive the same expected learning speed. Formally, these learning speed modulations make it possible to prove the limit behaviour of the algorithms in self-play converges to Nash dis- tributions in zero-sum games (Leslie and Collins, 2005), with convergence points shown to approach Nash equilibria as the exploration temperature decreases in two-agent two-action games (Kaisers and Tuyls, 2011; Kianercy and Galstyan, 2012). Bloembergen et al. (2010) proposed lenient frequency adjusted Q-learning (LFAQ) for cooperative multi-agent environments. This extension incorporates the concept of leniency (Panait et al., 2006) to account for initial mis-coordination, which enables LFAQ to obtain high convergence to Pareto optimal equilibria in cooperative games. Note that convergence results of this type of algorithms require the assumption of infinite interactions and/or infinitesimal learning rates. Effectively, the action-value estimates of frequently selected actions are expected to be more recent and accurate, receiving only small updates based on each new observation. In contrast, scarcely selected actions are likely to have older action-value estimates, which in non-stationary environments may become less accurate with age, and therefore more weight is put into the new observation -- the value estimator is updated with a larger learning rate towards the new observation. As a consequence, these algorithms can be said to implement a dynamic strategy to forget outdated action-value estimates. The win or learn fast (WoLF) principle was introduced to make an algorithm that (i) con- verges to a stationary policy in multiagent systems and (ii) if other players' policies converge to stationary policies then the algorithm should converge to a best response (Bowling and Veloso, 2002). The intuition of WoLF is to learn quickly when losing and cautiously when winning. One proposed algorithm that uses this principle is WoLF-IGA (infinitesimal gradient ascent). The algorithm at each interaction updates its strategy (in the direc- tion of the gradient) to increase its expected payoffs with some fixed step size. WoLF-IGA has been proved theoretically to converge in self-play in a two-person, two-action repeated matrix games. However, WoLF-IGA assumes to know an equilibrium from the start which can be complicated in many games. Generalized IGA (GIGA)-WoLF (Bowling, 2004) improves on WoLF-IGA in two aspects. First, it does not need to known an equilibrium strategy. Second, it also addresses the challenge of not being exploited by an opponent by showing no-regret in the limit (Bowling, 2004). Finally, another practical variant of the WoLF principle is WoLF policy-hill climbing (WoLF-PHC) (Bowling and Veloso, 2002), which is based on Q-learning and performs hill-climbing in the space of mixed poli- cies. To cope with non-stationary behaviour WoLF-PHC changes between two learning rates depending on how the algorithm sees the interaction is happening, i.e., by comparing whether the current expected value is greater than the current expected value of the average policy. 26 A Survey of Learning in Multiagent Environments In the context of cooperative game theory it is common to look for Pareto efficient solu- tions. On one side, recall that WoLF algorithms aim converge to the Nash equilibrium, thus they are not the best candidate for this different type of problems (Stimpson and Goodrich, 2003). On the other side, using simple Q-learning algorithms results in suboptimal solu- tions due to the parallel learning process which makes the environment non-stationary. To overcome this issue, CoLF (change of learn fast) (Munoz de Cote et al., 2006) is another algorithm inspired by the WoLF principle, but with the objective of promoting cooperation of self-interested agents to achieve a Pareto efficient solution in repeated games. CoLF pro- poses to adjust the learning rate of the algorithm depending on the received rewards: slow when unexpected or changing (i.e., non-stationary) and fast when they are stable, near- stationary. Note that changing the learning rates is a common method to keep up with non-stationary environments. In the end this adaptation results in updating information and forgetting outdated estimates. Weighted policy learner (WPL) (Abdallah and Lesser, 2008) is another algorithm designed to converge to a Nash equilibrium. However, in contrast to previous algorithms it can do so with limited knowledge observing only local rewards (the agent neither knows the underlying game nor observes other agents actions). WPL share some similarities with WoLF-IGA since it also has two modes for adjusting its learning rate, however there are also some key differences: (i) WPL needs considerably less information and (ii) WPL uses a continuous spectrum of learning rates (WOLF-IGA uses two fixed ones). Fast adaptive learner (FAL) (Elidrisi et al., 2012) is designed to learn quickly in two- player repeated games. The algorithm is based on two components: (i) to predict the next action of the opponent the entropy learning pruned hypothesis space (ELPH) algorithm is used, ELPH is an online learning algorithm that maintains a set of hypotheses according to a fixed window of the history of observations (Jensen et al., 2005). The frequency count of each hypothesis is used to obtain the entropy which is used as an indicator of the quality of the prediction. (ii) To obtain a strategy against the opponent the authors use a modified version of the Godfather strategy.9 An extension of FAL for stochastic games is FAL-SG (Elidrisi et al., 2014). To deal with this different setting, FAL-SG abstracts the stochastic game into a meta-game matrix via clustering, after which the original FAL approach can be used. 5.3 Respond to target opponents Previous approaches updated their behaviour according to the newest information avail- able, in contrast, algorithms in this group have a pre-defined target of opponents. This is the category with the largest number of algorithms. The reason is that easier to provide guarantees against specific opponents than against general classes; to better understand the different approaches we made subdivision for this category into model-free and model-based approaches. Model-free approaches 9. The Godfather strategy gives the opponent the opportunity to cooperate with an action that is beneficial for both players. If the opponent does not accept the offer, Godfather will force the opponent to obtain its security level (Littman and Stone, 2001). 27 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote In the context of multi-armed bandits one extension of Exp3 is Exp3.S (Auer, 2002) which targets a specific adversarial bandit scenario in which the bandits are allowed to shift S times (a parameter of the algorithm). The algorithm keeps track of the alternative which gives highest reward even if this best alternative changes over time. The algorithm guarantees low regret assuming the number of shifts (S) and the number of rounds in the interaction is known in advance. In the traditional single-agent version of Q-learning the objective is to maximise the sum of rewards in an environment. In contrast, Minimax-Q proposes to extend Q-learning to zero-sum stochastic games, assuming an opponent which has a diametrically opposed ob- jective to the agent. The algorithm uses the minimax operator to take into account the opponent actions (Littman, 1994). This allows the agent to converge to a fixed strategy that is guaranteed to be safe in that it does as well as possible against the worst possi- ble opponent (the agent tries to maximize its rewards and the opponent aims to minimise those). The algorithm is guaranteed to converge in self-play to a stationary policy. Never- theless, there are cases when minimax-Q does not converge to the best response, i.e., is not rational (Bowling and Veloso, 2002). Hyper-Q (Tesauro, 2003) is another extension of Q-learning designed for multiagent systems (specifically for stochastic games). The main difference that the Q function depends on three parameters: the state, the estimated joint mixed strategy of all other agents, and the current mixed strategy of the agent. Hyper-Q assumes that only the opponents' actions (not the payoffs) are observable. To obtain an approximation of the mixed strategies a discretisation has to be performed and the Q-table could easily grow exponentially in the number of discretisation points. Hyper-Q is guaranteed to converge to the optimal value function against the following three groups of opponents: (i) stationary opponents, (ii) non-stationary opponents that define its history-independent strategy depending only on themselves and not on the Hyper-Q player (e.g., replicator dynamics model, see Borgers and Sarin, 1997) and (iii) non-stationary opponents that accurately estimate the Hyper-Q agent strategy and then adapt using a fixed history-independent rule. M-Qubed (Max or Minimax Q-learning) (Crandall and Goodrich, 2011) is a RL algorithm designed for two-player repeated games. The authors mention several compro- mises which an algorithm needs to balance: bounding loses (safety), playing optimally (best respond) and taking risks for ensuring cooperation and coordination. To achieve this, the algorithm targets two groups of opponents and proposes different behaviours (best-response and cautious) against each group. M-Qubed typically selects actions based on its Q-values updated via SARSA (best-response), but triggers to a minimax strategy when its total loss exceeds a pre-determine threshold (cautious). Another targeted set of opponents consist of agents using non-stationary policies with a limit (i.e., decreasing possibly infinite changes). The non-stationary converging poli- cies (NSCP) algorithm (Weinberg and Rosenschein, 2004) it is based on Q-learning and computes a best response to opponents in which the probability that the strategy would be far away from the limit gets smaller as the rounds increase. For this, Weinberg and Rosenschein (2004) define a distance between two stage game strategies as the distance between the probability vectors of the strategies. An example of this type of opponent is start with a uniform distribution over a set of actions and at each time-step the probability slowly moves towards one action with probability 1 and the rest with 0. 28 A Survey of Learning in Multiagent Environments Previous algorithms aim to best-respond to target opponents, however, another common approach is to respond with the aim of converging to a Nash equilibrium. Nash-Q (Hu and Wellman, 1998, 2003) is a variation of Q-learning that needs to observe the opponent actions and rewards to converge in some cases. The algorithm update Q-values over joint actions rather than a single-agent Q function. Another main difference with respect to Q-learning is that it updates with future payoffs assuming all agents will use a NE strategy. Friend-or- foe Q-learning (FF-Q) (Littman, 2001) generalise Nash-Q and Minimax-Q algorithms. FF-Q treat each opponent either as friend or foe and can converge in two cases: adversarial (minimax) equilibrium or in coordination games with unique equilibrium. Furthermore, a generalization of FF-Q is Correlated Q-learning (Greenwald and Hall, 2003) which instead of converging to a Nash equilibrium, it looks for a correlated equilibrium10 which is more general than a NE (Aumann, 1974). A common problem regarding NE is the selection when there are multiple options, to deal with this issue Correlated-Q uses four equilibrium selection functions which depending on the objective to maximise (e.g., each individual reward, the sum of the players rewards). However, to compute any of those it needs to observe opponents' actions and rewards. A limitation of previous approaches is that they target only one group (class) of op- ponents. Therefore, some algorithms improve on that regard, one example is EXORL (extended optimal response) (Suematsu and Hayashi, 2002) which has two main act- ing behaviours: best response or Nash equilibrium. EXORL starts learning a best response to the opponent (using on-policy learning), but if the opponent adapts (determined by a parameter) then it will look for a Nash equilibrium. Replicator dynamics with a vari- able learning rate (ReDVaLeR) (Banerjee and Peng, 2004) builds on the same ideas of EXORL: best response against stationary opponents and NE against adaptive opponents. Moreover, ReDVaLeR adds another characteristic, constant bounded expected regret at any time against any number of opponents (Banerjee and Peng, 2004). This makes the algorithm more robust since it is implicitly targeting opponents that are neither stationary nor using the same learning algorithm. ReDVaLeR needs to observe opponent actions, if this is not possible then AWESOME (adapt when everybody is stationary otherwise move to equilibrium) (Conitzer and Sandholm, 2006) is designed for this case. AWESOME converges to a Nash-equilibrium in self-play and when the opponents seem stationary it will learn a best response and can do so with limited information (i.e., only local rewards). We noted that algorithms basically target three main behaviours depending on the opponents: convergence (against adaptive opponents), best response (against stationary opponents) and bound the loss (against other types of opponents). In this regard, Powers and Shoham (2004) formalised these three properties as compatibility, targeted optimality and safety. Moreover, they proposed the MetaStrategy (Powers and Shoham, 2004) algo- rithm that achieves those three properties by alternate among the strategies: fictitious play, minimax and a modified Bully.11 A slightly different algorithm is Manipulator12 (Powers and Shoham, 2005) which alternates among: best response, minimax and a modified God- father strategy. Moreover, Manipulator has the same guarantees as MetaStrategy against 10. In these games it is assumed a public signal from the environment which is observed by all agents, a real-world example is a traffic signal, the agents decide its strategy based on that signal. 11. Littman and Stone (2001) proposed the Bully strategy which is an example of a Stackelberg leader. 12. PCM(A) is an extension of Manipulator to multiplayer games (Powers et al., 2007). 29 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote Figure 6: An example of an HM-MDP with 3 modes (large circles) and 4 states (smaller shaded circles). The value Xmn represents a transition probability between modes m and n, and ym(s, a, s(cid:48)) represents a state transition probability in mode m. a richer class of target opponents, memory-bounded opponents.13 These are defined as op- ponents that play a conditional strategy where actions can only depend on recent periods, this is its distribution over actions can only depend on the most recent k periods of past history. We note that the approach of how MetaStrategy and Manipulator decide on which strategy to use is the same: (i) first to explore (ii) to determine how the opponent reacts and possibly act with a best response; (iii) otherwise the algorithms opt for a safe option (minimax strategy). Model-based approaches Hidden-mode Markov decision processes (HM-MDPs) are a model-based technique to deal with non-stationary environments (Choi et al., 1999). They assume the environment can be represented in a small number of modes. Each mode is a stationary environment, which has different dynamics and needs a different policy. It is assumed that at each time-step there is only one active mode. The modes are hidden, which means that cannot be directly observed, they are only estimated by past observations. Moreover, transitions between modes are stochastic events. Each mode is modelled as an MDP. Different MDPs along with its transition probabilities form an HM-MDP which can be seen as a special case of a POMDP (Choi et al., 2001). Figure 6 depicts an example of an HM-MDP with 3 modes and 4 states. Each of the three large circles represent a mode, shaded circles inside the modes represent states. Thick arrows indicate stochastic transitions between modes and thinner arrows represent state-action-next state probabilities. A limitation of HM-MDPs is that they need to fix the number of modes from the start and do not provide any form of online learning. 13. In the same context of bounded memory adversaries, but in the bandit setting Arora et al. (2012) showed that no bandit algorithm can guarantee a sublinear policy regret against an adaptive adversary with unbounded memory. However, if the adversary's memory is bounded, they propose a technique which converts any bandit algorithm with sublinear regret bound into a sublinear policy regret bound. 30 Mode m state sXmnMode n state s'ym(s,a,s')A Survey of Learning in Multiagent Environments HM-MDPs assume the environment may change at every timestep, which may not hold in many environments. Hidden-semi-Markov-mode Markov decision processes (HS3MDPs) (Hadoux et al., 2014b) take inspiration from HM-MDPs to represent prob- lems in non-stationary environments but their difference is that HS3MSPs assume these changes evolve according to a semi-Markov chain (i.e., when the environment stochastically changes to a new environment it stays in that environment during a stochastically drawn duration). HS3MDPs are equivalent to HM-MDPs and form a subclass of POMDPs. To solve large-sized HS3MDPs, Hadoux et al. proposed an adaptation of POMCP (Silver and Veness, 2010). Learning an opponent model is usually a way to obtain an acting policy (see Section 5.4). However, some algorithms assume to start with a set of policies to act and the problem now becomes which policy to select in an non-stationary environment. For example, Mahmud and Ramamoorthy (2013) propose a scenario in which a latent variable changes rarely, but when it happens it modifies the optimal policy. Thus, the agent at the beginning of each round selects one policy from a known (and predefined) set Π. The goal of the agent is thus to select policies to minimise the total regret incurred in the limited task duration with respect to the performance of the best alternative from Π in hindsight. MDP-A (Mahmud and Ramamoorthy, 2013) was designed for single agent scenarios with a set of tasks in which an agent needs to perform similar tasks (the same state and action space), but with different policies for each task. MDP-A uses a transfer learning approach in which given a collection of source behaviour policies, eliminates the policies that do not apply in the new task using a statistical test in an online fashion. Similarly, Bayesian policy reuse (BPR) (Rosman et al., 2016) is another approach that draws inspiration from MDP-A since they work under the same scenario. However, BPR computes a belief distribution over the tasks and with every step of interaction it receives a signal which is used to update that belief using the Bayes rule. A limitation of BPR is the assumption of knowing a priori "performance models" (probability distributions) describing how policies behave on different tasks. A similar problem has been studied in the context of repeated games. Hernandez-Leal et al. (2014b) analysed a scenario where the opponent has a set of stationary strategies and changes among them during the interaction. Moreover, they assumed to know those strate- gies (represented as MDPs, see Banerjee and Peng, 2005) before the interaction. A priori MDP-CL is an algorithm designed to quickly detect the strategy used by a non-stationary opponent (Hernandez-Leal et al., 2014b). A priori MDP-CL explores with different actions for a period of rounds to learn an opponent model in the form of an MDP which is com- pared to the initially known strategies. If the learned model matches one of the prior known opponent strategies then the exploration phase finishes and the agent can solve the MDP (that represent the opponent) to obtain an policy against it. Inspired by the paradigm of optimism in face of uncertainty (Brafman and Tennenholtz, 2003), Crandall (2012) proposed the potential exploration with pseudo stationary restarts (Pepper) algorithm to learn in repeated stochastic games. Pepper creates a family of new algorithms when plugged together with learning algorithms for repeated matrix games (e.g., M-Qubed, see Crandall and Goodrich, 2011; fictitious play, see Brown, 1951). Hernandez-Leal and Kaisers (2017a) proposed a variation of repeated stochastic games in which the opponent may change constantly and its identity is unknown to the 31 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote learning agent. Their opponent learning in sequential interactions (OLSI) algorithm is generalization of Pepper to learn from different opponents by keeping a belief over a hypothesised opponent set. Note that most approaches respond optimally only to targeted opponents, but remain silent to what happen against other opponents. Against this background, Johanson et al. (2007) analysed how robust are counter strategies (learned to act against a single opponent) across different opponents in a poker domain. They performed this analysis using the mod- elling technique called frequentist best response (FBR). Their analysis showed that FBR is very successful at exploiting the opponent it was designed to exploit. However, when FBR strategies play against other opponents their performance is poor. To solve this issue, the authors propose the restricted Nash response (RNR) algorithm to generate robust strategies against a specific opponent, but at the same time they assume the opponent may slightly change (Johanson et al., 2007). The strategy obtained by RNR is based on defining a probability p for the opponent to act as the learned model and with probability 1−p it will act different than the model. RNR requires a large number of observations and sometimes can over fit the opponent model. Later, data biased response (DBR) (Johanson and Bowling, 2009) which extends from RNR was proposed to overcome those problems. Re- cently, restricted Stackelberg response with safety (RSRS) (Damer and Gini, 2017) was proposed to find a robust response against an opponent in normal-form games. As RNR, RSRS uses the confidence in its prediction over the opponent, however, RSRS adds a safety margin which reflects the level of risk it is willing to tolerate, which results in a trade-off between best-responding to the prediction and providing a guarantee of worst-case performance. Another algorithm which put emphasis on robustness is the online robust dynamic programming (ORDP). Yu and Mannor (2009b) presented ORDP for an extreme case of non-stationary behaviour; instead of assuming a adversary with a fixed objective ORDP assumes the opponent may play an arbitrary sequence of actions. This translates into arbi- trary variations in the reward function and arbitrary, but bounded, variations in the tran- sition probabilities. Since solving this problem is computationally expensive, ORDP has another (lazy) version which provides a trade-off between performance and computational complexity (Yu and Mannor, 2009b). 5.4 Learn opponent models Algorithms in the previous category share as deficiency that they target a specific opponent, but with limited adaptability if the opponent does not follow their assumptions. To cope with this problem algorithms in this category learn an opponent model and use it to derived an acting policy. Updating that model (and therefore the policy) is the way to keep up with against a non-stationary opponent. Recall the scenario where the environment changes (infrequently) among several station- ary modes and the agent needs to update its policy accordingly. In this scenario, Da Silva et al. (2006) proposed the the reinforcement learning with context detection (RL- CD) algorithm where the stationary environments are called contexts for which a partial model is learned (for example, using Dyna-Q; Sutton and Barto, 1998). At each time-step RL-CD decides which partial model to use according to a quality measure and when all par- 32 A Survey of Learning in Multiagent Environments tial models seem far then it starts learning a new model. Hadoux et al. (2014a) proposed an adaptation of RL-CD, replacing the quality measure by statistical tests for change-point detection, yielding RL-CD with sequential change-point detection. In a similar vein, Banerjee et al. (2017) proposed the quickest change detection (QCD) approach based on a two-threshold strategy to detect model changes in MDPs (with changes in transition and/or rewards). Previously we presented the adversarial bandit scenario (see Section 5.1). Later, we presented algorithms that use either a forget mechanism (see Section 5.2) or target a specific bandit switching scenario (see Section 5.3). Lastly, restless Markov bandits (Ortner et al., 2014) are another specific bandit scenario in which the stochastic process governing each arm does not depend on the actions of the learner, instead it depends on a Markov chain which transitions independently whether the learner pulls that arm or not. Note that in this case the problem becomes a partially observable setting (Ortner et al., 2014). Also one main characteristic of the setting is that the optimal policy cannot always be expressed in terms of arm indexes. Ortner et al. (2014) proposed to treat this problem as learning an MDP, in particular they use a modification of the URCL2 algorithm (Jaksch et al., 2010) for which they provide regret bounds. In the context of efficient adversarial exploration, the ζ-R-max algorithm (Lopes et al., 2012) extends from the classical R-max. Recall that R-max fixes a state-action pair after sufficient visitations. This has the drawback of not consider the actual empirical prediction performance or learning rate of the learner w.r.t. the data seen so far (Lopes et al., 2012). In contrast, ζ-R-max estimate the learning progress in terms of the loss over the training data used for model learning. The idea is to compute a ζ function which is based on the leave- one-out cross validation error. ζ-R-max handles changes in the environment better than R-max while also having a PAC-MDP efficient guarantee. A limitation of this approach is the computational cost of computing ζ, since it depends on the number of states and actions at every iteration. Memory-bounded opponents have been of interest in the MAL community (see Sec- tion 5.3; Powers and Shoham, 2005; Powers et al., 2007). However, previous approaches dot not actively seek to learn an opponent model. In contrast, Banerjee and Peng (2005) proposed to learn a model of those opponents whose policy is a (fixed) function of some historical window of past joint-actions by all the agents. The adversary induced MDP (AIM) (Banerjee and Peng, 2005) is a technique for repeated games which induces an MDP that implicitly has modelled the opponent (stationary) strategy. The learning agent, by knowing the MDP that the opponent induces, can compute an optimal policy π∗. These types of players can be thought of as a finite automata that take the most recent actions of the opponent and use this history to compute their policy (Munoz de Cote and Jennings, 2010). These AIM models have been used as basis to derive other learning algorithms. One of those is the convergence with model learning and safety (CMLeS) (Chakraborty and Stone, 2013). CMLeS achieves three results: (i) convergence to following a Nash equi- librium joint-policy in self-play; (ii) targeted optimality (close to best response) against memory-bounded agents whose memory size is upper bounded by a known value; and (iii) safety (ensures an individual return that is very close to its security value). Another approach that uses AIMs to model opponents is the MDP-CL (continuous learning) algorithm (Hernandez-Leal et al., 2014a). The algorithm was proposed to act 33 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote optimally against non-stationary opponents that switch among several stationary strategies. MDP-CL starts without prior models or polices and uses an exploratory phase (random actions) for a determined number of rounds. After this phase, it computes a model of the opponent in the form of an MDP which yields an optimal policy. In this point it starts learning another model (which will be used to detect changes) and after some rounds (de- fined by a parameter) the MDP-CL agent make comparisons between the learned models to evaluate their similarity. If the distance between models is greater than a given threshold, it is determined that the opponent has changed strategy and the modelling agent must restart the learning phase, resetting both models and starting from scratch with a random exploratory strategy. Otherwise, it means that the opponent has not switched strategies and the optimal policy is being used. DriftER (drift based on error rate) (Hernandez- Leal et al., 2017a) is another algorithm designed for acting against switching non-stationary opponents. DriftER uses R-max as exploratory policy instead of a random exploration and to detect switches it draws inspiration from concept drift (Widmer and Kubat, 1996). DriftER uses the learned MDP to predict the opponent actions and to keep track of their model quality. Moreover, DriftER provides guarantees of switch detection with high prob- ability (Hernandez-Leal et al., 2017a). A limitation of both DriftER and MDP-CL is that they assume a period of rounds where the opponent will remain stationary in which the model learning take place. Finally, it is worth mentioning a scenario where a switching opponent either can use a new strategy (unknown to the other agent) or a return to a previously used one, in this cases it will be useful only to learn the unknown strategy and quickly detect when it is a known strategy. BPR+ (Hernandez-Leal et al., 2016a,b) which is extension of BPR (Rosman et al., 2016) is designed for these scenarios. BPR+ assumes a non-stationary opponent that switches among stationary strategies. The algorithm starts without prior models or policies, therefore during the interaction it learns an opponent model and when the opponent changes (detected by low performance) it is stored it its memory which might be eventually useful if the opponent returns to that same strategy. 5.5 Theory of mind Approaches in the previous category learned models of other agents in the environment in order to derive an acting policy. In this last category of sophistication we present algorithms that do not only model opponents' behaviour, but also assume a strategic reasoning about the opponent, which represents a nested (or recursive) reasoning. In this category we distinguish algorithms which either are inspired by two main areas behavioural game theory and planning (see Section 6.4). In the former category we found the level-k and cognitive hierarchy models which have been used to model human in- teractions (Camerer et al., 2004; Stahl and Wilson, 1995; Costa Gomes et al., 2001). These are also known as iterative reasoning models, which refers to approach they take to make decisions. The general concept involves an initial set of zero level strategies, this is with- out strategic behaviour (for example, randomizing across all actions). Computing a best response against the lower level forms the base of the next level. However, most of these approaches have been studied only in the context of one-shot games. One exception is the work by Wunder et al. (2009) in which they a model populations consisting of agents 34 A Survey of Learning in Multiagent Environments with different reasoning levels in the iterated prisoner's dilemma. The way to act optimally against the population was obtained by best responding using a cognitive hierarchy model (Camerer et al., 2004) which was modelled as a POMDP (Littman, 1996). Sophisticated experience-weighted attraction (s-EWA) (Camerer et al., 2002) is another behavioural game theory algorithm inspired by fictitious play. It assumes two types of opponents, (simple) adaptive opponents (using the EWA, see Camerer and Ho, 1999) and sophisticated opponents that rationally best-responds to her forecasts of all other behaviours (they use the s-EWA algorithm). A limitations is that it has been only studied in the context of short repeated games (less than 10 rounds). In the planning category, one of the earliest approaches proposed by Gmytrasiewicz and Durfee (2000) is the recursive modelling method (RMM). They propose a specialized knowledge representation in the form of reward matrices that allows using a recursive rea- soning to obtain the best coordinated action in a MAS system. An approach inspired in RMM, but with a formal decision theoretic background are the interactive POMDPs (I-POMDPs) (Gmytrasiewicz and Doshi, 2005). They are called interactive because the model considers what an agent knows and believes about what another agent knows and believes (Aumann, 1999). This means that an agent will have a model of how it believes an- other agent reasons. I-POMDPs extend POMDPs incorporating models of other agents into the regular state space. The main limitation of these models is its inherent complexity, since solving one I-POMDP with M number of models considered in each level, with (cid:96) maximum reasoning levels, is equivalent to solving O(M (cid:96)) POMDPs (Seuken and Zilberstein, 2008). Despite these issues, there are recent algorithms for online learning (Ng et al., 2012). Also there are works using I-POMDPs with more than a thousand of agents (Sonu et al., 2015) and even in experiments with humans (Doshi et al., 2010). Parametrized I-POMDPs (PI-POMDPs) (Wunder et al., 2011, 2012) are an approach which combines I-POMDPs with the iterative reasoning models. The idea is to compute a policy that maximizes the score against either a distribution over previous levels, or a selection of agents from those levels, by solving the POMDP formed by them. While computationally expensive it pro- vides a clear formalism to work showing good results in highly adaptive domains, such as the lemonade stand game (Zinkevich et al., 2011). However, further work is needed to show the applicability to other domains. Lastly, another theory of mind model was proposed by de Weerd et al. (2013). Here, the zero-level is composed of beliefs indicating the likelihood of the opponent taking any action at any state, higher order models are generated based on the information from lower levels. Additionally, they use a confidence value which helps the agent to adapt to different opponents (with different levels of reasoning). Recently, an extension to more than one opponent, multiagent ToM (MToM), was proposed by Van der Osten et al. (2017). To cope with this challenge the authors propose a stereotyping mechanism (clustering), which segments the agent population into sub-groups of agents with similar behaviour; these groups are then treated as single agents. We have presented the five categories of how algorithms deal with non-stationary and classified state of the art algorithms with different characteristics. The next section presents the strengths and limitations of each category, related areas to multiagent learning, and pinpoints open avenues for future research. 35 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote 6. Discussion We have presented five categories of how learning algorithms deal with non-stationary be- haviour. In this section we start by discussing their strengths and limitations (see Sec- tion 6.1). Then, we mention the most common experimental domains that have been used (see Section 6.2), we outline the current theoretical results (see Section 6.3), and we describe related areas of research (see Section 6.4). We conclude with exploring promising avenues of future research (see Section 6.5). 6.1 Strengths and limitations of each category We briefly mention some advantages and limitations for each category and provide pointers for when each category is especially useful. Ignore. These algorithms are widely known in the community and most of them do not need to known extra information from the opponent (opponents payoffs). However, a large drawback is that most of them lose their theoretical guarantees when used in non-stationary environments (e.g., Q-learning). We advise to use this algorithms where no extra information can be obtained from the environment. Forget. One advantage of these algorithms, in contrast to the previous category, is that they do take into account the non-stationarity of a multiagent system. In general, these algorithms are model-free approaches with the limitation that they might take longer periods to converge to a solution (Suematsu and Hayashi, 2002). These algo- rithms could be used when no a priori information is known about the opponent and there are no constraints in the learning time. Respond to target opponents. If the opponent is restricted to a single class (i.e., worst case opponent, see Littman, 1994; stochastically changing among models, see Choi et al., 1999; converging to a Nash equilibrium, see Hu and Wellman, 1998) then algorithms in this category offer an efficient solution. A limitation is the constrained adaptability of these algorithms. For example, if we expect the opponent to use a wider set of strategies then the solution is to directly learn a model of the opponent. Learn opponent models. A main advantage of these algorithms is that the learned model of the opponent can be reused if the opponent returns to the same strategy (Da Silva et al., 2006; Hernandez-Leal et al., 2016a). Since these algorithms are model-based, they usually learn faster than other approaches. One limitation is that they need the opponent to remain stationary for a long enough period to model them, which can be unrealistic in some scenarios. Theory of mind. An interesting feature of algorithms in this category is that they are readily available to model populations (more than 2 agents) since that is the intrinsic way they obtain an acting policy (Camerer et al., 2004; Wunder et al., 2009, 2012). Another characteristic of these algorithms is that they perform a complex strategic reasoning process, which necessitates high computational costs to solve them (e.g., I-POMDPs, see Seuken and Zilberstein, 2008). Also, these approaches have been 36 A Survey of Learning in Multiagent Environments studied mostly for predicting behaviour in unrepeated games (Wright and Leyton- Brown, 2014). 6.2 Experimental domains and applications Most testing scenarios for multiagent interactions use the formal models of game theory, from extensive-form games, repeated games to stochastic games. However, there is also another category of specific applications, such as negotiation, smart grids, and routing problems. Extensive-form games The classic game of poker has different variations ranging from simple to complex (in terms of the state space and action space) which have been used to evaluate different opponents. Kuhn poker is a tiny, toy variant of poker. The game involves two players, two actions and a three card deck. This game has been studied previously since the two players strategies can be summarized in two or three parameters (Hoehn et al., 2005; Bard and Bowling, 2007). Leduc hold'em Poker is a larger version than Khun Poker in which the deck consists of six cards (Bard et al., 2015). Heads-up limit Texas hold'em is more complex variation, where the game tree consists of approximately 9.17×1017 states (Johanson et al., 2007). Given the size of the domain, algorithms have focused on dealing with this problem (Bard et al., 2013). Repeated games It is common to use repeated games as a setting with non-stationary opponents (Suematsu and Hayashi, 2002; Bowling and Veloso, 2002; Tesauro, 2003; Wein- berg and Rosenschein, 2004; Powers and Shoham, 2005; Conitzer and Sandholm, 2006; Ab- dallah and Lesser, 2008; Crandall and Goodrich, 2011; Elidrisi et al., 2012; Hernandez-Leal et al., 2013, 2014a, 2016a; Damer and Gini, 2017). The most simple games have two players and two actions (2x2); a 3x3 example is rock-scissors-paper. Also, it is common to eval- uate learning algorithms in randomly generated games according to certain specifications such as zero-sum games or, constant-sum games (Nudelman et al., 2004). Previous works have performed experimental comparisons among different multiagent learning algorithms in repeated games (Bouzy and M´etivier, 2010). One interesting competition which can be represented as a repeated game is the lemon- ade stand game (Zinkevich et al., 2011). Here, three agents (vendors) interact by choosing a position (12 different actions) on an "island" in order to sell lemonade to the island's population. The rewards depend on the actions of all the agents and several interesting al- gorithms were developed in this context where fast adaptation was needed (Wunder et al., 2010; Munoz de Cote et al., 2010; Sykulski et al., 2010; Wunder et al., 2011). Stochastic games This type of games generally poses a more difficult challenge than repeated games since there are different states (games) with probabilistic transitions (see Section 2.3). Many stochastic games represent grid-worlds, where agents need to take strategic decisions. For example, a mini-version of the sports game soccer was proposed as a stochastic game played on a 4x5 grid with five actions and two players, an attacker and the goal keeper (Littman, 1994). In this game, agents must use a probabilistic policy to obtain higher rewards (Littman, 1994; Bowling and Veloso, 2002). Other interesting games are stochastic versions of well-known games such as PD, coordination, and chicken (Munoz de Cote and Littman, 2008; Elidrisi et al., 2014). 37 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote Other domains Lastly, many algorithms have been evaluated in specific applications ranging from aerospace to security and surveillance; for a complete survey about the impact of MAS applications refer to Muller and Fischer (2014). A typical situation where non-stationary multi-agent learning plays an important role is automated negotiation and e-negotiation systems (Jennings et al., 2001; Kraus, 2001). Recent examples include the setting of e-commerce (He et al., 2003; Kowalczyk et al., 2003), virtual agents (DeVault et al., 2015; Gratch et al., 2015) and games such as diplomacy (Fab- regues et al., 2010) and coloured trails (Gal et al., 2005; Lin and Kraus, 2010). As with human negotiations, automated negotiation between agents is a non-stationary game with incomplete information, where the agents initially do not know their opponents prefer- ences and where strategies can change over time (for a survey on learning in negotiation, see Baarslag et al., 2016). As a result, they need to derive information from the exchange of offers with each other. Although rarely framed in the context of non-stationary learning, many automated negotiation strategies have been formulated that take advantage of non-stationary learning mechanisms. An important category is preference learning, in which agents aim to learn aspects of the opponent's preference profile by engaging in online opponent model learning in an effort to reach Pareto optimal (win-win) outcomes (e.g., Coehoorn and Jennings, 2004; Hindriks et al., 2009; Baarslag et al., 2013b; Zhang et al., 2015). Learning the opponent's negotiation strategy is another important aspect, which boils down to determining counter- offers in subsequent negotiation states. The agents face the challenge of a wide diversity of possible negotiation strategies and the fact that the opponent can change behaviour dynamically according to the offers received (Hou, 2004; Baarslag et al., 2011). That is, learning the opponent's strategy is a moving target problem, where the agent simultaneously seeks to acquire new knowledge about the opponent while the agent needs to optimize its negotiation actions based on the current model. In the negotiation literature, responding to target opponents is a opponent model classification problem, where the type of the opponent needs to be determined from a range of possibilities given its negotiation behaviour (Lin et al., 2008). There also exist simple ignore and forget strategies that either assume a stationary environment or only employ recent data, for example negotiation tactics that take into account elapsed time only (Faratin et al., 1998). More recently, automated negotiators have even been endowed with (second-order) theory of mind, so that agents can reason about what the opponent believes about their beliefs (de Weerd et al., 2015; Pynadath et al., 2013). An important negotiation domain involves smart energy grids and their trading markets used to buy and sell energy. The Power TAC simulator (Ketter et al., 2013) models a complex a dynamic energy system in this context, where different brokers can take actions in three markets. One of those is the wholesale market, which is a particular type of auction. The non-stationary behaviour appears when there are brokers that switch among different strategies through time (Hernandez-Leal et al., 2015). Another example is the problem of predicting the energy demand of users, which involves randomness and changes in behaviour (Marinescu et al., 2014, 2015). Routing problems have been also treated as a domain with non-stationary behaviour. In domain routing, an ISP operator has the opportunity to increase its revenue by charging external domains for the traffic transiting on its links. Moreover, agents must be able to deal with a non-stationary environment when the optimal price setting varies according to 38 A Survey of Learning in Multiagent Environments other ISPs' strategies and the network load (Vrancx et al., 2015). In the context of smart- cities, there are different routing problems that model non-stationary behaviour, such as traffic networks. In this case, the world is represented as a grid, with traffic lights on each junction and patterns of traffic representing different stationary environments (Choi et al., 1999; Da Silva et al., 2006). This section presented experimental domains commonly used in non-stationary environ- ments while the next section focuses on theoretical results. 6.3 Theoretical results In this section we outline different theoretical results presented in the context of learning in non-stationary environments. Regret bounds. Multi-armed bandits algorithms, usually provide regret bounds for dif- ferent algorithms and different types of scenarios (adversarial, stochastic, Markov chain; see Auer et al., 2002a; Auer, 2002; Auer et al., 2002b; Yu and Mannor, 2009a; Garivier and Moulines, 2011; Ortner et al., 2014; Besbes et al., 2014). Few algorithms provide regret bounds for sequential decision problems (Yu and Mannor, 2009b) or multiagent scenar- ios (Bowling, 2004). Efficient exploration guarantees. Another category of theoretical results comprises those algorithms which provide efficient exploration guarantees (for example, using sample complexity results; see Kakade, 2003) in adversarial stationary environments (Brafman and Tennenholtz, 2003) and non-stationary ones (Lopes et al., 2012; Hernandez-Leal et al., 2017b). Convergence to Nash equilibrium. A large group of algorithms has provided guaran- tees to converge to a NE under slightly different conditions: only local rewards (Abdallah and Lesser, 2008), partial observations (Conitzer and Sandholm, 2006), complete informa- tion settings (Bowling and Veloso, 2002; Hu and Wellman, 1998; Littman, 2001; Suematsu and Hayashi, 2002). Most of these algorithms assume NE only in self-play (Hu and Well- man, 1998; Bowling and Veloso, 2002; Banerjee and Peng, 2004; Chakraborty and Stone, 2013) or variations of self-play (Bowling, 2004). Best response. Q-learning loses its guarantees (convergence to an optimal policy) in non-stationary environments. Because of that, most algorithms try to improve on that re- gard. For example, by still having guarantees in stationary environments, but also better suited for non-stationary environments (Abdallah and Kaisers, 2016). Other address di- rectly non-stationary opponents and prove that will learn a best response policy (Tesauro, 2003; Weinberg and Rosenschein, 2004; Chakraborty and Stone, 2013). Robustness guarantees. Another common result is to assess the robustness of an algo- rithms by providing guarantees of safety, security or no-exploitability in the form of expected rewards (Littman, 1994; Johanson et al., 2007; Johanson and Bowling, 2009; Powers et al., 2007; Crandall and Goodrich, 2011; Chakraborty and Stone, 2013; Elidrisi et al., 2014; Damer and Gini, 2017) or regret bounds (Yu and Mannor, 2009b; Besbes et al., 2014). A different class of results is to provide switch detection guarantees against non-stationary opponents (Hernandez-Leal et al., 2017a) which makes the method robust. 39 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote Figure 7: Types of concept drift that change over time (Gama et al., 2014): (i) sudden, (ii) incre- mental, (iii) outliers, (iv) gradual and (v) recurrent. Next, we introduce different areas and paradigms that share a connection with learning in the presence of non-stationary behaviour. Later, we present open avenues of future research. 6.4 Related areas This section presents concepts and areas that might be useful to take into consideration when developing new algorithms. Supervised learning and concept drift. The machine learning community has devel- oped an area related to non-stationary environments and online learning which is called concept drift (Widmer and Kubat, 1996). The approach is similar to a supervised learn- ing scenario where the relation between the input data and the target variable changes over time. Gama et al. (2014) presented an survey of this problem where different types of concept drift where categorized as depicted in Figure 7 (using a one-dimensional data where changes happen in the data mean). (i) A change may happen suddenly/abruptly (from one time-step to the next). (ii) Incrementally, where there is a window of time where intermediate concepts appear. (iii) Outliers or noise, which refers to random deviation or anomaly, in which case no adaptation is needed. (iv) Gradually, where the concepts al- ternate one to another until finally converging to a different one. (v) Recurring, where previously seen concepts may reappear after some time. Concept drift scenarios are related to non-stationary environments, however they need to be adapted to a multiagent setting where there is a need for exploration in the form of action selection and uncertainty due to opponent's actions. However, some work in multiagent learning have drawn inspiration from concept drift (Hernandez-Leal et al., 2017a). Transfer learning. RL has been shown successful in many domains when a single agent is performing a single task (with the appropriate learning time). However, when having different tasks the basic approach is to learn a completely new model. To reduce this time consuming process, transfer learning algorithms use the experience gained in learning to perform one task to improve learning performance in a related, but different, task (Taylor and Stone, 2009). This is especially important in some types of non-stationary environ- ments. For example, in case of recurring changes (see Figure 7), previous information (for example, in the form of models or policies) will be useful to quickly have an acting 40 TimeData meansudden/abruptincrementalgradualoutlier (no concept drift)recurring concepts(i)(iii)(ii)(iv)(v)A Survey of Learning in Multiagent Environments policy. These ideas (e.g., reusing past policies) have inspired recent works on multiagent systems (Hernandez-Leal et al., 2016a; Hernandez-Leal and Kaisers, 2017b). Multiagent interaction without prior coordination. Stone et al. (2010) presented the challenge of ad-hoc teamworks, this is, to create an autonomous agent that is able to effi- ciently and robustly collaborate with previously unknown teammates on tasks to which they are all individually capable of contributing as team members. Similarly, ad-hoc coordination is the problem of designing an agent that is able to be flexible and efficient in a multiagent system that admits no prior coordination among the agents (Albrecht and Ramamoorthy, 2013). This active line of research (Barrett and Stone, 2014; Melo and Sardinha, 2016; Albrecht et al., 2016b; Liemhetcharat and Veloso, 2017; Chakraborty et al., 2017) is related since the agents involved can be of different types (heterogeneous agents) and they can have different adaptation behaviours which posses a problem since prior coordination is restricted. Partial observability and planning MDPs are the main model used by RL algorithms. However, there are other related models which are particularly relevant to the multiagent community. When cooperative teams of agents are planning in uncertain domains, they must coordinate to maximise their (joint) team value, in this scenario the multiagent Markov decision processes (MMDPs) (Boutilier, 1996) are useful. This model is a n−person stochas- tic game where the payoff function is the same for all agents. Currently there is undergoing research for reducing the costs related to computing these models (Scharpff et al., 2016). POMDPs, partially observable MDPs (Kaelbling et al., 1998) are models where it is no longer the case that the agent has full perception capabilities. Instead, there is probability distribution over observations. In this way, it is possible to model problems in a more real- istic way, the downside is that solving a POMDP is computationally more expensive than an MDP (Papadimitriou and Tsitsiklis, 1987). Note that a particular case of a POMDP is the HM-MDP (see Section 5.3). A generalization of POMDPs to a multiagent scenario with cooperative agents (since they need to share they utility function) are decentralized POMDPs (Seuken and Zilberstein, 2008). One limitation is its complexity which is NEXP- complete (Seuken and Zilberstein, 2008). Recent works have proposed different methods to overcome this limitation, for example by searching in the influence space (i.e., the space that represents probabilistic effects that agent policies may exert on one another, see Witwicki et al., 2012; Oliehoek et al., 2015). Evolutionary game theory. The application of game theoretic reasoning to the study of populations, initially to understand biological processes such as evolution, has received its own designation as evolutionary game theory (Weibull, 1995). Initial work bringing this field towards multi-agent learning algorithms has established the formal link between the simple reinforcement learning algorithm cross learning and the replicator dynamics, a central con- cept in evolutionary game theory (Borgers and Sarin, 1997). This has inspired a stream of follow-up work that links stochastic multi-agent reinforcement learning algorithms to vari- eties of deterministic dynamical systems, as summarized in a related survey (Bloembergen et al., 2015). The principle methodology is taking the limit of infinitesimal learning rates, and studying the resulting dynamical system to gain insight into the emergent behaviour of the multi-agent system, such as its convergence, stability and resilience. Additional in- 41 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote terest is given to each equilibrium's basin of attraction and resulting welfare, providing an assessment of the anticipated joint interaction outcome. Behavioural game theory. Many models proposed from a game theoretic approach do not accurately predict human-behaviour in many experiments (Kahneman and Tversky, 1979; Goeree and Holt, 2001). New models that take into account human characteristics (e.g., fairness, reciprocity, deception) were grouped under the name of behavioural game theory (Camerer et al., 2004). Even when these models tend to obtain good results in one-shot games with human populations (Wright and Leyton-Brown, 2010) these models are still not well studied in repeated games or sequential decisions problems. Having mentioned closely related areas, we now present some interesting avenues for future research. 6.5 Open questions and promising avenues of future research Although learning in multiagent systems has been an active research area in the past years there are still many open questions. In this section, we present four promising lines of research and we give example research questions that fall within each line. In a previous survey, Tuyls and Weiss (2012) presented three main challenges in MAS. We pinpoint some connections between those challenges and our proposed lines of research. In particular, for the "extending the scope of MAL" challenge, we propose ideas in the context of diversity in opponents (see Line 1), dynamic interactions (see Line 2) and ap- plications (see Line 4). Similarly, for the "classification limitations" challenge (a lack of classifications of what is missing in MAL), we proposed two ideas related to learning objec- tives (see Line 3). Line 1: Diversity in opponents • Heterogeneous learning agents. In real settings, one might encounter several agents with different learning characteristics, objectives, actuators, and representation of the world (including sensors). This heterogeneity is one of the most (if not the most) important complicating factors in acting optimally. One way to cope with such rich and complex environments is to characterise them (i.e., the set of learning opponents) across different labels, like diversity (i.e., how many types of learning agents), type distribution (i.e., the density distribution function) and set of learning techniques; e.g., if the learning agents are mostly using regret minimization or reinforcement learning or Bayesian non-parametrics, to name a few examples. This is an important strategy that has been used in the negotiation literature where agents can establish optimal bidding strategies against specific types of opponents encountered in the environment (Matos et al., 1998; Baarslag et al., 2013a). In this way, one can constrain solutions to some well-defined subset of multiagent environments. We encourage new algorithms to frame their work in the context of our proposed framework (see Section 3.2) which naturally accounts for heterogeneous opponents. • Modelling populations. There are many complications when interacting with many agents, and for this reason, most algorithms use few agents in the environment. How- ever, using those same algorithms could become intractable in large multiagent do- mains. To obtain efficient and scalable algorithms one would need to sacrifice detail 42 A Survey of Learning in Multiagent Environments by generalising the system to a population level, in a way to best respond to classes of populations rather than individuals (Wunder et al., 2011; Bard et al., 2015; Hernandez- Leal and Kaisers, 2017a; Van der Osten et al., 2017). A different approach is to deter- mine the degree of interaction among agents, this could help in defining whether to interact with an agent or ignore it and take it as part of the environment (De Hauwere et al., 2010; Yu et al., 2015). • Unknown world knowledge by opponents. Algorithms in the learn category assume the agent is aware of the knowledge of all opponents, i.e., attributes or features that correctly describe the opponents' observations of the world. However, in most real situations this information is not really accessible (Chakraborty et al., 2013). To re- lax this assumption, the agent needs to learn the model and at the same time the correct knowledge representation (Maillard et al., 2013). A possibility to learn with- out putting effort into designing the correct representation is to use deep learning techniques (Deng and Yu, 2013; Mnih et al., 2015). Another option to dealing with uncertain world representations by the opponents is to keep a set of known represen- tations, as in (Hernandez-Leal et al., 2016a), and infer the correct one by maintaining (and updating) a probability distribution over the set. This can be naturally modelled in the proposed framework (see Section 3.2) where beliefs over opponent behaviors are two main components. Line 2: Dynamic interactions • Learning in multiple concurrent interactions. Many multiagent learning algorithms assume interactions occur synchronously and among all agents. However, in real-world scenarios this is not always the case where interactions are usually asynchronous with different agents taking different response times. This holds especially true in large multi-agent coordination and negotiation systems where multiple, concurrent threads have to be coordinated. Communication protocols for committing and decommitting to deals have only been studied recently (Ito et al., 2009; Williams et al., 2012). It is still an open questions whether current learning algorithms will work under these slightly different conditions. • Intelligent reuse of information to reduce learning times. Learning a model of the other agents in the environment is a way to solve the non-stationarity problem. How- ever, this learning process usually requires a large period of repeated interactions, which is unreasonable in many scenarios. To alleviate this problem, information from previous interactions can be reused. For example, by generating a "portfolio" of the possible opponents (in an offline phase) and during the interaction estimate which is the most similar and act with a respective policy (Bard et al., 2013; Barrett et al., 2013; Albrecht et al., 2016a; Hernandez-Leal et al., 2016a), please refer to our pro- posed framework which naturally models this approach (see Section 3.2). Similarly, areas derived from transfer learning (Taylor and Stone, 2009) could be extrapolated to multiagent scenarios such as curriculum learning (Svetlik et al., 2016; Narvekar et al., 2017) where existing techniques work for a single agents (independently) and therefore an open question is to reuse information from different agents. Giving advice 43 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote to agents (Torrey and Taylor, 2013; Zhan et al., 2016) is another example in which multiagent algorithms are still in an early stage of development (da Silva et al., 2017). • Interaction against a dynamic number of opponents. In multiagent systems, the num- ber of agents in the environment is usually fixed before the interaction and remains constant during the interaction. However, it is possible to consider the opponents may come and go during the interaction (e.g., dynamic coalition formation) which will affect the environment and most probably the acting policy. One idea on how to model these type of scenarios is to model each (dis)appearance of the opponents as a switch in the environment and use algorithms designed for these cases (Da Silva et al., 2006; Hernandez-Leal et al., 2016a, 2017a). • Exploratory learning noise. The assumption to not explicitly model other agents and just consider them as part of the environment presents different problems. One of those happens when many learning agents explore at the same time creating noise to the rest, this is called exploratory action noise (HolmesParker et al., 2014; Munoz de Cote et al., 2006). To alleviate this problem different methods have been proposed (Tumer and Agogino, 2007; HolmesParker et al., 2014). Recently, the same problem appeared in a deep multiagent RL setting (Foerster et al., 2017a) and the proposed solution was based on previous work in the area (Tumer and Agogino, 2007). However, this problem is more complicated to solve in scenarios where no coordination or cooperation is possible. Line 3: Learning objectives • Tracking vs convergent algorithms -- transient performance. One way to categorise learning algorithms is to divide those that aim to converge to the best result (to an optimal policy) and those that only track the payoff of different solutions (no convergence guarantees). An analysis of these two approaches in a stationary task found that in certain cases a tracking algorithm obtains better results than one that converges to the optimal policy (Sutton et al., 2007). This is especially important when dealing with non-stationary environments. When a change in the environment occurs a converging algorithm may take longer to overcome this issue (Claus and Boutilier, 1998) and one that is only tracking will be able to adapt faster. This relates to the transient performance, where usually algorithms are more concerned with the results of learning than with the ongoing process of learning (Sutton et al., 2007). • Tolerated and induced non-stationarity. There are different types of theoretical re- sults in multiagent learning (e.g., convergence, optimality, non-exploitability; see Sec- tion 6.3). However, more general convergence results are needed and we propose two concepts that are worth analysing in MAS: tolerated non-stationarity, this is, how much non-stationarity does an algorithm accepts without sacrificing optimality; and induced non-stationarity, this is, how much non-stationarity an algorithm induces in the system. 44 A Survey of Learning in Multiagent Environments Line 4: Applications • Negotiation and MAS. As described in Section 6.2, negotiation is an interesting and real-world scenario to model multiagent interactions. However, generic negotiation using reinforcement learning seems an understudied subject with few works in the intersection (e.g., Lazaric et al., 2007), as most research in this area seems to have focused on Q-learning for trading agents in competitive market places so far (Hsu and Soo, 2001; Tesauro and Kephart, 2002). It would be interesting to employ a number of techniques mentioned in this survey (e.g., Johanson et al., 2007; Crandall and Goodrich, 2011; Babes et al., 2009; Hernandez-Leal et al., 2016b) in order to improve generic preference learning and strategy estimation in automated negotiation, both in bilateral and multilateral settings. This remains an unsolved challenge in a non- stationary setting in which preference evolution can occur, for example with regard to risk tolerance or fairness attitudes (Baarslag et al., 2017). • Deep RL and MAS. Deep learning (Bengio, 2009) has shown outstanding results when combined with reinforcement learning (Mnih et al., 2015; Silver et al., 2016). Even though most works assume a single-agent setting, problems with non-stationarity have already appeared, proposing extensions of existing algorithms that handle non- stationary environments in the deep learning setting. In particular, since deep learning approaches require large numbers of samples, common techniques such as experience replay have been adapted to handle non-stationarity (Foerster et al., 2017b; Cas- taneda, 2016). Moreover, deep multi-agent RL works are on the rise (He et al., 2016; Foerster et al., 2016, 2017a; Leibo et al., 2017; Gupta et al., 2017; Tampuu et al., 2017) with the obvious challenge of handling non-stationary environments (i.e., multi- ple learning agents). While these initial works have transferred a number of individual techniques to the deep setting, it remains an open challenge to provide a conceptual framework for deep multi-agent learning. Above, we have presented relevant open problems with potential impact on the multiagent community. The next section presents the conclusions drawn from this survey. 7. Conclusions Non-stationary environments in sequential decision making tasks have received attention from research in the domains of game theory, reinforcement learning and multi-armed ban- dits. This survey has reviewed a wide range of algorithms from these fields, and contributes a structure to think clearly about otherwise often implicit assumptions, characteristics and concepts related to the challenges of multiagent learning (see Section 3). First, we pro- posed a new framework for reasoning about multiagent systems (see Section 3.2). Then, we identified several principled approaches that algorithms take to deal with non-stationarity: ignore, forget, respond to target opponents, learn opponent models and theory of mind (see Section 3.3). For each category we provide an illustrative example (see Section 4) and later we present an extensive list of state-of-the-art algorithms classified into these categories (see Section 5). Moreover, we identified the strengths and limitations of each category and provide guideline scenarios when they should be applied (see Section 6.1). 45 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote We observed that most experimental results are formalised in terms of repeated games and stochastic games (see Section 6.2). Theoretical results are diverse and include: guaran- tees to learn optimal policies, non-exploitability guarantees and convergence to equilibria, to name a few (see Section 6.3). Following the coherent review of the state of the art, this survey pinpoints the remaining open questions and presents them clustered into four open avenues for promising future research: diversity in opponents, dynamic interactions, learning objectives and applications (see Section 6.5). While much progress has been achieved over the last decades, further fundamental research is required for the breakthrough guarantees and demonstration of algorithmic per- formance in non-stationary environments. This survey seeks to facilitate this future work by highlighting current gaps in the literature and providing the guideline taxonomy to position future work within it. Acknowledgments We would like to thank Frans Oliehoek and Daan Bloembergen for useful discussions and suggestions. This work is part of the Veni research programme with project num- ber 639.021.751, which is financed by the Netherlands Organisation for Scientific Research (NWO). References Sherief Abdallah and Michael Kaisers. Addressing the Policy-bias of Q-learning by Repeat- ing Updates. In Proceedings of the 12th International Conference on Autonomous Agents and Multi-agent Systems, pages 1045 -- 1052, Saint Paul, MN, USA, 2013. Sherief Abdallah and Michael Kaisers. Addressing Environment Non-Stationarity by Re- peating Q-learning Updates. The Journal of Machine Learning Research, 17:1 -- 31, April 2016. Sherief Abdallah and Victor Lesser. A multiagent reinforcement learning algorithm with non-linear dynamics. Journal of Artificial Intelligence Research, 33(1):521 -- 549, 2008. Stefano V. Albrecht and Subramanian Ramamoorthy. A game-theoretic model and best- response learning method for ad hoc coordination in multiagent systems. In Proceedings of the 12th International Conference on Autonomous Agents and Multi-agent Systems, pages 1155 -- 1156, Saint Paul, MN, USA, May 2013. Stefano V. Albrecht, Jacob W. Crandall, and Subramanian Ramamoorthy. Belief and truth in hypothesised behaviours. Artificial Intelligence, 235:63 -- 94, 2016a. Stefano V. Albrecht, Somchaya Liemhetcharat, and Peter Stone. Special issue on multiagent interaction without prior coordination: guest editorial. Autonomous Agents and Multi- Agent Systems, 31(4):765 -- 766, December 2016b. 46 A Survey of Learning in Multiagent Environments Raman Arora, Ofer Dekel, and Ambuj Tewari. Online Bandit Learning against an Adap- tive Adversary: from Regret to Policy Regret. In Proceedings of the 29th International Conference on Machine Learning, pages 1503 -- 1510, Edinburgh, Scotland, June 2012. Peter Auer. Using confidence bounds for exploitation-exploration trade-offs. Journal of Machine Learning Research, 3:397 -- 422, 2002. Peter Auer, Nicol`o Cesa-Bianchi, and Paul Fischer. Finite-time Analysis of the Multiarmed Bandit Problem. Machine Learning, 47(2/3):235 -- 256, 2002a. Peter Auer, Nicol`o Cesa-Bianchi, Y Freund, and Robert E. Schapire. The nonstochastic multiarmed bandit problem. SIAM Journal on Computing, 32(1):48 -- 77, 2002b. Robert J. Aumann. Subjectivity and correlation in randomized strategies. Journal of Mathematical Economics, 1(1):67 -- 96, March 1974. Robert J. Aumann. Interactive epistemology I: knowledge. International Journal of Game Theory, 28(3):263 -- 300, 1999. Robert Axelrod and William D. Hamilton. The evolution of cooperation. Science, 211(27): 1390 -- 1396, 1981. Tim Baarslag, Koen V. Hindriks, and Catholijn M. Jonker. Towards a quantitative concession-based classification method of negotiation strategies. In David Kinny, Jane Yung-jen Hsu, Guido Governatori, and Aditya K. Ghose, editors, Agents in Principle, Agents in Practice, volume 7047 of Lecture Notes in Computer Science, pages 143 -- 158, Berlin, Heidelberg, 2011. Tim Baarslag, Katsuhide Fujita, Enrico H. Gerding, Koen V. Hindriks, Takayuki Ito, Nicholas R. Jennings, Catholijn M. Jonker, Sarit Kraus, Raz Lin, Valentin Robu, and Colin R. Williams. Evaluating practical negotiating agents: Results and analysis of the 2011 international competition. Artificial Intelligence, 198:73 -- 103, May 2013a. Tim Baarslag, Mark J.C. Hendrikx, Koen V. Hindriks, and Catholijn M. Jonker. Predicting the performance of opponent models in automated negotiation. In International Joint Conferences on Web Intelligence (WI) and Intelligent Agent Technologies (IAT), 2013 IEEE/WIC/ACM, volume 2, pages 59 -- 66, Nov 2013b. Tim Baarslag, Mark J.C. Hendrikx, Koen V. Hindriks, and Catholijn M. Jonker. Learning about the opponent in automated bilateral negotiation: a comprehensive survey of oppo- nent modeling techniques. Autonomous Agents and Multi-Agent Systems, 30(5):849 -- 898, 2016. Tim Baarslag, Michael Kaisers, Enrico H. Gerding, Catholijn M. Jonker, and Jonathan Gratch. When will negotiation agents be able to represent us? the challenges and op- portunities for autonomous negotiators. In Proceedings of the 26th International Joint Conference on Artificial Intelligence, 2017. 47 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote Monica Babes, Michael Wunder, and Michael L. Littman. Q-learning in two-player two- action games. In Proceedings of the 8th International Conference on Autonomous Agents and Multiagent Systems, Budapest, Hungary, 2009. Bikramjit Banerjee and Jing Peng. Performance bounded reinforcement learning in strategic interactions. In Proceedings of the 19th Conference on Artificial Intelligence, pages 2 -- 7, San Jose, CA, USA, 2004. Bikramjit Banerjee and Jing Peng. Efficient learning of multi-step best response. In Proceed- ings of the 4th International Conference on Autonomous Agents and Multiagent Systems, pages 60 -- 66, Utretch, Netherlands, 2005. Taposh Banerjee, Miao Liu, and Jonathan P How. Quickest Change Detection Approach to Optimal Control in Markov Decision Processes with Model Changes. In Proceedings of American Control Conference, 2017. Nolan Bard and Michael Bowling. Particle filtering for dynamic agent modelling in simplified poker. In Proceedings of the 22nd Conference on Artificial Intelligence, pages 515 -- 521, Vancouver, Canada, 2007. Nolan Bard, Michael Johanson, Neil Burch, and Michael Bowling. Online implicit agent modelling. In Proceedings of the 12th International Conference on Autonomous Agents and Multiagent Systems, pages 255 -- 262, Saint Paul, MN, USA, May 2013. Nolan Bard, Deon Nicholas, Csaba Szepesv´ari, and Michael Bowling. Decision-theoretic In Proceedings of the 14th International Conference on Au- Clustering of Strategies. tonomous Agents and Multiagent Systems, pages 17 -- 25, Istanbul, Turkey, May 2015. Samuel Barrett and Peter Stone. Cooperating with Unknown Teammates in Complex Domains: A Robot Soccer Case Study of Ad Hoc Teamwork. In Proceedings of the 29th Conference on Artificial Intelligence, pages 2010 -- 2016, Austin, Texas, USA, December 2014. Samuel Barrett, Peter Stone, Sarit Kraus, and Avi Rosenfeld. Teamwork with Limited Knowledge of Teammates. In Proceedings of the Twenty-Seventh AAAI Conference on Artificial Intelligence, pages 102 -- 108, Bellevue, WS, USA, 2013. Richard Bellman. A Markovian decision process. Journal of Mathematics and Mechanics, 6(5):679 -- 684, 1957. Yoshua Bengio. Learning Deep Architectures for AI. Foundations and Trends R(cid:13) in Machine Learning, 2(1):1 -- 127, 2009. Daniel S. Bernstein, Eric A Hansen, Shlomo Zilberstein, and Christopher Amato. Dynamic programming for partially observable stochastic games. In Proceedings of the National Conference on Artificial Intelligence, pages 709 -- 715, December 2004. Omar Besbes, Yonatan Gur, and Assaf Zeevi. Stochastic multi-armed-bandit problem with non-stationary rewards. Advances in Neural Information Processing Systems, pages 199 -- 207, 2014. 48 A Survey of Learning in Multiagent Environments Alina Beygelzimer, John Langford, Lihong Li, Lev Reyzin, and Robert E. Schapire. Con- textual Bandit Algorithms with Supervised Learning Guarantees. In Proceedings of the 14th International Conference on Artificial Intelligence and Statistics, pages 19 -- 26, Fort Lauderdale, FL, USA, 2011. Daan Bloembergen, Michael Kaisers, and Karl Tuyls. Lenient frequency adjusted Q- In Proceedings of the 22nd Belgian/Netherlands Artificial Intelligence Con- learning. ference, December 2010. Daan Bloembergen, Karl Tuyls, Daniel Hennes, and Michael Kaisers. Evolutionary Dynam- ics of Multi-Agent Learning: A Survey. Journal of Artificial Intelligence Research, 53: 659 -- 697, 2015. Tilman Borgers and Rajiv Sarin. Learning Through Reinforcement and Replicator Dynam- ics. Journal of Economic Theory, 77(1):1 -- 14, November 1997. Craig Boutilier. Planning, learning and coordination in multiagent decision processes. In Proceedings of the 6th conference on Theoretical aspects of rationality and knowledge, pages 195 -- 210, De Zeeuwse Stromen, The Netherlands, 1996. Bruno Bouzy and Marc M´etivier. Multi-agent learning experiments on repeated matrix games. In Proceedings of the 27th International Conference on Machine Learning, pages 119 -- 126, Haifa, Israel, 2010. Michael Bowling. Convergence and no-regret in multiagent learning. In Advances in Neural Information Processing Systems, pages 209 -- 216, Vancouver, Canada, 2004. Michael Bowling and Manuela Veloso. Multiagent learning using a variable learning rate. Artificial Intelligence, 136(2):215 -- 250, 2002. Ronen I. Brafman and Moshe Tennenholtz. R-MAX a general polynomial time algorithm for near-optimal reinforcement learning. The Journal of Machine Learning Research, 3: 213 -- 231, 2003. George W. Brown. Iterative solution of games by fictitious play. Activity analysis of production and allocation, 13(1):374 -- 376, 1951. S´ebastien Bubeck and Nicol`o Cesa-Bianchi. Regret Analysis of Stochastic and Nonstochastic Multi-armed Bandit Problems. Foundations and Trends R(cid:13) in Machine Learning, 5(1):1 -- 122, 2012. S´ebastien Bubeck and Aleksandrs Slivkins. The best of both worlds: Stochastic and adver- sarial bandits. In Proceedings of the 25th Conference on Learning Theory, pages 1 -- 23, Edinburgh, Scotland, January 2012. Lucian Busoniu, Robert Babuska, and Bart De Schutter. A Comprehensive Survey of Mul- tiagent Reinforcement Learning. IEEE Transactions on Systems, Man and Cybernetics, Part C (Applications and Reviews), 38(2):156 -- 172, 2008. 49 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote Lucian Busoniu, Robert Babuska, and Bart De Schutter. Multi-agent reinforcement learn- ing: An overview. In Dipti Srinivasan and Lakhmi C Jain, editors, Innovations in Multi- Agent Systems and Applications - 1, pages 183 -- 221. Springer Berlin Heidelberg, Berlin, Heidelberg, 2010. Colin F. Camerer. Behavioral Game Theory: Experiments in Strategic Interaction (Roundtable Series in Behavioral Economics). Princeton University Press, February 2003. Colin F. Camerer and T Hua Ho. Experience-weighted attraction learning in normal form games. Econometrica, 67(4):827 -- 874, 1999. Colin F. Camerer, Teck-Hua Ho, and Juin-Kuan Chong. Sophisticated Experience-Weighted Attraction Learning and Strategic Teaching in Repeated Games. Journal of Economic Theory, 104(1):137 -- 188, May 2002. Colin F. Camerer, Teck-Hua Ho, and Juin-Kuan Chong. A cognitive hierarchy model of games. The Quarterly Journal of Economics, 119(3):861, 2004. Alvaro O. Castaneda. Deep Reinforcement Learning Variants of Multi-Agent Learning Algorithms. Master's thesis, University of Edinburgh, 2016. Doran Chakraborty and Peter Stone. Multiagent learning in the presence of memory- bounded agents. Autonomous Agents and Multi-Agent Systems, 28(2):182 -- 213, 2013. Doran Chakraborty, Noa Agmon, and Peter Stone. Targeted opponent modeling of memory- bounded agents. In Proceedings of the Adaptive Learning Agents Workshop (ALA), Saint Paul, MN, USA, 2013. Mithun Chakraborty, Sanmay Das, Brendan Juba, and Kai Yee Phoebe Chua. Coordinated Versus Decentralized Exploration In Multi-Agent Multi-Armed Bandits. In Proceedings of the 26th International Joint Conference on Artificial Intelligence, Melbourne, Australia, July 2017. Samuel P. M. Choi, Dit-Yan Yeung, and Nevin L. Zhang. An Environment Model for Nonstationary Reinforcement Learning. In Advances in Neural Information Processing Systems, pages 987 -- 993, Denver, Colorado, USA, 1999. Samuel P. M. Choi, Dit-Yan Yeung, and Nevin L. Zhang. Hidden-mode markov decision In Ron Sun and C Lee Giles, processes for nonstationary sequential decision making. editors, Sequence Learning, pages 264 -- 287. Springer, 2001. Caroline Claus and Craig Boutilier. The dynamics of reinforcement learning in coopera- In Proceedings of the 15th National Conference on Artificial tive multiagent systems. Intelligence, pages 746 -- 752, Madison, Wisconsin, USA, July 1998. Robert M. Coehoorn and Nicholas R. Jennings. Learning an opponent's preferences to make effective multi-issue negotiation trade-offs. In Proceedings of the 6th international conference on Electronic commerce, pages 59 -- 68, New York, NY, USA, 2004. 50 A Survey of Learning in Multiagent Environments Vincent Conitzer and Tuomas Sandholm. AWESOME: A general multiagent learning algo- rithm that converges in self-play and learns a best response against stationary opponents. Machine Learning, 67(1-2):23 -- 43, 2006. Miguel Costa Gomes, Vincent P. Crawford, and B. Broseta. Cognition and Behavior in Normal -- Form Games: An Experimental Study. Econometrica, 69(5):1193 -- 1235, 2001. Jacob W. Crandall. Just add Pepper: extending learning algorithms for repeated matrix games to repeated markov games. In Proceedings of the 11th International Conference on Autonomous Agents and Multiagent Systems, pages 399 -- 406, Valencia, Spain, 2012. Jacob W. Crandall. Towards minimizing disappointment in repeated games. Journal of Artificial Intelligence Research, 49(1):111 -- 142, January 2014. Jacob W. Crandall and Michael A. Goodrich. Learning to compete, coordinate, and coop- erate in repeated games using reinforcement learning. Machine Learning, 82(3):281 -- 314, March 2011. Robert H. Crites and Andrew G. Barto. Elevator Group Control Using Multiple Reinforce- ment Learning Agents. Machine Learning, 33:235 -- 262, December 1998. Bruno C. Da Silva, Eduardo W. Basso, Ana L.C. Bazzan, and Paulo M. Engel. Dealing In Proceedings of the 23rd with non-stationary environments using context detection. International Conference on Machine Learnig, pages 217 -- 224, Pittsburgh, Pennsylvania, 2006. Felipe Leno da Silva, Ruben Glatt, and Anna Helena Reali Costa. Simultaneously Learning and Advising in Multiagent Reinforcement Learning. In Proceedings of the 16th Confer- ence on Autonomous Agents and Multiagent Systems, Sao Paulo, Brazil, 2017. Steven Damer and Maria Gini. Safely using predictions in general-sum normal form games. In Proceedings of the 16th Conference on Autonomous Agents and Multiagent Systems, Sao Paulo, 2017. Yann-Michael De Hauwere, Peter Vrancx, and Ann Nowe. Learning multi-agent state space representations. In Proceedings of the 9th International Conference on Autonomous Agents and Multiagent Systems, pages 715 -- 722, Toronto, Canada, 2010. Harmen de Weerd, Rineke Verbrugge, and Bart Verheij. How much does it help to know what she knows you know? An agent-based simulation study. Artificial Intelligence, 199-200(C):67 -- 92, June 2013. Harmen de Weerd, Rineke Verbrugge, and Bart Verheij. Negotiating with other minds: the role of recursive theory of mind in negotiation with incomplete information. Autonomous Agents and Multi-Agent Systems, pages 1 -- 38, 2015. Li Deng and Dong Yu. Deep Learning Methods and Applications. Foundations and Trends in Signal Processing, 7(3-4):197 -- 387, June 2013. 51 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote David DeVault, Johnathan Mell, and Jonathan Gratch. Toward natural turn-taking in In AAAI Spring Symposium on Turn-taking and a virtual human negotiation agent. Coordination in Human-Machine Interaction, Standford, CA, USA, 2015. Prashant Doshi and Piotr J. Gmytrasiewicz. On the Difficulty of Achieving Equilibrium in Interactive POMDPs. In Twenty-first National Conference on Artificial Intelligence, pages 1131 -- 1136, Boston, MA, USA, 2006. Prashant Doshi, Xia Qu, A. Goodie, and Diana Young. Modeling recursive reasoning by humans using empirically informed interactive POMDPs. In Proceedings of the 9th International Conference on Autonomous Agents and Multiagent Systems, pages 1223 -- 1230, Toronto, Canada, 2010. Eric Eaton, Peter Stone, Toby Walsh, Michael Wooldridge, Tom Dietterich, Maria Gini, Barbara J. Grosz, Charles L Isbell, Subbarao Kambhampati, Michael L. Littman, Francesca Rossi, and Stuart J. Russell. Who speaks for AI? AI Matters, 2(2):4 -- 14, January 2016. Mohamed Elidrisi, Nicholas Johnson, and Maria Gini. Fast Learning against Adaptive Adversarial Opponents. In Proceedings of the Adaptive Learning Agents Workshop (ALA), Valencia, Spain, November 2012. Mohamed Elidrisi, Nicholas Johnson, Maria Gini, and Jacob W. Crandall. Fast adaptive learning in repeated stochastic games by game abstraction. In Proceedings of the 13th International Conference on Autonomous Agents and Multiagent Systems, pages 1141 -- 1148, Paris, France, 2014. Angela Fabregues, David Navarro, Alejandro Serrano, and Carles Sierra. Dipgame: A testbed for multiagent systems. In Proceedings of the 9th International Conference on Autonomous Agents and Multiagent Systems, pages 1619 -- 1620, Richland, SC, 2010. Fei Fang, Peter Stone, and Milind Tambe. Defender strategies in domains involving fre- In Proceedings of the 2015 International Conference on quent adversary interaction. Autonomous Agents and Multiagent Systems, pages 1663 -- 1664, 2015. Peyman Faratin, Carles Sierra, and Nicholas R. Jennings. Negotiation decision functions for autonomous agents. Robotics and Autonomous Systems, 24(3-4):159 -- 182, 1998. Jakob N. Foerster, Yannis M Assael, Nando De Freitas, and Shimon Whiteson. Learning In Advances in Neural to communicate with deep multi-agent reinforcement learning. Information Processing Systems, pages 2145 -- 2153, January 2016. Jakob N. Foerster, Gregory Farquhar, Triantafyllos Afouras, Nantas Nardelli, and Shimon Whiteson. Counterfactual Multi-Agent Policy Gradients. arXiv.org, 1705.08926v1, 2017a. Jakob N. Foerster, Nantas Nardelli, Gregory Farquhar, and Philip H.S. Torr. Stabilising experience replay for deep multi-agent reinforcement learning. arXiv.org, 1702.08887v1, 2017b. Drew Fudenberg and Jean Tirole. Game Theory. The MIT Press, August 1991. 52 A Survey of Learning in Multiagent Environments Nancy Fulda and Dan Ventura. Predicting and Preventing Coordination Problems in Co- operative Q-learning Systems. In Proceedings of the Twentieth International Joint Con- ference on Artificial Intelligence, pages 780 -- 785, Hyderabad, India, 2007. Yaakov Gal, Barbara J Grosz, Sarit Kraus, Avi Pfeffer, and Stuart Shieber. Colored trails: a formalism for investigating decision-making in strategic environments. In Proceedings of the 2005 IJCAI workshop on reasoning, representation, and learning in computer games, pages 25 -- 30, 2005. Joao Gama, Indre Zliobaite, Albert Bifet, Mykola Pechenizkiy, and Abdelhamid Bouchachia. A survey on concept drift adaptation. ACM Computing Surveys (CSUR), 46(4), April 2014. Aur´elien Garivier and Eric Moulines. On Upper-Confidence Bound Policies for Switch- ing Bandit Problems. In Algorithmic Learning Theory, pages 174 -- 188. Springer Berlin Heidelberg, Espoo, Finland, October 2011. Mohammed Ghavamzadeh, Shie Mannor, Joelle Pineau, and Aviv Tamar. Bayesian Rein- forcement Learning: A Survey. Foundations and Trends R(cid:13) in Machine Learning, 8(5-6): 359 -- 483, 2015. Piotr J. Gmytrasiewicz and Prashant Doshi. A framework for sequential planning in mul- tiagent settings. Journal of Artificial Intelligence Research, 24(1):49 -- 79, 2005. Piotr J. Gmytrasiewicz and Edmund H. Durfee. Rational Coordination in Multi-Agent Environments. Autonomous Agents and Multi-Agent Systems, 3(4):319 -- 350, December 2000. Jacob K. Goeree and C.A. Holt. Ten little treasures of game theory and ten intuitive contradictions. American Economic Review, pages 1402 -- 1422, 2001. Jonathan Gratch, David DeVault, Gale M. Lucas, and Stacy Marsella. Negotiation as a challenge problem for virtual humans. In Proceedings of the International Conference on Intelligent Virtual Agents, pages 201 -- 215, Delft, The Netherlands, 2015. Amy Greenwald and Keith Hall. Correlated Q-learning. In Proceedings of the 22nd Con- ference on Artificial Intelligence, pages 242 -- 249, Washington, DC, USA, 2003. Jayesh K Gupta, Maxim Egorov, and Mykel J Kochenderfer. Cooperative Multi-agent Control using deep reinforcement learning. In Adaptive Learning Agents at AAMAS, Sao Paulo, 2017. Emmanuel Hadoux, Aur´elie Beynier, and Paul Weng. Sequential decision-making under In Learning over non-stationary environments via sequential change-point detection. Multiple Contexts, Nancy, France, 2014a. Emmanuel Hadoux, Aur´elie Beynier, and Paul Weng. Solving Hidden-Semi-Markov-Mode Markov Decision Problems. In Scalable Uncertainty Management, pages 176 -- 189, Septem- ber 2014b. 53 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote John C. Harsanyi and Reinhard Selten. A general theory of equilibrium selection in games. MIT Press, 1988. C´edric Hartland, Sylvain Gelly, Nicolas Baskiotis, Olivier Teytaud, and Mich`ele Sebag. Multi-armed Bandit, Dynamic Environments and Meta-Bandits. HAL, hal-00113668, November 2006. He He, Jordan Boyd-Graber, Kevin Kwok, and Hal Daume. Opponent modeling in deep reinforcement learning. In 33rd International Conference on Machine Learning, ICML 2016, pages 2675 -- 2684, January 2016. Minghua He, Nicholas R. Jennings, and Ho-fung Leung. On agent-mediated electronic IEEE Transactions on Knowledge and Data Engineering, 15(4):985 -- 1003, commerce. 2003. Pablo Hernandez-Leal and Michael Kaisers. Learning against sequential opponents in re- In The 3rd Multi-disciplinary Conference on Reinforcement peated stochastic games. Learning and Decision Making, Ann Arbor, April 2017a. Pablo Hernandez-Leal and Michael Kaisers. Towards a fast detection of opponents in re- peated stochastic games. In Proceedings of the 1st Workshop on Transfer in Reinforcement Learning at AAMAS, Sao Paulo, Brazil, May 2017b. Pablo Hernandez-Leal, Enrique Munoz de Cote, and L. Enrique Sucar. Modeling non- stationary opponents. In Proceedings of the 12th International Conference on Autonomous Agents and Multiagent Systems, pages 1135 -- 1136, Saint Paul, MN, USA, May 2013. Pablo Hernandez-Leal, Enrique Munoz de Cote, and L. Enrique Sucar. A framework for learning and planning against switching strategies in repeated games. Connection Science, 26(2):103 -- 122, March 2014a. Pablo Hernandez-Leal, Enrique Munoz de Cote, and L. Enrique Sucar. Using a priori information for fast learning against non-stationary opponents. In Ana L.C. Bazzan and K. Pichara, editors, Advances in Artificial Intelligence -- IBERAMIA 2014, pages 536 -- 547, Santiago de Chile, 2014b. Pablo Hernandez-Leal, Matthew E. Taylor, L. Enrique Sucar, and Enrique Munoz de Cote. Bidding in Non-Stationary Energy Markets. In Proceedings of the 14th International Conference on Autonomous Agents and Multiagent Systems, pages 1709 -- 1710, Istanbul, Turkey, May 2015. Pablo Hernandez-Leal, Benjamin Rosman, Matthew E. Taylor, L. Enrique Sucar, and En- rique Munoz de Cote. A Bayesian Approach for Learning and Tracking Switching, Non- stationary Opponents (Extended Abstract). In Proceedings of 15th International Con- ference on Autonomous Agents and Multiagent Systems, pages 1315 -- 1316, Singapore, 2016a. Pablo Hernandez-Leal, Matthew E. Taylor, Benjamin Rosman, L. Enrique Sucar, and En- rique Munoz de Cote. Identifying and Tracking Switching, Non-stationary Opponents: a 54 A Survey of Learning in Multiagent Environments Bayesian Approach. In Multiagent Interaction without Prior Coordination Workshop at AAAI, Phoenix, AZ, USA, 2016b. Pablo Hernandez-Leal, Yusen Zhan, Matthew E. Taylor, L. Enrique Sucar, and Enrique Munoz de Cote. Efficiently detecting switches against non-stationary opponents. Au- tonomous Agents and Multi-Agent Systems, 31(4):767 -- 789, 2017a. Pablo Hernandez-Leal, Yusen Zhan, Matthew E. Taylor, L. Enrique Sucar, and Enrique Munoz de Cote. An exploration strategy for non-stationary opponents. Autonomous Agents and Multi-Agent Systems, 31(5):971 -- 1002, 2017b. Pablo Hernandez-Leal, Bilal Kartal, and Matthew E Taylor. inforcement learning the answer or the question? arXiv:1810.05587, 2018. a brief survey. Is multiagent deep re- arXiv preprint Koen V. Hindriks, Catholijn M. Jonker, and Dmytro Tykhonov. The benefits of opponent models in negotiation. In Proceedings of the 2009 IEEE/WIC/ACM International Joint Conference on Web Intelligence and Intelligent Agent Technology, volume 2, pages 439 -- 444, Milan, Italy, Sep 2009. Bret Hoehn, Finnegan Southey, Robert C. Holte, and Valeriy Bulitko. Effective Short-Term Opponent Exploitation in Simplified Poker. In The 20th National Conference on Artificial Intelligence, pages 783 -- 788, Pittsburgh, Pennsylvania, USA, 2005. Chris HolmesParker, Matthew E. Taylor, Adrian Agogino, and Kagan Tumer. CLEAN- ing the reward: counterfactual actions to remove exploratory action noise in multiagent learning. In Proceedings of the 13th International Conference on Autonomous Agents and Multiagent Systems, pages 1353 -- 1354, Paris, France, May 2014. Chongming Hou. Predicting agents tactics in automated negotiation. In Proceedings of the IEEE/WIC/ACM International Conference on Intelligent Agent Technology, pages 127 -- 133, Sep 2004. Wei-Tek Hsu and Von-Wun Soo. Market performance of adaptive trading agents in syn- chronous double auctions. In Proceedings of the 4th Pacific Rim International Workshop on Multi-Agents, Intelligent Agents: Specification, Modeling, and Applications, PRIMA 2001, pages 108 -- 121, London, UK, UK, 2001. Junling Hu and Michael P. Wellman. Multiagent Reinforcement Learning: Theoretical Framework and an Algorithm. In Proceedings of the 15th International Conference on Machine Learning, pages 242 -- 250, Madison, Wisconsin, USA, July 1998. Junling Hu and Michael P. Wellman. Nash Q-learning for general-sum stochastic games. Journal of Machine Learning Research, 4:1039 -- 1069, 2003. Takayuki Ito, Minjie Zhang, Valentin Robu, Shaheen Fatima, and Tokuro Matsuo. Advances in agent-based complex automated negotiations, volume 233. Springer, 2009. Thomas Jaksch, Ronald Ortner, and Peter Auer. Near-optimal regret bounds for reinforce- ment learning. Journal of Machine Learning Research, 11:1563 -- 1600, April 2010. 55 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote Nicholas R. Jennings, Peyman Faratin, Alessio R. Lomuscio, Simon Parsons, Michael J. Wooldridge, and Carles Sierra. Automated negotiation: Prospects, methods and chal- lenges. Group Decision and Negotiation, 10(2):199 -- 215, 2001. Steven Jensen, Daniel Boley, Maria Gini, and Paul Schrater. Non-stationary Policy Learning In Proceedings of the National Conference on Artificial in 2-player Zero Sum Games. Intelligence, pages 789 -- 794, Pittsburgh, Pennsylvania, USA, 2005. Michael Johanson and Michael Bowling. Data Biased Robust Counter Strategies. In Pro- ceedings of the 22nd Conference on Artificial Intelligence, pages 264 -- 271, Clearwater Beach, Florida USA, 2009. Michael Johanson, Martin A. Zinkevich, and Michael Bowling. Computing Robust Counter- Strategies. In Advances in Neural Information Processing Systems, pages 721 -- 728, Van- couver, BC, Canada, 2007. Leslie P. Kaelbling, Michael L. Littman, and Anthony R. Cassandra. Planning and acting in partially observable stochastic domains. Artificial Intelligence, 101(1-2):99 -- 134, 1998. Daniel Kahneman and Amos Tversky. Prospect theory: An analysis of decision under risk. Econometrica, 47(2):263 -- 291, 1979. Michael Kaisers and Karl Tuyls. Frequency adjusted multi-agent Q-learning. In Proceed- ings of the 9th International Conference on Autonomous Agents and Multiagent Systems, pages 309 -- 315, Toronto, Canada, 2010. Michael Kaisers and Karl Tuyls. FAQ-learning in matrix games: demonstrating convergence near Nash equilibria, and bifurcation of attractors in the battle of sexes. In AAAI Work- shop on Interactive Decision Theory and Game Theory, pages 309 -- 316, San Francisco, CA, USA, January 2011. Sham Machandranath Kakade. On the sample complexity of reinforcement learning. PhD thesis, Gatsby Computational Neuroscience Unit, University College London, 2003. Wolfgang Ketter, John Collins, and Prashant P. Reddy. Power TAC: A competitive eco- nomic simulation of the smart grid. Energy Economics, 39:262 -- 270, September 2013. Ardeshir Kianercy and Aram Galstyan. Dynamics of Boltzmann Q-learning in two-player two-action games. Physical Review E, 85(4):041145, April 2012. Ryszard Kowalczyk, Mihaela Ulieru, and Rainer Unland. Integrating mobile and intelli- gent agents in advanced e-commerce: A survey. In Jaime G. Carbonell, Jorg Siekmann, Ryszard Kowalczyk, Jorg P. Muller, Huaglory Tianfield, and Rainer Unland, editors, Agent Technologies, Infrastructures, Tools, and Applications for E-Services, volume 2592 of Lecture Notes in Computer Science, pages 295 -- 313. Springer Berlin Heidelberg, 2003. Sarit Kraus. Strategic Negotiation in Multiagent Environments. MIT Press, Oct 2001. Himabindu Lakkaraju, Ece Kamar, Rich Caruana, and Eric Horvitz. Identifying Unknown Unknowns in the Open World: Representations and Policies for Guided Exploration. In Thirty-First AAAI Conference on Artificial Intelligence, 2017. 56 A Survey of Learning in Multiagent Environments Alessandro Lazaric, Enrique Munoz de Cote, and Nicola Gatti. Reinforcement learning in extensive form games with incomplete information: the bargaining case study. In Proceedings of the 6th International Conference on Autonomous Agents and Multiagent Systems, Honolulu, Hawai, USA, May 2007. Joel Z. Leibo, V Zambaldi, M Lanctot, and J Marecki. Multi-agent Reinforcement Learning in Sequential Social Dilemmas. In Proceedings of the 16th Conference on Autonomous Agents and Multiagent Systems, Sao Paulo, 2017. David S. Leslie and E. J. Collins. Individual Q-learning in normal form games. SIAM Journal on Control and Optimization, 44(2):495 -- 514, 2005. Somchaya Liemhetcharat and Manuela Veloso. Allocating training instances to learning agents for team formation. Autonomous Agents and Multi-Agent Systems, 31(4):905 -- 940, 2017. Raz Lin and Sarit Kraus. Can automated agents proficiently negotiate with humans? Com- mun. ACM, 53(1):78 -- 88, January 2010. Raz Lin, Sarit Kraus, Jonathan Wilkenfeld, and James Barry. Negotiating with bounded rational agents in environments with incomplete information using an automated agent. Artificial Intelligence, 172(6-7):823 -- 851, 2008. Michael L. Littman. Markov games as a framework for multi-agent reinforcement learning. In Proceedings of the 11th International Conference on Machine Learning, pages 157 -- 163, New Brunswick, NJ, USA, 1994. Michael L. Littman. Algorithms for sequential decision making. PhD thesis, Department of Computer Science, Brown University, 1996. Michael L. Littman. Friend-or-foe Q-learning in general-sum games. In Proceedings of the 22nd Conference on Artificial Intelligence, pages 322 -- 328, Williamstown, MA, USA, 2001. Michael L. Littman and Peter Stone. Implicit Negotiation in Repeated Games. ATAL '01: Revised Papers from the 8th International Workshop on Intelligent Agents VIII, August 2001. J. Liu, P. Dolan, and E. R. Pedersen. Personalized news recommendation based on click behavior. In Proceedings of the th international conference on Intelligent user interfaces, pages 31 -- 40, Hong Kong, 2010. Manuel Lopes, Tobias Lang, Marc Toussaint, and Pierre-Yves Oudeyer. Exploration in Model-based Reinforcement Learning by Empirically Estimating Learning Progress. In Advances in Neural Information Processing Systems, pages 206 -- 214, Lake Tahoe, Nevada, USA, 2012. Patrick MacAlpine, Daniel Urieli, Samuel Barrett, Shivaram Kalyanakrishnan, Francisco Barrera, Adrian Lopez-Mobilia, Nicolae S¸tiurca, Victor Vu, and Peter Stone. UT Austin 57 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote Villa 2011: a Champion Agent in the RoboCup 3D Soccer Simulation Competition. In Proceedings of the 11th International Conference on Autonomous Agents and Multiagent Systems, pages 129 -- 136, Valencia, Spain, June 2012. M. M. Hassan Mahmud and Subramanian Ramamoorthy. Learning in non-stationary MDPs as transfer learning. In Proceedings of the 12th International Conference on Autonomous Agents and Multi-agent Systems, pages 1259 -- 1260, Saint Paul, MN, USA, May 2013. Odalric-Ambrym Maillard, Phuong Nguyen, Ronald Ortner, and Daniil Ryabko. Opti- mal Regret Bounds for Selecting the State Representation in Reinforcement Learning. Proceedings of the 30th International Conference on Machine Learning, 28(1):543 -- 551, February 2013. Andrei Marinescu, Ivana Dusparic, Adam Taylor, Vinny Cahill, and Siobh´an Clarke. Decen- tralised Multi-Agent Reinforcement Learning for Dynamic and Uncertain Environments. arXiv.org, 1409.4561, 2014. Andrei Marinescu, Ivana Dusparic, Adam Taylor, Vinny Cahill, and Siobh´an Clarke. P- MARL: Prediction-Based Multi-Agent Reinforcement Learning for Non-Stationary En- vironments. In Proceedings of the 14th International Conference on Autonomous Agents and Multiagent Systems, pages 1897 -- 1898, Istanbul, Turkey, May 2015. Laetitia Matignon, Guillaume J Laurent, and Nadine Le Fort-Piat. Independent reinforce- ment learners in cooperative Markov games: a survey regarding coordination problems. Knowledge Engineering Review, 27(1):1 -- 31, February 2012. Noyda Matos, Carles Sierra, and Nicholas R. Jennings. Determining successful negotiation strategies: an evolutionary approach. In Proceedings International Conference on Multi Agent Systems, pages 182 -- 189, Paris, France, 1998. Francisco S. Melo and Alberto Sardinha. Ad hoc teamwork by learning teammates' task. Autonomous Agents and Multi-Agent Systems, 30(2):175 -- 219, January 2016. Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, Stig Petersen, Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Ku- maran, Daan Wierstra, Shane Legg, and Demis Hassabis. Human-level control through deep reinforcement learning. Nature, 518(7540):529 -- 533, February 2015. J. P. Muller and K. Fischer. Application impact of multi-agent systems and technologies: a survey. In Onn Shehory and Arnon Sturm, editors, Agent-Oriented Software Engineering, pages 27 -- 53. Agent-oriented software engineering, Heidelberg, Germany, 2014. Enrique Munoz de Cote and Nicholas R. Jennings. Planning against fictitious players in repeated normal form games. In Proceedings of the 9th International Conference on Autonomous Agents and Multiagent Systems, pages 1073 -- 1080, Toronto, Canada, 2010. Enrique Munoz de Cote and Michael L. Littman. A Polynomial-time Nash Equilibrium Algorithm for Repeated Stochastic Games. In Uncertainty in Artificial Intelligence, pages 419 -- 426, Helsinki, Finland, 2008. 58 A Survey of Learning in Multiagent Environments Enrique Munoz de Cote, Alessandro Lazaric, and Marcello Restelli. Learning to cooperate in multi-agent social dilemmas. In Proceedings of the 5th International Conference on Autonomous Agents and Multiagent Systems, pages 783 -- 785, Hakodate, Hokkaido, Japan, May 2006. Enrique Munoz de Cote, Archie C. Chapman, Adam M. Sykulski, and Nicholas R. Jen- nings. Automated Planning in Repeated Adversarial Games. In Uncertainty in Artificial Intelligence, pages 376 -- 383, Catalina Island, California, 2010. Sanmit Narvekar, Jivko Sinapov, and Peter Stone. Autonomous Task Sequencing for Cus- tomized Curriculum Design in Reinforcement Learning. In Proceedings of the 26th Inter- national Joint Conference on Artificial Intelligence, Melbourne, Australia, July 2017. John F. Nash. The Bargaining Problem. Econometrica, 18(2):155 -- 162, April 1950a. John F. Nash. Equilibrium points in n-person games. Proceedings of the National Academy of Sciences, 36(1):48 -- 49, 1950b. Brenda Ng, Kofi Boakye, Carol Meyers, and Andrew Wang. Bayes-Adaptive Interactive POMDPs. In Proceedings of the Twenty-Sixth AAAI Conference on Artificial Intelligence, pages 1408 -- 1414, Toronto, Canada, May 2012. Thanh Thi Nguyen, Ngoc Duy Nguyen, and Saeid Nahavandi. Deep reinforcement learning for multi-agent systems: A review of challenges, solutions and applications. arXiv preprint arXiv:1812.11794, 2018. Eugene Nudelman, Jennifer Wortman, Yoav Shoham, and Kevin Leyton-Brown. Run the GAMUT: a comprehensive approach to evaluating game-theoretic algorithms. In Proceed- ings of the 3rd International Conference on Autonomous Agents and Multiagent Systems, pages 880 -- 887, New York City, NY, USA, 2004. Frans A. Oliehoek, Matthijs T. J. Spaan, and Stefan J. Witwicki. Influence-optimistic local values for multiagent planning. In Proceedings of the 14th International Conference on Autonomous Agents and Multiagent Systems, pages 1703 -- 1704, Istanbul, Turkey, 2015. Ronald Ortner, Daniil Ryabko, Peter Auer, and R´emi Munos. Regret bounds for restless Markov bandits. Theoretical Computer Science, 558(C):62 -- 76, November 2014. Liviu Panait and Sean Luke. Cooperative Multi-Agent Learning: The State of the Art. Autonomous Agents and Multi-Agent Systems, 11(3), November 2005. Liviu Panait, Keith Sullivan, and Sean Luke. Lenience towards teammates helps in co- In Proceedings of the 5th International Conference on operative multiagent learning. Autonomous Agents and Multiagent Systems, Hakodate, Japan, 2006. Sandeep Pandey, Deepayan Chakrabarti, and Deepak Agarwal. Multi-armed bandit prob- lems with dependent arms. In Proceedings of the 24th International Conference on Ma- chine Learning, pages 721 -- 728, Corvallis, OR, USA, 2007. 59 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote Christos H. Papadimitriou and John N. Tsitsiklis. The complexity of Markov decision processes. Mathematics of Operations Research, 12(3):441 -- 450, 1987. M. Pipattanasomporn, H. Feroze, and Saifur Rahman. Multi-agent systems in a distributed smart grid: Design and implementation. In Proceedings of IEEE Power Systems Confer- ence and Exposition, pages 1 -- 8, Seattle, Washington, USA, 2009. James Pita, M. Jain, F. Ord´onez, C Portway, M. Tambe, C Western, P Paruchuri, and Sarit Kraus. Using game theory for Los Angeles airport security. AI Magazine, 30(1):43 -- 57, 2009. Rob Powers and Yoav Shoham. New criteria and a new algorithm for learning in multi- agent systems. In Advances in Neural Information Processing Systems, pages 1089 -- 1096, Vancouver, Canada, 2004. Rob Powers and Yoav Shoham. Learning against opponents with bounded memory. In Proceedings of the 19th International Joint Conference on Artificial Intelligence, pages 817 -- 822, Edinburg, Scotland, UK, 2005. Rob Powers, Yoav Shoham, and Thuc Vu. A general criterion and an algorithmic framework for learning in multi-agent systems. Machine Learning, 67(1-2):45 -- 76, 2007. Martin L. Puterman. Markov decision processes: Discrete stochastic dynamic programming. John Wiley & Sons, Inc., 1994. David V. Pynadath, Ning Wang, and Stacy C. Marsella. Are You Thinking What I'm Thinking? An Evaluation of a Simplified Theory of Mind, pages 44 -- 57. Springer Berlin Heidelberg, Berlin, Heidelberg, 2013. Sarvapali D. Ramchurn, Alessandro Farinelli, Kathryn S. Macarthur, and Nicholas R. Jen- nings. Decentralized Coordination in RoboCup Rescue. The Computer Journal, 53(9), November 2010. Mathias Risse. What is rational about Nash equilibria? Synthese, 124(3):361 -- 384, 2000. Herbert Robbins. Some aspects of the sequential design of experiments. In Herbert Robbins Selected Papers, pages 527 -- 535. Springer, February 1985. Benjamin Rosman, Majd Hawasly, and Subramanian Ramamoorthy. Bayesian Policy Reuse. Machine Learning, 104(1):99 -- 127, 2016. Joris Scharpff, Diedrerik M. Roijers, Frans A. Oliehoek, Matthijs T. J. Spaan, and Math- ijs de Weerdt. Solving transition-independent multi-agent MDPs with sparse interactions. In Thirtieth AAAI Conference on Artificial Intelligence, Phoenix, AZ, USA, 2016. Sandip Sen, Mahendra Sekaran, and John Hale. Learning to coordinate without sharing information. In Proceedings of the Twelfth AAAI National Conference on Artificial In- telligence, pages 426 -- 431, Seattle, WA, USA, 1994. 60 A Survey of Learning in Multiagent Environments Sven Seuken and Shlomo Zilberstein. Formal models and algorithms for decentralized de- cision making under uncertainty. Autonomous Agents and Multi-Agent Systems, 17(2): 190 -- 250, February 2008. Yoav Shoham and Kevin Leyton-Brown. Multiagent Systems: Algorithmic, Game- Theoretic, and Logical Foundations. Cambridge University Press, December 2008. Yoav Shoham, Rob Powers, and T. Grenager. If multi-agent learning is the answer, what is the question? Artificial Intelligence, 171(7):365 -- 377, 2007. David Silver and J Veness. Monte-Carlo planning in large POMDPs. Advances in Neural Information Processing Systems, 46, 2010. David Silver, A Huang, C J Maddison, A Guez, L Sifre, George van den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, Sander Diele- man, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy Lillicrap, Madeleine Leach, Koray Kavukcuoglu, Thore Graepel, and Demis Hassabis. Mastering the game of Go with deep neural networks and tree search. Nature, 529(7587):484 -- 489, 2016. Satinder Singh, Tommi Jaakkola, Michael L. Littman, and Csaba Szepesv´ari. Convergence results for single-step on-policy reinforcement-learning algorithms. Machine Learning, 38 (287), 2000. Ekhlas Sonu, Yingke Chen, and Prashant Doshi. Individual Planning in Agent Populations: Exploiting Anonymity and Frame-Action Hypergraphs. In Proceedings of the 23rd In- ternational Joint Conference on Artificial Intelligence, pages 202 -- 210, Jerusalem, Israel, 2015. Dale O. Stahl and P.W. Wilson. On Players' Models of Other Players: Theory and Exper- imental Evidence. Games and Economic Behavior, 10(1):218 -- 254, 1995. Jeffrey L. Stimpson and Michael A. Goodrich. Learning To Cooperate in a Social Dilemma: A Satisficing Approach to Bargaining. Proceedings of the Twentieth International Con- ference on Machine Learning, pages 728 -- 735, September 2003. Peter Stone, G.A. Kaminka, Sarit Kraus, and Jeffrey S. Rosenschein. Ad Hoc Autonomous Agent Teams: Collaboration without Pre-Coordination. In The 20th National Conference on Artificial Intelligence, pages 1504 -- 1509, Atlanta, Georgia, USA, 2010. Nobuo Suematsu and Akira Hayashi. A multiagent reinforcement learning algorithm us- In Proceedings of the 1st International Conference on ing extended optimal response. Autonomous Agents and Multiagent Systems, pages 370 -- 377, Bologna, Italy, 2002. Richard S. Sutton and Andrew G. Barto. Reinforcement Learning An Introduction. MIT Press, Cambridge, MA, 1998. Richard S. Sutton, Anna Koop, and David Silver. On the role of tracking in stationary environments. In Proceedings of the 24th International Conference on Machine Learning, pages 871 -- 878, Corvallis, OR, USA, 2007. 61 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote Maxwell Svetlik, Matteo Leonetti, Jivko Sinapov, Rishi Shah, Nick Walker, and Peter Stone. Automatic Curriculum Graph Generation for Reinforcement Learning Agents. In Proceedings of the 31st AAAI Conference on Artificial Intelligence, San Francisco, CA, 2016. Adam M. Sykulski, Archie C. Chapman, Enrique Munoz de Cote, and Nicholas R. Jennings. EA2: The Winning Strategy for the Inaugural Lemonade Stand Game Tournament. In Proceeding of the 19th European Conference on Artificial Intelligence, pages 209 -- 214. IOS Press, August 2010. Ardi Tampuu, Tambet Matiisen, Dorian Kodelja, Ilya Kuzovkin, Kristjan Korjus, Juhan Aru, Jaan Aru, and Raul Vicente. Multiagent cooperation and competition with deep reinforcement learning. PLOS ONE, 12(4):e0172395 -- 15, April 2017. Ming Tan. Multi-Agent Reinforcement Learning: Independent vs. Cooperative Agents. In Machine Learning Proceedings 1993 Proceedings of the Tenth International Conference, University of Massachusetts, Amherst, June 27 -- 29, 1993, pages 330 -- 337. Elsevier, Jan- uary 1993. Matthew E. Taylor and Peter Stone. Transfer learning for reinforcement learning domains: A survey. The Journal of Machine Learning Research, 10:1633 -- 1685, 2009. Gerald Tesauro. Extending Q-learning to general adaptive multi-agent systems. In Advances in Neural Information Processing Systems, pages 871 -- 878, Vancouver, Canada, 2003. Gerald Tesauro and Jeffrey O. Kephart. Pricing in agent economies using multi-agent q-learning. Autonomous Agents and Multi-Agent Systems, 5(3):289 -- 304, 2002. Lisa Torrey and Matthew E. Taylor. Teaching on a Budget: Agents advising agents in In 12th International Conference on Autonomous Agents and reinforcement learning. Multiagent Systems 2013, AAMAS 2013, pages 1053 -- 1060, January 2013. Long Tran-Thanh, Archie C. Chapman, A. Rogers, and Nicholas R. Jennings. Knap- sack based optimal policies for budget-limited multi-armed bandits. In Proceedings of the Twenty-Sixth AAAI Conference on Artificial Intelligence, pages 1134 -- 1140, Toronto, Canada, 2012. Kagan Tumer and Adrian Agogino. Distributed agent-based air traffic flow management. In Proceedings of the 6th International Conference on Autonomous Agents and Multiagent Systems, Honolulu, Hawaii, 2007. Karl Tuyls and Gerhard Weiss. Multiagent learning: Basics, challenges, and prospects. AI Magazine, 33(3):41 -- 52, 2012. Konstantina Valogianni, Wolfgang Ketter, and John Collins. A Multiagent Approach to Variable-Rate Electric Vehicle Charging Coordination. In Proceedings of the 14th Inter- national Conference on Autonomous Agents and Multiagent Systems, pages 1131 -- 1139, Istanbul, Turkey, May 2015. 62 A Survey of Learning in Multiagent Environments Friedrich Van der Osten, Michael Kirley, and Tim Miller. The minds of many: opponent modelling in a stochastic game. In Proceedings of the 26th International Joint Conference on Artificial Intelligence, Melbourne, June 2017. Harm Van Seijen, Hado Van Hasselt, Shimon Whiteson, and Marco Wiering. A theoretical In IEEE Symposium on Adaptive Dynamic and empirical analysis of Expected Sarsa. Programming and Reinforcement Learning, pages 177 -- 184, Nashville, TN, USA, 2009. Vladimir Naumovich Vapnik. Statistical learning theory. Wiley-Interscience, 1st edition, September 1998. Peter Vrancx, Pasquale Gurzi, Abdel Rodriguez, Kris Steenhaut, and Ann Nowe. A Rein- forcement Learning Approach for Interdomain Routing with Link Prices. ACM Transac- tions on Autonomous and Adaptive Systems, 10(1):1 -- 26, March 2015. William E. Walsh, Rajarshi Das, Gerald Tesauro, and Jeffrey O Kephart. Analyzing complex strategic interactions in multi-agent systems. AAAI-02 Workshop on Game- Theoretic and Decision-Theoretic Agents, pages 109 -- 118, 2002. Christopher Watkins and Peter Dayan. Q-learning. Machine Learning, 8:279 -- 292, 1992. John Watkins. Learning from delayed rewards. PhD thesis, King's College, Cambridge, UK, April 1989. Jorgen W. Weibull. Evolutionary game theory. MIT press, 1995. Michael Weinberg and Jeffrey S. Rosenschein. Best-response multiagent learning in non- In Proceedings of the 3rd International Conference on Au- stationary environments. tonomous Agents and Multiagent Systems, pages 506 -- 513, New York, NY, USA, 2004. Gerhard Weiss, editor. Multiagent Systems. (Intelligent Robotics and Autonomous Agents series). MIT Press, Cambridge, MA, USA, 2nd edition, February 2013. Gerhard Widmer and Miroslav Kubat. Learning in the presence of concept drift and hidden contexts. Machine Learning, 23(1):69 -- 101, 1996. Colin R. Williams, Valentin Robu, Enrico H. Gerding, and Nicholas R. Jennings. Nego- In 20th tiating concurrently with unknown opponents in complex, real-time domains. European Conference on Artificial Intelligence, volume 242, pages 834 -- 839, May 2012. Stefan J. Witwicki, Frans A. Oliehoek, and Leslie P. Kaelbling. Heuristic search of multia- gent influence space. In Proceedings of the 11th International Conference on Autonomous Agents and Multiagent Systems, Valencia, Spain, 2012. James Robert Wright and Kevin Leyton-Brown. Beyond equilibrium: Predicting human behavior in normal-form games. In Twenty-Fourth Conference on Artificial Intelligence (AAAI-10), pages 901 -- 907, Atlanta, Georgia, 2010. James Robert Wright and Kevin Leyton-Brown. Level-0 meta-models for predicting human behavior in games. In Proceedings of the fifteenth ACM conference on Economics and computation, pages 857 -- 874, Palo Alto, CA, USA, 2014. 63 Hernandez-Leal, Kaisers, Baarslag and Munoz de Cote Michael Wunder, Michael L. Littman, and Matthew Stone. Communication, Credibility and Negotiation Using a Cognitive Hierarchy Model. In AAMAS Workshop# 19: Multi-agent Sequential Decision Making 2009, pages 73 -- 80, Budapest, Hungary, 2009. Michael Wunder, Michael Kaisers, Michael L. Littman, and John Robert Yaros. A cogni- tive hierarchy model applied to the lemonade game. In AAAI Workshop on Interactive Decision Theory and Game Theory, 2010. Michael Wunder, Michael Kaisers, John Robert Yaros, and Michael L. Littman. Using iterated reasoning to predict opponent strategies. In Proceedings of 10th International Conference on Autonomous Agents and Multiagent Systems, pages 593 -- 600, Taipei, Tai- wan, 2011. Michael Wunder, John Robert Yaros, Michael Kaisers, and Michael L. Littman. A frame- work for modeling population strategies by depth of reasoning. In Proceedings of the 11th International Conference on Autonomous Agents and Multiagent Systems, pages 947 -- 954, Valencia, Spain, 2012. Chao Yu, Minjie Zhang, Fenghui Ren, and Guozhen Tan. Multiagent Learning of Coordi- nation in Loosely Coupled Multiagent Systems. IEEE Transactions on Cybernetics, 45 (12):2853 -- 2867, January 2015. Jia Yuan Yu and Shie Mannor. Piecewise-stationary bandit problems with side observa- tions. In ICML '09: Proceedings of the 26th Annual International Conference on Machine Learning, Montreal, Canada, 2009a. Jia Yuan Yu and Shie Mannor. Online learning in Markov decision processes with arbitrarily changing rewards and transitions. In Proceedings of the First International Conference on Game Theory for Networks, pages 314 -- 322, Istanbul, Turkey, 2009b. Yusen Zhan, Haitham Bou Ammar, and Matthew E. Taylor. Theoretically grounded pol- icy advice from multiple teachr in reinfocement learning settings with Applications to negative transfer. In Proceedings of the Twenty-Fifth International Joint Conference on Artificial Intelligence, New York, NY, USA, April 2016. Jihang Zhang, Fenghui Ren, and Minjie Zhang. Bayesian-based preference prediction in bilateral multi-issue negotiation between intelligent agents. Knowledge-Based Systems, 84:108 -- 120, 2015. Martin A. Zinkevich, Michael Bowling, and Michael Wunder. The lemonade stand game competition: solving unsolvable games. SIGecom Exchanges, 10(1):35 -- 38, March 2011. 64
0810.0532	2	0810	2008-10-17T14:02:54	Three New Complexity Results for Resource Allocation Problems	[ "cs.MA", "cs.AI", "cs.CC", "cs.GT" ]	We prove the following results for task allocation of indivisible resources: - The problem of finding a leximin-maximal resource allocation is in P if the agents have max-utility functions and atomic demands. - Deciding whether a resource allocation is Pareto-optimal is coNP-complete for agents with (1-)additive utility functions. - Deciding whether there exists a Pareto-optimal and envy-free resource allocation is Sigma_2^p-complete for agents with (1-)additive utility functions.	cs.MA	cs	Three new complexity results for resource allocation problems Bart de Keijzer ([email protected]) October 24, 2018 Abstract We prove the following results for task allocation of indivisible resources: • The problem of finding a leximin-maximal resource allocation is in P if the agents have max-utility functions and atomic demands. • Deciding whether a resource allocation is Pareto-optimal is coNP-complete for agents with (1-)additive utility functions. • Deciding whether there exists a Pareto-optimal and envy-free resource allocation is Σp 2-complete for agents with (1-)additive utility functions. 1 Introduction In this text we prove complexity bounds for various problems in the field of resource allocation. These results come forth from an attempt to prove two open problems that were stated in the work of Bouveret, Lang et al ([1] and [2]). The problems are about resource allocation. In a resource allocation problem we have a set of agents (or alternatively, players) and a set of resources (or equivalently, goods, tasks, items, etc.). The goal is to allocate the resources to the agents such that some requirements are satisfied. These requirements may vary. In our case we are interested in finding fair allocations. The concept of fairness is not clear, and there are different criteria for deciding whether or not an allocation is fair. Two of these are envy-freeness and leximin-maximality. We will define these criteria (formally) later on. In the problems we consider, the resources are indivisible and a resource can not be shared by two or more agents. The two open problems of the aforementioned papers that we consider are: 1. In [1]: The problem of finding a leximin-maximal resource allocation for agents with max-utility functions and atomic demands is in NP. Could it be that it's in NPC (i.e. NP-complete), or is it perhaps in P? 2. In [2]: What is the complexity of deciding whether there exists a Pareto-efficient and envy-free resource allocation, when the agents have additive utility functions? Some of the more technical notions we just mentioned will be defined and explained later in this text. We do, however, assume that the reader is acquainted with computational complexity theory (especially the classes P, NP, coNP, and the classes of the polynomial hierarchy), the matching problem for bipartite graphs, logic, and the satisfiability problem. The first of these two problems is part of a quite an extensive series of problems and subproblems. The authors show for all of these problems that they are either in P or in NPC. 1 The only problem for which it remained an open question whether it is in P or in NPC (or possibly in between) is this one, where the agents have max-utility and a leximin-optimal allocation must be found. In section 2 we fill in the last open question of this series: we give a polynomial time algorithm for finding such an allocation, hence we prove that this problem is in P1. The second problem is also part of a collection of problems that the authors prove complete for various complexity classes. This particular problem is again the last open problem in this series. We prove in section 4 that this problem is Σp 2-complete (a class in the second level of the polynomial hierarchy) by a reduction from the complement of the language ∀∃3CNF (that is a restriction of the more well-known problem known as 2QSAT∀ or 2TQBF∀): a complete problem for Πp 2, which is naturally the complement of Σp 2. In the process of trying to prove the Σp 2-completeness of the second problem, we stum- bled on another interesting result, namely that the problem of deciding whether an alloca- tion of resources to agents is Pareto-efficient (also called: Pareto-optimal, efficient) is coNP- complete for agents with additive utility functions. We will give this proof in section 3. coNP-completeness of this problem has already been proved in the case of agents with ≥ 2- additive utility functions (implied from [3]), but not yet in the case of (1-)additive utility functions. 2 Leximin-maximal allocations with max-utility and atomic demands In this section, first, we make some definitions. After that we define the problem. Finally we give a polynomial time algorithm to solve the problem. 2.1 Preliminaries We first define formally the problem to solve. In a resource allocation problem, a set of resources must be divided among a set of agents. Such a division of resources to agents we call an allocation. The allocation must satisfy a certain set of constraints. Each agent has preferences on bundles of resources it may receive. The way these preferences are represented varies from setting to setting. In our case we use a cardinal preference structure: We represent the extent to which an agent values the bundle of resources he gets as real numbers. See for example [4] for examples of preference structures. Formally, we use the following definition for resource allocation settings: Definition 1 ((Indivisible) resource allocation setting). An indivisible resource allocation problem instance is a 5-tuple (cid:104)A,O, U,C, uc(cid:105), where A = {a1, . . . , an} is a set agents, O = {o1, . . . , om} is a finite set of resources. U = {u1, . . . , un} is a set of utility functions, ui is the utility function of agent ai. For all u ∈ U, u : 2O → R. C is a finite set of constraints, and uc is a collective utility function to be defined later. Definition 2 (Allocation of indivisible resources). Given a resource allocation problem setting (cid:104)A,O, U,C, uc(cid:105), an allocation is a mapping a : A → 2O. 1Of course we're talking about complexity classes for decision problems here. In [1], only the decision variant of this problem is considered. An algorithm from the decision variant of this problem is easily obtained if we have an algorithm for the optimization variant. 2 Definition 3 (Admissability of an allocation). Given a resource allocation setting (cid:104)A,O, U,C, uc(cid:105), an allocation a is admissable if it satisfies all constraints in C. For the specific case of the resource allocation problem that we are interested in, there is only one constraint in C, namely the preemption constraint. Also, we restrict ourselves to a special case of max-utility functions. The definitions of these concepts are as follows. Definition 4 (Preemption constraint). Given a resource allocation setting (cid:104)A,O, U,C, uc(cid:105) and an allocation a, then a satisfies the preemption constraint cpreempt iff ∀i ∈ A : ∀j ∈ A : (j (cid:54)= i) → (a(i) ∩ a(j) = ∅). We write a (cid:15) cpreempt. In words, the preemption constraint requires that an item is allocated to no more than one agent. Definition 5 (max-utility function). In a resource allocation setting (cid:104)A,O, U,C, uc(cid:105), a utility function u ∈ U is a max-utility function if u(O(cid:48) ∈ 2O) = max{du(o)o ⊆ O(cid:48)}, where du : 2O → R. In words, a max-utility function has an associated demand function d. The max-utility of a set of resources O(cid:48) is the subset of O(cid:48) for which the demand is the highest. We are interested in the following special case of max-utility functions Definition 6 (max-utility function with atomic demands). u is a max-utility function with atomic demands if u is a max-utility function as defined in definition 5, and du has an associated atomic demand set Du = {ri, . . . , rm} ⊂ R such that (cid:40) ri 0 du(O(cid:48) ∈ 2O) = if O(cid:48) = {oi} for 1 ≤ i ≤ m otherwise . This means: agents only express demands for single resources. Their utility for a set of resources is the highest demand they have for each of the individual resources of that set. Note that a max-utility function is completely represented by its associated atomic demand set. Now we are ready to discuss the collective utility function mentioned in definition 1. The purpose of the collective utility function uc is to express the quality of an allocation. For this we need to be able to compare the answers that uc gives for any two different allocations. This implies: • uc : (A → 2O) → X, • we need to specify X, • we need to define a transitive comparison relation ≺X over X. In a lot of cases we can say for example X = R or X = N. The comparison relation is then simply ≤. This is the case for classical utilitarian collective utility functions or egalitarian collective utility functions [4]. For us, the relation is a bit more complex. We are concerned with leximin-egalitarian collective utility functions. 3 Definition 7 (Leximin-egalitarian collective utility). Given a resource allocation setting (cid:104)A,O, U,C, uc(cid:105). uc : (A → 2O) → X is a leximin-egalitarian collective utility function iff X = Rn and for all allocations a: uc(al) = (cid:126)x, where u1(al(1)) ...  . (cid:126)x = un(al(n)) Definition 8 (Leximin-egalitarian comparison relation). The leximin-egalitarian comparison relation ≺leximin is defined as follows: Let (cid:126)u ∈ Rn and (cid:126)v ∈ Rn and let (cid:126)u↑ and (cid:126)v↑ be the sorted versions of (cid:126)u and (cid:126)v respectively. Now, it holds that ↑ ↑ ↑ (cid:126)v ≺leximin (cid:126)u ⇔ ∃i : ∀j < i : (cid:126) j ∧ (cid:126) j = (cid:126) v u v i < ↑ (cid:126) u i . Definition 9 (Leximin-maximality). Given a resource allocation setting (cid:104)A,O, U,C, uc(cid:105), with uc being a leximin-egalitarian collective utility function. An admissable allocation a is leximin- maximal if there exists no admissable allocation a(cid:48) such that uc(a) ≺leximin uc(a(cid:48)). A leximin-maximal allocation has a desirable 'fairness'-property to it: The most important priority in a leximin-maximal allocation, is that the lowest utility among all the agents is as high as possible. As a second most important priority, the second-lowest utility among all the agents is made as high as possible, etcetera. Finally we are ready to state the problem that we will prove to be in P. Definition 10 (LMMUAB-ALLOCATION (i.e. Leximin-maximal max-utility atomic bids resource allocation)). A problem instance of LMMUAB-ALLOCATION is a resource alloca- tion problem setting (cid:104)A,O, U,C, uc(cid:105) and a vector K, where • uc is a leximin-egalitarian collective utility function, • C = {cpreempt}, • ∀u ∈ U : u is a max-utility function with atomic demands. • K ∈ Rn It is sufficient to represent a LMMUAB-ALLOCATION-instance as the triple (cid:104)A,O, D(cid:105), where D = {D1, . . . , Dn} is a set of atomic demand sets, and for 1 ≤ i ≤ n, Di is the atomic demand set associated with ai and ui. The task is to determine if there exists an admissable allocation a such that K ≺leximin uc(a). We prove LMMUAB-ALLOCATION in P by giving a polynomial time algorithm for its optimization variant. Definition 11 (LMMUAB-ALLOCATION-OPT (i.e. Leximin-maximal max-utility atomic bids resource allocation, optimization variant)). A problem instance of LMMUAB-ALLOCATION- OPT is the same as a problem instance of LMMUAB-ALLOCATION, but without the vector K. The task is to find a leximin-maximal, admissable allocation. 4 2.2 A polynomial time algorithm for LMMUAB-ALLOCATION-OPT Consider the following algorithm for LMMUAB-ALLOCATION-OPT: Algorithm A: I, an instance of LMMUAB-ALLOCATION-OPT. That is, I = (cid:104)A = {a1, . . . , an},O = {o1, . . . , om}, D = {D1, . . . , Dn}(cid:105), and for 1 ≤ i ≤ n, Di = {ri,1, . . . , ri,m}. a, a leximin-maximal allocation for I. 1. Create a complete weighted bipartite graph G = (V = (L ∪ R), E), where L and R are the left and right parts of the graph respectively. We set L := O, R := A. 2. Generate weights (cid:96)i,j for all {ai, oj} ∈ E such that (cid:96)i,j ≥(cid:80){(i(cid:48),j(cid:48))ri(cid:48),j(cid:48) >ri,j} (cid:96)i(cid:48),j(cid:48). 3. Find with the Hungarian algorithm [5] a minimum weighted bipartite 4. For all i, j ∈ M, set a(ai) := {oj}. matching M on G, using the weights computed in step 2. Input: Output: Begin End First please note: a minimum weighted bipartite matching is a maximum matching in a weighted bipartite graph such that the cumulative weight of the matching (i.e. the sum of the weights of the edges in the matching) is minimal. See for example [6]. We will now prove that this algorithm is correct and runs in polynomial time. From these two facts it follows that the decision variant of this problem also runs in polynomial time and hence is in P Theorem 12. Algorithm A is a correct algorithm for LMMUAB-ALLOCATION-OPT, i.e. the allocation that algorithm A outputs on an LMMUAB-ALLOCATION-OPT-instance as input, is leximin-maximal. Proof. First note that there exists a leximin-maximal allocation in which every agent gets at most one resource. This is due to the combination of max-utility functions with atomic demands: of a bundle allocated to an agent, only a single resource in that bundle decides the agent's utility of that bundle, so we could just as well remove all the other items from the bundle. Step 4 allocates an item to an agent if the corresponding edge is in M. Because M is a minimum weighted matching, an agent is allocated at most 1 item. What remains is proving that if our algorithm has found a minimum weighted matching M, then the algorithm constructs a leximin-maximal a. Suppose that is not the case: call the leximin-maximal allocation aOPT, and assume our algorithm returns an a such that uc(a) ≺leximin uc(aOPT). By the definition of the leximin order ≺leximin this means that ∃i : ∀j < i : uc(a)↑ j = uc(aOPT)↑ j ∧ uc(a)↑ i < uc(aOPT)↑ i . We will now prove that there exists not such an i, resulting in a contradiction. We prove i . For the remainder of the proof, let by induction that for all 1 ≤ i ≤ n : uc(a)↑ MOPT be the matching that corresponds to aOPT, in the same way as M corresponds to a. i = uc(aOPT)↑ 5 1 = uc(aOPT)↑ Base case uc(a)↑ 1. First of all, by construction of the weights in step 3, for all 1 ≤ i ≤ n, 1 ≤ i(cid:48) ≤ n, 1 ≤ j ≤ m, 1 ≤ j(cid:48) ≤ m : ri,j < ri(cid:48),j(cid:48) ⇔ (cid:96)i,j > (cid:96)i(cid:48),j(cid:48). So the edge with highest weight in M corresponds to the agent with the lowest utility of the allocation, hence this utility corresponds to uc(a)↑ 1. Secondly, let e and eOPT be the edges with the highest weight that are in M and MOPT respectively. Now, consider the set of edges E> with weights that are strictly greater than the weight of eOPT. By construction of the weights, it follows that any matching in which an e(cid:48) ∈ E> is included, always has a greater cumulative weight than a matching in which eOPT is included as the edge with the highest weight. Step 4 of the algorithm returns the matching with minimum cumulative weight, so the weight of e must be the weight of eOPT. Induction hypothesis ∀j < i : uc(a)↑ i = uc(aOPT)↑ Induction step uc(a)↑ i . This follows more or less trivially from the same OPT be the edges with the i'th highest arguments as given for the base case: weight that are in M and MOPT respectively. Now, consider the set of i'th highest edges Ei > OPT and strictly less than the weight with weights that are strictly greater than the weight of ei of edge ei−j OPT, 1 ≤ j ≤ n − 1. By construction of the weights, it follows that any matching in which an e(cid:48) ∈ Ei > is included as an i'th highest edge, always has a greater cumulative weight OPT is included as an i'th highest edge. Step 3 of the algorithm than a matching in which ei returns the matching with minimum cumulative weight, so the weight of ei must be the weight of ei j = uc(aOPT)↑ j. let ei and ei OPT. Theorem 13. Algorithm A runs in polynomial time. Proof. The complexities of the individual steps of the algorithm are2: • In step 1, m + n nodes and mn edges are constructed. This takes O(mn) time. • In step 2 mn weights are computed. This step is not described in a very constructive way, but it can be easily seen that it can be done by first sorting the union of all the demand vectors, and subsequently constructing the weights from the highest to the lowest element in the sorted array. In this step, the sorting is the most intensive part and takes O(mn log mn) time. • In step 3 the Hungarian algorithm for minimum weighted bipartite matchings is ran. This algorithm needs a helper shortest-path algorithm. If we use Dijkstra's algorithm as a helper algorithm for the Hungarian algorithm, then this step can be done in O((m + n) log(m + n) + (m + n)(m2n2)) time [6]. • Step 4 is clearly done in O(m + n) time. Adding up the complexities of these steps, we conclude that the algorithm can run in O((m + n) log(m + n) + (m + n)(m2n2)) time. Corollary 14 (from theorems 12 and 13). LMMUAB-ALLOCATION is in P. 2We assume a RAM-model where the elementary arithmetic operations take unit time. 6 3 Complexity of deciding whether an allocation is pareto op- timal for agents with additive utility In this section we prove that deciding whether an allocation of resources among a set of agents is coNP-complete if the agents have additive utility functions. We will make use of the definitions given in section 2.1. As said in the introduction of this paper, coNP-completeness has already been proved for the case where agents have k-additive utility functions and k ≥ 2. Definition 15 (k-additive utility). In a resource allocation setting (cid:104)A,O, U,C, uc(cid:105), a utility function ui of an agent ai is k-additive if for each set T ⊆ O with T = k there exists a coefficient αT and for all R ⊆ O it holds that ui(R) = (cid:88) αT . T⊆R k-additive utility functions are a generalisation of additive utility functions. Definition 16 (additive utility). An additive utility function is a k-additive utility function with k = 1, i.e. a 1-additive utility function. An additive utility function can be represented as a set of coefficients: one coefficient for each item in O. Next, we define the notion of Pareto-efficiency. Definition 17 (Pareto-efficiency). In a resource allocation setting (cid:104)A,O, U,C, uc(cid:105), an admiss- able allocation a is Pareto-efficient (also called: Pareto-optimal, or simply efficient) if there exists not a different admissable allocation a(cid:48) where the utility of at least one agent is higher than in allocation a, and the utilities of all other agents are not lower than in allocation a. More formal: allocation a is Pareto-optimal if there exists no allocation a(cid:48) such that ∃ai ∈ A : ui(a(cid:48)(ai)) > ui(a(ai)) ∧ (∀aj ∈ A : uj(a(cid:48)(aj)) ≥ uj(a(aj))). If such an allocation a(cid:48) does exist, then a is not Pareto-optimal and we say that a(cid:48) Pareto- dominates a. Also we say that a can be Pareto-improved to a(cid:48) if a(cid:48) is an allocation that Pareto-dominates a. The process of reallocating items to get from a to a(cid:48) is called a Pareto- improvement. If for a there is no Pareto-improvent possible, then clearly a is Pareto-optimal. Now we state the problem and prove it coNP-complete. Definition 18 (PO-ALLOCATION-ADDITIVE (i.e. Pareto-Optimal Allocation with Addi- tive utility functions)). A problem instance of PO-ALLOCATION-ADDITIVE is a resource allocation problem setting (cid:104)A,O, U,C, uc(cid:105) and an associated admissable allocation a : A → 2O, where • C = {cpreempt}, • ∀u ∈ U : u is an additive utility function. The problem is to decide whether a is Pareto-optimal. The collective utility function uc can be disregarded here, so the problem is representable as the 4-tuple (cid:104)A,O, V, a(cid:105). In this 4-tuple, V = {v1, . . . , vn} represents the utility functions of U. For all 1 ≤ i ≤ n, vi is the representation of ui as described in definition 16. 7 Theorem 19. PO-ALLOCATION-ADDITIVE is coNP-complete. Proof. Showing membership of coNP is easy: If the allocation a of a PO-ALLOCATION- ADDITIVE-instance is not Pareto-optimal, then a certificate would be an allocation that Pareto-dominates a. Proving coNP-hardness for this problem is very difficult. We do it by a Karp reduction from 3-UNSAT. 3-UNSAT is the problem of deciding whether a propositional formula in 3CNF is unsatisfiable. Because satisfiable instances of such a formula are easy to verify, the complement of 3-UNSAT is in NP. Hence 3-UNSAT is in coNP. The reduction is as follows. We are given an instance of 3-UNSAT I with variables {x1, . . . , xw} and clauses {c1, . . . , cw(cid:48)}. A clause is given as a set of at most 3 literals. We transform this instance to a PO-ALLOCATION-ADDITIVE instance I(cid:48) in the following way. As in the definition, I(cid:48) is represented as the 4-tuple (cid:104)A,O, V, a(cid:105). • In I(cid:48), A = 2w + w(cid:48) + 2: For each variable xi in I, two agents are introduced: aset(xi) and aset(¬xi). aset(xi) represents the set of clauses in which the literal xi occurs. aset(¬xi) represents the set of clauses in which the literal ¬xi occurs. For each clause ci in I, one agent aci is introduced in I(cid:48). Lastly, 2 additional agents are introduced: aunassigned and asatisfied. • In I(cid:48), O = w + w(cid:48) + L + 1, where L is the total number of literals in the formula. For each clause ci we introduce for each literal l in that clause the resource oci,l. For each variable xi we introduce the resource oxi. For each clause ci we introduce the resource oci. Lastly, the resource osatisfied is added. • The additive utility functions V of the agents are specified as follows. Remember that we use the following names: V = {vset(x1), . . . , vset(xw)} ∪ {vset(¬x1), . . . , vset(¬xw)} ∪ {vc1, . . . , vc(cid:48) w} ∪ {vunassigned, vsatisfied}. All v ∈ V are vectors of coefficients. We name these coefficients as follows. Let ai ∈ A, and let oj ∈ O. Thus, i and j stand not for numbers in this case, but for subscripts. Then the coefficient for resource j in the additive utility function of agent i goes by the name of αi,j (and hence αi,j ∈ vi). The coefficients for all resources for all agents are set to zero, with the following excep- tions: -- All coefficients in {αunassigned,x1, . . . , αunassigned,xw} are set to 1. -- All coefficients in {αsatisfied,c1, . . . , αsatisfied,cw(cid:48)} are set to 1. -- All coefficients in {αc1,c1, αc2,c2, . . . , αcw(cid:48) ,cw(cid:48)} are set to 1. -- For all coefficients αset(l),xi in {αset(x1),x1, αset(x2),x2, . . . , αset(xw),xw} ∪ {αset(¬x1),x1, αset(¬x2),x2, . . . , αset(¬xw),xw}, αset(l),xi is set to the number of times that l occurs in the formula of I. 8 -- All coefficients in are set to 1. -- All coefficients in {αset(l),(ci,l)1 ≤ i ≤ w(cid:48) ∧ l ∈ ci} {αci,(ci,l)1 ≤ i ≤ w(cid:48) ∧ l ∈ ci} are set to 1. -- αsatisfied,satisfied is set to w(cid:48) and αsatisfied,unassigned is set to w + 1. • Lastly, we must specify the allocation a. -- All resources {ox1, . . . , oxw} are allocated to aunassigned. -- For all resources oci, 1 ≤ i ≤ w(cid:48) we allocate oci to aci. -- All resources oci,l, 1 ≤ i ≤ w(cid:48), l ∈ ci, are allocated to aset(l). -- The resource osatisfied is allocated to agent asatisfied. That completes the reduction. It can clearly be done in polynomial time. Before contin- uing with the correctness proof of this reduction, an example would be appropriate, due to the complexity of the reduction. Consider the 3-UNSAT instance given by the formula (x1 ∨ x2 ∨ ¬x3) ∧ (¬x1 ∨ ¬x2 ∨ ¬x3). We represent this instance as the tuple (cid:104) {x1, x2, x3}, {c1 = {x1, x2,¬x3}, c2 = {¬x1,¬x2,¬x3}} (cid:105) Now if we run the reduction process on this instance, we get a PO-ALLOCATION- ADDITIVE instance that is displayed in the table below. The columns of the table represent the agents and the rows of the table represent the items. The entries in the table are the coefficients. An entry is displayed in italic if the item of the corresponding row is allocated to the agent of the corresponding column. Empty cells in the table should be regarded as zero entries. ox1 ox2 ox3 oc1 oc2 oc1,x1 oc1,x2 oc1,¬x3 oc2,¬x1 oc2,¬x2 oc2,¬x3 osatisfied ac1 ac2 aset(x1) 1 aset(¬x1) 1 aset(x2) aset(¬x2) aset(x3) aset(¬x3) 1 1 1 1 1 1 1 1 1 1 1 1 1 1 9 2 1 1 asatisfied aunassigned 1 1 1 1 1 2 4 Now we will continue with the correctness proof. We must show that there only exists a Pareto-dominating allocation if the formula of the 3-UNSAT instance is satisfiable. This follows from the following two lemmas and concludes the proof. Lemma 20. If the 3-UNSAT instance I is a NO-instance, i.e. the formula is satisfiable, then the allocation a in I(cid:48) is not Pareto-optimal. Proof. First have to explain the function of all agents and resources with respect to the 3- UNSAT instance I. The allocations of resources {ox1, . . . , oxw} represent to which truth-value the variables are set. If oxi is allocated to aunassigned, this means that xi is set to no truth- value. If oxi is allocated to aset(xi), this means that xi is set to true, and the clauses in which the literal xi occurs are made true. If oxi is allocated to aset(¬xi), this means that xi is set to false, and clauses in which the literal ¬xi occurs are made true. The agents {ac1, . . . , acw(cid:48)} represent the clauses of the formula. If resource oci ∈ {oc1, . . . , ocw(cid:48)} is allocated to aci, it means that clause ci is not satisfied. If resource oci ∈ {oc1, . . . , ocw(cid:48)} is allocated to asatisfied, it means that clause ci is satisfied. In allocation a, all clauses are unsatisfied and all variables are not assigned a truth-value. If in allocation a, we reallocate some oxi ∈ {x1, . . . , xw} to one of the agents aset(lxi ) ∈ {aset(xi), aset(¬xi)}, then by construction we can move all of the , lxi ∈ cj to acj without lowering the utility of aset(lxi ). Now, because cj gets 1 resources ocj ,lxi extra utility, we are able to reallocate ocj to asatisfied. The key thing to see here is that the procedure we just described is, from the viewpoint of I, equivalent to assigning xi some truth value, and making all clauses true in which the In I(cid:48) this is the same as reallocating literal occurs that corresponds to that truth-value. some specific resources to some specific agents, and this reallocation can be done without lowering anyone's utility except for the utility of aunassigned. The utility of aunassigned can only be compensated if aunassigned gets allocated the resource osatisfied. If that happens, then by construction the utility of aunassigned gets suddenly strictly higher than in allocation a. But we can only reallocate osatisfied to aunassigned if all resources {oc1, . . . , ocw(cid:48)} are allocated to asatisfied, otherwise the utility of asatisfied would be too low. Reallocating all of these resources is clearly equivalent with finding a satisfying truth-assignment for the formula. Now we wil describe the reallocation process in a more systematic way: When the propo- sitional CNF formula denoted by instance I is satisfiable, there is an allocation a(cid:48) that Pareto- dominates a. It can be obtained in the following way. 1. Take allocation a and reallocate the resources {ox1, . . . , oxw} to the allocation that corresponds to the assignment that satisfies the formula of I. By doing this, the utility of aunassigned becomes lower than the utility it has in allocation a. This problem will be dealt with in step 4. 2. By construction, all of the other resources of the agents that obtained a resource in step 1 can now all be reallocated so that the utility of those agents is not decreased below the utility they have in allocation a. (The resource they received in step 1 gets them high enough utility to maintain at least the same utility as in a, even if they lose all of their other resources.) So we reallocate all those resources 'appropriately' to the agents {ac1, . . . , acw(cid:48)}. By appropriately we mean that a reallocated resource is reallocated to the single other agent that has non-zero utility for it. By construction, there is precisely one such agent for each item that is reallocated in this step. 10 3. Because, in step 2, the utility of agents {ac1, . . . , acw(cid:48)} is increased, we can reallocate the items {oc1, . . . , ocw(cid:48)} to agent asatisfied. Without giving the agents {ac1, . . . , acw(cid:48)} a lower utility than in allocation a. Now it is the case that each agent except aunassigned has a utility that is at least as high as allocation a. aunassigned has no items allocated, so his utility is 0. The utility of asatisfied is 2w(cid:48) in our current allocation, while in allocation a it was w(cid:48). 4. So, as a last step, we can reallocate osatisfied to aunassigned. The utility of aunassigned is then w + 1 in our new allocation a(cid:48), while it was only w in allocation a. By performing this last step, the utility of asatisfied decreases to w, but this is not a problem since the utility of asatisfied was also w in allocation a. Lemma 21. If the 3-UNSAT instance I is a YES-instance, i.e. the formula is unsatisfiable, then the allocation a in I(cid:48) is Pareto-optimal. Proof. In an allocation a(cid:48) that Pareto-dominates allocation a, at least one agent has strictly greater utility in a(cid:48) than he has in a, and all the other agents have a utility that is at least as great. We divide the proof up in cases, and show that in a(cid:48) no agent can be the agent that has strictly greater utility than he has in a, while all other agents don't have a lower utility than they have in a. Agent aunassigned: In a(cid:48), the utility of agent aunassigned can only be greater than in a if he gets the resource osatisfied. Because the other agents may not have lower utility than they have in a, agent asatisfied needs then be allocated the set of items {oc1, . . . , ocw(cid:48)}. By the same argument, every agent aci ∈ {ac1, . . . , acw(cid:48)} needs to get allocated at least one of the resources {oci,ll ∈ ci}. If we allocate such a resource oci,l to aci, then the utility of aset(l) gets too low, and we must compensate by allocating the resource oxj , xj ∈ l to aset(l). As explained in the previous lemma, regarding I this is equivalent to setting the variable xi to a truth value such that clause cj gets satisfied. We must do this for all clauses, so then there must be an assignment where all of the clauses are satisfied, i.e. I must be a satisfiable instance. Which it isn't. All other cases: It is also impossible to create an allocation a(cid:48) that Pareto-dominates a, where some agent ai (cid:54)= aunassigned has strictly greater utility than in a, while all the other agents have a utility that is at least as high as the utility that they had in a: no matter what agent we choose for the role of ai, it is always neccessary to allocate at least one of the resources in {ox1, . . . , oxw} to an agent other than aunassigned. This means that we are required to allocate osatisfied to aunassigned, and we fall back to the case we just proved for agent aunassigned. It is easy to check that this is true for any ai that we pick. 11 4 Complexity of finding an efficient and envy-free allocation for agents with additive utility The proof given in the previous section was somewhat of an intermediate result that we came across in the process of finding a proof for our next theorem. We first make an additional definition. Definition 22 (Envy-freeness). Given a resource allocation setting (cid:104)A = {a1, . . . , an},O, U = {u1, . . . , un},C, uc(cid:105) and an admissable allocation a, a is called envy-free iff ∀ai ∈ A : ∀aj ∈ A : ui(a(ai)) ≥ ui(a(aj)). not admissable if a is not envy-free. We can define an envy-freeness constraint cenvyfree so that we can add it to C. a then is If there exists an i and there exists a j for which ui(a(aj)) > ui(a(ai)) and i (cid:54)= j, then a is not envy-free and we say that ai envies aj in allocation a. Now we state the problem and give a proof that this problem is Σp 2-complete. Definition 23 (EEF-EXISTENCE-ADDITIVE). In the problem EEF-EXISTENCE-ADDITIVE we must decide whether there exists a Pareto-efficient and envy-free admissible allocation in the resource allocation setting (cid:104)A,O, U,C, uc(cid:105), where • C = {cpreempt, cenvyfree}, • ∀u ∈ U : u is an additive utility function. The collective utility function uc can be disregarded here, so the problem is representable as the 3-tuple (cid:104)A,O, V (cid:105). In this 3-tuple, V = {v1, . . . , vn} represents the utility functions of U. For all 1 ≤ i ≤ n, vi is the representation of ui as described in definition 16. Theorem 24. EEF-EXISTENCE-ADDITIVE is Σp 2-complete. Proof. Membership of Σp 2 is easily shown. The problem can be decided by an alternating turing machine that makes 1 alternation and starts in an existential state: In the existential state, an allocation a is guessed, and it is checked if this allocation is envy-free. The turing machine then enters the universal state. In this universal state it is checked for all possible allocations if an allocation Pareto-dominates a. If this is not the case, then a is Pareto-efficient and envy-free. We prove hardness by a Karp reduction from the complement of the problem ∀∃3CNF. ∀∃3CNF is Πp 2. It is perhaps the most well known complete problem in the second level of the polynomial hierarchy. We selected this problem from [7], a list of complete problems in the polynomial hierarchy. An instance of ∀∃3CNF consists of two disjoint sets of propositional variables X∀ = {x∀ X∃} and a propositional formula in 3CNF over the 1, . . . , x∀ variables in X∀ ∪ X∃. This propositional formula is represented as the set of clauses C = {c1, . . . , cC}. A clause ci ∈ C is a set of at most 3 literals. The problem for a ∀∃3CNF- instance is to decide whether for every possible assignment of the variables in X∀, there exists some assignment of the variables of X∃ that makes the formula true3. 2-complete, that is, complete for the complement of Σp X∀} and X∃ = {x∃ 1, . . . , x∃ 3To remove ambiguity: please note that the assignment of the variables in X∃ needs not be the same for every assignment of the variables in X∀. 12 For this proof we must introduce some additional terminology: given a set of propositional variables, in a partial truth-assignment, or simply partial assignment to these variables, only a part of the variables are assigned a truth value, and the other part is left unassigned. Also, given a partial assignment s on a set of propositional variables and a propositional formula on these propositional variables, we say that the formula is satisfiable on s iff we can transform s into a full assignment s(cid:48) by assigning in s a truth-value to the unassigned variables, such that s(cid:48) satisfies the formula. We make a minor assumption on the ∀∃3CNF instances. For every variable x ∈ X∀ ∪ X∃, both the literals xi and ¬xi must appear at least once in the formula C. Fortunately, if we have a ∀∃3CNF instance this assumption can be made without loss of generality: where the assumption doesn't hold for some variable x ∈ X∀ ∪ X∃, then we can simply add the tautological clause {x,¬x} to C. We make this assumption in order to reduce the complicatedness of our reduction. In this proof we use the following notational conventions. We will use the symbol l to refer to a literal and we will use for any variable xi ∈ X∃ ∪ X∀ the symbol lxi to refer to a literal in which xi occurs. Also, if we use the notation ¬lxi, then by that we mean the positive literal xi if lxi is a negative literal, and we mean the negative literal ¬xi if lxi is a positive for each literal of each variable xi ∈ X∃ ∪ X∀ as the set literal. Lastly, We define the set Clxi of clauses in which lxi occurs. The reduction in this proof resembles the reduction in the proof of theorem 19: we reuse a lot of the same ideas and tricks. The reduction for this proof however, is more complex. We have to deal this time with universally quantified variables and envy-freeness. Moreover, we cannot "set" an allocation in advance, as we could in the reduction of the proof of theorem 19. We will now describe the entire reduction. We advise the reader to work out an example for a small ∀∃3CNF-instance in the table format as we did in the proof of theorem 19. This is because we won't give an example in this proof: the table format size of the EEF- EXISTENCE-ADDITIVE instance is too large to put on this sheet, even for small instances. Given a ∀∃3CNF-instance I = (cid:104)X∀ = {x∀ 1, . . . , x∀ X∀}, X∃ = {x∃ 1, . . . , x∃ X∃}, C = {c1, . . . , cC}(cid:105), we reduce it to a EEF-EXISTENCE-ADDITIVE-instance I(cid:48) = (cid:104)A,O, V (cid:105) in the following way. • A = 4X∀ + 2X∃ +C + L∀ + 3, where L∀ is the total number of literal occurences in C of variables in X∀. For each variable x∀ i ), ahelper set(x∀ i ) and ahelper i ) and aset(¬x∃ set(¬x∀ i ) i ) are introduced. For each clause ci ∈ C, the agent aci is introduced. For all ci ∈ C, for each literal l ∈ ci wherein a variable of X∀ occurs, we introduce the agent aenvyprotection . The remaining three agents are aunassigned, aenvyprotection For ease of explaining and understanding the rest of the proof, we introduce the following symbols and terminology: are introduced. For each variable x∃ i ∈ X∀, four agents aset(x∀ i ∈ X∃, two agents aset(x∃ , and asatisfied. i ), aset(¬x∀ unassigned ci,l -- We refer to the set {ac1, . . . , acC} as Aca. Alternatively, we may refer to those i ∈ l} as Aevaa Alternatively, we may refer to those -- We refer to the set {aset(l)x∃ agents as clause agents. agents as existential variable assignment agents. 13 -- We refer to the set {aset(l)x∀ set(l) x∀ -- We refer to the set {ahelper agents as universal variable assignment agents. i ∈ l} as Auvaa Alternatively, we may refer to those i ∈ l} as Auvaha Alternatively, we may refer to those c ∈ C ∧ l ∈ c} as Aulepa. Alternatively, we may agents as universal variable assignment helper agents. -- We refer to the set {aenvyprotection c,l refer to those resources as universal literal envy-protection agents. Using these definitions, we have A = Aca ∪ Aevaa ∪ Auvaa ∪ Auvaha ∪ Aulepa ∪ {aunassigned, aenvyprotection unassigned , asatisfied}. . For all variables x∃ i ∈ X∀, we introduce the resources ox∀ • O = 4X∀+X∃+2C+L+L∀+3, where L is the total number of literal occurences in the 3CNF formula C, and L∀ is the total number of literal occurences in C of variables in X∀. For all variables x∀ , ohelper set(x∀ i ) and ohelper . For each clause set(¬x∀ i ) . For all ci ∈ C, for each literal ci ∈ C, we introduce the resources oci and ocompensation l ∈ ci, we introduce the resource oci,l. For all ci ∈ C, for each literal l ∈ ci wherein a variable of X∀ occurs, we introduce the resource oenvyprotection . The remaining three resources are osatisfied, oenvy1 and oenvy2. For ease of explaining and understanding the rest of the proof, we introduce the following symbols and terminology: x∀ i ∈ X∃, we introduce the resource ox∃ , ocompensation ci,l ci i i i -- We refer to the set {oc1, . . . , ocC} as Ocr. Alternatively, we may refer to those } as Occr. Alternatively, we may -- We refer to the set {ocompensation resources as clause resources. , . . . , ocompensation c1 cC refer to those resources as clause compensation resources. -- We refer to the set {oc,lc ∈ C ∧ l ∈ c ∧ x∀ ∈ l ∧ x∀ ∈ X∀} as Oulr. Alternatively, we may refer to those resources as universal literal resources. -- We refer to the set {oc,lc ∈ C ∧ l ∈ c ∧ x ∈ l ∧ x ∈ X∃} as Oelr. Alternatively, we may refer to those resources as existential literal resources. -- We refer to the set Oulr ∪Oelr as Olr. Alternatively, we may refer to those resources -- We refer to the set {ox∀ } as Ouvr. Alternatively, we may refer to those as literal resources. , . . . , ox∀ } as Oevr. Alternatively, we may refer to those resources as universal variable resources. -- We refer to the set {ox∃ , . . . , ox∃ X∃ resources as existential variable resources. X∀ as variable resources. -- We refer to the set {ocompensation 1 1 x∀ 1 -- We refer to the set Ouvr∪Oevr as Ovr. Alternatively, we may refer to those resources , . . . , ocompensation x∀ X∀ } as Ouvcr. Alternatively, we may refer to those resources as universal variable compensation resources. -- We refer to the set {ohelper set(x∀ i ) } as Ouvahr. Alternatively, we may refer to those resources as universal variable assignment helper resources. } ∪ {ohelper set(¬x∀ i ) , . . . , ohelper set(¬x∀ , . . . , ohelper set(x∀ X∀) X∀) 14 -- We refer to the set {oenvyprotection c ∈ C ∧ l ∈ c} as Oulepr. Alternatively, we may refer to those resources as universal literal envy-protection resources. c,l Using these definitions, we have O = Ocr∪Occr∪Oulr∪Oelr∪Ouvr∪Oevr∪Ouvcr∪Ouvahr∪Oulepr∪{osatisfied, oenvy1, oenvy2}. • To complete the reduction, we specify the additive utility functions. Due to the extensive use of subscripts and superscripts for the agents and resources, we don't use the same notation for this as we did in the proof for theorem 19. All members of V are vectors of coefficients. vi ∈ V is the vector representing the additive utility function of agent ai. The members of vi are coefficients. In vi there is one coefficient for each resource in O. We name these coefficients as follows. Let a ∈ A, and let o ∈ O. Then we simply denote the utility-coefficient of agent a for resource o as α[a, o]. In the list below, let M be an extremely large number. By default all coefficients of all agents are set to zero, with the following exceptions: -- For all oci ∈ Ocr: α[aci, oci] α[asatisfied, oci] , oci] ∀l ∈ ci : α[aenvyprotection ci,l := M, := 1, := M. -- For all ocompensation ci ∈ Occr: α[aci, ocompensation α[aunassigned, ocompensation ci ci := M − 1, := 1. ] ] -- For all oc,l ∈ Oulr: -- For all oc,l ∈ Oelr: -- For all ox∀ i ∈ Ouvr: -- For all ox∃ i ∈ Oevr: α[ac, oc,l] set(l) , oc,l] , oc,l] α[ahelper α[aenvyprotection c,l := 1, := 1, := 1. α[ac, oc,l] α[aset(l), oc,l] := 1, := 1. α[aset(x∀ α[aset(¬x∀ i i ), ox∀ i ), ox∀ i ] ] := 1, := 1. α[aset(x∃ i ), ox∃ α[aset(¬x∃ i ), ox∃ α[aunassigned, ox∃ i i i i , := Cx∃ := C¬x∃ , := 1. i ] ] ] 15 -- For all ocompensation x∀ i ∈ Ouvcr: i x∀ i ), ocompensation α[aset(x∀ i ), ocompensation α[aset(¬x∀ α[aunassigned, ocompensation x∀ i x∀ i ] ] ] := 1, := 1, := 1. -- For all ohelper x∀ set(l i ) ∈ Ouvahr: α[ahelper x∀ set(l i α[aset(l α[aset(¬l x∀ i x∀ i i set(l ), ohelper x∀ ), ohelper x∀ set(l ), ohelper x∀ set(l i i := Cl x∀ i , := 1, := 1. )] )] )] -- For all oenvyprotection c,l ∈ Oulepr: α[aenvyprotection c,l , oenvyprotection c,l ] := M. α[aunassigned, osatisfied] α[asatisfied, osatisfied] := X∃ + X∀ + C + 1, := C. α[aunassigned, oenvy1] α[asatisfied, oenvy1] := 2 × α[aunassigned, osatisfied], := 1 2 . -- For osatisfied: -- For oenvy1: -- For oenvy2: α[aunassigned, oenvy2] , oenvy2] α[aenvyprotection unassigned := α[aunassigned, oenvy1] + X∃ + X∀ + C, := M. That completes the reduction. It should be obvious that generating this EEF-EXISTENCE- ADDITIVE-instance from the ∀∃3CNF instance takes polynomial time. We now continue with the correctness proof. ∀∃3CNF is a Πp 2-complete problem, and we want to prove EEF-EXISTENCE-ADDITIVE 2-complete. Therefore we need to show that in I(cid:48) there is only a Pareto-efficient, envy-free is Σp (EEF) allocation if there exists some assignment to the variables in X∀ for which there is no assignment to the variables in X∃ which makes the 3CNF-formula C true. by giving the definition and lemmas that are ommitted in the outline. Now we will outline the correctness-proof for this reduction. After that we finish the proof We define in definition 25 the specific set of allocations for I(cid:48), that correspond to a specific type of partial truth-assignment to the variables in I. Namely, assignments that satisfy the following two conditions: 16 1. All universally quantified variables are set to either true or false, and 2. all existential variables are left unassigned. In lemma 26 we prove that all allocations that correspond to such truth-assignments are envy-free. We call these allocations X∀-allocations. We will show in lemma 27 that in I(cid:48), any EEF allocation must be an X∀-allocation. Next, we will show in lemmas 28 and 29 that for an X∀-allocation, a Pareto-improvement is possible only if in I the formula can get satisfied on the partial truth-assignment that corresponds to this X∀-allocation. Now if I is a YES- instance of ∀∃3CNF, then clearly the formula is satisfiable on all partial assignments with the two aforementioned conditions, hence a pareto-improvement is possible on all envy-free allocations. So then I(cid:48) is a NO-instance of EEF-EXISTENCE-ADDITIVE. On the other hand, if I is a NO-instance of ∀∃3CNF, then clearly there must be a partial assignment satisfying the 2 aforementioned conditions for which the formula is not satisfiable. Hence there is in this case an envy-free allocation that is pareto-optimal. The remainder of the proof consists of definition 25 and lemmas 26, 27, 28 and 29. Definition 25 (X∀-assignments and X∀-allocations (corrected)). For I, we define an X∀- assignment as a partial assignment to the variables in X∀ ∪ X∃ where all variables in X∀ are set to either true or false, and all variables in X∃ are not assigned to a truth value. Given an X∀-assignment s, we define the corresponding X∀-allocation in the following way: 1. All agents aci ∈ Aca get allocated the resource oci. 2. For all x∃ 3. For all x∀ i x∃ ) ∈ Aevaa get allocated the resources {oc,l i ∈ X∃, all agents aset(l i ∈ X∀, for all pairs of agents aset(x∀ ∈ c}. i ) ∈ Auvaa. Allocate ox∀ i ) ∈ Auvaa, aset(¬x∀ to ). Now, if xi is true in s, one of the two agents, it doesn't matter which one, say aset(l x∀ allocate ohelper ). Otherwise, if xi is false in set(x∀ i ) ) and s, allocate these two resources the other way around: allocate ohelper set(x∀ i ) ) and allocate ohelper set(¬x∀ i ) to ahelper set(¬l to ahelper set(¬l to aset(¬l lx∃ x∀ x∀ x∃ i i i i i i x∀ i allocate ohelper set(¬x∀ i ) to aset(¬l ). i x∀ ∈ Aulepa get the resource oenvyprotection c,l 4. All agents aenvyprotection 5. aunassigned gets allocated all of the resources Oevr ∪ Occr ∪ Ouvcr ∪ {oenvy1}. 6. aenvyprotection gets allocated the resource oenvy2 c,l . unassigned 7. asatisfied gets allocated the resource osatisfied. 8. 9. The only resources that have not been allocated up to this point are the universal literal or set(l) . It doesn't matter which of the two. If l is true in s, then resources oc,l. If l is not true in s, then oc,l can be allocated to either aenvyprotection they are allocated to ahelper oc,l must be allocated to ahelper set(l) , and thus may not be allocated to aenvyprotection c,l c,l . Lemma 26. All X∀-allocations are envy-free. 17 Proof. Let a be any X∀-allocation for I(cid:48) and let s be the corresponding X∀-assignment for I. For every agent we will show that he doesn't envy any other agent. In this proof we say that an agent wants a resource if the agent has a non-zero utility-coefficient for that resource. For simplicity we also say that an agent has a resource if he is allocated that resource. • aenvyprotection unassigned a non-zero utility-coefficient. doesn't envy any agent because he has the single resource for which he has • asatisfied doesn't envy any agent. Its utility in allocation a is C; the total utility of the C + 1 resources that he wants but doesn't have is C + 1. For all of these C + 1 resources, asatisfied has a utility-coefficient of 1. So asatisfied would only envy an agent if there is an agent in a that has all of these C + 1 resources, and that's not the case. • aunassigned doesn't envy any other agent because the only items he wants but doesn't have are oenvy2 and osatisfied. The former is allocated to aenvyprotection and the latter is unassigned allocated to asatisfied. aunassigned doesn't envy aenvyprotection because the utility-coefficient that aunassigned has for oenvy2 is equal to (and not higher than) the utility that aunassigned currently has in a. aunassigned also doesn't envy asatisfied because the utility-coefficient that aunassigned has for osatisfied is lower than the utility that aunassigned currently has in a. unassigned • For all aenvyprotection ∈ Aulepa: aenvyprotection c,l c,l c,l has an item for which he has a utility coeffi- cient of M. For aenvyprotection , there are two more items that he wants. For one of those items he has a utility-coefficient of M. For the other item he has a utility-coefficient of 1. These items are not both allocated to the same agent, so aenvyprotection envies no-one. • All aci ∈ Aca have no envy: aci has a utility of M. The total utility of all items that aci wants but doesn't have is M − 1 + ci. For the resource ocompensation , aci has a utility coefficient of M − 1. For the other resources that aci wants but doesn't have (at most 3), aci has a utility coefficient of 1. These are literal resources. Literal resources and ocompensation are not all allocated to the same agent in allocation a, so aci doesn't envy ci any agent. c,l ci • For all aset(l) ∈ Aevaa, aset(l) has a utility of Cl in a. The maximal utility they can have is 2Cl, so aset(l) doesn't envy anyone because he already has half of his total possible utility. • For all ahelper set(l) ∈ Auvaha, ahelper set(l) has a utility of at least Cl in a. The maximal utility they can have is 2Cl, so aset(l) doesn't envy anyone because he already has half of his total possible utility. • All aset(l) ∈ Auvaa have a utility of 1 in a. The maximal utility they can have is 4. There are 3 items that aset(l) wants but doesn't have. For all of these 3 items, aset(l) has a utility-coefficient of 1. aset(l) doesn't envy anyone because each of these 3 items is allocated to a different agent: one of these 3 items is allocated to aunassigned, one is allocated to aset(¬l), and one is allocated to either ahelper set(l) or ahelper set(¬l). Lemma 27. All EEF-allocations must be X∀-allocations. 18 Proof. We show this by reasoning about how the resources must be allocated in order to achieve envy-freeness and Pareto-optimality. After having done this, it turns out that the set of allocations that are possibly EEF is exactly the set of all X∀-allocations. First of all, it doesn't make sense to allocate a resource to an agent whose utility-coefficient is zero for that resource. A Pareto-improvement is always possible in such an allocation, by simply reallocating the resource to an agent that has a positive utility-coefficient for it. This is why we will only consider allocating resources to agents who have positive utility-coefficients ∈ Oulepr must for the resources. By this argument it immediately follows that all oenvyprotection be allocated to aenvyprotection c,l . c,l unassigned oenvy2 must be allocated to aenvyprotection , or else he would envy agent aunassigned. Also, we see that aunassigned always envies asatisfied if oenvy1 isn't allocated to aunassigned, because aunassigned has a utility-coefficient of 2(X∃ + X∀) + 2 for oenvy1. This is more than half of the maximal utility it is still able to get (given that oenvy2 is allocated to aenvyprotection unassigned ). unassigned Next, it follows that osatisfied must be allocated to asatisfied, since if it would be allocated to aunassigned, then asatisfied always envies aunassigned because aunassigned then has the items osatisfied and oenvy1. If asatisfied would get this bundle of items, then he has a utility that's more than half of his total possible utility, so asatisfied would envy aunassigned in that case. Given our current set of EEF-allocation-requirements up till now, it's clear that aunassigned must get allocated all of the resources Oevr ∪ Ouvcr ∪ Occr. Only if we allocate all of these resources to aunassigned, then the utility of aunassigned is high enough to not envy aenvyprotection . At this point, it is certain that for all oci ∈ Ocr, oci must be allocated to aci. This must be the case because: firstly, aci has a utility-coefficient of M for this resource; secondly, aci has a utility of M − 1 for ocompensation , but according to our current set of EEF-allocation- ci requirements, ocompensation must already be allocated to aunassigned; and thirdly, aci has a utility-coefficient of 1 for all other resources that aci wants. That is very low compared to M, so even if aci would get all of these resources instead of oci, aci would still envy the agent that gets oci. ∈ Oevr must be allocated to aunassigned, the agents aset(x∃ get allocated all of the resources that aset(x∃ i ) wants, except for ox∃ set of resources {oc,lxi ∈ l}. Allocating these resources to aset(x∃ to α[aset(x∃ reasoning holds for the agents aset(¬x∃ aset(¬x∃ to α[aset(¬x∃ i ) must . These are exactly the i ) makes his utility equal i ) doesn't envy anyone. Analogous i ): They must get allocated all of the resources that i ) makes his utility equal ], and therefore it is ensured that aset(x∃ ], and therefore it is ensured that aset(¬x∃ . Allocating these resources to aset(¬x∃ i ) wants, except for ox∃ Because all items ox∃ i ) doesn't envy anyone. set(¬xi) and ocompensation For all pairs of universal variable assignment agents aset(xi) and aset(¬xi), we have the following situation: the total possible utility that both agents can get is 4: they both have four resources that they want, and they both have a utility of 1 for each resource. Also they both want exactly the same four resources. However, we already concluded that the resources ocompensation must be allocated to aunassigned. According to this requirement, set(xi) the total possible utility that both agents can still get is 3. aset(xi) and aset(¬xi) are the only , so we can only allocate agents that can have a positive utility-coefficient for the resource ox∀ this resource to one of these two agents. If we allocate it to either agent, say aset(lxi ), then the other agent aset(¬lxi ) will envy aset(lxi ) unless he gets allocated one of the other two resources that are left (xhelper set(¬xi)). We can choose either one to allocate to aset(¬lxi ). After we have done this, our only possibility is to allocate the other resource to ahelper set(¬lxi ) (if we allocate set(xi) and xhelper i ), ox∃ i ), ox∃ i i i ci i i i 19 it to aset(lxi ) or aset(¬lxi ) then there will be envy among aset(lxi ) and aset(¬lxi )). For the universal literal resources, the following holds. A universal literal resource oc,l i i i set(l ), or else ahelper x∀ x∀ ) will envy either is assigned to ahelper ), we have the possibility x∀ set(l )}. But if we would allocate , ahelper x∀ set(l }. x∀ has currently only M to , and aenvyprotection x∀ x∀ c,l c,l i i i i x∀ i to ac, and the only possibilities left are to assign oc,l would envy ac because ac has the bundle of items {oc, oc,l i must be allocated to ahelper set(x∀ i ) if ohelper set(x∀ i ) is not assigned to ahelper x∀ set(l i ) or aset(¬x∀ aset(x∀ to allocate oc,l i ). In the case that ohelper set(x∀ i ) to one of the agents in {ac, aenvyprotection x∀ i i i c,l x∀ x∀ to ac, then aenvyprotection oc,l Having this bundle would give M + 1 to aenvyprotection utility. So we cannot allocate oc,l either aenvyprotection x∀ x∀ c,l i i c,l x∀ i or ahelper x∀ set(l ). i The requirements we just described clearly restrict the set of allocations that are possibly EEF, to the set of X∀-allocations. Lemma 28. Given an X∀-assignment s for I, and the X∀-allocation a in I(cid:48) that corresponds to s. If the propositional 3CNF-formula C is satisfiable on s, then there is an allocation a(cid:48) that Pareto-dominates a. Proof. Let s be the X∀-assignment and a be the corresponding X∀-allocation. Given a, it is possible to reallocate some resources to yield a Pareto-dominating allocation a(cid:48) where the utility of aunassigned is increased, and the utility of the other agents is at least as high as in a. First note that the only way to increase the utility of aunassigned is to reallocate the resource osatisfied from asatisfied to aunassigned. If this happens, then aunassigned gets X∀ + X∃ + C + 1 extra utility, so in that case aunassigned can lose X∀ + X∃ + C utility, and he will still have higher utility than in a. We can only move osatisfied to aunassigned if we reallocate all of the clause resources to asatisfied, otherwise the utility of asatisfied would be too low. If we reallocate all of these clause resources, then all clause agents would lose M utility. We can compensate this by reallocating all of the clause compensation resources to the clause agents (this gives M − 1 utility to each clause agent). There are two problems with this move: first of all, by doing this, aunassigned loses C utility; and secondly each clause resource only gets M − 1 utility, so we need to allocate each clause resource at least 1 more utility in order to compensate for the loss of M utility of each clause agent. The first problem turns out not to be a problem at all, because aunassigned has a "surplus" of X∀ + X∃ + C utility, and by reallocating all clause compensation resources, aunassigned loses only C utility, so aunassigned is still allowed to lose X∀ + X∃ utility. The second problem can be remedied by reallocating at least 1 literal resource to each clause agent. A clause agent aci has a utility-coefficient of 1 for a literal resource oci,l and a utility-coefficient of 0 for all other literal resources. Literal resources can be either existential literal resources or universal literal resources: i ∈ Ouvcr, if x∀ is assigned to true (false) in s, reallocate ocompensation 1. For any universal variable we can execute the following procedure. Step 1: for all ocompensation from x∀ aunassigned to the universal variable assignment agent that has currently got the resource set(¬xi)). aunassigned loses X∀ utility by this move, so there is still X∃ utility ohelper set(xi) (ohelper to "spend" for aunassigned. Step 2: if x∀ is assigned to true (false) in s, reallocate the set(xi) (ahelper item ohelper set(¬xi)). Note that by executing steps set(¬xi)) from aset(xi) to ahelper set(xi) (ohelper x∀ i i i 20 i i ∈ {oc,¬x∀ x∀ i ∈ c}), we move oc,x∀ if x∀ 1 and 2, no universal variable assignment agent loses any utility. Step 3: is ∈ {oc,x∀ xi ∈ c} assigned to true (false) in s, then for all of the literal resources oc,x∀ (ahelper (oc,¬x∀ ) to ac. Note set(¬x∀ i ) that by this last step, no universal variable assignment helper agent loses any utility. When we execute the procedure we just described, we can move a certain set of clause resources to asatisfied without lowering anyone's utility. By construction, this set of clause resources corresponds exactly the set of clauses that are satisfied by the X∀-assignment s. The clause resources that correspond to clauses that still need to get satisfied, still need to get reallocated. We will see how to do this in the next step: )from ahelper set(x∀ i ) (oc,¬x∀ i i i i i 2. For any existential variable x∃ i i i x∃ x∃ i , an existential literal resource oc,l x∃ ). If we reallocate oc,l is allocated in a to an to ac, then we have existential variable assignment agent aset(l to compensate this by moving an oxi to aset(x∃ i ). In a, oxi is allocated to aunassigned. So if we reallocate all of the existential variable resources, aunassigned loses X∃ utility. So after reallocating the existential variable resources, aunassigned may not lose any utility anymore, since we still want aunassigned to have strictly greater utility than in a. We can reallocate an existential variable resource ox∃ only to aset(x∃ i ). Altogether this means that for any existential variable x∃ i , we can give extra utility to either the clause agents {acxi ∈ c} or to the clause agents {ac¬xi ∈ c}. Remember that this extra utility is needed for the clause agents in order to be able to reallocate the clause resources to asatisfied. In this sense, reallocating all of the clause resources to asatisfied is equivalent to finding a truth-assignment for the unassigned variables of s such that C is satisfied. Such an assignment exists by our assumption, hence from a, a Pareto- improvement to a(cid:48) is possible. a(cid:48) is clearly not EEF because it is not an X∀-allocation. (More concretely, in a(cid:48), asatisfied envies aunassigned because aunassigned has the bundle of items {aenvy1, asatisfied}.) i ) or aset(¬x∃ i Lemma 29. Given an X∀-assignment s for I, and the X∀-allocation a in I(cid:48) that corresponds to s. If the propositional 3CNF-formula C is unsatisfiable on s, then a is EEF. Proof. By lemma 26, a is envy-free, so we only need to show that a is Pareto-optimal. We do this by proving the following two things: 1. There doesn't exist an allocation a(cid:48) that Pareto-dominates a in which all clause-resources are allocated to asatisfied. 2. Any allocation a(cid:48) that Pareto-dominates a must have all clause-resources allocated to asatisfied. Proof for 1: This is a lot like our story in the previous lemma. We will try to make a Pareto-dominating allocation a(cid:48) where all clause resources are allocated to asatisfied. We do this by trying to transform a into a(cid:48), and we will see that this is not possible. If we take a, and reallocate the clause resources to asatisfied, then all of the clause agents lose M utility. The only way to compensate is reallocating all of the clause compensation 21 resources to the clause agents and reallocating to every clause agent at least one literal resource. If we reallocate the clause compensation resources then the utility of aunassigned is lowered by C and needs to be compensated. The only way to do so is to reallocate osatisfied to aunassigned. This is no problem: the utility of asatisfied was C in allocation a, and now it is still C. The reallocation of at least one literal resource to every clause agent is going to be the problem. There are literal resources allocated to four types of agents: • Some universal literal resources may in a be allocated to universal literal envy- protection agents. It is impossible to reallocate such literals because it is impossible to compensate the utility of these agents by giving them another resource. The only resources these agents want but don't have are the clause resources, but the are already reallocated to asatisfied. • Some universal literal resources may in a be allocated to universal variable assign- ment helper agents ahelper set(l) . It is impossible to reallocate these literal resources: the only resources that ahelper set(l) wants but doesn't have can be literal resources that are allocated to a universal literal envy-protection agent. We can not reallocate these literal resources, as we argued in the previous item of this list. set(l) who already have resource ohelper set(l) set(l) who do not have resource ohelper • Some universal literal resources may in a be allocated to universal variable as- signment helper agents ahelper set(l) . In this case, it is possible to reallocate these literal resources to the clause agents. The only way to compensate the loss of utility of agent ahelper set(l) by reallocating the resource ohelper from one of the two universal variable assignment agents to ahelper set(l) . Subsequently we can compensate the loss of the universal variable assignment agent by reallocat- ing a universal variable compensation resource from aunassigned to him. aunassigned may lose all of its universal variable compensation resources. Its utility will still remain higher than it was in a because it has received the resource osatisfied. Just as in the previous lemma, the procedure we just mentioned will add at least 1 extra utility to a certain set of clause agents. This set of clause agents are the clause agents that correspond to the clauses that are satisfied by the partial truth-assignment s. • The existential literal resources are allocated to the existential variable assignment agents. It is possible to reallocate some of these existential literal resources to the clause agents. As we already pointed out in the previous lemma, for any existential i , we can give extra utility to either the clause agents {acxi ∈ c} or to variable x∃ the clause agents {ac¬xi ∈ c}. So, we can give a literal to the clause agents that correspond to clauses satisfied by s. And the remaining clause agents we can give a literal if it is possible to reallocate an existential literal resource to these clause agent. This is obviously equivalent to finding a truth-assignment to the variables in X∃ that satisfies formula C on s. By our assumption such a truth-assignment doesn't exist, so there exists no allocation a(cid:48) that Pareto-dominates a in which all clause-resources are allocated to asatisfied. 22 Proof for 2: For each agent, we show that we can only transform a to a Pareto- dominating allocation a(cid:48) and increase that agent's utility if we allocate all clause- resources to asatisfied. For agent aunassigned: The only way to improve the utility of aunassigned is to reallocate osatisfied from asatisfied to aunassigned. But then sat would lose C utility. The only way to remedy this is to reallocate all of the C clause-resources to asatisfied. For all existential variable assignment assignment agents aset(l ): The only way to increase the utility of aset(l ). Now, aunassigned loses 1 utility, so we need to increase aunassigned's utility by allocating him the resource osatisfied. So we fall back to the case for aunassigned. from aunassigned to aset(l ) is to reallocate ox∃ x∃ x∃ i i i x∃ i For all universal variable assignment agents aset(l i i i i i i i i x∀ x∀ x∀ x∀ x∀ x∀ x∀ x∀ set(x∀ i ) ) to aset(l ): Reallocating acompensation ) is the same problem as needing to improve the utility of aset(l ) or aset(l to ) is not possible. In that scenario we would again fall back to the case for aset(l ) does not help. This ac- aunassigned. Reallocating an item from aset(¬l ). Hence we would need to be able to increase tion removes 1 utility from aset(¬l x∀ ) have exactly the same utility co- ) and aset(¬l ), but aset(l the utility of aset(¬l efficients, i.e. they are clones of each other. So, needing to improve the utility of ). aset(¬l ). Given We can try one more thing to improve the utility of aset(l that x∀ is false in s, the reasoning is analogous), we can try ) by reallocating to either agent the to increase the utility of aset(l x∀ resource ohelper . Because we remove half of the total possible utility set(¬x∀ i ) with this move, this move can only possibly be done if in a, ohelper of ahelper is set(¬x∀ set(¬x∀ i ) i ) the only resource that ahelper has. But even in this case it will turn out that it's set(¬x∀ i ) impossible: it is easily seen that it would require reallocating a clause resource to a universal literal envy-protection agent, without lowering anyone's utility below the utility he has in allocation a. We will show that this is not possible when we arrive at the case for the universal literal envy-protection agents. is true in s (if x∀ from ahelper set(¬x∀ i ) ) or aset(¬l x∀ x∀ x∀ x∀ i i i i i i i i i i i i set(l set(l ) or ahelper x∀ For all universal variable assignment helper agents ahelper x∀ set(l here: either ahelper x∀ set(l ) has resource ohelper x∀ ): There are two cases ) doesn't have resource ohelper ). x∀ In the former case, it is possible to try to increase the utility of ahelper ) by reallo- x∀ set(l cating the resource ohelper ) gets into trouble x∀ set(l and we have to increase his utility. Therefore we fall back to the case for aset(l ). x∀ In the second case we can only try to increase the utility of ahelper ) by allocating x∀ set(l to him a literal resource for which he has a non-zero utility-coefficient. If there is such a literal resource, then it is allocated in a to a universal literal envy-protection agent. This would require reallocating a clause resource to a universal literal envy- protection agent, without lowering anyone's utility below the utility he has in ). If we do this, then aset(l ) to ahelper x∀ set(l set(l x∀ i i i i i i i 23 allocation a. We will show that this is not possible when we arrive the case for the universal literal envy-protection agents. For all clause agents aci: The utility of a clause agent aci can only be improved by reallocating 1 or more of the literal-resources to him for which aci has non-zero utility. Let oci,l be this literal resource. If we reallocate oci,l to aci, then we would in turn need to improve the utility of an existential variable assignment agent, universal variable assignment helper agent or a universal literal envy-protection agent. For the first two, we refer back to their cases, that we already handled in this list. For the last one, the universal literal envy-protection agent, we will show that it's impossible to increase his utility. We arrive at this case now: For all universal literal envy-protection agents aenvyprotection : The only way to in- crease the utility of aenvyprotection is to reallocate the clause resource oc from ac to aenvyprotection . By this move, ac would lose M utility. To compensate it we need at c,l least to allocate ocompensation from aunassigned to ac. This implies that we will need to increase the utility of aunassigned. From the case we already handled for agent aunassigned, we conclude that we would need to assign all of the clause resources to asatisfied. c,l c,l c For agent asatisfied: We can try to reallocate one or more clause-resources to asatisfied. If we do that, then we need to improve the utility of at least one clause agent. As shown as a previous case in this list, improving the utility of this clause agent implies that we need to move all of the clause resources to asatisfied. Another possibility is to try to reallocate aenvy1 to asatisfied. But then we would need to improve the utility of aunassigned. As shown, this implies that we would need to move all of the clause resources to asatisfied. For agent aenvyprotection unassigned : Obviously we can not improve the utility for this agent, be- cause in a it already has gotten allocated the single resource that he wants. References [1] Sylvain Bouveret, H´el`ene Fargier, J´erome Lang, and Michel Lemaıtre. Allocation of indi- visible goods: a general model and some complexity results. In F. Dignum, V. Dignum, S. Koenig, S. Kraus, M. P. Singh, and M. Wooldridge, editors, Proceedings of AAMAS'05, Utrecht, The Nederlands, July 2005. ACM Press. [2] Sylvain Bouveret and J´erome Lang. Efficiency and envy-freeness in fair division of in- divisible goods: Logical representation and complexity. Journal of Artificial Intelligence Research, 32:525 -- 564, June 2008. [3] Y. Chevaleyre, U. Endriss, S. Estivie, and N. Maudet. Multiagent resource allocation with k-additive utility functions. In Proceeding of the DIMACS-LAMSADE Workshop on Computer Science and Decision Theory (Annales du LAMSADE 3), pages 83 -- 100, 2004. 24 [4] Yann Chevaleyre, Paul E. Dunne, Ulle Endriss, J´erome Lang, Michel Lemaıtre, Nicolas Maudet, Julian Padget, Steve Phelps, Juan A. Rodr´ıguez Aguilar, and Paulo Sousa. Issues in multiagent resource allocation. Informatica, 30:3 -- 31, 2006. Survey paper. [5] H. W. Kuhn. The Hungarian method for the assignment problem. Naval Research Logistic Quarterly, 2:83 -- 97, 1955. [6] Christos H. Papadimitriou and Kenneth Steiglitz. Combinatorial Optimization : Algo- rithms and Complexity. Dover Publications, July 1998. [7] M. Schaefer and C. Umans. Completeness in the polynomial-time hierarchy: a com- pendium. SIGACT News, September 2002. 25
1807.04118	1	1807	2018-07-10T04:40:38	Emergence of Altruism Behavior for Multi Feeding Areas in Army Ant Social Evolutionary System	[ "cs.MA", "cs.ET", "cs.NE" ]	Army ants perform the altruism that an ant sacrifices its own well-being for the benefit of another ants. Army ants build bridges using their own bodies along the path from a food to the nest. We developed the army ant inspired social evolutionary system which can perform the altruism. The system has 2 kinds of ant agents, `Major ant' and `Minor ant' and the ants communicate with each other via pheromones. One ants can recognize them as the signals from the other ants. The pheromones evaporate with the certain ratio and diffused into the space of neighbors stochastically. If the optimal bridge is found, the path through the bridge is the shortest route from foods to the nest. We define the probability for an ant to leave a bridge at a low occupancy condition of ants and propose the constructing method of the optimal route. In this paper, the behaviors of ant under the environment with two or more feeding spots are observed. Some experimental results show the behaviors of great interest with respect to altruism of ants. The description in some computer simulation is reported in this paper.	cs.MA	cs	Emergence of Altruism Behavior for Multi Feeding Areas in Army Ant Social Evolutionary System 8 1 0 2 l u J 0 1 ] A M . s c [ 1 v 8 1 1 4 0 . 7 0 8 1 : v i X r a Takumi Ichimura and Takuya Uemoto Department of Management and Systems, Prefecture University of Hiroshima, Hiroshima, 734-8558 Japan E-mail: {[email protected], [email protected]} Akira Hara Graduate School of Information Sciences, Hiroshima City University, Hiroshima, 731-3194 Japan E-Mail: [email protected] Abstract-Army ants perform the altruism that an ant sac- rifices its own well-being for the benefit of another ants. Army ants build bridges using their own bodies along the path from a food to the nest. We developed the army ant inspired social evolutionary system which can perform the altruism. The system has 2 kinds of ant agents, 'Major ant' and 'Minor ant' and the ants communicate with each other via pheromones. One ants can recognize them as the signals from the other ants. The pheromones evaporate with the certain ratio and diffused into the space of neighbors stochastically. If the optimal bridge is found, the path through the bridge is the shortest route from foods to the nest. We define the probability for an ant to leave a bridge at a low occupancy condition of ants and propose the constructing method of the optimal route. In this paper, the behaviors of ant under the environment with two or more feeding spots are observed. Some experimental results show the behaviors of great interest with respect to altruism of ants. The description in some computer simulation is reported in this paper. Keywords-Army Ants, Altruism Behavior, Multi Agent System, Evolutionary Simulation, Swarm library I. INTRODUCTION In animal societies, self-organization is the theory of how minimal complexity in the individual can generate greater com- plexity at the population. The rules specifying the interactions among the components in the system are implemented by using only local information without global information. Deneubourg et al. [1] developed the model of collective decision making without any form of centralized control. The model was devel- oped to show the characteristic patterns of self-organization by Monte Carlo simulation. In the study of social evolution, army ant performs altruism as one behavior of complexities, where each individual reduces its own fitness but increases the fitness of other individuals in the population. Such behaviors seem to be involved acts of self-sacrifice in order to aid the others. In evolutionary biology, such a behavior is called reciprocal altruism. The concept was initially developed to explain the evolution of cooperation as mutually altruistic acts[2]. The basic idea is close to the strategy of "equivalent relation" in the study of strategic decision making. Army ants are characterized by their two different phases of activities, a nomadic phase and a stationary phase. During c(cid:13)2014 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. the nomadic phase, army ants move during the day to capture insects, spiders, and so on. The stationary phase starts when the larvae pupate for a few weeks. Moreover, army ants build a living nest with their bodies instead of building a nest like other ants. Each ant will hold on to the other legs and form a linked chain or a ball structure. This behavior is known as a bivouac. This allows the bridging of an empty space. In order to address the self-assembled structure as a particular type of aggregation, Deneubourg et al. defined the probability of an ant entering or leaving chain in [3]. Moreover, they showed that the gregarious behavior facilitates cooperation by Blattella germanica in shelters during the resting period. The probability to leave the shelter was defined. Ishiwata et al. [4] developed the simulation system for the foraging behavior and the altruism of army ants by using Swarm library, Swarm-2.2[5]. (The original website www.swarm.org is in the process of being rebuilt.) The prob- abilities to form the chain defined in [6], [7] was used in their simulation experiments. The number of neighboring active ants is considered as the condition for altruistic behavior. Their simulation results show a mimic altruistic behavior. By inspiring Ishiwata's study, Douzono et al. developed the multi-agent simulation system to execute more realistic altruistic behavior where two or more kinds of agents re- alize the sub-tasks of army ants [8], [9]. According to the environment in [8], the simulation results reported that the optimal path from the food to the nest cannot be always found, because two or more chains in the environment were formed. Although more emergence of altruistic behaviors was observed, but the capabilities of forming chain was dispersed. As a result, the performance of foraging decreases and some ants took a circuitous route. On the contrary, Ichimura et al. defined the evaporation rate dues to normal distribution probability and the probability to leave from the chain when the ants in its neighbor region depart gradually in [9]. The altruism simulation results are reported to find more optimal paths from food to the nest. In this paper, we observed the behaviors of ant agents under the multi feeding spots in the same environment of [9]. Some experiments with different ratio of feed size were investigated. In general, ant agents take an action to be concentrated in the largest feeding spot. The shortest path from the spot to the nest is constructed and the ants bring feed to the nest. Then, the feeding spots will be disappeared in the order of larger spot. However, it has turned out that there is a certain tendency without regard to the size of feed. The altruism behavior does not work well and the bridge will be broken, if enough ant agents are not gathered into the ditch. As a result, the food at the spots remains to the end of simulation. We report the experimental results for the emergence of altruism behavior for multi feeding spots in this paper. The remainder of this paper is organized as follows. Section II describes the simulation environment with Swam library. Section III defines the behaviors of agents such as search phase, homing phase (return to the nest), and altruism phase. Section IV describes the proposed method related to pheromone and the leaving probability from chain. Experi- mental results for simulations are described in Section V. In Section VI, we give some discussions to conclude this paper. II. SIMULATION ENVIRONMENT The swarm is the basic unit of simulation for a collection of agents executing a schedule of actions. The Swarm provides object oriented libraries of reusable components for building models and analyzing, displaying, and controlling experiments on those models. We executed the altruism simulation system by pheromone evaporation and its diffusion in army ant multi agent systems. The developed system is depicted in 100× 100 2D space as shown in Fig.1. The solid-filled rectangle, which consists of 3 kind of bars: 'leftbar', 'rightbar', and 'centerbar', divides the space into 2 parts. The inner part is the nest region and the bottom part under the rectangle is the food source part. The 4 coordinates (x,y) of leftbar, rightbar, and centerbar are {(30, 30), (33, 30), (30, 70), and (33, 70)}, {(70, 30), (73, 30), (70, 70), and (73, 70)}, and {(30, 70), (73, 70), (30, 73), and (73, 73)}, respectively. Each bar represents a ditch and the width of ditch is 3. The center rectangle of the space is 'nest' and the bottom rectangle under the bar is 'food source.' Since an ant cannot cross the ditch by itself, some ants begin altruistic behavior to cooperate with each other. The two hypotheses were proposed as the judgment criteria for altruistic activity, Model 1: Based on the Presence of Neighboring Ants and Model 2: Based on the Presence of Pheromone [4]. In Model 1, an ant will start formation of living bridge over a gully only when neighboring ants are present. Hypothetically, this approach will be more efficient compared to forming a bridge blindly. In Model 2, the places where pheromone concentrations are higher than a fixed level are the locations that many ants have passed and/or will pass through in future. Fig.2 shows the area of activities and the visual field by an army ant. In this paper, ant agents can move in the diagonal direction, but the scattered pheromone diminishes compared to the adjacent positions on up and down, left and right. The distance from a position to the neighbor is defined Distance. For example, the distance to 'A' and 'C' in Fig.2 are 1 Distance and 2 Distance, respectively. Practically,'B' is √2Distance. However, for the sake of ease, we define 'B' as 1 Distance in the diagonal direction. III. AGENT BEHAVIORS Fig. 1. Environment in Army Ant Simulation System C B A Fig. 2. The area of activities finds food. Once the pheromone is attenuated and is dispersed, the information about the food position is disseminated among the ants. The system has 2 kinds of ant agents, 'Major ant' and 'Minor ant' and the ants communicate with each other via pheromones. Major ant scatters pheromones and moves throughout in the environment. On the other hand, Minor ant makes a mimic altruistic behavior to foraging and transporting. Douzono et al. show the numerical superiority in case of the 2 variants of ants [8]. In this paper, the experimental simulation related to the altruistic behaviors has the 2 kinds of ant agents. A minor ant agent aims to find a food source and then to return to the nest. If there is a ditch in the path among them, the ant will build a bridge. The 3 kind of states are defined according to the behavior of ants [4]. Search State Search state is an initial condition of agents to seek the food source by random walk. Once an agent reaches the food, it moves into Return state. The ant takes a food on the way back to the food until the food source becomes empty. The actions include foraging for foods and transport of them and communications with neighboring ants using pheromone. The pheromone is released by an agent when it Return State In Return state, an agent comes back to the nest carrying the food. After reaching the nest, the agent moves into Search state. Altruism State Some agents stop walking before a ditch and come together as flock. Two or more agents will build a bridge. A. Search State 1) Search a food source First, search a food source within 1 Distance. If the agent finds a food at the destination, moving into Return state. 2) Perception of pheromone Perceiving pheromone within 2 Distances. 3) Search the other agents Checking the other agents within 2 Distances. If the other agents stays, go to 4). Otherwise, go to 6). 4) Search a ditch Search a ditch within 1 Distance. If there is a ditch, transit to Altruism State. Otherwise, go to 6). 5) Move Move to the other position according to the scattered pheromone described in section III-D. 6) Random Selection of Walking Direction Check whether the other agent stays or a ditch exists at the next position except going straight ahead. The next position is selected with an arbitrary probability α. If the position is empty, go back to 1). Otherwise, select another position. Moreover, if the agent is surrounded to other agents or a ditch, it stays at the same position until the neighbor becomes empty. Fig. 3. The search algorithm B. Return State 1) Current Position Check the current position of an agent. If it is in the nest, go to Search State. Otherwise, it goes to next step to move to the nest. 2) Search a ditch If there is a ditch within 1 Distance, go to 3). Otherwise, it moves a next position to the nest and go to 1). 3) Random Walk Check whether the other agent stays or a ditch exists at the randomly selected next position except going straight ahead. If the position is empty, go back to 1). Otherwise, select another position. Moreover, if the agent is surrounded to other agents or a ditch, it stays at the same position until the neighbor becomes empty. Fig. 4. The homing algorithm C. Altruism State 1) Search the other agents If there are n agents within 2 Distances, the agent stays with an arbitrary probability 1 − Pi described in the section IV-B and continues to check its surrounded situation. Otherwise, go to 2) with the probability Pi. 2) Go to Search State Select a position within 1 Distance in the part of chain. If the position is empty, go to the position, and then make the transition to Search State. Otherwise, the ant is embedded in the chain. Fig. 5. The altruism algorithm D. Pheromone Update In many works related ant systems, the ants communicate with each other via the pheromone dissemination. However, the researchers have discussed only about the pheromone on the ground. We consider that the pheromone evaporates and spreads into the space in order to take into consideration the influence distributed in the air. The ant in this study can recognize the volatilization of pheromone in the space, but not know the pheromone on the ground. Based on such an idea, pheromone update process is executed by Eq.(1) and Eq.(2). space′ (x,y)(t) = rA × space(x,y)(t) +rB × (X +rC × (X p space(ip,jp)(t) − 4space(x,y)(t)) space(iq ,jq)(t) − 4space(x,y)(t)) q (xip , yjp ) = {(x, y + 1), (x, y − 1), (x + 1, y), (x − 1, y)} (xiq , yjq ) = {(x + 1, y + 1), (x + 1, y − 1), (x − 1, y + 1), (x − 1, y − 1)}, (1) space(x,y)(t + 1) = space′ (x,y)(t) + re ∗ ground(x,y)(t), (2) ground(x,y)(t + 1) = ground(x,y)(t) − re ∗ ground(x,y)(t), (3) where space(x,y) means the amount of pheromone in the space over the position (x, y) in Eq.(1). rA is a decay rate. rB is the diffusion rate in the direction of up and down, left and right. rc is the diffusion rate in the direction of the diagonal. groundx,y(t) means the pheromone amount on the ground at the position (x, y) in Eq.(2). re is the evaporation rate. E. Multi Feeding Spots Fig. 6 shows the environment with multi feeding spots to extend the simulation system. As shown in Table I, we investigate the behaviors of ant for some ratios of food size. TABLE I. THE SIZE OF FEEDING SPOTS Env. 1-1 1-2 1-3 1-4 Agents Major:3 Minor:100 A:B:C 2:1:1 1:2:1 1:1:2 4:2:1 Env. 2-1 2-2 2-3 2-4 Agents Major:3 Minor:50 A:B:C 2:1:1 1:2:1 1:1:2 4:2:1 r C r r B A Fig. 8. The evaporation rate of pheromone these probabilities as follows. Pe = Ce0 + Pl = Cs0 + Ce1Xi Ce2 + Xi Cs1 Cs2 + X ν i , , (4) (5) where Xi is the number of ants in the chain i. Ce0, Ce1, and Ce2 are parameters for entering the chain. Cs0, Cs1, and Cs2 are parameters for leaving the chain. The function Pe expresses the idea that the probability for an ant to join the chain grows the number of nest mates already presented and reaches a plateau value equal to Ce0 + Ce1. Ce0 is the value of spontaneous hanging when Xi = 0. The function Pl expresses the probability for an ant to leave the chain decreases with Xi. The ant in the chain does not always stay in the same chain. A certain probability for leaving from the chain is required to realize Altruism Status. Due to interaction between ants, the probability decreases with the number of con-specifics in the chain. The phenomena is ruled by empirical equation very similar to that for Oecophylla[3]. The probability for leaving from chain is given by Eq.(6). Pi = a 1 + bX 2 i , (6) where Xi is the number of ants in the chain i. a is the probability of leaving a chain under a disregard for the number of other agents. b is the parameter of depending the amount of pheromone in chain i: b = min{η(log(space(x,y) + 1)) + ǫ, 1}. A theoretical model suggests that these basic mechanisms account for the clustering of insects[3], [6], [7]. Ichimura et al. [9] report the agents in altruism situation perform the shortest path by construction of the bridge. Fig. 9 show the constructed bridge on the way from the feeding spot to the nest. In this paper, we investigate the altruism situation by using the condition in section III-C for entering a chain and Eq.(6) for leaving the chain. V. EXPERIMENTAL RESULTS The behavior of army ants at each environment as shown in Table I was observed. Parameter settings are given as follows by preliminary experiments: n = 2, a = 0.4, and re = 0.05. At each environment in this paper, 10 trials for each set were executed and the behavior of ants were recorded as the motion video. Each trial is continued till all food in the feeding spots eats up. There are 2 kinds of ants at each environment, Major ants to make a random search and Minor ants to follow the scattered pheromone. In this paper, for almost trials we can Fig. 6. The environment for multi feeding spots Fig. 7. The distribution of pheromone IV. PROPOSED METHOD The simulation system mainly focuses two parts, 'Pheromone Evaporation and Its Diffusion' and 'Probability for leaving from chain'. A. Pheromone Evaporation and Its Diffusion As for the former part, Pheromone Evaporation and Its Diffusion, the method in [8] assumed the improper rate of evaporation and diffusion of pheromone in the agent and its behavior. The parameter setting causes a bias for the flock. That is, there is much pheromone in simulation environment partially. The situation increases the agents swarming around them. As a result, it becomes easy to enter Altruism State and two or more bridges are built without the shortest path from a food source to the nest. In order to avoid such a situation, the ratio of pheromone evaporation is defined based on the normal distribution prob- ability as shown in Fig.7. Fig.8 shows the transverse plane of Fig.7. The rate of pheromone in 3D space is set to 'rA':'rB':'rC '=0.788:0.043:0.010 in Fig.8. B. The model of Army Ant The probabilities of an ant entering a chain(Pe) or leaving a chain (Pl) are depending on the size of the chain. The chain is a small part of constructing bridge. Lioni et al.[6], [7] defined Fig. 9. The discover of shortest path observe the following simulation results. During the initial phase as the search of an area for prey, the movement of Major agent will be a key in the change of course to search the subspace, because the Major agent scatters the pheromone while moving in a space. After the discovery of food, the Minor agent catches the food and scatters the pheromone on the way from the spot to the nest. The path becomes congested since there is an obstacle of a ditch. Such situation causes the construction of the bridge on the ditch, since the ants search the shortest path from the feeding spot to the nest. Moreover, the agent swarming around the food increases with the size of food, because more pheromone is scattered while the agent brings a food to the nest. That is, the larger the food in the environment has, the more pheromone the agents scattered. 1 : 2 : 1, Figs. 10 show the number of agents in each area, A, B, and C, with 100 agents in the environment. In case of 100 agents are in the environment, they are divided into some groups and each group can search the area, respectively. Fig. 10(a), Fig. 10(b), and Fig. 10(d) show the transition of agents with A : 4 : 2 : 1}, respectively. B : C = {2 : 1 : 1, As shown in these figures, we can observe some characteristic behavior related to the altruism in the search space. Note that even if the food size of A is the largest, the food of A remains until B or C is disappeared. Because ants find B or C and bring all food from B or C, respectively. On the contrary, Fig. 10(c) (A : B : C = 1 : 1 : 2) shows that the result is beyond our expectations. In the experiment, more agents gather to the area B than the area A at the initial phase, but it took longer time than the area A until they finish carrying the food. Because each simulation was recorded as a motion video, we investigated the detailed behavior of ants. In the almost cases, the constructed bridge was not in the shortest path on the way to the nest, the bridge in the lower area of B was constructed. For this reason, the remaining ant without the participation to the bridge took a roundabout route to avoid the ditch and to search of the shortest path in the area B. Figs. 11 also show the transition of agents under 50 agents in the environment. The situation has 50 agents in the environment, they cannot search the space sufficiently. The ants did not divide into some search group and then the search of area was processed sequentially such as A → B → C. Fig. 11(a) and Fig. 11(d) show the transition of agents in case of 4 : 2 : 1}, respectively. As A : B : C = {2 : 1 : 1, shown in these figures, we can observe some characteristic behavior related to the altruism in the search space. That is, the feeding spots are disappeared in the order of larger spot. (a) A:B:C=2:1:1 (b) A:B:C=1:2:1 (c) A:B:C=1:1:2 (d) A:B:C=4:2:1 Fig. 10. Transition of 100 Agents On the contrary, Fig. 11(b) (A : B : C = 1 : 2 : 1) and Fig. 11(c) (A : B : C = 1 : 1 : 2) show that the result is beyond our expectations. In case of Fig. 11(b), the constructed bridges were not in the shortest path on the way to the nest as shown in Fig. 10(c). In case of Fig. 11(c), it is an interesting case, and almost agents will make bridges place to place in the ditch from area C to the nest. The remaining agents should deliver the food to the nest, however very few agents cannot take all of them. The scattered pheromone was smaller than the evaporated pheromone. Therefore, the pheromone around the bridge disappears and then the bridge was destroyed, because the agents in altruism situation depart the bridge in the condition of less pheromone. The agent leaving from the bridge moves to area A, but there are only a few agents in the area. From such results, only few agents in the environment with large size of food cannot get the altruism situation easily. VI. CONCLUSIVE DISCUSSION We developed the army ant inspired social evolutionary system which can perform the altruism. There are 2 kinds of ant agents communicated with each other via pheromones. Moreover, the pheromones evaporate with the certain ratio and diffused into the space of neighbors stochastically. In order to avoid the over-concentration in the chain, the probability of leaving from a chain is introduced. The system with the facilities can find the optimal place of bridge. The path through the bridge is the shortest from foods to the nest. In this paper, the behaviors of ant under the environment with multi feeding spots and the adequate number of agents were observed. The altruism behavior in the few agents to the size of food spot is hard to keep its situation. Such observations of behaviors in the computer simulation strongly will contribute to the shift to knowledge and power from the individual to the collective. We will develop the autonomous intelligent robots with the altruism behavior and investigate the collective intelligence system in near future. REFERENCES [1] J.L.Deneubourg, S.Goss, N.Franks, J.M.Pasteels, "The blind leading the blind: Modeling chemically mediated army ant raid patterns", Journal of Insect Behavior, Vol.2, No.5, pp.719-725 (1989) [2] R.L.Trivers, "The evolution of reciprocal altruism", Quarterly review of biology, Vol.46, pp.35-67(1971) [3] J.L.Deneubourg, A.Lioni and C.Detrain, "Dynamics of Aggregation and Emergence of Cooperation", The Biological Bulletin, Vol.202, No.3, pp.262-267(2002) [4] H.Ishiwata, N.Noman and H.Iba, "Emergence of Cooperation in a Bio- inspired Multi-agent System", Proc. of AI 2010: Advances in Artificial Intelligence: 23rd Australasian Joint Conference, Lecture Notes in Com- puter Science, Vol. 6464/2011, pp.364-374 (2011) [5] Alex Lancaster, et al., Swarm Project, http://savannah.nongnu.org/projects/swarm Retrieved 2014-01-4. [6] A.Lioni, C.Sauwens, G.Theraulaz, "Chain Formation in cophylla longin- oda", Journal of Insect Behavior, Vol.14, No.5, pp.679-696(2001) [7] A.Lioni, J.L.Deneubourg, "Collective decision through self-assembling", Naturwissenschaften, Vol.91, No.5, pp.237-241(2004) [8] Y.Douzono, A.Hara, and T.Takahama, "The multi-agent simulation of the army ant by introducing division of roles and an altruism", Proc. of 2012 IEEE SMC Hiroshima Chapter Young Researchers' Workshop(to appear in 2012). [9] T.Ichimura and Y.Douzono, "Altruism Simulation based on Pheromone Evaporation and Its Diffusion in Army Ant Inspired Social Evolutionary System", Proc. of SCIS-ISIS2012, pp.1357-1362 (2012) (a) A:B:C=2:1:1 (b) A:B:C=1:2:1 (c) A:B:C=1:1:2 (d) A:B:C=4:2:1 Fig. 11. Transition of 50 Agents
1610.02943	2	1610	2017-06-19T09:29:20	Diffusion LMS for Multitask Problems with Local Linear Equality Constraints	[ "cs.MA" ]	We consider distributed multitask learning problems over a network of agents where each agent is interested in estimating its own parameter vector, also called task, and where the tasks at neighboring agents are related according to a set of linear equality constraints. Each agent possesses its own convex cost function of its parameter vector and a set of linear equality constraints involving its own parameter vector and the parameter vectors of its neighboring agents. We propose an adaptive stochastic algorithm based on the projection gradient method and diffusion strategies in order to allow the network to optimize the individual costs subject to all constraints. Although the derivation is carried out for linear equality constraints, the technique can be applied to other forms of convex constraints. We conduct a detailed mean-square-error analysis of the proposed algorithm and derive closed-form expressions to predict its learning behavior. We provide simulations to illustrate the theoretical findings. Finally, the algorithm is employed for solving two problems in a distributed manner: a minimum-cost flow problem over a network and a space-time varying field reconstruction problem.	cs.MA	cs	Diffusion LMS for Multitask Problems with Local Linear Equality Constraints Roula Nassif, C´edric Richard, Senior Member, IEEE Andr´e Ferrari, Member, IEEE, Ali H. Sayed, Fellow Member, IEEE 1 7 1 0 2 n u J 9 1 ] A M . s c [ 2 v 3 4 9 2 0 . 0 1 6 1 : v i X r a Abstract-We consider distributed multitask learning problems over a network of agents where each agent is interested in estimating its own parameter vector, also called task, and where the tasks at neighboring agents are related according to a set of linear equality constraints. Each agent possesses its own convex cost function of its parameter vector and a set of linear equality constraints involving its own parameter vector and the parameter vectors of its neighboring agents. We propose an adaptive stochastic algorithm based on the projection gradient method and diffusion strategies in order to allow the network to optimize the individual costs subject to all constraints. Although the derivation is carried out for linear equality constraints, the technique can be applied to other forms of convex constraints. We conduct a detailed mean-square-error analysis of the proposed algorithm and derive closed-form expressions to predict its learn- ing behavior. We provide simulations to illustrate the theoretical findings. Finally, the algorithm is employed for solving two problems in a distributed manner: a minimum-cost flow problem over a network and a space-time varying field reconstruction problem. I. INTRODUCTION Single-task distributed optimization over networks allows to minimize the aggregate sum of convex cost functions, each available at an agent, subject to convex constraints that are also distributed across the agents. Each learner seeks to estimate the minimizer through local computations and communications among neighboring agents without the need to know any of the constraints or costs besides their own. Several useful strategies have been proposed to solve constrained and unconstrained versions of this problem in a fully decentralized manner [1]– [13]. Diffusion strategies [3], [8]–[12] are attractive since they are scalable, robust, and enable continuous learning and adaptation in response to drifts in the location of the minimizer due to changes in the costs or in the constraints. Multitask distributed learning over networks is particularly well-suited for applications where several parameter vectors need to be estimated simultaneously from successive noisy measurements using in-network processing [8], [14]–[33]. The network is decomposed into clusters of agents and each cluster estimates its own parameter vector [22]. Distributed strategies for solving multitask problems have been addressed in two The work of C. Richard and A. Ferrari was partly supported by ANR and DGA grant ANR-13-ASTR-0030 (ODISSEE project). The work of A. H. Sayed was supported in part by NSF grants CIF-1524250 and ECCS-1407712. R. Nassif, C. Richard, and A. Ferrari are with the Universit´e de Nice Sophia- Antipolis, France (email: [email protected]; [email protected]; an- [email protected]). A. H. Sayed is with the department of electrical engineering, University of California, Los Angeles, USA (email: [email protected]). main ways. In a first scenario, agents do not know the cluster they belong to and no prior information on possible relationships between tasks is assumed. In this case, all agents cooperate with each other as dictated by the network topology. It is shown in [8] that, in this case, the diffusion iterates will converge to a Pareto optimal solution corresponding to a multi-objective optimization problem. To avoid cooperation with neighbors seeking different objectives, automatic cluster- ing techniques based on diffusion strategies have also been proposed [27]–[29]. In a second scenario, it is assumed that agents know which cluster they belong to. Multitask diffusion strategies are then derived by exploiting prior information about relationships among the tasks. For example, one way to model and exploit relationships among tasks is to formulate convex optimization problems with appropriate co-regularizers between the agents [16], [22]–[25]. While [16] deals with deterministic optimization problems, [22]–[25] are concerned with adaptive estimation problems. In [15], distributed al- gorithms are derived to estimate node-specific signals that lie in a common latent signal subspace in the presence of node-specific linear equality constraints. Several useful works consider stochastic [17]–[20] and deterministic [21] multitask estimation problems with overlapping parameter vectors. They assume that each agent is interested in estimating its own parameter vector, and that the local parameter vectors at neigh- boring agents have some entries that are equal. Unsupervised strategies are also considered in [30], [31] to address multitask overlapping problems. In [26], a diffusion algorithm is derived to solve multitask estimation problem where the parameter space is decomposed into two orthogonal subspaces, with one of the subspaces being common to all agents. In some applications, it happens that the optimum parameter vectors to be estimated at neighboring agents are related according to a set of constraints. This observation motivates us to consider in this work multitask estimation problems subject to linear equality constraints of the form: Dp(cid:96) w(cid:96) + bp = 0, p = 1, . . . , P. (1a) (1b) J glob(w1, . . . , wN ) (cid:44) N(cid:88) (cid:88) k=1 Jk(wk), minimize w1,...,wN subject to (cid:96)∈Ip Each agent k in the network seeks to estimate its own Mk × 1 parameter vector wk, and has knowledge of its cost function Jk(·) and the set of linear equality constraints that agent k is involved in. Each constraint is indexed by p, and defined by the Lp × M(cid:96) matrices Dp(cid:96), the Lp × 1 vector bp, and the set Ip of agent indices involved in this constraint. Note that, by properly selecting the matrices Dp(cid:96) and setting the vectors bp to 0 in (1), the single-task estimation problem [2]–[4] and the multitask overlapping estimation problem [17]–[21] can be recast as problem (1). Assumption. In the current work, it is assumed that each agent k in Ip can collect estimates from all agents in Ip in order to satisfy the p-th constraint, i.e., Ip ⊆ Nk for all k ∈ Ip where Nk denotes the neighborhood of agent k. This assump- tion is reasonable in many applications, for instance, in remote monitoring of physical phenomena involving discretization of spatial differential equations [34], and in network monitoring involving conservation laws at each junction [35]. For illustration purposes, consider a minimum-cost flow problem over the network shown in Fig. 1. This network consists of 10 nodes, 1 destination sink D, and 15 commu- nication links. With each link j, we associate a directed arc and we let fj denote the flow or traffic on this link, with fj > 0 meaning that the flow is in the direction of the arc, and fj < 0 otherwise. At each node k, an external source flow sk enters and flows through the network to the destination sink. The flow must satisfy a conservation equation, which states that at each node k, the sum of flows entering the node, plus the external source sk, is equal to the sum of flows leaving node k. Given the external sources sk and the network topology, a number of studies have been devoted to finding the optimal flows f (cid:63) j that minimize a total flow transmission cost and satisfy the conservation equations [35]–[37]. Problems of this type arise in applications such as electrical networks, telecommunication networks, pipeline networks [35]. In some of these applications, it happens that node k has only access to noisy measurements sk(i) of the external source at each time instant i. For example, in electrical networks, the agents may not be able to collect the exact values of the current sources (or the current demands). Denoting by wk the Mk × 1 vector containing the flows fj entering and leaving node k, we are interested in distributed online learning settings where each node k seeks to estimate wk from noisy measurements sk(i) by relying only on local computations and communi- cations with its neighbors. This problem can be recast in the form (1a)–(1b) and addressed with the multitask strategy proposed in this paper. This example will be considered further in the numerical experiments section. We shall propose a primal adaptive technique (based on propagating and estimating the primal variable) for solving problem (1) in a distributed manner. The technique relies on combining diffusion adaptation with a stochastic gradient projection step, and on the use of constant step-sizes to enable continuous adaptation and learning from streaming data. Since we are learning from streaming data, the dual function cannot be computed exactly and the use of primal-dual methods may result in stability problems as already shown in [12]. For this reason, we focus on primal techniques. Our current work is able to cope with the following two scenarios: 1) multi- task problems with prior information on linear relationships between tasks, and 2) constrained multitask problems with distributed information access. We analyze the behavior of our algorithm in the mean and mean-square-error sense (w.r.t. the 2 Fig. 1: Flow network topology with 10 nodes, 1 destination sink D, and 15 communication links. minimizers of the local costs and w.r.t. the solution of the constrained multitask problem) and we derive expressions to predict its transient and steady-state behavior. Some simulation results show that, for small constant step-sizes, the expected distance between the estimates at each agent and the optimal value can be made arbitrarily small. Notation. Normal font letters denote scalars, boldface low- ercase letters denote column vectors, and boldface uppercase letters denote matrices. We use the symbol (·)(cid:62) to denote matrix transpose, the symbol (·)−1 to denote matrix inverse, the symbol (·)† to denote the pseudo-inverse of a full row- rank matrix, and the symbol tr(·) to denote the 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 symbols ⊗ and ⊗b denote the Kronecker product and the block Kronecker product, respectively. The symbol vec(·) refers to the standard vectorization operator that stacks the columns of a matrix on top of each other and the symbol bvec(·) refers to the block vectorization operation that vectorizes each block of a matrix and stacks the vectors on top of each other. The identity matrix of size N × N is denoted by I N . The N × 1 vector of ones is denoted by 1N×1. For a P × N block matrix A, the 1× N k-th block row is denoted by [A]k,• and the P × 1 k-th block column is denoted by [A]•,k. The notation PΩ(w) denotes the projection of w onto the manifold Ω. II. PROBLEM FORMULATION AND CENTRALIZED A. Problem formulation and assumptions SOLUTION Consider a network of N agents, labeled k = 1, . . . , N. At each time instant i, each agent k has access to a zero-mean real-valued observation dk(i), and a zero-mean real-valued Mk × 1 regression vector xk(i), with positive covariance matrix Rx,k = E{xk(i)x(cid:62) k (i)} > 0. We assume the data to be related via the linear data model: dk(i) = x(cid:62) (2) k is an Mk×1 unknown parameter vector, and zk(i) is z,k, independent where wo a zero-mean measurement noise of variance σ2 k + zk(i), k (i)wo 1122334455667788991010DDs1s1s2s2s3s3s4s4s5s5s6s6s7s7s8s8s9s9s10s10f1f1f2f2f3f3f4f4f5f5f6f6f7f7f8f8f9f9f10f10f12f12f11f11f13f13f14f14f15f15d,k of x(cid:96)(j) for all (cid:96) and j, and independent of z(cid:96)(j) for (cid:96) (cid:54)= k or i (cid:54)= j. We let rdx,k (cid:44) E{dk(i)xk(i)} and σ2 (cid:44) Edk(i)2. Let wk denote some generic Mk×1 vector that is associated with agent k. The objective of agent k is to find an estimate for k, and we associate with this agent the mean-square-error wo criterion: Jk(wk) = Edk(i) − x(cid:62) which is strongly convex, second-order differentiable, and minimized at wo k. In addition, P linear equality constraints of the form (1b) are imposed on the parameter vectors {wk} at each time instant i. Let us collect the parameter vectors {wk} and {wo k} from across the network into N × 1 block column vectors w and wo, respectively: w (cid:44) col{w1, . . . , wN}, and let us write the P linear equality constraints in (1b) more compactly as: wo (cid:44) col{wo k (i)wk2, 1, . . . , wo N}, (3) (4) Dw + b = 0, where D is a P ×N block matrix, with each block Dp(cid:96) having dimensions Lp × M(cid:96), and b is a P × 1 block column vector where each block bp has dimensions Lp × 1. Combining (5) and (3), the network optimization problem becomes: N(cid:88) Edk(i) − x(cid:62) k (i)wk2, w minimize subject to Dw + b = 0, k=1 (5) (6) where each agent k is in charge of estimating the k-th sub- vector wk of w. Since the mean-square-error criterion in (6) is separable, we shall assume without loss of generality that each parameter vector wk is involved in at least one constraint so that cooperation is justified. We shall also assume that D is full row-rank to ensure that equation Dw + b = 0 has at least one solution. We also introduce an assumption on the availability of the constraints. Let Ip be the set of agent indices involved in the p-th constraint. We shall assume that every agent k in Ip is aware of the p-th constraint, and that the network topology permits this agent to collect estimates from all agents in Ip, that is, Ip ⊆ Nk, so it can apply this constraint to its own estimate. This assumption is reasonable in many applications, for instance, in remote monitoring of physical phenomena [34], and in network distribution system monitoring (as described in the introduction) [35]. These examples will be considered in numerical experiments section. Before proceeding, note that problem (6) can be recast as a quadratic program (QP) [36], and any algorithm that solves QPs can solve it. We are interested instead in distributed adaptive solutions that can operate in real-time on streaming data. As we will see later, the traditional constrained LMS algorithm [38] can solve (6) in a centralized manner. In this centralized solution, each agent at each iteration sends its data to a fusion center, which in turn processes the data and sends the estimates back to the agents. The entire matrix D and the entire vector b then need to be available at the fusion center. While centralized solutions can be powerful, decentralized solutions are more attractive since they are more robust and respect the privacy policy of each agent [9], [39], [40]. 3 B. Centralized solution Let us first describe the centralized solution. We assume that the agents transmit the collected data {dk(i), xk(i)} to a fusion center for processing. Problem (6) can be written equivalently as: w(cid:62)Rxw − 2r(cid:62) minimize subject to Dw + b = 0, w dxw + r(cid:62) d 1N×1, (7) where the N × N block diagonal matrix Rx, the N × 1 block column vector rdx, and the N × 1 column vector rd are given by: d,1, . . . , σ2 Rx (cid:44) diag{Rx,1, . . . , Rx,N}, rdx (cid:44) col{rdx,1, . . . , rdx,N}, rd (cid:44) col{σ2 d,N}. (8) (9) (10) Since Rx is positive definite, problem (7) is a positive definite quadratic program with equality constraints. It has a unique global minimum given by: x D(cid:62) −1(Dwo + b). ) w(cid:63) = wo − R−1 (DR−1 Let Ω denote the linear manifold: x D(cid:62) (11) Ω (cid:44) {w : Dw + b = 0}. (12) If wo ∈ Ω, the optimum w(cid:63) coincides with wo. In this case, the constrained optimization problem (6) can be thought as k given prior estimating the unknown parameter vectors wo information about relationships between tasks of the form (1b). Exploiting such prior information may improve the estimation as we will see in the experiments. Let M denote the dimension k=1 Mk. The of the network parameter vector w, i.e., M =(cid:80)N projection of any vector y ∈ IRM onto Ω is given by: PΩ(y) = Py − f , where Let w(i) denotes the estimate of w(cid:63) at iteration i. In order to solve (7) iteratively, the gradient projection method [41] can be applied on top of a gradient-descent iteration: f (cid:44) D†b. P (cid:44) I M − D†D, (cid:0)w(i) + µ[rdx − Rxw(i)](cid:1), w(i + 1) = PΩ i ≥ 0. (15) In order to run recursion (15), we need to have access to the second-order moments {Rx,k, rdx,k}. Since these mo- ments are rarely available beforehand, the agents use their instantaneous data {dk(i), xk(i)} to approximate the second- order moments, namely, Rx,k ≈ xk(i)x(cid:62) k (i) and rdx,k ≈ dk(i)xk(i). Doing so and replacing PΩ(·) by (13), we obtain the following stochastic-gradient algorithm in lieu of (15): w(i+1) = P·col(cid:8)wk(i)+µxk(i)[dk(i)−x(cid:62) k=1−f . (16) Collecting the regression vectors into the M × N matrix X(i) (cid:44) diag{xk(i)}N k=1 and the observations into the N × 1 vector d(i) (cid:44) col{dk(i)}N k=1, algorithm (16) becomes the k (i)wk(i)](cid:9)N Constrained Least-Mean-Squares (CLMS) algorithm: w(i + 1) = P(cid:0)w(i) + µX(i)[d(i)− X(cid:62) (i)w(i)](cid:1) − f . (17) (13) (14) This procedure was originally proposed in [38] as an online linearly constrained minimum variance (LCMV) filter for solv- ing mean-square-error estimation problems subject to linear constraints; the motivation there was not concerned with multi- task problems. In this section, we showed that the centralized multitask constrained problem reduces to a similar problem, for which algorithm (17) can be applied. The performance of such stand-alone centralized solutions was studied in [38], [42], [43]. III. PROBLEM REFORMULATION AND DISTRIBUTED SOLUTION A. Problem reformulation We move on to develop a distributed solution with a con- tinuous adaptation mechanism. First, note that several works for solving problems of the form (6) with possible distributed information access already exist in the literature [4], [6], [7], [10], [13], [21], [33], [44], [45]. However, except for [21], [33], these other works solve single-task estimation problems where the entire network is employed to estimate the minimizer of (6). Furthermore, compared to [6], [21], [33], [44], [45], we shall assume stochastic errors in the evaluation of the gradients of local cost functions. To proceed with the analysis, one of the challenges we now face is that any given agent k may be involved in several constraints. Our strategy is to transform (6) into an equivalent optimization problem exhibiting structure amenable to distributed optimization with separable constraints. Let jk denote the number of constraints that agent k is involved in. We expand each node k into a cluster Ck of jk virtual sub- nodes, namely, Ck (cid:44) {km}jk m=1. Each one of these sub-nodes is involved in a single constraint. Let wkm denote the Mk × 1 auxiliary vector associated with sub-node km. In order to ensure that agent k satisfies simultaneously all the constraints at convergence, we will allow all sub-nodes at agent k to run diffusion learning to reach agreement on their estimates {wkm} asymptotically. We denote by Ie,p the set of sub-nodes which are involved in the p-th constraint. In order to clarify the presentation, an illustrative example is provided in Fig. 2. On the left of this panel is the original network topology with N = 6 agents and P = 3 constraints. On the right is the network topology model with clusters of sub-nodes shown in grey color. Observe that I2 = {1, k}, I3 = {3, k} and I4 = {4, k, (cid:96)}, which means that agent k is involved in constraints 2, 3, and 4. Agent k is thus expanded into a cluster Ck = {k1, k2, k3} of 3 sub-nodes. Sub-nodes k1, k2, and k3 are assigned to constraints 2, 3, and 4, respectively. Each other agent, say (cid:96), involved in a single constraint is renamed (cid:96)1 and assigned to a single-node cluster C(cid:96) = {(cid:96)1} for consistency of notation. This leads to the sets Ie,2 = {12, k1}, Ie,3 = {31, k2} and Ie,4 = {41, (cid:96)1, k3} where all sub-nodes are involved in a single constraint. All sub-nodes km in cluster Ck can share data since they refer to the same agent k. In the sequel, we shall propose a general algorithm for strongly-connected clusters (see (35) below) and show how the designer can simplify the algorithm by choosing fully-connected clusters (see (40) below). 4 Fig. 2: (Left) Network topology with constraints identified by the subsets of nodes I1, I2, I3, and I4. (Right) Network topology model with fully-connected clusters shown in grey color and constraints now identified by the subsets of sub- nodes Ie,1, Ie,2, Ie,3, and Ie,4. All sub-nodes in this model are involved in a single constraint. Diffusion learning is run in clusters with more than one sub-node to reach agreement on local estimates while satisfying their respective constraints. Accordingly, we can now reformulate problem (6). We start by collecting the vectors wkm into the Ne × 1 network block column vector: (cid:110) col(cid:8)wkm (cid:9)jk (cid:111)N , (18) m=1 k=1 we (cid:44) col where Ne (cid:44)(cid:80)N k=1 jk. Throughout this work, a subscript "e" below a symbol indicates an extended version associated with sub-nodes (auxiliary variables). For example, while the symbol N represents the number of nodes, the symbol Ne represents the number of sub-nodes. Likewise, the vector we in (18) corresponds to the extended version of the vector w in (4). We introduce for each agent k a set of jk coefficients {ckm} that satisfy two conditions: ckm > 0, for m = 1, . . . , jk, and ckm = 1. (19) jk(cid:88) m=1 N(cid:88) jk(cid:88) The coefficients {ckm} are free parameters that are chosen by for all m. The global the user. A natural choice is ckm = 1 jk cost in (1a) can be written as: J glob(w1, . . . , wN ) (cid:44) N(cid:88) Jk(wk) = ckm Jk(wk). k=1 k=1 m=1 (20) We reformulate problem (1) in the following equivalent form by introducing the auxiliary variables {wkm}: minimize we subject to ckmJk(wkm) (21a) Dp(cid:96)n w(cid:96)n + bp = 0, p = 1, . . . , P, jk(cid:88) N(cid:88) (cid:88) k=1 m=1 (cid:96)n∈Ie,p wk1 = . . . = wkjk , k = 1, . . . , N . In the following, we shall address the equality constraints (21c) with a diffusion algorithm within each cluster of sub-nodes with the objective of reaching an agreement within each (21b) (21c) 213k4`I1I4I2I2I3I3213141`11112k3k2Ie,4Ie,1Ie,3k1k1Ie,2Ie,2cluster (all sub-nodes converge to the same estimate). Since the diffusion strategy in a single-task network allows the agents to converge to the same limit point asymptotically for sufficiently small constant step-sizes when the network is strongly connected [9], we allow the sub-nodes in cluster Ck to be connected such that the resultant cluster Ck is strongly connected. This does not lead to a change in the network topology since each sub-node in a cluster refers to the same agent. We refer to the virtual set of neighboring sub-nodes of km in Ck by Nkm∩ Ck. N(cid:88) The cost function in (21a) can be written as: jk(cid:88) ckm Jk(wkm ) = w(cid:62) e Rx,ewe−2r(cid:62) dx,ewe+r(cid:62) d,e k=1 m=1 where the Ne × Ne block diagonal matrix Rx,e, the Ne × 1 block column vector rdx,e, and the Ne× 1 column vector rd,e are given by: 1Ne×1, (22) Rx,e (cid:44) diag(cid:8)Ck ⊗ Rx,k (cid:9)N (cid:9)N rdx,e (cid:44) col(cid:8)ck ⊗ rdx,k rd,e (cid:44) col(cid:8)σ2 (cid:9)N d,kck k=1, k=1, k=1, (23) (24) (25) (cid:48) = 0 (cid:21) with m=1. ewe + b compactly as: The equality constraints in (21b)–(21c) can be written more with Ck (cid:44) diag{ckm}jk m=1 and ck (cid:44) col{ckm}jk D(cid:48) (cid:21) (cid:20) b where De is a P × Ne block matrix constructed according P × N block matrix D, and H is a(cid:80)N to (21b) which can be viewed as an expanded form of the k=1(jk − 1)× Ne block Using similar arguments as in Section II-B, we find that the matrix constructed according to (21c). (cid:20) DeH D(cid:48) (27) (26) e = = 0 b (cid:48) , , (cid:48) e + b ewo ), e is given by: (28) e = wo eR−1 x,eD(cid:48)(cid:62) x,eD(cid:48)(cid:62) e ) e − R−1 solution of (21) is given by: e (D(cid:48) −1(D(cid:48) w(cid:63) (cid:9)N where the Ne × 1 block column vector wo k=1. (cid:9)N (cid:44) col(cid:8)1jk×1 ⊗ wo e = col(cid:8)1jk×1 ⊗ w(cid:63) e can be written alternatively as: Let w(cid:63) vector w(cid:63) k=1. w(cid:63) k wo e (29) k denote the k-th block of w(cid:63) in (11). The optimum k (30) B. Distributed solution To solve problem (21) with distributed information access, we propose an iterative algorithm based on diffusion strategies and gradient-projection principle. First, we present the algo- rithm when the second order moments of the observations are assumed to be known by each sub-node. Although cluster Ck and agent k refer to the same entity, we shall use the notion of cluster and sub-nodes in order to simplify the presentation. Let we,p denote the ip × 1 block column vector given by we,p = col{w(cid:96)n}(cid:96)n∈Ie,p where ip is the number of 5 nodes involved in the p-th constraint. Also, note that ip is the cardinality of Ip and Ie,p. Let Ωp denote the linear manifold corresponding to the p-th constraint in (21b), namely, Ωp (cid:44) {Dpwe,p + bp = 0} where Dp is a 1× ip block matrix. k at sub-node km and time Let wkm(i) be the estimate of w(cid:63) instant i. We assume that km ∈ Ie,p. Following the same line of reasoning as [11] in the single-task case, and extending the argument to our multitask problem, we arrive at the following diffusion algorithm consisting of three steps: ψkm(i + 1) = wkm (i) + µ ckm [rdx,k − Rx,kwkm(i)] (31a) (31b) φkm (i + 1) = akn,km ψkn (i + 1) (cid:88) (cid:0)col(cid:8)φ(cid:96)n (i + 1)(cid:9) kn∈Nkm∩ Ck PΩp (cid:104) (cid:1)(cid:105) wkm (i + 1) = (cid:96)n∈Ie,p km (31c) where µ > 0 is a constant step-size parameter, [x]km is the block of x corresponding to sub-node km, and wkm (0) = wk(0) for all m. In the first step (31a), also called adaptation step, each sub-node km in the network adapts its estimate wkm(i) via gradient descent on ckm Jk(·). This step results in the intermediate estimate ψkm (i + 1). In the combination step (31b), each sub-node km combines its estimate ψkm(i + 1) with the estimates ψkn (i + 1) of its intra-cluster neighbors Nkm∩ Ck. This step results in the intermediate estimate φkm(i+1). The nonnegative coefficients {akn,km} are chosen to satisfy the following conditions: akn,km ≥ 0, akn,km = 1, km∈Nkn∩ Ck kn∈Nkm∩ Ck and akn,km = 0 if kn /∈ Nkm∩ Ck. akn,km = 1, (32) (cid:88) (cid:88) Collecting these coefficients into a jk × jk matrix Ak for each (cid:80) cluster Ck, it follows that Ak is doubly stochastic. Let Mp denote the dimension of the vector we,p, i.e., Mp = (cid:96)n∈Ie,p M(cid:96). Before describing the third step, we recall that the projection of any point y onto Ωp has the form: (cid:0)y(cid:1) = P p y − f p (33) PΩp where km f p and (34) pbp. pDp (cid:44) D† P p (cid:44) I Mp − D† To evaluate the block(cid:2)PΩp (y)(cid:3) , even if sub-node km is only in charge of estimating wkm, it needs the entire vector y, the Mk × Mp matrix [P p]km,•, and the Mk × 1 vector [f p]km. In the projection step (31c), each sub-node km ∈ Ie,p collects the intermediate estimates φ(cid:96)n (i+1) from all sub-nodes (cid:96)n ∈ Ie,p and combines them according to (31c). This step results in the estimate wkm(i + 1) of w(cid:63) k at sub-node km and iteration i + 1. The adaptation step (31a) requires knowledge of the second- order moments of data. Proceeding as in the centralized case, and replacing the moments by instantaneous approximations, we obtain algorithm (35) for solving (21) in a distributed way: (35a) (35b) (35c) (36) (37) (38) (39) (40b) (40c) 6 . k=1 m=1 (41) (cid:111)N Subtracting wo φkm (i). We further introduce the Ne × 1 block network error (cid:101)we(i) (cid:44) col vector: Let Me denote the length of the network error vector (cid:101)we(i), that is, Me (cid:44) (cid:80)N (cid:110) col(cid:8)(cid:101)wkm(i)(cid:9)jk k=1 jkMk. Using the linear model (2), the estimation error in the adaptation step (121a) can be written as: k (i)(cid:101)wkm(i) + zk(i). dk(i) − x(cid:62) k (i)wkm(i) = x(cid:62) (42) using (42), and collecting the error vectors (cid:101)ψkm (i) into the (cid:111)N (cid:110) k from both sides of the adaptation step (121a), col(cid:8)(cid:101)ψkm(i)(cid:9)jk Ne × 1 block vector (cid:101)ψe(i) (cid:44) col obtain:(cid:101)ψe(i + 1) =(cid:2)I Me − µRx,e(i)(cid:3)(cid:101)we(i) − µ pzx,e(i), (cid:110) (cid:111)N (cid:110) (cid:111)N Rx,e(i) (cid:44) diag Ck ⊗ xk(i)x(cid:62) pzx,e(i) (cid:44) col ck ⊗ xk(i)zk(i) where k (i) , we (43) (44) (45) m=1 . k=1 , k=1 k=1 Projecting ψe(i + 1) onto the sets Ωp in (33), we obtain from (121c): where φe(i + 1) = P eψe(i + 1) − f e, (cid:110) (cid:111)N col(cid:8)ψkm(i)(cid:9)jk (cid:111)N (cid:110) col(cid:8)φkm (i)(cid:9)jk ψe(i) (cid:44) col φe(i) (cid:44) col m=1 m=1 k=1 k=1 (46) (47) (48) , , k=1 m=1 f e −1De (49) (50) (cid:44) D† Subtracting wo P e is an Me × Me orthogonal projection matrix, and f e is an Me × 1 vector given by (see Appendix A): e (DeD(cid:62) e ) P e (cid:44) I Me − D† eDe = I Me − D(cid:62) eb = D(cid:62) e (DeD(cid:62) −1b. e ) e in (29) from both sides of recursion (46), we (cid:111)N (cid:110) col(cid:8)(cid:101)φkm (i + 1)(cid:9)jk obtain:(cid:101)φe(i + 1) (cid:44) col = P e(cid:101)ψe(i + 1) +(cid:0)I Me − P e (cid:1)wo (51) Subtracting wo k from both sides of the combination step (121b) and using (32), we obtain that the network error vector for the diffusion strategy (35) evolves according to the following recursion: (cid:2)I Me − µRx,e(i)(cid:3)(cid:101)we(i)− (I Me − P e)wo (52) k=1. Before proceeding, let us µA(cid:62)P epzx,e(i) + A(cid:62) where A (cid:44) diag{Ak ⊗ I Mk}N introduce the following assumption on the regression data. Assumption 1. (Independent regressors) The regression vec- tors xk(i) arise from a zero-mean random process that is temporally white and spatially independent. (cid:101)we(i + 1) = A(cid:62)P e e + A(cid:62)f e, e + f e. (cid:88) ψkm (i + 1) = wkm(i) + µ ckmxk(i)[dk(i) − x(cid:62) φkm(i + 1) = [P p]km,• · col(cid:8)ψ(cid:96)n(i + 1)(cid:9) (cid:96)n∈Ie,p− [f p]km , k (i)wkm(i)], wkm(i + 1) = akn,km φkn (i + 1). kn∈Nkm∩ Ck Compared to (31), observe in (35) that each sub-node km projects its intermediate estimate before combining it. We recommend this permutation since it allows, with the ap- propriate parameter settings described below, to reduce the algorithm complexity without compromising its convergence, as confirmed in the sequel. Consider any agent k. By setting factors ckm to 1 for all m = 1, . . . , jk, and combining the jk intermediate estimate φkm (i + 1) at each sub-node km with the estimates of all other sub-nodes available at node k using uniform combination coefficients, i.e., Nkm ∩ Ck = Ck and for n = 1, . . . , jk, (121a) and (121b) reduce to: akn,km = 1 jk ψkm(i + 1) = ψk(i + 1), wkm(i + 1) = wk(i + 1), for m = 1, . . . , jk, for m = 1, . . . , jk, where ψk(i + 1) and wk(i + 1) are given by: xk(i)(cid:2)dk(i) − x(cid:62) k (i)wk(i)(cid:3), µ jk ψk(i + 1) = wk(i) + jk(cid:88) m=1 1 jk In this case, at each agent k, the algorithm (35) becomes: wk(i + 1) = φkm(i + 1). xk(i)(cid:2)dk(i) − x(cid:62) φkm(i + 1) = [P p]km,• · col(cid:8)ψ(cid:96)(i + 1)(cid:9) ψk(i + 1) = wk(i) + µ jk k (i)wk(i)(cid:3) (40a) (cid:96)∈Ip − [f p]km , jk(cid:88) m=1 km ∈ Ie,p, m = 1, . . . , jk, 1 jk φkm (i + 1). wk(i + 1) = Instead of maintaining and updating jk coefficient vectors ψkm (i+1), agent k maintains and updates only one parameter vector ψk(i + 1). Then, it transmits the vector ψk(i + 1) to its neighbors, receives {ψ(cid:96)(i + 1)} from its neighborhood, and generates jk parameter vectors φkm(i + 1) by projecting onto its constraints. Finally, it combines these vectors to obtain wk(i+1), i.e., the estimate of w(cid:63) k at iteration i+1. Therefore, with this setting, the computational and communication com- plexity of our distributed algorithm is significantly reduced. IV. PERFORMANCE ANALYSIS RELATIVE TO wo e A. Network error vector recursion with respect to the optimal parameter vector wo the error vector (cid:101)wkm(i) (cid:44) wo error vectors (cid:101)ψkm(i) (cid:44) wo We shall first study the stochastic behavior of algorithm (35) e. We introduce k − wkm(i) and the intermediate k − k − ψkm (i) and (cid:101)φkm (i) (cid:44) wo Under this assumption, xk(i) is independent of w(cid:96)m (j) for i ≥ j and for all (cid:96)m. This assumption is commonly used in the adaptive filtering literature since it helps simplify the analysis, and the performance results obtained under this assumption match well the actual performance of stand-alone filters for sufficiently small step-sizes [43]. B. Mean behavior analysis Recursion (52) can be rewritten in a more compact form: where we introduced the following notations: (cid:101)we(i + 1) = B(i)(cid:101)we(i) − µg(i) + r, (cid:2)I Me − µRx,e(i)(cid:3), (I Me − P e)wo B(i) (cid:44) A(cid:62)P e g(i) (cid:44) A(cid:62)P epzx,e(i), e + A(cid:62)f e. r (cid:44) A(cid:62) (53) (54) (55) (56) (57) C. Mean-square-error behavior analysis 7 Σ (cid:62) (cid:44) (cid:101)w To perform the mean-square-error analysis, we shall use the block Kronecker product operator [46] and the block vectorization operator bvec(·). As explained in [9], these block operators preserve the locality of the blocks in the original matrix arguments. To analyze the convergence in mean-square- vector (cid:101)we(i), weighted by any positive-definite matrix Σ, error sense, we consider the variance of the weight error that is, E(cid:107)(cid:101)we(i)(cid:107)2 e (i)Σ(cid:101)we(i). The E{(cid:107)(cid:101)we(i + 1)(cid:107)2 freedom in selecting Σ allows us to extract various types of information about the network and the sub-nodes. From (53) and Assumption 1, we obtain: Σ, where (cid:107)(cid:101)we(i)(cid:107)2 Σ} = E{(cid:107)(cid:101)we(i)(cid:107)2 Σ + 2r(cid:62)ΣBE(cid:101)we(i), Σ(cid:48)} + µ2E{(cid:107)g(i)(cid:107)2 (cid:107)r(cid:107)2 where matrix Σ(cid:48) is given by: Σ(cid:48) (cid:44) E{B(cid:62) (64) e × 1 vector representation of Σ that is obtained by the block vectorization operator, namely, σ (cid:44) bvec(Σ). In the sequel, it will be more convenient to work with σ than with Σ itself. We will use the notations (cid:107)x(cid:107)2 Σ and σ to denote the same quantity x(cid:62)Σx. Let σ(cid:48) = bvec(Σ(cid:48) ). (cid:107)x(cid:107)2 Using the property bvec(U ΣW ) = (W (cid:62) ⊗b U )σ, the vector σ(cid:48) can be related to σ: (i)ΣB(i)}. Let σ denotes the M 2 Σ}+ (63) σ(cid:48) = F σ, (65) where F is an M 2 e matrix given by: e × M 2 F (cid:44) E{B(cid:62) (i) ⊗b B(cid:62) (i)}. (66) The evaluation of the matrix F requires knowledge of the fourth-order moments of the regression vectors. In practice, when E{Rx,e(i) ⊗b Rx,e(i)} can be computed, as in the case of zero-mean Gaussian regressors (see Appendix B), the matrix F can be calculated in closed form and its stability can be checked for a given µ.2 The second term on the RHS of relation (63) can be written as: (i)Σg(i)} µ2E{(cid:107)g(i)(cid:107)2 )(cid:3)(cid:62) σ, (67) = µ2tr(ΣG) = µ2(cid:2)bvec(G(cid:62) Σ} = µ2E{g(cid:62) (i)} = A(cid:62)P ediag(cid:8)ckc(cid:62) where G is the Me × Me matrix given by: G (cid:44) E{g(i)g(cid:62) k ⊗σ2 Σ =(cid:2)bvec(rr(cid:62) )(cid:3)(cid:62) P eA. (68) Similarly, the third term on the RHS of relation (63) can be written as: (cid:9)N z,kRx,k (69) σ, k=1 (cid:107)r(cid:107)2 2When E{Rx,e(i)⊗bRx,e(i)} cannot be evaluated, a common alternative is to use the approximation F ≈ B(cid:62)⊗b B(cid:62) for sufficiently small step-sizes (see [9], [11]). In this case, we have ρ(F ) ≈ ρ(B(cid:62) ⊗b B(cid:62)) = ρ(B)2. As long as this approximation is reasonable, the stability of F is ensured if ρ(B) < 1, i.e., if the step-size is chosen according to condition (62). where Taking the expectation of both sides of recursion (53), using Assumption 1, and E g(i) = 0, we find that the mean error vector evolves according to the recursion: E(cid:101)we(i + 1) = B E(cid:101)we(i) + r, (cid:0)I Me − µRx,e B (cid:44) EB(i) = A(cid:62)P e (cid:1), (58) with Rx,e = ERx,e(i) given in (23)1. Recursion (57) con- verges as i → ∞ if the matrix B is stable. If we let i → ∞ on both sides of (57), we find that the asymptotic mean bias is given by: E(cid:101)we(∞) = lim i→∞ E(cid:101)we(i) = (I Me − B) −1r. (59) It is known that any induced matrix norm is lower bounded by the spectral radius of the matrix. We can thus write in terms of the 2-induced matrix norm: ρ(B) ≤ (cid:107)A(cid:62) (cid:107)2 · (cid:107)P e(cid:107)2 · (cid:107)I Me − µRx,e(cid:107)2, (60) where we used the sub-multiplicative property of the 2-induced norm. Since P e is an orthogonal projection matrix and A(cid:62) is a doubly-stochastic matrix, their 2-induced norms are equal to one. Since the matrix I Me − µRx,e is a symmetric block diagonal matrix, its 2-induced norm agrees with its spectral radius: (cid:107)I Me − µRx,e(cid:107)2 = ρ(I Me − µRx,e) max 1≤m≤jk = max 1≤k≤N ρ(I Mk − µ ckmRx,k). (61) Thus, the stability of B is ensured by choosing µ such that: 0 < µ < 2 ck,max · λmax(Rx,k) , ∀k = 1, . . . , N. (62) where ck,max (cid:44) max e = 1≤m≤jk e, i.e., perfect model scenario where wo satisfies the linear wo equality constraints, the bias reduces to 0. ckm. We observe that when w(cid:63) 1If U (i) is a random matrix, its expected value E U (i) is denoted by U. (cid:62) σ. by: (71) (70) + rr(cid:62) )(cid:3)(cid:62) Y(i) (cid:44) µ2G(cid:62) and the fourth term can be written as: (cid:62)B(cid:62) . Then, the variance relation (63) can be expressed as: σ. (72) Provided that F is stable, recursion (72) is stable. Since G, r, B, σ, and µ are constant and finite terms, the boundedness 2r(cid:62)ΣBE(cid:101)we(i) = 2tr(r(cid:62)ΣBE(cid:101)we(i)) = 2(cid:2)bvec(rE{(cid:101)w e (i)}B(cid:62) Let us define the Me × Me time dependent matrix Y(i) given + 2rE{(cid:101)we(i)} σ} = E{(cid:107)(cid:101)we(i)(cid:107)2F σ} +(cid:2)bvec(Y(i))(cid:3)(cid:62) E{(cid:107)(cid:101)we(i + 1)(cid:107)2 of (cid:2)bvec(Y(i))(cid:3)(cid:62) σ depends on E(cid:101)we(i) being bounded. We know from (57) that E(cid:101)we(i) is bounded if the step-size µ a bounded driving term r. It follows that(cid:2)bvec(Y(i))(cid:3)(cid:62) is chosen according to condition (62) because (57) is a Bounded-Input Bounded-Output (BIBO) stable recursion with σ is error stable, i.e., E{(cid:107)(cid:101)we(i + 1)(cid:107)2 uniformly bounded. As a result, the algorithm is mean-square- σ} converges to a bounded value as i → ∞, if µ is chosen such that F in (66) is stable in addition to condition (62) that ensures mean stability. As explained above, step-sizes that ensure stability in the mean and that are sufficiently small will also ensure stability in the mean-square. e ), (74) (73) σ + Γ(i)σ, defined as: e )F iσ+ with Γ(0) = 0. (F − I M 2 E{(cid:107)(cid:101)we(i+1)(cid:107)2 vector that can be evaluated from Γ(i) according to: σ} = E{(cid:107)(cid:101)we(i)(cid:107)2 e (0)})(cid:3)(cid:62) Following similar arguments as in [22], [23], [26] and doing the required adjustments, we find that the weighted variance σ} evolves according to the following recursion: E{(cid:107)(cid:101)we(i + 1)(cid:107)2 (cid:2)bvec(E{(cid:101)we(0)(cid:101)w σ}+ (cid:2)bvec(Y(i))(cid:3)(cid:62) (F − I M 2 where (cid:101)we(0) is the initial condition and Γ(i + 1) is a 1× M 2 Γ(i + 1) = Γ(i)F +(cid:2)bvec(Y(i))(cid:3)(cid:62) E(cid:107)(cid:101)we(i)(cid:107)2 (75) If the matrix F is stable, from the recursive expression (72), we obtain as i → ∞: The steady-state network performance with metric σss is −F )σ} = [bvec(Y(∞))] + 2rE{(cid:101)we(∞)} (cid:62)B(cid:62) e − F )−1σss. (cid:33) To obtain (75), we replace σ in (76) by (I M 2 The theoretical findings (57), (59), (73), and (76) allow us to predict the behavior in the mean and in the mean-square- error sense of the stochastic algorithm (35) w.r.t. the parameter vector wo e given by: e. Note that, the network MSD w.r.t. wo E{(cid:107)(cid:101)we(i)(cid:107)2 Y(∞) = µ2G(cid:62) where, from (71) and (59), we have: ζ (cid:63) = lim i→∞ + rr(cid:62) (cid:32) lim i→∞ (cid:62)σ, (77) (76) (IM 2 e σss . (cid:62) . e N(cid:88) k=1 jk(cid:88) m=1 1 jk E(cid:107)(cid:101)wkm (i)(cid:107)2 (cid:110) 1 (78) , (cid:111)N k=1 . MSDnet(i) (cid:44) 1 N can be obtained by setting Σ = 1 N diag I jk·Mk jk V. PERFORMANCE ANALYSIS RELATIVE TO w(cid:63) e 8 We shall now study the convergence behavior of algo- rithm (35) toward the solution w(cid:63) e of the optimization problem with constraints (21). To this end, we introduce for each sub- node km the weight error vector: and the Ne × 1 network block error vector: (cid:48) km(i) (cid:44) w(cid:63) (cid:110) col(cid:8)(cid:101)w (cid:48) e(i) (cid:44) col (cid:101)w k − wkm (i), km (i)(cid:9)jk (cid:48) m=1 (cid:111)N k=1 (79) (80) (cid:101)w We note that the behavior of algorithm (35) with respect to e using w(cid:63) the following relation: e can be deduced from its behavior with respect to wo (cid:101)w (cid:48) e(i + 1) = (cid:101)we(i + 1) − wδ (81) e. Using (81) with (52), the fact that w(cid:63) e. e where wδ e (cid:44) wo e − w(cid:63) P ew(cid:63) e − f e = w(cid:63) e, verifies the constraints {Dewe + b = 0}, namely, and the fact that A(cid:62)1 = 1, we obtain that (cid:101)w (cid:2)I Me − µRx,e(i)(cid:3)(cid:101)w e(i + 1) = A(cid:62)P e (cid:48) µA(cid:62)P epzx,e(i) − µA(cid:62)P eRx,e(i)wδ according to the following recursion: (cid:101)w (cid:48) e(i)− e. (82) (cid:48) e(i + 1) evolves (83) Taking the expectation of both sides of recursion (83), using Assumption 1, and E pzx,e(i) = 0, the mean error vector evolves according to: E(cid:101)w (cid:48) (cid:48) e(i + 1) = B E(cid:101)w r(cid:48) (cid:44) A(cid:62)P eRx,ewδ e. e(i) − µr(cid:48) , (84) (85) where B is given by (58) and Using arguments similar to subsection IV-B, we find that the multitask diffusion algorithm (35) is stable in the mean if the step-size is chosen such that the matrix B is stable. The asymptotic mean bias is given by: (cid:48) e(i) = −µ[I Me − B] −1r(cid:48) . (86) E(cid:101)w lim i→∞ e = wo e − w(cid:63) simulation results that limi→∞ (cid:107)E(cid:101)w Note that the bias depends on the step-size µ and the vector e. In the next section, we shall illustrate with wδ (cid:48) e(i)(cid:107)2 is on the order of µ2. The bias (86) is 0 in two cases: 1) in the perfect model scenario where wo e = 0); 2) if each e = D). agent is involved in at most one constraint (De = D(cid:48) In this second case, consider (85) and observe that A = I Me. e by its expression obtained from (28), and P e Replacing wδ by (49), yields r(cid:48) = 0. e in the To obtain the behavior of algorithm (35) toward w(cid:63) k satisfy the constraints (wδ mean-square sense, we use (81) to write: E{(cid:107)(cid:101)w 2 E{(cid:101)w Σ} = E{(cid:107)(cid:101)we(i + 1)(cid:107)2 The transient and steady-state behaviors of E{(cid:107)(cid:101)w be derived from the models derived for (cid:101)we(i) in the mean Σ} can and mean-square sense. We shall show with simulation results (cid:62) e (i + 1)}Σwδ Σ}− e(cid:107)2 Σ. (cid:48) e(i + 1)(cid:107)2 e + (cid:107)wδ (cid:48) e(i)(cid:107)2 (87) (cid:48) e(i)(cid:107)2 that the steady-state limi→∞ E(cid:107)(cid:101)w (cid:48) e(i)(cid:107)2 is on the order of E{(cid:107)(cid:101)w µ. We observed experimentally that modeling the behavior of Σ} accurately needs the exact expression of F. For zero-mean real valued regressors with Mk = M0 ∀ k, the evaluation of F leads to (see Appendix B): (cid:104)(cid:0) N(cid:88) ⊗b B(cid:62) F = B(cid:62) k (I Ne ⊗ Rx,k) ⊗b Sk(I Ne ⊗ Rx,k)(cid:1) + k ⊗b (Sk(I Ne ⊗ Rx,k))(cid:1) (cid:0) S (cid:62)(cid:1)(cid:105) (cid:0)I N 2 S (P eA ⊗b P eA), e ⊗ vec(I M0) ⊗ [vec(Rx,k)] (88) µ2 k=1 + (cid:62) (cid:62) where Sk is the N × N block diagonal matrix whose (k, k)-th block is equal to Ck ⊗ I M0. VI. SIMULATION RESULTS Throughout this section, the factors ckm were set to 1 jk , and Nkm ∩Ck = Ck for all m. We run algorithm (35) with uniform combination coefficients akn,km = 1 jk for all n. A. Theoretical model validation We considered a network consisting of 15 agents with the topology shown in Fig. 3. The regression vectors xk(i) were 2 × 1 zero-mean Gaussian distributed with covariance matrices Rx,k = σ2 x,kI 2. The noises zk(i) were zero-mean i.i.d. Gaussian random variables, independent of any other signal with variances σ2 z,k are shown in Fig. 3. We randomly sampled 9 linear equality z,k. The variances σ2 x,k and σ2 constraints of the form:(cid:88) dp(cid:96)w(cid:96) = bp · 12×1, (89) (cid:96)∈Ip where the scalars dp(cid:96) and bp were randomly chosen from the set {−3,−2,−1, 1, 2, 3}. We used a constant step-size µ = 0.025 for all agents. The results were averaged over 200 Monte-Carlo runs. First, we considered the case of a perfect model sce- nario where the observation parameter vector wo satisfies the equality constraints, i.e., w(cid:63) = wo. In Fig. 4 (left), we compare three algorithms: the non-cooperative LMS algorithm (obtained from (17) by setting P = I M and f = 0), the centralized CLMS algorithm (17) which assumes that the con- straints are available at the fusion center, and algorithm (35). For each algorithm, we report the theoretical transient MSD, the theoretical steady-state MSD, and the simulated MSD. We observe that the simulation results match well the actual performance. Furthermore, the network MSD is improved by promoting relationships between tasks. Finally, our algorithm performs well compared to the centralized solution. Next, we perturbed the optimum parameter vector wo as follows: wo pert = wo + uo, (90) pert does not satisfy the constraints (89). The entries of so wo uo were sampled from Gaussian distribution N (0, σ2). We evaluated algorithm (35) on 6 different setups characterized 9 by σ ∈ {0, 0.01, 0.05, 0.1, 0.2, 0.5, 1}. The theoretical and simulated learning curves with respect to wo e are reported in Fig. 4. Observe that the performance with respect to wo e highly deteriorates when σ increases. However, even for the largest values of σ = 1, algorithm (35) still performs well with respect to the solution w(cid:63) e of the optimization problem with constraints. e and w(cid:63) For comparison purposes, we illustrate in Fig. 5 the theo- retical and simulated learning curves with respect to w(cid:63) for the settings where σ = 0.5 (left) and σ = 1 (right) of the cen- tralized CLMS algorithm (17), algorithm (35) where the sub- nodes "project-then-combine", and the stochastic version (ob- tained by replacing the moments by instantaneous approxima- tions) of algorithm (31) where the sub-nodes "combine-then- project" (Appendix C explains how the performance of this algorithm can be obtained). Observe that both algorithms (35) and the stochastic version of (31) have approximately the same performance. However, with the settings considered in this section, algorithm (35) is less complex than algorithm (31) as explained in subsection III-B. Furthermore, we observe that the e is, the larger the performance gap between larger the vector wδ the centralized solution and the distributed solutions is. This is due to the bias (86) induced in the distributed solution which does not exist in the centralized CLMS algorithm (see Appendix C). In order to characterize the constraints violation at the sub-nodes for the setting where σ = 0.5, we evaluate the steady-state quantity (cid:107)D(cid:48) (cid:107)2 where D(cid:48) (cid:48) are given by (27) and we(∞) (cid:44) limi→∞ we(i). e and b When the sub-nodes project first and then combine, we obtain (cid:107)2 = (cid:107)Dewe(∞) + b(cid:107)2 = 0.0264. On the (cid:107)D(cid:48) ewe(∞) + b (cid:107)2 = (cid:107)Hwe(∞)(cid:107)2 = 0.0072. we obtain (cid:107)D(cid:48) (cid:48) ewe(∞) + b Thus, at the expense of a higher computational complexity, the constraints violation, measured by (cid:107)D(cid:48) (cid:107)2, is smaller when the projection step is performed after the combination step3. contrary, when the sub-nodes combine first and then project, (cid:48) ewe(∞) + b ewe(∞) + b (cid:48) (cid:48) In order to characterize the influence of the step-size µ on the performance of algorithm (35), Fig. 6 (left) reports the e for different theoretical steady-state MSD with respect to w(cid:63) values of µ. We observe that the network MSD increases 10 dB per decade (when the step-size goes from µ1 to 10µ1). This means that the steady-state MSD is on the order of µ. Fig. 6 (right) reports the squared norm of the bias (86) for different values of µ. We note that it increases approximately 20 dB per decade. This shows that, as expected, this quantity is on the order of µ2. Next, we considered the case of non-diagonal matrices Dp(cid:96) defined as: Dp(cid:96) = dp(cid:96)I 2 + ∆p(cid:96) (91) Parameters dp(cid:96) were randomly selected as in (89). The entries of the 2 × 2 matrix ∆p(cid:96) were sampled from Gaussian distri- bution N (0, σ2 D was set to 0.01 (left) and 1 (right). To test the tracking ability of algorithm (35), we also perturbed the parameter vector wo as in (90) by increasing σ2 every 500 iterations. In both cases, D). As shown in Fig. 7, the variance σ2 3We show in Appendix D that, for the perfect model scenario, the steady- state MSD is lower when the combination step is the last step in the algorithm. 10 Fig. 3: Experimental setup. (Left) Network topology with constraints. (Right) Regression and noise variances. Fig. 4: (Left) MSD comparison of the non-cooperative LMS, the centralized CLMS, and our multitask algorithm for the perfect model scenario. Learning curves of algorithm (35) with respect to wo e (right) for 6 different values of σ. e (middle) and w(cid:63) Fig. 5: Network MSD comparison with respect to w(cid:63), w(cid:63) of algorithm (31) for σ = 0.5 (left) and σ = 1 (right). e of the centralized CLMS (17), algorithm (35), and stochastic version D = 0.01 and σ2 D = 1, wo in (90) was set to satisfy i.e., σ2 the equality constraints defined by Dp(cid:96). We observe that the theoretical models match well the actual performance whatever the constraints are. Furthermore, algorithm (35) adapts its response to drifts in the location of w(cid:63) when wo changes over time. B. Optimal network flow As briefly discussed in the Introduction, we shall now consider the minimum-cost flow problem over the network with topology shown in Fig. 1. We are interested in online distributed learning where each node k seeks to estimate the entering and leaving flows fj from noisy measurement sk(i) of the external source, by relying only on local computations and communications with its neighbors. Let Mk be the number of flows to be estimated at node k. We denote by wk the Mk × 1 parameter vector containing the flows fj entering and leaving node k, negatively and positively signed, respectively. For instance, for nodes 1 and 2, we have: (92) w1 (cid:44) [f1 f2] (cid:62) w2 (cid:44) [−f1 f3 f4 f5] (cid:62) From the flow conservation principle, the noisy measurement sk(i) can be related to wk(i) as follows: sk(i) = 1(cid:62) (93) with zk(i) a zero-mean measurement noise, and 1Mk×1 an Mk × 1 vector of ones. We consider the bi-objective problem Mk×1wk + zk(i), 112233I1I14455I2I266I3I37788I4I4991010I5I5I6I61111I7I712121313I8I8I9I914141515Nodek2468101214<2x;k11.21.4Nodek2468101214<2z;k0.10.110.120.130.140.151.2221.3291.021.1881.0771.1841.2961.0281.1711.3751.1041.3111.1111.3231.0420.1140.12710.12880.13340.1190.13690.13820.12890.1280.130.13430.11860.13410.10320.1322Iterationi01002003004005006007008009001000MSDindB-35-30-25-20-15-10-50510SimulatedtransientMSDTheoreticaltransientMSDTheoreticalsteady-stateMSDNon-cooperativeLMSProposedAlgorithmCentralizedCLMSIterationi01002003004005006007008009001000MSDindB-35-30-25-20-15-10-5051015SimulatedtransientMSDTheoreticaltransientMSDTheoreticalsteady-stateMSD<=0:1<=0<=0:01<=0:05<=0:2<=1<=0:5Iterationi01002003004005006007008009001000MSDindB-35-30-25-20-15-10-5051015SimulatedtransientMSDTheoreticaltransientMSDTheoreticalsteady-stateMSD<=0:1<=0:05<=0:01<=0<=0:2<=0:5<=1Iterationi01002003004005006007008009001000MSDindB-25-20-15-10-5051015SimulatedtransientMSDTheoreticaltransientMSDTheoreticalsteady-stateMSD500600700800-22-21.9-21.8-21.7-21.6-21.5-21.4CentralizedCLMSCombinethenProjectProjectthenCombineIterationi01002003004005006007008009001000MSDindB-25-20-15-10-50510SimulatedtransientMSDTheoreticaltransientMSDTheoreticalsteady-stateMSD500600700800-20.6-20.5-20.4-20.3-20.2-20.1-20-19.9ProjectthenCombineCentralizedCLMSCombinethenProject11 Fig. 6: Influence of the step-size µ on the performance of the algorithm. (Left) Network steady-state MSD for different values of µ. (Right) Squared norm of the bias, i.e, lim (cid:48) (i)(cid:107)2, for different values of µ. i→∞(cid:107)E(cid:101)w Fig. 7: Tracking ability of the algorithm for two sets of linear equality constraints. (Left) σ2 D = 0.01. (Right) σ2 D = 1. so-called weight-drift problem of the LMS algorithm [43]. In order to test the tracking ability of the algorithm, the external flow sk at each node k was re-generated from U(0, 3) after 45000 iterations. The MSD learning curve with respect to the solution of problem (94) is reported in Fig. 8. This result was obtained by averaging over 150 Monte-Carlo runs. This figure shows that our strategy was able to solve the minimum-cost flow problem in a fully distributed manner. The estimated flows over the network for both settings considered in the tracking experiment are showed in Fig. 9 (left and middle). Note that the direction of the estimated flow between nodes 3 and 4 is reversed. The true and estimated flows are reported in Fig. 9 (right) for both settings. (cid:16)E sk(i) − 1(cid:62) N(cid:88) k=1 Mk×1wk2 + 2(cid:107)wk(cid:107)2(cid:17) η , consisting of minimizing Ezk(i)2 and the cost network flow. We shall assume that the cost for flow through an arc is quadratic in the flow, as in applications such as electrical network monitoring and urban traffic control [35], [37]. We formulate the estimation problem as follows: minimize w1,...,wN subject to [wk]f (k,(cid:96)) + [w(cid:96)]f ((cid:96),k) = 0, (cid:96) ∈ Nk, for all k, (94) where [wp]f (p,q) returns the flow entry in wp that node p has in common with node q, and η is a tuning parameter to trade off between both objectives. For each agent k, the external flow sk and the variance σ2 z,k of the Gaussian noise zk(i) were randomly generated from the uniform distributions U(0, 3) and U(0.1, 0.14), respectively. In order to solve the multitask problem (94) in a fully distributed manner, we applied algorithm (35) by modifying the adaptation step according to: ψkm (i + 1) = wkm(i) + µ ckm1Mk×1[sk(i) − 1(cid:62) Mk×1wkm (i)] µ 2 − ckmη wkm (i), (95) and setting µ = 0.2 and η = 0.002. Note that equation (95) leads to a leaky-LMS version of the proposed algorithm. It is well-known that the leaky-LMS algorithm introduces a bias compared to the LMS, but improves its robustness against the Fig. 8: MSD performance and tracking ability of algo- rithm (35) for the minimum cost network flow problem. Step-size710-310-210-1MSDindB-40-35-30-25-20-15-10Step-size710-310-210-1Squared`2-normofthebiasindB-50-45-40-35-30-25-20-15-10-5Iterationi05001000150020002500MSDindB-35-30-25-20-15-10-50510SimulatedMSDw.r.t.w?Theoreticalsteady-stateMSDw.r.t.w?SimulatedMSDw.r.t.woTheoreticalsteady-stateMSDw.r.t.wo<=0<=0:1<=0:25<=0:5<=1Iterationi05001000150020002500MSDindB-35-30-25-20-15-10-50510SimulatedMSDw.r.t.w?Theoreticalsteady-stateMSDw.r.t.w?SimulatedMSDw.r.t.woTheoreticalsteady-stateMSDw.r.t.wo<=0<=0:1<=0:25<=0:5<=1Iterationi#1040123456789MSDindB-30-25-20-15-10-5051012 Fig. 9: Estimated network flows. (Left) First experiment. (Middle) Second experiment. A rounding to 2 decimal places is adopted when visualizing the estimated flows. (Right) Comparison of the true and estimated flows sk (top: first experiment, bottom: second experiment). C. Numerical solution of a two-dimensional process Consider now the problem of estimating a two-dimensional process driven by a partial differential equation (PDE) with a sensor network. To see how our distributed algorithm can be tuned to address this issue, we shall focus on the Poisson's PDE defined by: ∂2f (x, y) ∂2f (x, y) + ∂y2 ∂x2 = g(x, y), (x, y) ∈ [0, 1]2, (96) with g : [0, 1]2 → R an input function, and on a two dimen- sional network of (n−2)2 sensor nodes and 4(n−1) boundary points equally spaced over the unit square (x, y) ∈ [0, 1]2 with ∆x = ∆y = ∆ = 1 We introduce the grid point (xk, y(cid:96)) (cid:44) (k∆, (cid:96)∆) and the sampled values at this point fk,(cid:96) (cid:44) f (k∆, (cid:96)∆) and gk,(cid:96) (cid:44) g(k∆, (cid:96)∆) with 0 ≤ k, (cid:96) ≤ n−1. We use the central difference approximation for the second derivative [34]: n−1, as illustrated in Fig. 10 (a). known a priori as they correspond to boundary conditions. We denote by f o k(cid:96) the value at (xk, y(cid:96)) of the function f (x, y) that satisfies (96), and by fk(cid:96) the estimated value of f o k(cid:96). To each node (k, (cid:96)) we associate an Mk(cid:96) × 1 parameter vector wk(cid:96) to estimate, an Mk(cid:96) × 1 regression vector xk(cid:96) and a scalar vo k(cid:96), defined in Table I depending on the node location on the grid. Given the values of f (x, y) at the boundary points, and according to (99), the linear regression model can be written as follows: gk(cid:96)(i) = x(cid:62) k(cid:96)wk(cid:96) + vo k(cid:96) + zk(cid:96)(i). (102) As can be seen in Table I, equality constraints of the form (1b) need to be imposed on the parameter vectors of neighboring sensor nodes in order to achieve equality between common entries. For instance, let us consider neighboring nodes (k, (cid:96)) and (k + 1, (cid:96)) with 2 ≤ k ≤ n − 4 and 2 ≤ (cid:96) ≤ n − 3. Since these nodes are jointly estimating fk,(cid:96) and fk+1,(cid:96), the following equality constraint is required: (cid:20) 0 −1 0 0 0 −1 0 0 0 0 (cid:21) (cid:2)gk(cid:96)(i)−x(cid:62) wk(cid:96) + w(k+1)(cid:96) = 0. (103) (cid:3), Algorithm (35) can be used to address this problem by replacing the adaptation step (121a) by: ψk(cid:96)m (i+1) = wk(cid:96)m(i)+µ ck(cid:96)m xk(cid:96) k(cid:96)wk(cid:96)m (i)−vo k(cid:96) (104) the m-th where wk(cid:96)m(i) denotes the estimate of wk(cid:96) at sub-node of (k, (cid:96)). The noises zk,(cid:96)(i) were zero-mean i.i.d. Gaussian distributed with variances σ2 z,k(cid:96) randomly generated from the uniform distribution U(0.1, 0.14). We used a constant step-size µ = 7 · 10−5 for all nodes. Figure 11 shows the network MSD learning curves for n = 9. The simulated curves were obtained by averaging over 100 independent runs. Figure 12 shows the true (left) and estimated (right) process after convergence of our algorithm. VII. CONCLUSION AND PERSPECTIVES In this work, we proposed a multitask LMS algorithm for solving problems that require the simultaneous estimation of multiple parameter vectors that are related locally via linear (cid:17) (cid:17) (97) (98) (cid:20) 1 0 0 0 0 (cid:21) 0 0 0 0 1 (cid:16) (cid:16) fk+1,(cid:96) − 2fk,(cid:96) + fk−1,(cid:96) fk,(cid:96)+1 − 2fk,(cid:96) + fk,(cid:96)−1 (cid:17) ∂2f (k∆, (cid:96)∆) ∂x2 ∂2f (k∆, (cid:96)∆) ∂y2 1 ∆2 1 ∆2 ≈ ≈ which leads to: (cid:16) 1 ∆2 −4fk,(cid:96)+fk−1,(cid:96)+fk,(cid:96)−1+fk,(cid:96)+1+fk+1,(cid:96) = gk,(cid:96). (99) In this experiment, we shall consider the unknown physical process f and the input function g given by: (100) f (x, y) = (1 − x2)(2y3 − 3y2 + 1), (101) g(x, y) = −2(2y3 − 3y2 + 1) + 6(1 − x2)(2y − 1), for (x, y) ∈ [0, 1]2 with boundary conditions f (0, y) = 2y3 − 3y2 + 1, f (x, 0) = 1 − x2, and f (1, y) = f (x, 1) = 0. These functions are illustrated in Fig. 10 (b), (c). The objective is to estimate f (x, y) at the interior grid points (xk, y(cid:96)) with 0 < k, (cid:96) < n − 1, given noisy measurements gk(cid:96)(i) = gk(cid:96) + zk(cid:96)(i) of g(x, y) collected by the sensors located at these interior grid points. The noise process zk(cid:96)(i) is assumed to be zero mean, temporally white, and spatially independent. The values of f (x, y) at the boundary points are 1122334455667788991010DD0.580.582.792.790.260.261.031.031.11.11.911.910.750.750.630.630.810.812.92.92.12.10.20.20.750.751.231.232.332.330.540.541.981.982.22.21.571.571.761.760.660.662.992.994.984.987.777.770.690.691122334455667788991010DD2.122.121.121.122.912.910.940.941.661.661.91.90.210.212.872.872.132.132.442.441.531.530.410.410.740.741.751.751.681.683.343.340.970.972.062.063.93.91.021.022.822.820.190.19448.098.0910.2210.22Nodek12345678910Ext..owssk00.511.522.53TruevalueEstimatedvalueNodek12345678910Ext..owssk00.511.522.53TruevalueEstimatedvalue13 (a) An n × n grid network for the solution of Poisson's equation (b) f (x, y) = (1 − x2)(2y3 − 3y2 + 1) (c) g(x, y) = −2(2y3 − 3y2 + 1) + 6(1 − x2)(2y − 1) Fig. 10: Network topology, function f (x, y) to estimate over the interior grid points, and input function g(x, y). k (cid:96) 1 2 ... n − 3 n − 2 1 2, . . . , n − 3 n − 2 [fk,(cid:96), fk,(cid:96)+1, fk+1,(cid:96)](cid:62) [−4, 1, 1](cid:62) f o 1,0 + f o 0,1 [fk,(cid:96), fk−1,(cid:96), fk,(cid:96)+1, fk+1,(cid:96)](cid:62) [−4, 1, 1, 1](cid:62) f o k,0 [fk,(cid:96), fk,(cid:96)−1, fk,(cid:96)+1, fk+1,(cid:96)](cid:62) [fk,(cid:96), fk−1,(cid:96), fk,(cid:96)−1, fk,(cid:96)+1, fk+1,(cid:96)](cid:62) [−4, 1, 1, 1](cid:62) f o 0,(cid:96) [fk,(cid:96), fk,(cid:96)−1, fk+1,(cid:96)](cid:62) [−4, 1, 1](cid:62) f o 0,n−2 + f o 1,n−1 [fk,(cid:96), fk−1,(cid:96), fk,(cid:96)−1, fk+1,(cid:96)](cid:62) [−4, 1, 1, 1, 1](cid:62) 0 [−4, 1, 1, 1](cid:62) f o k,n−1 [fk,(cid:96), fk−1,(cid:96), fk,(cid:96)+1](cid:62) [−4, 1, 1](cid:62) f o n−2,0 + f o n−1,1 [fk,(cid:96), fk−1,(cid:96), fk,(cid:96)−1, fk,(cid:96)+1](cid:62) [−4, 1, 1, 1](cid:62) f o n−1,(cid:96) [fk,(cid:96), fk−1,(cid:96), fk,(cid:96)−1](cid:62) [−4, 1, 1](cid:62) f o n−2,n−1 + f o n−1,n−2 TABLE I: Parameter vector wk(cid:96) (first row of each cell), regression vector ∆2xk(cid:96) (second row of each cell), and scalar value ∆2vo k(cid:96) (last row of each cell) at each node (k, (cid:96)). sense was studied. We checked with simulations that the agents are able to reach the optimal solution with good precision. In future work, we shall extend our approach to other types of constraints and also consider other constraints distribution over networks. APPENDIX A Fig. 11: Network MSD performance for n = 9. Fig. 12: Poisson process f (x, y) over the network grid. (Left) True process. (Right) Estimated process. constraints. Our primal technique was based on the stochastic gradient projection algorithm with constant step-sizes. The behavior of the algorithm in the mean and mean-square-error PROJECTION MATRIX STRUCTURE We denote by De,p the p-th block row in De and by [De]p,km the Lp × Mk block of De,p corresponding to the km-th sub-node. First, we show that the Mk × M(cid:96) (km, (cid:96)n)-th block of the Ne × Ne block matrix P e in (49) is equal to: e,p)−1[De]p,km, I Mk − [De](cid:62) −[De](cid:62) 0Mk×M(cid:96), p,km(De,pD(cid:62) if km = (cid:96)n and km ∈ Ie,p, if km (cid:54)= (cid:96)n and km, (cid:96)n ∈ Ie,p, e,p)−1[De]p,(cid:96)n , p,km(De,pD(cid:62) otherwise. [P e]km,(cid:96)n =  (cid:26) [De](cid:62) (105) Furthermore, we show that the km-th block of the Ne × 1 block column vector f e in (50) is equal to: p,km(De,pD(cid:62) e,p)−1bp, if km ∈ Ie,p, otherwise. 0Mk×1, [f e]km = It can be verified that DeD(cid:62) (106) e is a P × P block diagonal matrix whose (p, p)-th block is of dimension Lp × Lp and is given by: (107) e ]p,p = De,pD(cid:62) e,p = DpD(cid:62) p . [DeD(cid:62) (0,0)(0,0)(0,y)(0,y)(x,0)(x,0)(x,y)(x,y)((n1)x,(n1)y)((n1)x,(n1)y)BoundarypointsBoundarypointsSensornodesSensornodesxxyy=(1,1)=(1,1)Iterationi050001000015000MSDindB-60-50-40-30-20-10000.20.40.6x0.8110.8y0.60.40.2010.80.20.40.60Truef(x;y)00.10.20.30.40.50.60.70.80.9100.20.40.6x0.8110.80.6y0.40.2100.80.20.40.60Estimatedf(x;y)00.10.20.30.40.50.60.70.80.91(cid:110) (De,pD(cid:62) e,p) −1(cid:111)P The inverse of the block diagonal matrix DeD(cid:62) e is: (DeD(cid:62) e ) −1 = diag (108) By multiplying the matrix (DeD(cid:62) e )−1 from the left by D(cid:62) we obtain an Ne × P block matrix whose (km, p)-th block is of dimension Mk × Lp given by: p=1 . e [D(cid:62) (cid:26) [De](cid:62) e (DeD(cid:62) e ) p,km (De,pD(cid:62) −1]km,p = 0Mk×Lp , e,p)−1, if km ∈ Ie,p otherwise. (109) e (DeD(cid:62) When we multiply the matrix D(cid:62) e )−1 from the right by De, we obtain an Ne × Ne block matrix whose (km, (cid:96)n)- th block corresponding to sub-nodes km, (cid:96)n is of dimension Mk × M(cid:96) and is given by: (cid:26) [De](cid:62) e (DeD(cid:62) −1De]km,(cid:96)n [D(cid:62) e ) p,km(De,pD(cid:62) e,p)−1[De]p,(cid:96)n , if km, (cid:96)n ∈ Ie,p, otherwise. = 0Mk×M(cid:96), (110) From (50) and (109), we obtain (106). APPENDIX B EVALUATION OF THE MATRIX F Without loss of generality, we assume in the following that Mk is uniform across the network, i.e., Mk = M0 for all k. We note that for any symmetric matrix T , we have [47]: E{xk(i)x(cid:62) k (i)T x(cid:96)(i)x(cid:62) = Rx,kT Rx,(cid:96) + δk,(cid:96) (cid:0)Rx,kT Rx,k + Rx,ktr(Rx,kT )(cid:1). (cid:96) (i)} (111) = P eAΣA(cid:62)P e − µP eAΣA(cid:62)P eRx,e− From (54) and (64), we obtain: Σ(cid:48) µRx,eP eAΣA(cid:62)P e + µ2E{Rx,e(i)P eAΣA(cid:62)P eRx,e(i)}. (112) In order to evaluate Σ(cid:48) we need to evaluate the fourth term on the RHS of the above equation. Let: K (cid:44) E{Rx,e(i)P eAΣA(cid:62)P eRx,e(i)}, T (cid:44) P eAΣA(cid:62)P e. (113) (114) It can be verified that the (km, (cid:96)n)-th block of the matrix K corresponding to the (km, (cid:96)n)-th sub-node is given by: [K]km,(cid:96)n = ckmc(cid:96)n = ckmc(cid:96)n Rx,k[T ]km,(cid:96)n Rx,(cid:96)+ δk,(cid:96)ckmc(cid:96)n E(cid:8)xk(i)x(cid:62) (cid:0)Rx,k[T ]km,(cid:96)n Rx,k + Rx,ktr(Rx,k[T ]km,(cid:96)n )(cid:1), where the M0 × M0 matrix [T ]km,(cid:96)n is the (km, (cid:96)n)-th block of the matrix T . The matrix K in (113) can be written as: Sk(I Ne ⊗ Rx,k)T (I Ne ⊗ Rx,k)Sk+ K =Rx,eT Rx,e + k (i)[T ]km,(cid:96)n x(cid:96)(i)x(cid:62) (cid:96) (i)(cid:9) N(cid:88) (115) k=1 Sk(I Ne ⊗ Rx,k)Z kSk, (116) N(cid:88) k=1 14 hm,(cid:96)n (cid:3) = I M0[vec(Rx,k)] (cid:62)vec([T ]km,(cid:96)n ). where Sk is the N × N block diagonal matrix whose (k, k)-th block is equal to Ck ⊗ I M0, and Z k is the Ne × Ne block (cid:2)Z k matrix whose (km, (cid:96)n)-th block is given by: (117) Applying the block-vectorization operator to K and using the property bvec(ABC) = (C(cid:62) (cid:16) N(cid:88) bvec(K) = (Rx,e ⊗b Rx,e)bvec(T )+ (cid:16) N(cid:88) (cid:62) k (I Ne ⊗ Rx,k) ⊗b Sk(I Ne ⊗ Rx,k) S bvec(Z k), Sk[I Ne ⊗ Rx,k](cid:1)(cid:17) ⊗b A)bvec(B), we obtain: bvec(T )+ (cid:62) S k ⊗b (cid:17) (cid:0) k=1 k=1 (118) (cid:16) (cid:62)(cid:17) I N 2 e ⊗ vec(I M0) ⊗ [vec(Rx,k)] where bvec(Z k) can be expressed as: bvec(T ), bvec(Z k) = (119) where bvec(T ) = (P eA ⊗b P eA)σ. Finally, we conclude that the matrix F in (66) can be written as: (cid:16) N(cid:88) F = B(cid:62) N(cid:88) ⊗b B(cid:62) (cid:62) (cid:16) k (I Ne ⊗ Rx,k) ⊗b Sk(I Ne ⊗ Rx,k) S (cid:16) Sk(I Ne ⊗ Rx,k)(cid:1)(cid:17) (cid:62) k ⊗b S e ⊗ vec(I M0) ⊗ [vec(Rx,k)] (cid:62)(cid:17) (P eA ⊗b P eA). (P eA ⊗b P eA) + µ2 (cid:17) I N 2 (cid:0) µ2 k=1 k=1 + · (120) APPENDIX C PERFORMANCE OF COMPETING ALGORITHMS We compare in the simulation section algorithm (35) with the non-cooperative LMS algorithm (obtained from (17) by setting P = I M and f = 0), the centralized CLMS algorithm (17), and the following algorithm: ψkm (i + 1) = wkm (i) + (cid:88) µ ckmxk(i)[dk(i) − x(cid:62) kn∈Nkm∩ Ck φkm (i + 1) = wkm (i + 1) = [P p]km,• · col(cid:8)φ(cid:96)n (i + 1)(cid:9) akn,km ψkn (i + 1), k (i)wkm(i)], (121a) (121b) (cid:96)n∈Ie,p − [f p]km, (121c) where the sub-nodes "combine-then-project" instead of "project-then-combine". In the following, we show how the theoretical learning curves of these algorithms can be obtained from the analysis in Sections IV and V. Consider the central- (i) denote the ized CLMS algorithm (17). Let (cid:101)w(i) and (cid:101)w N × 1 block error vectors at the fusion center given by: (cid:48) (cid:101)w(i) (cid:44) wo − w(i), (cid:101)w (cid:48) (i) (cid:44) w(cid:63) − w(i). (122) 15 (123) Subtracting wo from both sides of recursion (17) and using the linear data model (2), we obtain: (cid:101)w(i + 1) = P (I M − µRx(i))(cid:101)w(i) − µPpxz(i)+ (124) (125) (I M − P) wo + f , where Rx(i) and pxz(i) are given by: Rx(i) (cid:44) diag(cid:8)xk(i)x(cid:62) k (i)(cid:9)N pxz(i) (cid:44) col{dk(i)xk(i)}N (i) = (cid:101)w(i) − wδ with (123) Let wδ (cid:44) wo − w(cid:63). Using (cid:101)w and the fact that w(cid:63) satisfies Pw(cid:63) − f = w(cid:63), we obtain: b(i+1) = P (I M − µRx(i))(cid:101)w (cid:101)w (i)−µPpxz(i)−µPRx(i)wδ. (126) Comparing recursions (123) and (126) with recursions (52) and (83), we observe that the learning curves of the centralized solution (17) can be deduced from those of the decentralized solution (35) by properly modifying the coefficient matrices and vectors. Note that, the centralized solution is unbiased with respect to w(cid:63) since µPERx(i)wδ = 0. k=1 , k=1 . (cid:48) (cid:48) (cid:48) Next, consider the distributed solution (121). Following the same line of reasoning as in Subsection IV-A, we obtain the following recursions for the block error vectors (41) and (80): (cid:101)we(i + 1) = P eA(cid:62) (cid:101)w [I Me − µRx,e(i)](cid:101)we(i)− µP eA(cid:62)pxz,e(i) + (I Me − P e) wo [I Me − µRx,e(i)](cid:101)w (cid:48) e(i)− µP eA(cid:62)pxz,e(i) − µP eA(cid:62)Rx,e(i)wδ e(i + 1) = P eA(cid:62) (cid:48) e. e + f e, (127) (128) Comparing recursions (127) and (128) with recursions (52) and (83), we observe that the learning curves of the distributed solution (121) can be deduced from the theoretical curves of the decentralized solution (35) by properly replacing the prod- uct A(cid:62)P e in the analysis of Sections IV and V by the product P e A(cid:62) and the vector r in (56) by r (cid:44) (I Me − P e) wo e +f e. APPENDIX D PERFORMANCE COMPARISON As explained after algorithm (35), this permutation of the projection and the aggregation steps is very useful since it allows to simplify the algorithm. We show hereafter that, for the perfect model scenario (i.e., wo = w(cid:63)), this permutation enhances the steady-state mean-square-error performance. First, let us consider the algorithm where agent k performs the projection step before the aggregation step. In the case of the perfect model scenario, we know that r = 0 (see (56)), and E(cid:101)we(∞) = 0 (see (59)). When matrix F is stable, Using A(cid:62) ζ (cid:63) 1 = µ2 from (75)–(77), we obtain the following steady-state network performance with metric σss (cid:44) bvec(Σss): ζ (cid:63) = [bvec(Y(∞))] = [bvec(Y(∞))] = [bvec(Y(∞))] +∞(cid:88) +∞(cid:88) j=0 = tr = tr (129) (cid:62) j=0 j=0 bvec )j]σss )jΣssBj(cid:17) (cid:62) +∞(cid:88) (I − F ) −1σss )j ⊗b (B(cid:62) [(B(cid:62) (cid:16) (cid:62) +∞(cid:88) (B(cid:62) )jΣssBj(cid:17) (cid:16)Y(∞)(B(cid:62) )j(cid:17) (cid:16) ΣssBjY(∞)(B(cid:62) (cid:16)A(cid:62)P e(I − µRx,e) (cid:17)j A(cid:62)P eSP eA· (I − µRx,e)P eA(cid:17)j(cid:17) (cid:111)N (cid:110) (cid:16) Σss , (130) Replacing B and Y(∞) in the above expression by (58) and (77), we obtain: ckc(cid:62) z,kRx,k k ⊗ σ2 where S (cid:44) diag Let us consider now that agent k performs the aggregation step before projecting. Let ζ (cid:63) 1 be the steady-state network performance with metric σss (cid:44) bvec(Σss). Following the same line of reasoning, we obtain: k=1 . j=0 (cid:16) tr +∞(cid:88) j=0 ζ (cid:63) = µ2 (cid:16) +∞(cid:88) j=0 ζ (cid:63) 1 = µ2 tr Σss (cid:16)P eA(cid:62) (cid:17)j P eA(cid:62) SAP e (I − µRx,e) (cid:16) (cid:17)j(cid:17) (I − µRx,e)AP e . tr j=0 1 jk (cid:16) ζ (cid:63) = µ2 +∞(cid:88) (131) , and Nkm ∩ Let us assume that the factors ckm are set to 1 jk Ck = Ck for all m. We further assume that akn,km are set to for all n. In this case, it can be verified that A = A(cid:62), SA = S. Thus, (I − µRx,e) = (I − µRx,e)A(cid:62), and A(cid:62) A(cid:62) (cid:17)j ζ (cid:63) can be written alternatively as: (cid:125) (cid:123)(cid:122) P e· (I − µRx,e) A(cid:17) (cid:17)j Wj (cid:123)(cid:122) (cid:125) ΣssA(cid:62)(cid:16)P eA(cid:62) (cid:124) (cid:16) (cid:124) (I − µRx,e)AP e SP e SA = S, ζ (cid:63) (cid:17)j (cid:16)P eA(cid:62) +∞(cid:88) (cid:124) (cid:123)(cid:122) (cid:125) (I − µRx,e) (cid:16) Wj (cid:123)(cid:122) (cid:124) SP e (I − µRx,e)AP e 1 can be written alternatively as: (cid:17) . (133) P e· (cid:17)j (cid:125) (132) (cid:16) Σss W(cid:62) j=0 tr . j W(cid:62) j Hence, we obtain: +∞(cid:88) ζ (cid:63) − ζ (cid:63) = µ2 (cid:16) Σss tr 1 (cid:16)A(cid:62) j=0 WjP eSP eW (cid:62) j A − WjP eSP eW (cid:62) j (cid:17)(cid:17) . (134) I, we obtain: When Σss = 1 Ne +∞(cid:88) ζ (cid:63) − ζ (cid:63) µ2 = Ne (cid:16)A(cid:62) tr 1 (cid:17) (cid:62) j=0 WjP eSP eW j A − WjP eSP eW ≤ 0, where we used the fact that tr(A(cid:62)HA) ≤ tr(H) for any doubly-stochastic matrix A and any non-negative matrix H of compatible dimensions (see Theorem C.3 in [11]). (cid:62) j REFERENCES [1] D. P. Bertsekas, "A new class of incremental gradient methods for least squares problems," SIAM J. on Optim., vol. 7, no. 4, pp. 913–926, Nov. 1997. [2] A. Nedic and D. P. Bertsekas, "Incremental subgradient methods for nondifferentiable optimization," SIAM J. on Optim., vol. 12, no. 1, pp. 109–138, Jul. 2001. [3] C. G. Lopes and A. H. Sayed, "Diffusion least-mean squares over adaptive networks: Formulation and performance analysis," IEEE Trans. Signal Process., vol. 56, no. 7, pp. 3122–3136, Jul. 2008. [4] 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. [5] K. Srivastava and A. Nedi´c, "Distributed asynchronous constrained stochastic optimization," IEEE J. Sel. Topics Signal Process., vol. 5, no. 4, pp. 772–790, Aug. 2011. [6] J. F. C. Mota, J. M. F. Xavier, P. M. Q. Aguiar, and M. Puschel, "D- ADMM: A communication-efficient distributed algorithm for separable optimization," IEEE Trans. Signal Process., vol. 61, no. 10, pp. 2718– 2723, May 2013. [7] S. Lee and A. Nedi´c, "Distributed random projection algorithm for convex optimization," IEEE J. Sel. Topics Signal Process., vol. 7, no. 2, pp. 221–229, 2013. [8] 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, Apr. 2013. [9] A. H. Sayed, "Adaptation, learning, and optimization over networks," Found. Trends Mach. Learn., vol. 7, no. 4-5, pp. 311–801, 2014. [10] Z. J. Towfic and A. H. Sayed, "Adaptive penalty-based distributed stochastic convex optimization," IEEE Trans. Signal Process.,, vol. 62, no. 15, pp. 3924–3938, Aug. 2014. [11] A. H. Sayed, "Diffusion adaptation over networks," in Academic Press Library in Signal Processing, vol. 3, pp. 322–454. Elsevier, 2014. [12] Z. J. Towfic and A. H. Sayed, "Stability and performance limits of adaptive primal-dual networks," IEEE Trans. Signal Process., vol. 63, no. 11, pp. 2888–2903, Jun. 2015. [13] H. Zhang, W. Shi, A. Mokhtari, A. Ribeiro, and Q. Ling, "Decentralized constrained consensus optimization with primal dual splitting projec- tion," in Proc. IEEE Glob. Conf. Signal Inf. Process., Washington, DC, USA, Dec 2016, pp. 565–569. [14] A. Bertrand and M. Moonen, "Distributed adaptive node-specific signal estimation in fully connected sensor networks – Part I: Sequential node updating," IEEE Trans. Signal Process., vol. 58, no. 10, pp. 5277–5291, Oct. 2010. [15] A. Bertrand and M. Moonen, "Distributed node-specific LCMV beam- forming in wireless sensor networks," IEEE Trans. Signal Process., vol. 60, no. 1, pp. 233–246, Jan. 2012. [16] C. Eksin and A. Ribeiro, "Distributed network optimization with heuristic rational agents," IEEE Trans. Signal Process., vol. 60, no. 10, pp. 5396–5411, Oct. 2012. [17] V. Kekatos and G. B. Giannakis, "Distributed robust power system state estimation," IEEE Trans. Power Syst., vol. 28, no. 2, pp. 1617–1626, May 2013. 16 [18] N. Bogdanovi´c, J. Plata-Chaves, and K. Berberidis, "Distributed incremental-based LMS for node-specific parameter estimation over adaptive networks," in Proc. Int. Conf. Acoust., Speech, Signal Process., May 2013, pp. 5425–5429. [19] N. Bogdanovi´c, J. Plata-Chaves, and K. Berberidis, "Distributed incremental-based LMS for node-specific adaptive parameter estima- tion," IEEE Trans. Signal Process., vol. 62, no. 20, pp. 5382–5397, Oct. 2014. [20] J. Plata-Chaves, N. Bogdanovic, 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. [21] 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. [22] 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. [23] R. Nassif, C. Richard, A. Ferrari, and A. H. Sayed, "Multitask diffusion adaptation over asynchronous networks," IEEE Trans. Signal Process., vol. 64, no. 11, pp. 2835–2850, Jun. 2016. [24] R. Nassif, C. Richard, A. Ferrari, and A. H. Sayed, "Proximal multitask learning over networks with sparsity-inducing coregularization," IEEE Trans. Signal Process., vol. 64, no. 23, pp. 6329–6344, Dec. 2016. [25] R. Nassif, C. Richard, A. Ferrari, and A. H. Sayed, "Multitask diffusion LMS with sparsity-based regularization," in Proc. Int. Conf. Acoust., Speech, Signal Process., Brisbane, Australia, Apr. 2015, pp. 3516–3520. [26] 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., Reims, France, Sept. 2014, pp. 1–6. [27] X. Zhao and A. H. Sayed, "Clustering via diffusion adaptation over in Proc. Int. Workshop Cogn. Inf. Process., Parador de networks," Baiona, Spain, May 2012, pp. 1–6. [28] J. Chen, C. Richard, and A. H. Sayed, "Diffusion LMS over multitask networks," IEEE Trans. Signal Process., vol. 63, no. 11, pp. 2733–2748, Jun. 2015. [29] X. Zhao and A. H. Sayed, "Distributed clustering and learning over networks," IEEE Trans. Signal Process., vol. 63, no. 13, pp. 3285– 3300, Jul. 2015. [30] J. Chen, S. K. Ting, C. Richard, and A. H. Sayed, "Group diffusion LMS," in Proc. Int. Conf. Acoust., Speech, Signal Process., Shangai, China, Mar. 2016. [31] J. Plata-Chaves, M. H. Bahari, M. Moonen, and A. Bertrand, "Unsuper- vised diffusion-based LMS for node-specific parameter estimation over wireless sensor networks," in Proc. Int. Conf. Acoust., Speech, Signal Process., Shangai, China, Mar. 2016, pp. 4159–4163. [32] R. Abdolee, B. Champagne, and A. H. Sayed, "Estimation of space- time varying parameters using a diffusion LMS algorithm," IEEE Trans. Signal Process., vol. 62, no. 2, pp. 403–418, Jan. 2014. [33] J. Chen, Z. J. Towfic, and A. H. Sayed, "Dictionary learning over distributed models," IEEE Trans. Signal Process., vol. 63, no. 4, pp. 1001–1016, Feb. 2015. [34] D. P. Bertsekas and J. N. Tsitsiklis, Parallel and Distributed Computa- tion: Numerical Methods, Prentice-Hall, Inc., NJ, USA, 1989. [35] R. K. Ahuja, T. L. Magnanti, and J. B. Orlin, Network Flows: Theory, Algorithms, and Applications, Prentice Hall, Inc. Upper Saddle River, NJ, USA, 1993. [36] S. Boyd and L. Vandenberghe, Convex Optimization, Cambridge University Press, 2004. [37] J. A. Ventura, "Computational development of a Lagrangian dual approach for quadratic networks," Networks, vol. 21, no. 4, pp. 469–485, 1991. [38] O. L. Frost III, "An algorithm for linearly constrained adaptive array processing," Proc. IEEE, vol. 60, no. 8, pp. 926–935, Aug. 1972. [39] D. Estrin, L. Girod, G. Pottie, and M. Srivastava, "Instrumenting the in Proc. Int. Conf. Acoust., world with wireless sensor networks," Speech, Signal Process., 2001, vol. 4, pp. 2033–2036. [40] I. K. Harrane, R. Flamary, and C. Richard, "Toward privacy-preserving diffusion strategies for adaptation and learning over networks," in Proc. 24th Eur. Signal Process. Conf., Budapest, Hungary, Aug. 2016. [41] D. P. Bertsekas, Nonlinear Programming, Athena scientific, 1999. [42] R. Arablouei, K. Doganc¸ay, and S. Werner, "On the mean-square Available as performance of arXiv:1412.2424, Feb. 2015. the constrained LMS algorithm," [43] A. H. Sayed, Adaptive Filters, Wiley, NY, 2008. 17 Ali H. Sayed (S'90-M'92-SM'99-F'01) is a profes- sor and former chairman of electrical engineering at the University of California, Los Angeles, where he directs the UCLA Adaptive Systems Laboratory. An author of over 480 scholarly publications and six books, his research involves several areas including adaptation and learning, statistical signal processing, distributed processing, network and data sciences, and biologically-inspired designs. Dr. Sayed has re- ceived several awards including the 2015 Education Award from the IEEE Signal Processing Society, the 2014 Athanasios Papoulis Award from the European Association for Signal Processing, the 2013 Meritorious Service Award, and the 2012 Technical Achievement Award from the IEEE Signal Processing Society. Also, the 2005 Terman Award from the American Society for Engineering Education, the 2003 Kuwait Prize, and the 1996 IEEE Donald G. Fink Prize. He served as Distinguished Lecturer for the IEEE Signal Processing Society in 2005 and as Editor-in-Chief of the IEEE TRANSACTIONS ON SIGNAL PROCESSING (2003?2005). His articles received several Best Paper Awards from the IEEE Signal Processing Society (2002, 2005, 2012, 2014). He is a Fellow of the American Association for the Advancement of Science (AAAS). He is recognized as a Highly Cited Researcher by Thomson Reuters. He is serving as President-Elect of the IEEE Signal Processing Society. [44] J. F. C. Mota, J. M. F. Xavier, P. M. Q. Aguiar, and M. Puschel, "Distributed basis pursuit," IEEE Trans. Signal Process., vol. 60, no. 4, pp. 1942–1956, Apr. 2012. [45] K. Yuan, Q. Ling, and W. Yin, "On the convergence of decentralized gradient descent," SIAM J. on Optim., vol. 26, no. 3, pp. 1835–1854, 2016. [46] R. H. Koning, H. Neudecker, and T. Wansbeek, "Block Kronecker products and the vecb operator," Linear Algebra and its Applications, vol. 149, pp. 165–184, Apr. 1991. [47] L. Isserlis, "On a formula for the product-moment coefficient of any order of a normal frequency distribution in any number of variables," Biometrika, vol. 12, no. 1/2, pp. 134–139, November 1918. Roula Nassif received the bachelor's degree in Electrical Engineering from the Lebanese University, Lebanon, in 2013. She received the M.S. degrees in Industrial Control and Intelligent Systems for Trans- port from the Lebanese University, Lebanon, and from Compi`egne University of Technology, France, in 2013. She received the Ph.D. degree in 2016, from the University of Cote d'Azur (UCA), France. She is currently a researcher and a teaching assistant at UCA. Her current research interests include adapta- tion and learning over networks. C´edric Richard (S'98–M'01–SM'07) received the Dipl.-Ing. and the M.S. degrees in 1994, and the Ph.D. degree in 1998, from Compi`egne University of Technology, France. He is a Full Professor at the University of Nice Sophia Antipolis, France. He was a junior member of the Institut Universitaire de France in 2010-2015. His current research interests include statisti- cal signal processing and machine learning. Prof. Richard is the author of over 250 papers. He was the General Co-Chair of the IEEE SSP'11 Workshop that was held in Nice, France. He was the Technical Co-Chair of EUSIPCO'15 that was held in Nice, France, and of the IEEE CAMSAP'15 Workshop that was held in Cancun, Mexico. Since 2015, he serves as a Senior Area Editor of the IEEE Transactions on Signal Processing, and as an Associate Editor of the IEEE Transactions on Signal and Information Processing over Networks. He is an Associate Editor of Signal Processing Elsevier since 2009. Prof. Richard is member of the IEEE Machine Learning for Signal Processing Technical Committee, and served as member of the IEEE Signal Processing Theory and Methods Technical Committee in 2009-2014. received Ferrari (SM'91-M'93) Andr´e the Ing´enieur degree from ´Ecole Centrale de Lyon, Lyon, France, in 1988 and the M.Sc. and Ph.D. from the University of Nice Sophia degrees in 1989 and 1992, Antipolis respectively, computer engineering. (UNS), France, electrical He is currently a Professor at UNS. He is a member of the Joseph-Louis Lagrange Laboratory (CNRS, OCA), where his research activity is centered around statistical signal processing and all in and modeling, with a particular interest in applications to astrophysics.
1601.08116	1	1601	2016-01-29T14:11:37	Densifying the sparse cloud SimSaaS: The need of a synergy among agent-directed simulation, SimSaaS and HLA	[ "cs.MA", "cs.DC" ]	Modelling & Simulation (M&S) is broadly used in real scenarios where making physical modifications could be highly expensive. With the so-called Simulation Software-as-a-Service (SimSaaS), researchers could take advantage of the huge amount of resource that cloud computing provides. Even so, studying and analysing a problem through simulation may need several simulation tools, hence raising interoperability issues. Having this in mind, IEEE developed a standard for interoperability among simulators named High Level Architecture (HLA). Moreover, the multi-agent system approach has become recognised as a convenient approach for modelling and simulating complex systems. Despite all the recent works and acceptance of these technologies, there is still a great lack of work regarding synergies among them. This paper shows by means of a literature review this lack of work or, in other words, the sparse Cloud SimSaaS. The literature review and the resulting taxonomy are the main contributions of this paper, as they provide a research agenda illustrating future research opportunities and trends.	cs.MA	cs	Densifying the sparse cloud SimSaaS: The need of a synergy among agent-directed simulation, SimSaaS and HLA Tiago Azevedo, Rosaldo J. F. Rossetti, Jorge G. Barbosa Artificial Intelligence and Computer Science Lab Faculty of Engineering, University of Porto, Portugal Department of Informatics Engineering {tiago.manuel, rossetti, jbarbosa}@fe.up.pt 6 1 0 2 n a J 9 2 ] A M . s c [ 1 v 6 1 1 8 0 . 1 0 6 1 : v i X r a Keywords: Agent-directed simulation, Agent-supported simulation, HLA, High Level Architecture, Cloud, SimSaaS, Simulation Software-as-a-service Abstract: Modelling & Simulation (M&S) is broadly used in real scenarios where making physical modifications could be highly expensive. With the so-called Simulation Software-as-a-Service (SimSaaS), researchers could take advantage of the huge amount of resource that cloud computing provides. Even so, studying and analysing a problem through simulation may need several simulation tools, hence raising interoperability issues. Having this in mind, IEEE developed a standard for interoperability among simulators named High Level Architecture (HLA). Moreover, the multi-agent system approach has become recognised as a convenient approach for modelling and simulating complex systems. Despite all the recent works and acceptance of these technologies, there is still a great lack of work regarding synergies among them. This paper shows by means of a literature review this lack of work or, in other words, the sparse Cloud SimSaaS. The literature review and the resulting taxonomy are the main contributions of this paper, as they provide a research agenda illustrating future research opportunities and trends. 1 INTRODUCTION Modelling & Simulation (M&S) is widely used in real scenarios such as traffic and transportation networks, where making physical modifications could be highly expensive, dependent on political decisions and very disruptive to the environment. Its uses could be deci- sion making and what-if analysis, performance opti- misations, testing and training, making M&S method- ologies a huge need for Universities and companies worldwide. Nowadays, there is a new paradigm called Sim- ulation Software-as-a-Service (SimSaaS) where sim- ulation software is used in the form of services, thanks to the latest evolutions in cloud computing and Software-as-a-Service (SaaS). Instead of having the simulation software installed on their own computers, researchers could take advantage of the huge amount of resource that cloud computing provides. Even so, studying and analysing a problem through simulation may need several simulation tools, with different resolutions and domain perspectives, hence raising interoperability issues that could not be trivial to solve. IEEE has already a standard for interoperabil- ity among simulators named High Level Architecture (HLA). HLA is covered by many works in the litera- ture. Moreover, the multi-agent system approach has become recognised as a convenient approach for mod- elling and simulating complex systems (Moya and Tolk, 2007). Indeed, many researchers have devel- oped work regarding agents. This paper shows that the M&S community did not make a complete jump from the simulations in the local machines to the simulations in the cloud offered in the form of services. To prove such an assertion, we conducted a literature review to make the body of knowledge of the current synergy among agent- directed simulation, SimSaaS and HLA. The literature review was conducted using the methodological and systematic framework proposed by (vom Brocke et al., 2009). It is not important in the context of this paper to explicitly describe all the phases. Yet, some considerations must be made. The databases sources selected were Scopus, Engi- neering Village and ACM. These databases are com- monly known to contain vast work and have been used by many researchers in software engineering. The queries made to the databases sources were based in four main keywords: SimSaaS, Cloud Computing, HLA and agents. It is considered a time frame from 2004 to 2015. The evolution of knowledge and tech- nology in the software engineering field is tremen- dous every year. Thus, a time frame of a decade seems enough. This paper will start to briefly explain some pre- liminary background concepts regarding the agent- oriented paradigm, HLA and cloud for a better under- standing of the scope of this work. After that, the re- sults of the literature review are broadly indicated and a taxonomy of the research work is presented. The literature review and resulting taxonomy are the main contributions of this paper, as they provide a research agenda illustrating future research opportunities and trends. 2 PRELIMINARY BACKGROUND that is, For the sake of clarification in future references, this section will briefly describe what is the agent- directed simulation paradigm, the HLA standard and the cloud paradigm. Yilmaz and Oren (Yilmaz and Oren, 2007) in- dicated that the agent-directed simulation paradigm consists in three main areas: (1) simulation for agents (simulation of agent systems, the simula- tion model is one or more agents), (2) agent-based simulation (model behaviour generation or monitor- ing of this process by using agents) and (3) agent- supported simulation (improving simulation by using agents as support facilities). There are several re- searchers which consider agent simulation and agent- based simulation the same principle as they do not take into account the contribution of agents in model generation. In this work, it is adopted the same per- spective in which the two principles are seen as the same. In order to have a structural basis for interoper- ability among simulators, IEEE developed HLA, a software standard that provides a common techni- cal architecture for distributed M&S. A federate is the name given to every participant of the simula- tion, whereas each one can interact within a federa- tion. Communication between simulators is possible thanks to a Run-Time Infrastructure (RTI). HLA base- line components are: (1) Federate Interface Specifica- tion (IEEE, 2010c); (2) Framework and Rules (IEEE, 2010a); and (3) Object Model Template (OMT) Spec- ification (IEEE, 2010b). The first is a definition of the services that each federate can use for communi- cation. The second is a set of rules that ensure the proper interaction within a federation. The latter is a specification of the format and syntax of the data that is exchanged among federates. Cloud computing is a fresh and on-going recent buzzword where more and more work is being done not only in the industry but also among academics. Nonetheless, there is no general consensus on an un- ambiguous definition (Geelan, 2009). A problem in defining cloud computing is that it overlaps with other domains in distributed systems. Foster et al. (Foster et al., 2008) define the fields of distributed systems according to scale and domain (application-oriented versus service-oriented). Web 2.0 covers the spec- trum of service-oriented applications, opposing to the Supercomputing and Cluster Computing, which have been more focused on traditional local applications. Cloud Computing lies at the large-scale side, being more scalable than Grid Computing. Grid Computing overlaps with all these fields, and because of that it is normal to exist wrong definitions. Despite the overlapping of cloud computing with other domains, it is possible to distinguish it from grid computing. Indeed, there are three aspects that are new in cloud computing (Armbrust et al., 2010): (1) the appearance of infinite computing resources available on demand; (2) the elimination of an up- front commitment by cloud users; and (3) the abil- ity to pay for use of computing resources on a short- term basis as needed. Two new buzzwords emerged from Cloud Computing, trying to extend it even fur- ther: Fog Computing (Bonomi et al., 2012) and Cloud 2.0 (Miluzzo, 2014). Concluding, we adopt the defi- nition provided by The National Institute of Standards and Technology (Mell and Grance, 2011): Cloud computing is a model for enabling ubiquitous, con- venient, on-demand network access to a shared pool of configurable computing resources (e.g., networks, servers, storage, applications, and services) that can be rapidly provisioned and released with minimal management effort or service provider interaction. 3 AN UNEXPLORED CLOUD SIMSAAS This section will show the unexplored Cloud SimSaaS by means of a literature review. It starts by referring SimSaaS and Cloud generically, following the HLA standard and the agent-directed simulation topics. In our research, it was clear that SimSaaS is a very modern topic. The great majority of the papers were published after 2011 and the first ones do not directly use the term SimSaaS, vaguely mentioning simulation and web. Although there is an increasing number of works per year, papers about SimSaaS are not many. Nevertheless, they are wide concerning the domains of application. It is possible to see works in the biomedical domain (Sawicki et al., 2012), in crowd and pedestrian field (Wang and Wainer, 2015) as well as works regarding ontology learning (Wang and Wainer, 2014), traffic and transportation (Harri et al., 2010), scheduling parallel discrete event simu- lation jobs (Liu et al., 2012a) and a cloud simulation in manufacturing (Taylor et al., 2014a), just to cite some. Beyond these specific domain works, there are also some generic ones concerning frameworks for development, for example (Tsai et al., 2011) (Guo et al., 2011). Moreover, Cayirci refers to SimSaaS using the term Modelling and Simulation-as-a- service (MSaaS). He clarifies MSaaS, including top threats (Cayirci, 2013c). He also talks about the no- tions and relations of accountability, risk and trust modelling (Cayirci, 2013a), as well as MSaaS com- position in multi-datacenter or multi-cloud scenar- ios (Cayirci, 2013b). Cayirci is not the only one using the term MSaaS: (Siegfried et al., 2014) illustrate po- tential benefits that may be achieved by MSaaS and challenges that remain to be solved. Since cloud simulation started to be studied, an overall picture took some time to arrive. So, in 2012 Liu et al. (Liu et al., 2012b) proposed a general ar- chitecture of cloud simulation in the form of a SaaS type cloud. Figure 1 summarises the main blocks of this architecture. The simulation services which are offered through a website to very different users, are divided into three self-explanatory groups: Modelling as a Service, Execution as a Service and Analysis as a Service. All these services are possible due to the physical and virtual resources in the bottom, as well as the so-called Cloud Operating System which man- ages and connects the baseline infrastructure to the top. Figure 1: The main blocks of the Cloud Simulation general architecture (Liu et al., 2012b) Although HLA has already been used in a variety of works like agent-based simulations, there is almost no clear references about interoperability among sim- ulators when talking about SimSaaS. A first approach in extending HLA to support grid-wide distributed simulation dates back to 2005 (Xie et al., 2005). A very relevant work from 2012 discusses how HLA can be integrated with Service Oriented Ar- chitecture (SOA) in the context of a smart building project (Dragoicea et al., 2012). The Simulation En- gine Service is exposed by RESTful services. In- side this Engine, there is also a RESTful API that ex- poses access to the RTI's federation management and deals with the creation, initialisation, deletion, start- ing, stopping, and execution of simulations. Still in this context, the authors refer another paper (Wang et al., 2008), where a comparison between HLA and SOA concluded that: • HLA has good interoperability, synchronization and effective and uniform information exchange mechanism between the communicating compo- nents (federates), but lacks several features of web services, such as: the integration of heterogeneous resources, web-wide accessibility across firewall boundaries; • SOA benefits from loose coupling, component reuse and scalability but lacks a uniform data ex- change format and time synchronization mecha- nisms; JAVA Agent DEvelopment Framework (JADE) is a common software framework that simplifies the im- plementation of multi-agent systems. Web Services Integration Gateway (WSIG) is an add-on for JADE • The combination of HLA and SOA can extend the capabilities of the two technologies and thus en- able integrated simulated and real services. Like HLA, simulation paradigm is used in a huge variety of fields but, when it comes to SimSaaS, few examples exists. Nevertheless, already in 2006, it was mentioned the importance of agents on simulation by exploring the relationship of software agents to simulation and games (Yilmaz et al., 2006). agent-directed the Some authors (Tolk and Diallo, 2010)(Tolk et al., 2011) defend that most of the current simulation in- teroperability standards are inadequate as they just support the federation by focusing on information ex- change without providing the necessary introspective. HLA provides more flexibility as it only standardises how to structure the data, not the exchanging of in- formation. However, the focus remains on the infor- mation exchanged within a system. Consequently, a formal approach to simulation interoperability using agent-supported simulation tries to solve this prob- lem. which performs two-way translations between service requests and responses and JADE agent requests and responses. Thanks to this, it was possible to design a service-oriented simulation software framework as part of a broader approach towards generating im- proved levels of actionable views of situation aware- ness (Shao and McGraw, 2009). Shao and McGraw referred that the great benefit of using JADE as the underlying agent development framework is that JADE agent entities can invoke web service functionality hosted outside the JADE run-time environment using normal JADE agent pro- tocols, and that external entities can invoke JADE agent functionality from outside the JADE environ- ment using normal web service protocols. Although the framework is very relevant, the applications were not in the cloud nor in a grid. A truly implementation of agents in the cloud showed that agent-based M&S can benefit a lot from cloud computing, making it easier to have more accu- rate and faster results, as well as timely experimenta- tion and optimisation (Taylor et al., 2014b). Even so, agent-based M&S in the cloud may be highly com- plex due to the very different clouds, cloud middle- wares and service approaches. Federated simulation environments have some limitations in supporting dynamic model and simulat- ing updating, as it was pointed in 2004 and 2006 (Yil- maz et al., 2006). An example is HLA federation development, as it requires complete specification of object models and information exchanges before the simulation begins. It was also argued that there is a fundamental roadblock because of a lack of machine processable formal annotations describing behaviour, assumptions and obligations of federates. 4 TAXONOMY OF THE RESEARCH WORK SimSaaS is a trendy term which has been growing considerably in recent years. Thus, it is the ideal time to take advantage of this hype. However, there are some concerns: there is a lack of automation and in- tegration of tools in M&S (Wang and Wainer, 2015), and research dissemination methods suffer as they do not allow publishing simulation code and scripts along with the published paper (Sliman et al., 2013). Herewith, HLA is another term referenced a lot in the literature since the first complete version (HLA 1.3) was published in 1998, but once again, when it comes to SimSaaS, almost nothing focuses on this and there is few work regarding extension of HLA to al- low simulation services in general and in the cloud. Indeed, HLA solely has some disadvantages (Yilmaz et al., 2006)(Tolk and Diallo, 2010). In a 2014 panel about the future of research in M&S (Yilmaz et al., 2014) it is referred as a future re- search topic the distribution of SimSaaS in the cloud. So, once more SimSaaS is still mentioned as an un- explored area, now specifically in distributed simula- tion. Although a truly implementation of agents in the cloud showed that agent-based M&S can benefit from cloud computing (Taylor et al., 2014b), there is a lack of work putting together agents and cloud in order to support SimSaaS. Summing up all the discoveries of the described literature review, it is possible to see a lot of gaps in the literature concerning SimSaaS, SimSaaS in spe- cific domains of application, SimSaaS in the cloud, HLA in the cloud, solutions to HLA restrictions, agents to support SimSaaS and agents in the cloud. As the metaphor in the title of this paper tries to ad- dress, SimSaaS in the cloud is currently too sparse since it has so many gaps in research. It is necessary to make it less sparse (densifying) in order to augment the scientific and technological knowledge among re- searchers in the field. Densifying the sparse cloud SimSaaS is not just putting together cloud and SimSaaS, but also the syn- ergies among them and agent-directed simulation and HLA, which could bring so many advantages. Fog Computing and Cloud 2.0, which were previously mentioned, could also help in this evolutionary pro- cess. Concluding, making these synergies a reality will be the front research opportunities for the next years. A taxonomy of the research work could make the gaps identified and the research agenda for the next years more clear. Figure 2 illustrates the taxonomy in the form of a Venn Diagram. Every work is about M&S, more precisely SimSaaS. So, there are works that simply mention SimSaaS. Then, inside SimSaaS topic, research can focus on Cloud, HLA or Agent- directed simulation. In the particular case of Agent- directed simulation, there is a subset regarding Agent- supported simulation. As some works can address more than just one term, the representation in the form of a Venn Diagram was chosen to illustrate these pos- sible synergies. The presented taxonomy could be used as a con- ceptual framework for future developments. Re- searchers should look to these opportunities in the sci- entific community to orient their work, taking advan- tage of the benefits that this trends can give to their daily investigations. M&S SimSaaS HLA Cloud AdSim AsSim (AdSim: Agent-directed Simulation, AsSim: Agent-supported Simulation) Figure 2: Diagram representing the taxonomy of the re- search work 5 CONCLUSIONS This paper started to briefly expose important con- cepts for a better understanding of its contents. After that, a literature review is presented, focusing in four distinct yet related topics: SimSaaS, Cloud Comput- ing, HLA and Agents. Finally, a taxonomy of the re- search work is presented in the form of a Venn Dia- gram for a clear visualisation about which topics can (and should) have synergies among them. This tax- onomy could be used as a conceptual framework for future developments. With the front research opportunities for the next years and current research work identified, we hope this paper can leverage the scientific activity in the field, with researchers actually finding it useful to make the jump to the cloud. That jump will bring advantages not only to each researcher in particular, but also to the overall simulation scientific commu- nity seeking for more knowledge. ACKNOWLEDGEMENT This work has been partially supported by MIEIC, Faculty of Engineering, University of Porto. REFERENCES Armbrust, M., Fox, A., Griffith, R., Joseph, A. D., Katz, R., Konwinski, A., Lee, G., Patterson, D., Rabkin, A., Stoica, I., et al. (2010). A view of cloud computing. Communications of the ACM, 53(4):50 -- 58. Proceedings of the first edition of the MCC workshop on Mobile cloud computing, pages 13 -- 16. ACM. Cayirci, E. (2013a). A joint trust and risk model for MSaaS mashups. pages 1347 -- 1358, Washington, DC. Cayirci, E. (2013b). Configuration schemes for modeling and simulation as a service federation. Simulation, 89(11):1388 -- 1399. Cayirci, E. (2013c). Modeling and simulation as a cloud service: A survey. pages 389 -- 400, Washington, DC. Dragoicea, M., Bucur, L., Tsai, W.-T., and Sarjoughian, H. (2012). Integrating HLA and Service-Oriented Archi- In Proceedings tecture in a Simulation Framework. of the 2012 12th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (Ccgrid 2012), CCGRID '12, pages 861 -- 866, Washington, DC, USA. IEEE Computer Society. Foster, I., Zhao, Y., Raicu, I., and Lu, S. (2008). Cloud computing and grid computing 360-degree compared. In Grid Computing Environments Workshop, 2008. GCE'08, pages 1 -- 10. IEEE. Geelan, J. (2009). Twenty-one experts define cloud com- puting. Cloud Computing Journal, 4:1 -- 5. Guo, S., Bai, F., and Hu, X. (2011). Simulation software as a service and Service-Oriented simulation experiment. pages 113 -- 116, Las Vegas, NV. Harri, J., Killat, M., Tielert, T., Mittag, J., and Hartenstein, H. b. (2010). DEMO: Simulation-as-a-service for ITS applications. Taipei. IEEE (2010a). IEEE Standard for Modeling and Simulation (M &S) High Level Architecture (HLA) -- Framework and Rules. IEEE Std 1516-2010 (Revision of IEEE Std 1516-2000), pages 1 -- 38. IEEE (2010b). IEEE Standard for Modeling and Simula- tion (M &S) High Level Architecture (HLA) -- Ob- IEEE ject Model Template (OMT) Specification. Std 1516.2-2010 (Revision of IEEE Std 1516.2-2000), pages 1 -- 110. IEEE (2010c). IEEE Standard for Modeling and Simulation (M&S) High Level Architecture (HLA) -- Federate In- terface Specification. IEEE Std 1516.1-2010 (Revision of IEEE Std 1516.1-2000), pages 1 -- 378. Liu, X., Qiu, X., Chen, B., He, Q., and Huang, K. (2012a). Scheduling parallel discrete event simulation jobs in the cloud. volume 2012, London. Liu, X., Qiu, X., Chen, B., and Huang, K. (2012b). Cloud-based simulation: The state-of-the-art com- puter simulation paradigm. In Proceedings of the 2012 ACM/IEEE/SCS 26th Workshop on Principles of Ad- vanced and Distributed Simulation, PADS '12, pages 71 -- 74, Zhangjiajie. IEEE Computer Society. Mell, P. and Grance, T. (2011). The NIST definition of cloud computing. Miluzzo, E. (2014). I'm Cloud 2.0, and I'm Not Just a Data Center. Internet Computing, IEEE, 18(3):73 -- 77. Bonomi, F., Milito, R., Zhu, J., and Addepalli, S. (2012). Fog computing and its role in the internet of things. In Moya, L. J. and Tolk, A. (2007). Towards a taxonomy of In Proceedings of agents and multi-agent systems. Theory of Modeling & Simulation - DEVS Integrative, DEVS '14, pages 25:1 -- 25:8, San Diego, CA, USA. Society for Computer Simulation International. Wang, S. and Wainer, G. (2015). A Simulation As a Service Methodology with Application for Crowd Modeling, Simulation and Visualization. Simulation, 91(1):71 -- 95. Wang, W., Yu, W., Li, Q., Wang, W., and Liu, X. (2008). In Pro- Service-oriented High Level Architecture. ceedings of the 2008 Summer Computer Simulation Conference, SCSC '08, pages 16:1 -- 16:12, Vista, CA. Society for Modeling & Simulation International. Xie, Y., Teo, Y. M., Cai, W., and Turner, S. J. (2005). Ser- vicing Provisioning for HLA-Based Distributed Sim- ulation on the Grid. In Proceedings of the 19th Work- shop on Principles of Advanced and Distributed Sim- ulation, PADS '05, pages 282 -- 291, Washington, DC, USA. IEEE Computer Society. Yilmaz, L., Oren, T., and Aghaee, N.-G. (2006). Intelligent agents, simulation, and gaming. Simulation & Gam- ing, 37(3):339 -- 349. Yilmaz, L. and Oren, T. I. (2007). Agent-directed sim- In Proceedings of ulation systems engineering. the 2007 Summer Computer Simulation Conference, pages 897 -- 904. Society for Computer Simulation In- ternational. Yilmaz, L., Taylor, S. J. E., Fujimoto, R., and Darema, F. (2014). Panel: The Future of Research in Modeling & Simulation. In Proceedings of the 2014 Winter Simu- lation Conference, WSC '14, pages 2797 -- 2811, Pis- cataway, NJ, USA. IEEE Press. the 2007 spring simulation multiconference-Volume 2, pages 11 -- 18. Society for Computer Simulation Inter- national. Sawicki, B., Chaber, B., Starzy´nski, J., and Szmuro, R. (2012). Internet application concept to trivialize EMF biomedical computing. COMPEL - The International Journal for Computation and Mathematics in Electri- cal and Electronic Engineering, 31(4):1190 -- 1197. Shao, G. and McGraw, R. (2009). Service-oriented Sim- ulations for Enhancing Situation Awareness. In Pro- ceedings of the 2009 Spring Simulation Multiconfer- ence, SpringSim '09, pages 48:1 -- 48:7, San Diego, CA, USA. Society for Computer Simulation Interna- tional. Siegfried, R., Van Den Berg, T., Cramp, A., and Huiskamp, W. (2014). M&S as a service: Expectations and chal- lenges. pages 248 -- 257. SISO - Simulation Interoper- ability Standards Organization. Sliman, L., Charroux, B., and Stroppa, Y. (2013). A new collaborative and cloud based simulation as a service platform: Towards a multidisciplinary research simu- lation support. pages 611 -- 616, Cambridge. Taylor, S., Kiss, T., Terstyanszky, G., Kacsuk, P., and Fan- tini, N. (2014a). Cloud computing for simulation in manufacturing and engineering: Introducing the CloudSME simulation platform. volume 46, pages 89 -- 96, Tampa, FL. The Society for Modeling and Simulation International. Taylor, S. J. E., Anagnostou, A., Kiss, T., Terstyanszky, G., Kacsuk, P., and Fantini, N. (2014b). A Tutorial on Cloud Computing for Agent-based Modeling & Sim- ulation with Repast. In Proceedings of the 2014 Win- ter Simulation Conference, WSC '14, pages 192 -- 206, Piscataway, NJ, USA. IEEE Press. Tolk, A. and Diallo, S. Y. (2010). Using a Formal Approach to Simulation Interoperability to Specify Languages for Ambassador Agents. In Proceedings of the Win- ter Simulation Conference, WSC '10, pages 359 -- 370. Winter Simulation Conference. Tolk, A., Diallo, S. Y., Padilla, J. J., and Herencia-Zapana, H. (2011). Model Theoretic Implications for Agent Languages in Support of Interoperability and Com- In Proceedings of the Winter Simulation posability. Conference, WSC '11, pages 309 -- 320. Winter Simu- lation Conference. Tsai, W.-T., Li, W., Sarjoughian, H., and Shao, Q. (2011). SimSaaS: Simulation Software-as-a-service. In Pro- ceedings of the 44th Annual Simulation Symposium, ANSS '11, pages 77 -- 86, San Diego, CA, USA. Soci- ety for Computer Simulation International. vom Brocke, J., Simons, A., Niehaves, B., Riemer, K., Plat- tfaut, R., and Cleven, A. (2009). Reconstructing the Giant: On the Importance of Rigour in Documenting the Literature Search Process. In Proceedings of the 17th European Conference on Information Systems, Verona, Italy. Wang, S. and Wainer, G. (2014). Semantic Mashups for Simulation As a Service with Tag Mining and Ontol- In Proceedings of the Symposium on ogy Learning.
1906.09029	1	1906	2019-06-21T09:44:29	Topology Inference over Networks with Nonlinear Coupling	[ "cs.MA", "cs.IT", "cs.IT" ]	This work examines the problem of topology inference over discrete-time nonlinear stochastic networked dynamical systems. The goal is to recover the underlying digraph linking the network agents, from observations of their state-evolution. The dynamical law governing the state-evolution of the interacting agents might be nonlinear, i.e., the next state of an agent can depend nonlinearly on its current state and on the states of its immediate neighbors. We establish sufficient conditions that allow consistent graph learning over a special class of networked systems, namely, logistic-type dynamical systems.	cs.MA	cs	Topology Inference over Networks with Nonlinear Coupling Augusto Santos(cid:63) , Vincenzo Matta†, and Ali H. Sayed(cid:63) 9 1 0 2 n u J 1 2 ] A M . s c [ 1 v 9 2 0 9 0 . 6 0 9 1 : v i X r a Abstract This work examines the problem of topology inference over discrete-time nonlinear stochastic net- worked dynamical systems. The goal is to recover the underlying digraph linking the network agents, from observations of their state-evolution. The dynamical law governing the state-evolution of the interacting agents might be nonlinear, i.e., the next state of an agent can depend nonlinearly on its current state and on the states of its immediate neighbors. We establish sufficient conditions that allow consistent graph learning over a special class of networked systems, namely, logistic-type dynamical systems. Index Terms Topology inference, causal inference, graph learning, structure estimation, nonlinear stochastic net- worked dynamical systems. I. INTRODUCTION T he evolution of a networked dynamical system is determined by the local interactions among its neighboring agents. Graph or structure learning refers to the problem of estimating the underlying graph from observations collected at the agents. This is a challenging inverse problem, which would allow us to understand more fully the evolution of systems arising across several application domains including, e.g., epidemics [1], social networks [2], and brain activity [3]. Figure 1 provides an illustration of the topology inference problem with reference to the last application. Very often, the dynamics of real-world phenomena is governed by highly nonlinear and possibly random forms of interaction. For this reason, one useful class of graph learning problems concerns (cid:63) A. Santos (email: [email protected]) and A. H. Sayed (email: [email protected]) are with the ´Ecole Polytechnique F´ed´erale de Lausanne (EPFL), CH-1015 Lausanne, Switzerland. The work of A. H. Sayed was also supported in part by US NSF grant CCF-1524250 and Swiss NSF (SNSF) grant 205121 184999. † V. Matta is with DIEM, University of Salerno, via Giovanni Paolo II, I-84084, Fisciano (SA), Italy (email: [email protected]). June 24, 2019 DRAFT 2 Fig. 1. Illustration of the topology inference paradigm: data are collected from a set of nodes in a networked dynamical system -- e.g., signals measured at regions of interest in the brain -- and the underlying connectivity among the nodes in the system -- or functional connectivity among the regions in the brain -- is then inferred. the case of nonlinear stochastic dynamical systems. In this article, we focus on the following class of logistic-type systems, which can model several forms of nonlinear coupling between locally-interacting (i.e., neighboring) agents: yi,n+1 = σi gi(yi,n) N(cid:88) j=1 aijhj(yj,n) + xi,n+1  (1) In the above formulation, we use an index n = 0, 1, . . . to denote the n-th time epoch, and an index i = 1, 2, . . . N to denote the i-th agent of the network. The relevant quantities in (1) are as follows. -- The random variable yi,n denotes the state of agent i at time n. -- The random variable xi,n represents an input source of randomness or noise, affecting agent i at time n. -- A local nonlinear coupling between agents i and j at time n is determined by the product gi(yi,n)hj(yj,n). -- We will be dealing with weighted directed graphs, where: i) the graph accounts for the topology linking pairs of nodes; ii) the interaction between pairs of nodes can be directional, e.g., a coupling effect can exist from i to j, but not from j to i; and iii) the graph weights are encoded in a combination or interaction matrix A, whose (i, j)-th entry, aij, quantifies the strength of interaction in the directional coupling from i to j. We observe from (1) that, if aij = 0, then information about the state of agent j June 24, 2019 DRAFT 3 does not flow to agent i. Otherwise, if aij (cid:54)= 0, then the state of agent j directly impacts the state of agent i, provided that the coupling functions g and h do not vanish. -- σi : R → R is a nonlinear invertible function that plays the role of an activation function (such as a sigmoidoscopy function). Depending on its particular shape, the function σi can emphasize or de- emphasize its input values. At two extremes, a constant σi kills the dynamics, whereas a linear σi is basically transparent to its input. Formulations of the type shown in (1) are often used to model interactions over nonlinear oscillators [4], [5], or population dynamics and epidemics over networks [6] -- [12]. For example, in the context of epidemics over networks, the state yi,n can model the likelihood of a node i being infected at time n at the individual/microscopical level [13]; at the aggregate/macroscopical level, the state can also represent the fraction of infected individuals within community i in an epidemics across communities, as can be established through a thermodynamic or fluid-limit analysis -- see [14] -- [17]. Likewise, in a general SIR (Susceptible Infected Recovered) formulation, the functions gi and hj in (1) can represent the so-called incidence rate of the infection, whereas the function σi is simply chosen as the identity function [11], [12]. Within the aforementioned frameworks, a natural question is whether the underlying network of interactions can be inferred given the evolution of the infection across nodes or communities. II. MAIN RESULTS Our main goal is to answer the following identifiability question: Is it possible to learn the network graph over this class of models? We will see that the answer is in the affirmative. To answer, we start by adopting a classic nonlinear regression approach to construct two nonlinear functions (see (20) and (21) further ahead for details) that depend on the node measurements. More specifically, stacking the observations yi,n of all agents at time n into the N × 1 vector yn, we will introduce a zero-lag function, F0(yn), which depends only on the current observation vector yn, and a one- lag function, F1(yn, yn+1), which depends on the interaction between observation vectors corresponding to adjacent time epochs. The expected values of these functions will play an important role: F0(n) (cid:44) E[F0(yn)], F1(n) (cid:44) E[F1(yn, yn+1)]. (2) Under suitable conditions on the various nonlinearities (σ, g, h), the combination matrix A admits the following closed-form representation: A = F1(n) [F0(n)]−1 [Our nonlinear case] (3) DRAFT June 24, 2019 4 (6) (7) n−1(cid:88) k=0 1 n with: F0(yk) a.s.−→ F0, 1 n F1(yk, yk+1) a.s.−→ F1 n−1(cid:88) k=0 A = F1 F−1 0 We shall call this relationship generalized Granger estimator for the following reason. In the special case where σi(y) = y, gi(y) ≡ 1, and hj(y) = y, the model in (1) degenerates into the classical linear model (a.k.a. first-order vector autoregressive model): yi,n+1 = aijyj,n + xi,n+1. (4) In this particular case, a well-known representation for A is given by the one-step linear predictor, sometimes called Granger estimator in the context of causal analysis: j=1 N(cid:88) [Classic linear case] (5) A = R1(n)[R0(n)]−1 n ] and R1(n) = E[yn+1y(cid:62) where R0(n) = E[yny(cid:62) n ] are the zero and one-lag correlation matrices of the samples at time n. The functions F0(n) and F1(n) depend on the nonlinear functions that characterize the system in (1), in a way that highlights the role that these nonlinearities play on topology estimation, as we will show in Sec. V The relationship in (3) actually suggests a strategy to estimate the topology. However, we see from (3) that A depends upon expected values, which in turn depend on the knowledge of the distribution of the data. This leads to a circular argument since the distribution would require knowledge of the matrix A. Accordingly, since only a sample path (i.e., one realization of the process) is observed, in order to show that (3) can be useful, we will prove that the aforementioned expected values can be consistently learned from the samples. To this aim, it will be critical to establish that the considered dynamical system a.s.−→ denotes almost-sure convergence as n → ∞): possesses the following ergodic property (the symbol The main contributions of this work are as follows. We first answer an identifiability question, namely, we establish whether graph learning is possible over the considered class of logistic nonlinear dynamical systems. To this aim, we exploit a closed-form relationship existing between the interaction matrix A and a pair of functionals of the samples that arise from a solution of a nonlinear regression problem. Through this closed-form representation, we characterize the various nonlinearities and attributes that determine the dynamics in (1), in order to ascertain under which conditions the graph can be learned consistently. This characterization is obtained through a set of transparent sufficient conditions, which allows relating the graph identifiability to the physical evolution of the dynamics. We note that the method and theory are consistent for arbitrary topologies (whether directed or undirected, and dense or sparse). June 24, 2019 DRAFT 5 Second, to show that the graph is consistently learned as the number of samples grows, we will prove that the consistency result in (6) holds true, through a limiting characterization that builds on powerful results available for Markov chains in the general state space. Finally, in order to enlarge the range of application, we explore numerically scenarios where some of the conditions used to carry out the technical analysis are relaxed. For example, while the aforementioned theoretical results apply in the regime of full observability (i.e., when all nodes of the network are accessible), in Sec. VIII we will show some examples that illustrate how the proposed method can work also in the case of partial observability, along the lines of what has been developed in [38] -- [43] for linear models. III. RELATED WORK The problem of graph learning arises in several disciplines, giving rise to different terminologies and relevant models, including structural equations, structural dynamical systems, graphical or vector autoregressive models. In all these models, the evolution of the observables is determined by some form of local interaction between neighboring nodes, and this structure of interaction is encoded in an underlying network graph. In the following, we provide a list of works that are relevant to the present treatment. A. Learning Graphical Models Graphical models conform to an important and well-studied class of systems in the framework of topology inference. In a graphical model, the state of each agent is represented by the realization of a random variable, and a joint distribution among these variables encodes the dependency among the agents, and, hence, encodes the underlying topology. Then, inference about the topology is carried out by assuming that independent samples from the joint distribution can be collected. There are several works on topology inference for specific classes of graphical models, including, e.g., Ising models [28], [30] and Gaussian graphical models [29]. These works deal with topology inference under the assumption that measurement from all network nodes can be gathered. Another relevant framework, especially over large networks, is that of partial observations. This paradigm corresponds to a challenging inverse problem, where one tries to figure out the graph connecting the observed nodes, despite the latent influence of the unobserved ones. With reference to the partial observations setting for graphical models, in [31] a technique to learn the topology over large-girth graphs (e.g., the bipartite Ramanujan graphs or the random Cayley graphs) is proposed. In [32], a consistent graph-learning algorithm is proposed under the assumption that the adjacency graph matrix is sparse, and that the error matrix June 24, 2019 DRAFT 6 associated to the latent-variables is low-rank. In [33], an approach based on influence maximization is adopted to establish when the graph learning problem is feasible over the class of restricted Boltzmann machines. However, the independence among samples is one fundamental assumption in the framework of graphical models, i.e., it is often assumed that the previous system state does not affect the next state. For this reason, graphical models do not natively match the dynamical system framework that we need here. Likewise, the shape of proper graph estimators can differ substantially from those suited to a dynamical system. For example, while in a Gaussian graphical model, the inverse of the correlation matrix (a.k.a. precision matrix) contains full information on the graph topology, the Granger estimator R1R−1 optimal solution for a first-order diffusion model must rely also on the dependency between subsequent samples as encoded in the one-lag correlation matrix R1. that is the 0 B. Graph Learning with Linear Dynamics Many works on graph learning or causal relationship identification focus on linear models, such as diffusion or Vector AutoRegressive (VAR) models [18]. These linear systems are of practical interest since they arise in several applications, for example, they are used in economics [53]; they may represent well the linearized dynamics of more general nonlinear systems [54]; they can describe the dynamics of biological systems [19] and the operation of distributed algorithms such consensus or diffusion [20] -- [23]. The problem of topology inference over linear systems has been actively studied in the past several years. A great emphasis lies on exploiting the natural regression formulation that these linear systems exhibit and on reinforcing priors on the network structure (e.g., sparsity, smoothness) when available. For example, recent techniques include: spectral-domain techniques based on optimization with sparsity constraints [24]; graph signal processing techniques applied to causal graph processes [25]; directed information graphs [27] to infer causal dynamics; and approaches based on Wiener filtering to infer the topology [26]. There are also works addressing topology inference over linear systems under partial observations. There are results established with reference to particular graphs, such as polytrees [34], [35], as well as results for more general graphs [36], [37]. For the case of large-scale graphs, asymptotic statistical approaches are exploited, where the conditions on the graph are encoded into average summary indicators, such as the connection probability between any two nodes. The works [38] -- [43] pursue this approach to show that the graph of the observed component can be faithfully reconstructed as the network dimension scales to infinity, under different regimes of connectivity and/or graph models. June 24, 2019 DRAFT 7 C. Graph Learning with Nonlinear Dynamics One common approach to tackle the nonlinear problem is to perform some form of linearization of the system. In some cases, by a proper change of coordinates, certain dynamics can be represented linearly. For example, linearization can be obtained by a variational characterization (under a small- noise assumption) [44]; by suitable augmentation of observable space dimension [45]; by appropriate exploitation of the vector-field Jacobian, in conjunction with a compressed sensing method to mitigate the course of dimensionality and the computational complexity [46]. However, linear or linearized models, while useful under some favorable (e.g., small-noise or small- deviations) conditions, in other cases offer only a convenient approximation of the real dynamics, failing to capture some important aspects thereof. One relevant class of genuinely nonlinear models is given by nonlinear vector autoregressive mod- els [47], [48]. In [47], [48], the nonlinear interaction is modeled through a sum of nonlinear univariate functions of the node variables, belonging to a certain reproducing kernel Hilbert space. Advanced methods are developed to learn the topology, reinforcing prior information in suitably formulated re- gression problems in order to boost sample-complexity performance. These methods are flexible enough to incorporate different regularization constraints (such as, e.g., smoothness or sparsity) as well as to cope with the presence of unknown nonlinearities. While fairly general, the class of dynamical systems treated in these references does not encompass the class of logistic systems. Another relevant class of nonlinear systems (more closely related to our approach) focuses on proper modeling of the nonlinear coupling between pairs of node variables. These models arise across many domains, especially in fundamental physics applications, with one notable model being the Kuramoto model. In [4], a general class of continuous-time nonlinear dynamical systems of Kuramoto-type is considered to model oscillators. The vector field of the dynamical law is approximated in a certain complete orthogonal basis of a Hilbert space (e.g., Legendre/Chebyshev polynomials or Fourier series). The coefficients of this expansion encompass information about the topology. Linear regression (i.e., linear in the entries of the coefficients) is used to extract the coefficients and hence the topology. Recent efforts aim at finding proper measures of influence or causality among the agents. The un- derlying structure of the directed network is then generally obtained under appropriate thresholding of the connectivity-measures. For example, functional dependency graphs are introduced in [49] for a fairly broad class of dynamical systems. In [50], a "causal information" measure is proposed to estimate the underlying network of causal relationships. In particular, a Kullback-Leibler based measure is devised for a neural network model. June 24, 2019 DRAFT 8 Along the line of the present work, other works have been devoted to identify the underlying network structure of dynamical systems that model natural phenomena such as oscillatory systems or spread of diseases. For example, in [51] the object of inference is the phylogenetic tree that accounts also for evolutionary elements concerning the disease spread under study and a Bayesian method is proposed. In [8], a log-MLE estimator is devised to infer the underlying structure. The dynamical model assumed is of SIR (susceptible-infected-recovered) type, driven by a Poisson process. The stochastic dynamical model yields a particular distribution parametrized by (among other parameters) the network structure. In [5] (for Kuramoto type of models) an estimator based on a heuristic influence function that maps the relative phase difference between nodes' outputs into an estimate of their pairwise connection is proposed. Notation. We use boldface letters to denote random variables, and normal font letters for their realizations. Matrices are denoted by capital letters, and vectors by small letters. Sets and events are denoted by calligraphic letters. We denote by P[A] the probability of event A. For a random variable y, the notation E[y] denotes the expected value of y. When we say that E[y] exists and is well-defined, we mean that E[y] < ∞. The symbol (cid:12) denotes the Hadamard product. The symbol Ni stands for the set of nodes that point to node i in a directed-graph. For a vector v ∈ RN , we denote by (cid:107)v(cid:107) a generic vector norm. When dealing with an N × N matrix M, the symbol (cid:107)M(cid:107) will denote the matrix norm induced by the particular vector norm (cid:107) · (cid:107), which is defined as: (cid:107)M(cid:107) = sup v(cid:54)=0 (cid:107)M v(cid:107) (cid:107)v(cid:107) . (8) We will be dealing with random variables defined at network nodes and evolving over time. The notation yi,n will generally denote a (random) variable defined at time n = 0, 1, . . ., and at node i ∈ {1, 2, . . . , N}. With the notation yn we shall denote the vector that collects the variables of all nodes at time n, namely, yn = [y1,n, y2,n, . . . , yN,n](cid:62). IV. NONLINEAR MODEL (9) In a networked dynamical system, the state of each agent evolves over time as a result of their peer-to- peer interactions. In particular, in the context of discrete-time systems, the state of an agent i at time n+1 depends upon its own current state at time n and also on the state of its immediate neighbors at time n. The underlying network defining the neighborhood plays critical role in the long-term properties of such dynamical systems. In its most general form, a discrete-time first-order time-homogeneous continuous June 24, 2019 DRAFT state-space nonlinear stochastic system can be described by the following law (refer to Proposition 7.6 9 in [52]) or in vector form yi,n+1 = Ti (yn, xn+1) , (10) yn+1 = T (yn, xn+1) , (11) : RN → R, is the vector-field (at node i) representing how the local interactions affect the where Ti evolution of node i; {xn} is an i.i.d random process modeling exogenous input/perturbations across nodes over time n. The process is time-homogeneous because the law T depends only on the values of the states and of the noise, and does not change over time. The process is with continuous state space because the range of output values can in general vary in RN . One characterizing property of such a networked dynamical system is its locality: the state yi,n+1 of agent i at time n + 1, only depends on its own state yi,n and the states of the agents within its neighborhood {yj,n}j∈Ni at time n. In other words, yi,n+1 is independent of the state of all nodes outside its neighborhood, given the state of the nodes in its neighborhood. This property is referred to as a local Markov property and it is naturally induced by the characterizing vector field, Ti, which must be sensitive only to the entries associated with the node i and its neighbors in Ni, formally, for any vectors y ∈ RN and x ∈ RN , the function Ti(y, x) depends only on yi,{yj}j∈Ni and x. The underlying network topology critically determines the evolution of the vector field via this local Markov property. In this paper, we are interested in the inverse problem of inferring the underlying network characterizing the local Markov property of T given the samples {yn}. It is however hard to devise a universal scheme to extract consistently information about the underlying network of interactions over such a general class of dynamical systems given by (11). It is obvious that not all models can allow consistent graph learning. For this reason, in this work we focus on a subclass of discrete-time continuous-state systems with the vector field defined by (1), which can be compactly represented in vector form as1: yn+1 = σ (g(yn) (cid:12) Ah(yn) + xn+1) having defined the vector-valued functions of vector argument y ∈ RN : σ(y) (cid:44) [σ1(y1), σ2(y2), . . . , σN (yN )](cid:62), g(y) (cid:44) [g1(y1), g2(y2), . . . , gN (yN )](cid:62), h(y) (cid:44) [h1(y1), h2(y2), . . . , hN (yN )](cid:62). (12) (13) (14) (15) 1In order to avoid confusion, we note explicitly that the notation g(yn)(cid:12) Ah(yn) used in (12) denotes the Hadamard product between g(yn) and the vector Ah(yn). June 24, 2019 DRAFT Throughout this work, we assume that {xn} are independent and identically distributed (i.i.d.) random vectors, with zero-mean and finite second moments, and independent of the initial condition y0. Finally, it is worth noting that, since σ is an invertible mapping, by setting zn = σ−1(yn), the system 10 in (12) can be also represented as: where we have introduced the composition of functions: zn+1 =(cid:101)g(zn) (cid:12) A(cid:101)h(zn) + xn+1, (cid:101)g(y) = g(σ(y)), (cid:101)h(y) = h(σ(y)). (16) (17) The type of model in (16) is commonly referred to as additive noise model, and is extensively studied, e.g., in the literature of stochastic dynamical systems [56] -- [58]. Throughout our treatment, we will generally work in terms of the original untransformed model in (12), since working in terms of the composed functions (17) might obfuscate the role of the different nonlinear functions. On the other hand, the representation in (16) will be useful to prove the technical results in Appendix C, where we will make appeal to existing results pertaining to the additive noise model [56] -- [58]. V. GENERALIZED GRANGER ESTIMATOR The proper scheme to process the samples with the goal of estimating either the graph-structure or more generally the interaction matrix relies critically on the nature of the samples. Our goal is to establish the proper structure retrieval scheme for the class of networked dynamical systems (1). We highlight that for the most general class of networked dynamical systems (11), one should not expect to have a closed-form expression for the underlying structure in terms of a functional of the samples -- in general, the inference is carried out by indirect means, e.g., as the solution of an optimization problem. By closed-form expression, we mean A = F (yn+1, yn), for some functional F (e.g., expectation) that can be written in closed-form and expressed only in terms of the observable variables. In what follows, we introduce a weighting function defined for any y ∈ R as: (cid:20) (cid:21)(cid:62) ω(y) (cid:44) 1 g1(y1) , 1 g2(y2) , . . . , 1 gN (yN ) . (18) Since in principle ω(y) can be singular when g(y) has some nonzero entries, it is useful to introduce the set: Z = {y ∈ RN : g(y) has at least one zero entry}. Let us also introduce the zero-lag matrix: F0(yn) (cid:44) h(yn)h(yn)(cid:62), June 24, 2019 (19) (20) DRAFT 11 and the one-lag matrix: F1(yn+1, yn) (cid:44)(cid:2)ω(yn) (cid:12) σ−1(yn+1)(cid:3) h(yn)(cid:62) I(yn /∈ Z), (21) where I[·] is the indicator function, which takes on the value 1 if the event under brackets is true, and 0 otherwise. The indicator has been included in the definition to assign a finite value (zero) to the one-lag matrix when the function ω is singular. The next lemma exploits the nonlinear auto-regressive structure of (12) to relate the matrix A to the expected value of two functions of the samples. Lemma 1 (Generalized Granger): If all the entries in the vector g(yn) are nonzero with probability one, and if the following expectations are well-defined: F0(n) = E[F0(yn)], E(cid:104) ω(yn)h(yn)(cid:62)(cid:105) (22) (23) (24) (25) (26) then the following expectation is well-defined: and we have that: F1(n) (cid:44) E[F1(yn+1, yn)] F1(n) = A F0(n) Moreover, if the matrix F0(n) is invertible, we have that: A = F1(n)[F0(n)]−1 Proof: From (12) we have the identity: ω(yn) (cid:12) σ−1(yn+1)I[yn /∈ Z] = [Ah(yn) + ω(yn) (cid:12) xn+1] I[yn /∈ Z]. Multiplying both sides of (26) by h(yn)(cid:62), from (21) we get: F1(yn+1, yn) = Ah(yn)h(yn)(cid:62)I[yn /∈ Z] + xn+1 (cid:12) ω(yn)h(yn)(cid:62)I[yn /∈ Z]. (27) Since by assumption all entries of vector g(yn) are nonzero with probability one, we have that P[yn /∈ Z] = 1. Thus, by taking expectations of both sides in (27), and using the definitions of F0(n) in (22), and of F1(n) in (23), we obtain: F1(n) = AF0(n) + E(cid:104) xn+1 (cid:12) ω(yn)h(yn)(cid:62)(cid:105) . (28) June 24, 2019 DRAFT = 0, (29) where the last equality holds since E [xn+1] = 0, whereas the second relationship holds in view of the integrability assumption in (22), and since x has finite first moment [52]. Remark 1: The solution F1(n)[F0(n)]−1 arises as the classical solution to the following nonlinear regression problem (where (cid:107) · (cid:107)2 is the Euclidean norm): Likewise, the empirical counterparts of F0(n) and F1(n): would give (cid:98)F1(n)[(cid:98)F0(n)]−1 as the solution of the least-squares fitting problem: k=0 k=0 n n F1(yk, yk+1) (cid:105) . E(cid:104)(cid:13)(cid:13)ω(yn) (cid:12) σ−1(yn+1) − Bh(yn)(cid:13)(cid:13)2 n−1(cid:88) F0(yk), (cid:98)F1(n) (cid:44) 1 n−1(cid:88) 2 argmin B (cid:98)F0(n) (cid:44) 1 n−1(cid:88) k=0 argmin B (cid:107)ω(yk) (cid:12) σ−1(yk+1) − Bh(yk)(cid:107)2 2. 12 (30) (31) (32) (cid:3) In view of the definition in (20), the first term on the RHS is automatically well-defined because we assume integrability of the random variable F0(yn) -- see the first relationship in (22). As to the second term, from the independence between xn+1 and yn we can write: xn+1 (cid:12) ω(yn)h(yn)(cid:62)(cid:105) E(cid:104) = E(cid:104) = E [diag(xn+1)] E(cid:104) diag(xn+1)ω(yn)h(yn)(cid:62)(cid:105) ω(yn)h(yn)(cid:62)(cid:105) The relationships in Lemma 1 are basically obtained through straightforward algebra. However, there are two fundamental issues that have been overlooked so far. First, the relationship in (25) contains expected values. These expected values are obviously unknown (since they depend on the unknown distribution of the samples, which in turn depends on the object of the topology inference, the unknown matrix A). Nevertheless, Eq. (25) would become very useful if we can show that the matrix functions F0(n) and F1(n) can be consistently estimated from the samples, e.g., using (31). This analysis requires addressing issues of stability and ergodicity, and will be carried out in (cid:3) Sec. VI. Second, the statement of Lemma 1 is in a sense optimistic, since it relies on two generic assumptions like "assume this is well-defined" and "assume this is invertible". More precisely, for the direct rela- tionship (24) to hold, one needs that the expected values in (22) are well-defined. Once this condition is ascertained, one more condition is needed. Indeed, to get the fundamental inverse relationship (25), June 24, 2019 DRAFT 13 Fig. 2. An example where the shape of g does impair integrability, resulting in a singular estimator of the graph. which allows retrieving the object of topology inference, A, from the zero-lag and one-lag functions, one (cid:3) needs F0(n) to be invertible. We conclude that the inference problem treated in this work is not tractable for all possible functions (σ, g, h) and statistical laws for the noise x characterizing the dynamical system. That is why establishing when the aforementioned conditions are met (we will provide next a set of sufficient conditions for that) is critical to establish whether or not full information about the topology is contained in the samples. Moreover, it is also relevant to understand the practical meaning of these assumptions in connection to the properties of the networked dynamical system under consideration. The forthcoming sections address all these fundamental concerns. A. Existence of F0(n) and F1(n) Examining (1), we see that the map g controls the flow of information among nodes. To give (an extreme) example, if g ≡ 0 then the information about the state of the neighboring agents does not flow or in particular, the network information (that is entailed in A) is lost. It is further natural to expect that if g is generally too small then, even though information is technically flowing, it is hard to extract it. Thus, first of all, we have to guarantee that the inverse vector 1/g is almost surely well defined, i.e., that the probability that g has some zero entries is zero. One critical difficulty here is that g is a function of yn, and that the distribution of yn depends in some intricate way on the evolution of the dynamical system. Such physical property reflects into the following mathematical requirement. Since the matrix function F1(n) in (22) contains the inverse vector function 1/g, one must guarantee that the latter behaves properly in terms of integrability. One example that shows how the result of Lemma 1 needs to be carefully checked, is shown in Fig. 2, where we used the simple linear function gi(y) = y for all i = 1, 2, . . . , N. June 24, 2019 DRAFT -3-2-10123-3-2-1012314 We evaluate empirically the matrix functions F0(n) and F1(n), and we observe that they basically blow up. The next assumption formally sets a condition on the aforementioned aspect. Assumption 1 (Integrability conditions): We assume bounded moments: E[(cid:107)y0(cid:107)2] < ∞, E[(cid:107)x(cid:107)2] < ∞, for the initial condition y0 and the noise term x. We assume also that: E[(cid:107)ω(σ(x + c))(cid:107)2] ≤ K, ∀c ∈ RN , and for some constant K. (33) (34) (cid:3) Moreover, since there are several functions involved in the definition of F1(n) as well as in the definition of F0(n), we need a set of sufficient conditions to guarantee the existence of the latter two matrix functions. Assumption 2 (Regularity of the nonlinearities): We assume that σ is a diffeomorphism2. We assume that g and h are continuous functions. Furthermore, let us introduce the diagonal matrix: Dg(y) (cid:44) diag(g(y)). We assume the following conditions on the nonlinearities (σ, g, h): (cid:107)σ(y)(cid:107) ≤ ασ(cid:107)y(cid:107) + βσ, (cid:107)Dg(y)(cid:107) ≤ αg(cid:107)y(cid:107)p + βg, (cid:107)h(y)(cid:107) ≤ αh(cid:107)y(cid:107)q + βh. (35) (36) (37) (38) for some nonnegative constants ασ, αg, αh, βσ, βg, βh with p ≥ 0, q ≥ 0, and p + q = 1. (cid:3) For instance, a condition like (36) holds when (cid:107)σ(y)(cid:107) grows not faster than (cid:107)y(cid:107) outside some compact set. Indeed, in this case the constant βσ can be given by the maximum of (cid:107)σ(y)(cid:107) inside that compact set. Likewise, when some of the functions σ, h, g are bounded, we assume the pertinent α-constants equal to zero. We have the following result. Proposition 1 (Existence of F0(n) and F1(n)): Under Assumptions 1 and 2, the expectations F0(n) = E[h(yn)h(yn)(cid:62)], F1(n) = E(cid:104)(cid:0)ω(yn) (cid:12) σ−1(yn+1)(cid:1) h(yn)(cid:62)(cid:105) 2Actually, to prove Proposition 1 we need just σ to be continuous and invertible. June 24, 2019 (39) DRAFT 15 Fig. 3. An example where the shape of h does impair invertibility of F0(n), resulting in a singular estimator of the graph. are well-defined, and the random variable g(yn) has all nonzero entries with probability 1. Proof: The proof is given in Appendix A. B. Invertibility of F0(n) A critical requirement for retrieving A from the matrix functions F0(n) and F1(n) is invertibility of F0(n). However, it is important to remark that there might be dynamical systems where this condition can be violated. In Fig. 3, we show an example where the function σ has different behaviors for different agents. In particular, for all agents we have that the component-wise functions have a hyperbolic tangent shape. For agents 1 and 2, however, they are shifted (upward and downward, respectively). The function g is chosen so as to fulfill the conditions for integrability. The function h is the limiter (i.e., linear saturating function) displayed in Fig. 3. We can give to these nonlinearities some physical meaning. The nonlinear shape is typical of several applications (for instance, neural networks). The saturation effects present both June 24, 2019 DRAFT -3-2-10123-3-2-10123-3-2-10123-1-0.8-0.6-0.4-0.200.20.40.60.8116 in σ and h are as well typical. For instance, the saturation in h takes on the practical meaning of limiting the interaction when the sample yi,n is too large. This behavior can be present, e.g., in a social learning problem where the agents might use saturation to filter observations that look like outliers. In the example reported in Fig. 3, we have computed empirically the matrices F0(n) and F1(n). In particular, we have verified that F0(n) is not invertible, which gives rise to the singular behavior of the topology estimator observed in the lowermost-rightmost panel. This behavior is due to the fact that the vector function h (uppermost-rightmost panel) becomes constant along the two coordinates corresponding to nodes 1 and 2, since the output range of σ1 and σ2 (uppermost-leftmost panel) forces h1 and h2 to operate only in the saturation region (i.e., y < −1 for node 1, and y > 1 for node 2). This yields singularity of the matrix: (40) Indeed, since the matrix E[h(yn)h(yn)(cid:62)] is always positive semi-definite, in order to grant invertibility we must exclude the condition that there exists some (deterministic) vector v such that: F0(n) = E[h(yn)h(yn)(cid:62)]. v(cid:62)E[h(yn)h(yn)(cid:62)]v = 0. However, the above condition would correspond to say that: h(yn)(cid:62)v = 0 almost surely, (41) (42) which basically means that h(yn) must not be a low-rank map. The following assumption (which is, e.g., usually encountered in ergodic theory [56]) makes this statement precise. Assumption 3 (Non-singularity of response h): Let µLeb be the Lebesgue measure in RN . We assume that h : RN → RN is such that µLeb(h−1(A)) = 0 for all sets A such that µLeb(A) = 0. (43) In words, this assumption states that if the input set A has full-dimension, then its image h(A) is non-degenerate as well. It can be verified that transformations that fulfill property (43) are: linear full- rank maps, open maps, diffeomorphisms, and differentiable maps with non-singular Jacobian almost everywhere [56]. On the other hand, a constant map does not have this property or, more generally, a low-rank linear map (e.g., a projection onto a lower-dimensional subspace) does not have this property as the image of any set necessarily lies in a lower-dimensional subspace (orthogonal to the kernel of the linear application). (cid:3) Proposition 2 (Invertibility of F0(n)): Under Assumptions 1 -- 3, and if σ is a diffeomorphism, then the matrix June 24, 2019 F0(n) = E[h(yn)h(yn)(cid:62)] (44) DRAFT 17 is invertible for all n ∈ N. Proof: The proof is given in Appendix B. VI. CONSISTENT GRAPH LEARNING Lemma 1 states that the interaction matrix A can be obtained in terms of two matrix functions, namely, the functions F0(n) and F1(n). However, this property alone would not be sufficient to guarantee that we can reliably estimate the topology from the samples {yn}. For this to be possible, we should demonstrate that the aforementioned functions can be consistently learned from the samples. The requirement of consistency means that we should be able to converge to the desired matrix functions as the number of available samples grows. This property would yield a practical scheme to estimate the interaction matrix A (and hence its support) from the samples. In other words, we are requiring the system to be stable and ergodic. And indeed, inference problems over dynamical systems often rely on the stability of the system: it is hard to perform faithful inference over systems that blow up. Technically speaking, the model in (12) corresponds to a Markov chain with a general state space, since, for instance, when the noise component x is absolutely continuous, the chain can walk over a continuous space [55]. For such type of Markov chains, the limiting results (e.g., stability and/or ergodicity) are much more involved than those corresponding to the classic case of finite or discrete state space. In the following, we will appeal to powerful results to prove that our dynamical system is in fact ergodic, which would be a critical property to enable consistent topology learning. We have the following result. Proposition 3 (Consistent Graph Learning): Assume that Assumptions 1 -- 3 are fulfilled. Assume further that the noise xn ∈ RN is absolutely continuous with almost-everywhere (with respect to the Lebesgue measure in RN ) positive density. Let us introduce the stability constant:  κs (cid:44) ασαgαh(cid:107)A(cid:107), ασαgβh(cid:107)A(cid:107), ασαhβg(cid:107)A(cid:107), if p > 0 and q > 0, if p = 1 and q = 0, if p = 0 and q = 1. Then, if κs < 1, we have that: A = F1 F−1 0 where June 24, 2019 F0 = lim n→∞ F0(n), F1 = lim n→∞ F1(n), (45) (46) (47) DRAFT and: (cid:98)F0(n) = (cid:98)F1(n) = n−1(cid:88) n−1(cid:88) k=0 k=0 1 n 1 n F0(yk) a.s.−→ F0, F1(yk, yk+1) a.s.−→ F1. 18 (48) (49) Proof: The proof is given in Appendix C. Proposition 3 provides consistency (i.e., faithful estimation as the number of samples grows) of the generalized Granger within the scope of the class of nonlinear dynamical systems (1). Based on this result, we are now in the position of proposing the following graph learning algorithm to reconstruct the underlying (directed) graph from each realization (or sample-path) of the nonlinear system. Algorithm 1 Empirical Generalized Granger (EGG) Input: Samples {yk}, for time epochs k = 0, 1, . . . , n. Dynamical-law triple (σ, g, h) Output: (cid:98)A(n) = Estimate of A n−1(cid:88) 1: (cid:98)F0(n) = n−1(cid:88) (cid:2)ω(yk) (cid:12) σ−1(yk+1)(cid:3) h(yk)(cid:62) 2: (cid:98)F1(n) = (cid:104)(cid:98)F0(n) (cid:105)−1 3: (cid:98)A(n) =(cid:98)F1(n) h(yk)h(yk)(cid:62) 1 n 1 n k=0 k=0 One special comment is deserved by the last step of the algorithm. Once we estimate the combination matrix A, we need to reconstruct its support graph. However, due to finiteness of the number of samples, also the zero entries in A would result in some (possibly small) nonzero entries. Accordingly, in the last step of the algorithm we apply a clustering algorithm (here the k-means with k = 2) to the entries of the estimated interaction matrix (cid:98)A. Such clustering algorithm is aimed at devising a boundary threshold: j affects i if the entry(cid:98)aij assumes a high(er) value, otherwise, if aij assumes a weak(er) value between causality and non-causality among agents, i.e., to provide an automated data-driven classification then j is deemed as not affecting i. In order for the clustering algorithm to work properly, one might question that the zero/nonzero entries of A should inherently possess some clustering property. Indeed, such form of clustering has been proved to hold (for sufficiently large network sizes) when the underlying graph is an Erdos-R´enyi random graph, for a relevant class of combination matrices [38] -- [43]. June 24, 2019 DRAFT 19 VII. ILLUSTRATIVE EXAMPLES We now present a number of examples aimed at illustrating the validity of the theoretical analysis conducted so far. In what follows, the ground-truth combination matrix, A, is constructed through the following steps. First, we generate a realization of a binomial graph with connection probability p, which means that any arrow (i.e., directed edge) exists with probability p, and independently from all the other edges [59]. Once the underlying graph has been generated, the combination weights are assigned to each arrow of the graph to yield the combination matrix A. We will consider the following assignment rule (a.k.a.  uniform averaging rule): aij = gij di , ρ ρ −(cid:80) k(cid:54)=i aik, for i (cid:54)= j for i = j , (50) where gij ∈ {0, 1} is equal to 1 only if there is a directed edge from j to i, and where di = Ni is the in-degree or number of agents directly influencing agent i. In the following experiments, we will consider a network of N = 50 nodes. We start by examining the full observation case. In Fig. 4, we consider the following nonlinearities to drive our dynamical system, for all i = 1, 2, . . . , N: σi(y) = sign(y)y0.5, gi(y) = sign(y)y0.3, hi(y) = sign(y)y0.7. (51) (52) (53) The parameter ρ of the combination matrix is set equal to 0.5. In the lowermost row of Fig. 4, we display the Empirical Generalized Granger (EGG) algorithm and contrast its performance against three standard estimators, namely, the (linear) Granger estimator, the precision matrix (i.e., the inverse of the correlation matrix) and the correlation matrix. For the sake of a neater data-visualization, and in order to display the matrices values in a one-dimensional plot, we proceeded as follows. First, in all panels the true matrix A is represented in black. In particular, we first vectorized the true matrix A and removed the entries associated to its diagonal. As a result, each element in the abscissa of each plot, say 10, corresponds to a particular entry index, say (2, 6). Then, we sorted the entries of the resulting vector in ascending order -- that is why the black curves are nondecreasing. Still in the lowermost row of Fig. 4, the blue curves are obtained by arranging the off-diagonal entries of the pertinent matrix estimators as induced by the aforementioned ordering of A. In this way, we are contrasting entry-by-entry the ground-truth June 24, 2019 DRAFT 20 Fig. 4. Nonlinear dynamical system described by the triple of functions in (51) -- (53), and depicted in the first three uppermost panels. The true graph is generated according to a binomial graph with connection probability p = 0.2. The combination matrix follows the uniform averaging rule in (50), and is displayed in the fourth uppermost panel. The lowermost panels display the performance of the different estimators as indicated in the panel titles. In the pertinent panels, the matrix entries are vectorized and sorted as described in the main text. The inset plot of the uppermost/rightmost panel displays the true graph corresponding to a network portion of 10 nodes. The inset plots of the lowermost panels display the sub-graph learned by the pertinent algorithms, with the red arrows representing edges that do not exist in the true topology, and that are erroneously detected by the learning algorithm. matrix A with the matrix-estimators in a one-dimensional ordered plot. The inset plots displayed in the various panels show just one portion (for the sake of clarity) of the whole network graph learned by the pertinent algorithm. We remark that in this analysis, the algorithm has access to the full network, whereas the panel is limited to a sub-graph just for a matter of visualization. In the inset plots, the red arrows represent edges erroneously detected by the learning algorithm (i.e., edges that are not present in the true graph). Three major observations arise by inspection of Fig. 4. First, we see that the EGG algorithm is able to estimate faithfully the combination weights. Second, the clustering algorithm is able to properly reconstruct the network skeleton from the estimated combination matrix. Third, and perhaps not unex- pectedly, the classical methods that work in the linear case (Granger), or in the Gaussian graphical model setting (precision), or over correlation-networks (correlation matrix), are essentially blind in our nonlinear framework. June 24, 2019 DRAFT -3-2-10123-2-1.5-1-0.500.511.52-3-2-10123-1.5-1-0.500.511.5-3-2-10123-2.5-2-1.5-1-0.500.511.522.521 Fig. 5. Nonlinear dynamical system described by the triple of functions in (54) -- (56), and depicted in the first three uppermost panels. The true graph is generated according to a binomial graph with connection probability p = 0.2. The combination matrix follows the uniform averaging rule in (50), and is displayed in the fourth uppermost panel. The lowermost panels display the performance of the different estimators as indicated in the panel titles. In the pertinent panels, the matrix entries are vectorized and sorted as described in the main text. The inset plot of the uppermost/rightmost panel displays the true graph corresponding to a network portion of 10 nodes. The inset plots of the lowermost panels display the sub-graph learned by the pertinent algorithms, with the red arrows representing edges that do not exist in the true topology, and that are erroneously detected by the learning algorithm. In Fig. 5, we repeat the experiment on another set of nonlinearities, namely, σi(y) = tanh y, gi(y) = sign(y)y0.4, hi(y) = sin(4y) + sign(y)y0.6. (54) (55) (56) We see that similar conclusions apply. One interesting difference is the better convergence of the EGG and of the one-lag-functional estimators, which could be ascribed to the fact that the activation function, σ, is now bounded, a feature that concurs to increase the system stability, and, hence, the speed of convergence of the various empirical estimators. VIII. BEYOND THE THEORETICAL RESULTS The results summarized in Proposition 3 allow consistent estimation of the network graph under the specific setting and under a set of assumptions that we have extensively discussed. On the other hand, June 24, 2019 DRAFT -3-2-10123-1-0.8-0.6-0.4-0.200.20.40.60.81-3-2-10123-2-1.5-1-0.500.511.52-3-2-10123-3-2-1012322 Fig. 6. Same example as in Fig. 2, but with a regularized weighting function(cid:101)ω(y). In the leftmost panel,(cid:101)ω(y) is equal to ω(y) outside the interval (−0.1, 0.1). In the rightmost panel, a stronger perturbation is applied to ω(y), since (cid:101)ω(y) is equal to ω(y) outside the interval (−0.2, 0.2). there exist relevant scenarios where the setting need to be enlarged and some of the assumptions relaxed. In this section, we show how the analysis can be helpful to address some of these scenarios. A. Regularization of the Weighting Function ω(y) From our analysis, it is apparent that one limitation is the integrability condition (34) imposed on the weighting function ω(y). For example, in Fig. 2 we have seen that a linear g(y) can lead to a weighting function ω(y) with infinite expectation (because of the too-fast singularity of g(y) = y around the origin), resulting in a singular estimator. In order to remedy this issue, one could replace ω(y) with a regularized version, (cid:101)ω(y), in the evaluation of the one-lag function F1(yn+1, yn). For example, we could set (cid:101)ω(y) = ω(y) outside some small neighborhoods of its singularities, and (cid:101)ω(y) equal to some constant within these neighborhoods. The rationale behind this regularization is that, on one hand, we rule out the pathological behavior of the estimator since we remove the singularities; on the other hand, for sufficiently small perturbations of the original ω(y), we expect that the deviations from the true combination matrices are small enough to let the clustering algorithm be still able to classify correctly the links between nodes. We apply the proposed regularization to the example of Fig. 2, and the result is shown in Fig. 6. Some notable features emerge. We start by examining the leftmost panel. First, we see that the regularized weighting function removes the singular behavior of the estimator (blue curve), which is now capable to follow the true profile (black curve) of the combination matrix entries. This behavior is reasonable June 24, 2019 DRAFT 23 because our regularization has in fact removed the singularity. Second, we observe that the estimator looks noisier as compared to the examples of Figs. 4 and 5. This augmented irregularity can be explained because we started from a singular weighting function. Third, we see that the estimator seems not to converge to the true combination matrix, which is expected because consistent reproduction of the combination matrix is not granted when we use a surrogate weighting function. More specifically, a trade-off arises between the degree of irregularity and the fidelity of reconstruction, and we expect that the smaller the perturbation of the original weighting function is, the higher the irregularity of the estimator and the fidelity of reconstruction will be. This trade-off is confirmed in the rightmost panel of Fig. 6, where we consider a stronger perturbation (i.e., we modify the weighting function in an ampler neighborhood). Comparing the rightmost panel against the leftmost panel, we see that the blue curve is now less wild, but that it is more distant from the true matrix (black curve). In summary, a sort of uncertainty principle is exhibited: one can either get a more precise knowledge (less oscillating curves) of a less precise matrix value; or a less precise knowledge (i.e., more oscillating curves) of a matrix value closer to the true value. This notwithstanding, we should keep in mind that the basic goal of graph learning is retrieving the skeleton of the network, i.e., the support graph of the combination matrix. For this reason, it is not so crucial that the estimator is not able to reproduce exactly the values of the combination matrix. What is critical is that the identifiability gap between the estimated matrix entries corresponding to disconnected or connected node pairs is still well recognizable from the estimator, which would allow to the clustering algorithm to recover the correct graph. The inset panels of Fig. 6 show that this can be in fact possible. This notion of identifiability gap has been introduced and extensively discussed in [42], [43], and will play a role especially in the partial observation setting, as we will see in the forthcoming section. B. Partial Observation Setting All the analysis conducted so far was based on a full-observability assumption, namely, on the assumption that samples from all nodes can be collected. We note however that in complex large-scale systems, we are often unable to probe the state-evolution of all nodes. Accordingly, in this section we consider the partial observation setting where the data can be collected from only a subset S of nodes. The objective of the learning becomes then reconstructing the support graph of the partial sub-matrix, AS, namely, of the sub-matrix corresponding to subset S. Likewise, in the partial observation case we June 24, 2019 DRAFT assume that the truncated matrix functions3: [F0]S, [F1]S, 24 (57) will replace the full matrices F0 and F1. However, since A = F1F−1 0 , in general the latent (unobserved) samples for nodes outside S affect the possibility of constructing AS from [F0]S and [F1]S. For example, we have: AS (cid:54)= [F1]S([F0]S)−1. (58) For the special case of linear networked dynamical systems, the generalized Granger boils down to the Granger estimator, and the partial construction in (58) is obtained by considering the zero-lag and one-lag correlation matrices. Recent works have in fact established the structural consistency of this partial (i.e., applied only to a subset of nodes) Granger estimator in the linear case, for a class of regular symmetric combination matrices A (that include, for instance the classic Laplacian and Metropolis matrices), and when the underlying graph is an undirected Erdos-R´enyi graph under various regimes of connectivity [38] -- [43]. In particular, in [39] feasibility of such inverse problem is proved for sparse Erdos-R´enyi graphs and increasing cardinality of the observable space; the analysis is extended in [40], [41] to cover the relevant case where the observed nodes have arbitrary connection structure and the cardinality of observed nodes is fixed (thus, the degree of observability is low); in [42], [43], the analysis is extended to cover the case of densely connected networks. Motivated by these results holding for the linear case, we now test the partial Generalized Granger estimator in (58) with the more general nonlinear dynamical systems addressed in this work. Devising technical guarantees for structural consistency of the generalized Granger under this latent scenario appears to be a highly nontrivial task. Therefore, we are not in the position here to prove that the proposed method is consistent under this particular latent regime. Nevertheless, it is useful to see whether, under the conditions used to examine the nonlinear model under full-observation, the algorithm applied to partial observation can work. In Fig. 7, we use the same settings adopted in Figs. 4 and 5, but for one essential detail. Now, only a subset S with 10 out of 50 nodes is accessible. According to the discussion in Sec. VIII, we implement two estimators, namely, the EGG with [(cid:98)F0(n)]S and [(cid:98)F1(n)]S, namely, with the matrix functions evaluated only on data coming from the observed network component. We see how both the proposed method is able to faithfully estimate the network graph. 3We remark that the sub-matrices [F0]S and [F1]S can be constructed from the samples {yn} because the i-th component of the functions σ(y), h(y), g(y), g(y) depends solely on the i-th entry yi. June 24, 2019 DRAFT 25 Fig. 7. Partial generalized Granger estimator in (58) applied when only 10 out of 50 nodes are probed. The inset plots display the sub-graph of probed nodes as reconstructed by the learning algorithm. Uppermost panels: Same example as in Fig. 4. Lowermost panels: Same example as in Fig. 5. We remark that the proposed analysis is by no means exhaustive, and is meant to show that there are cases where the generalized Granger method can work in the nonlinear regime. On the other hand, even if we do not have a definite answer, we have also evidences that it can fail, and that it is in particular sensitive to two features: i) heterogeneity, namely, when the nonlinear components at different nodes behave very differently, the role of some latent nodes might become dominant and corrupt the fidelity of the graph reconstruction; ii) level of noise, which is a distinguishing feature of the nonlinear setting, since in the linear case the noise variance played basically as a scale factor that does not affect the graph identifiability. Perhaps not unexpectedly, in the nonlinear case the level of noise can alter even substantially the overall qualitative behavior of the dynamics. June 24, 2019 DRAFT 26 APPENDIX A PROOF OF PROPOSITION 1 Lemma 2: Under Assumption 1, we have, for all n ∈ N: (59) Proof: Let k(y) = g(y) (cid:12) Ah(y). In view of (12), we have that yn = σ(k(yn−1) + xn). We can E[(cid:107)ω(yn)(cid:107)2] ≤ K. use the tower property to write: E(cid:2)(cid:107)ω(yn)(cid:107)2(cid:3) = E(cid:2) E(cid:2)(cid:107)ω(σ(k(yn−1) + xn))(cid:107)2 yn−1 (cid:3)(cid:3) However, since xn is statistically independent from yn−1, we have: E(cid:2)(cid:107)ω(σ(k(yn−1) + xn))(cid:107)2 yn−1 (cid:3) E(cid:2)(cid:107)ω(σ(c + x))(cid:107)2(cid:3) ≤ K, ≤ sup c∈R (60) (61) where in the first inequality we used the fact that whenever x and y are independent random variables, and for any measurable map q, the underlying function r (62) is given by r(y) = E [q(x, y)] for all y ∈ RN [52] and we have that r(y) ≤ supc∈RN r(c) almost surely (which yields the first inequality in (61) with the proper choice of q). The latter inequality in (61) holds r(y) (cid:44) E [q(x, y)y ] in view of (34). Remark 2: Note that the integrability condition on ω(yn) implies that the vector g(yn) has some zero (cid:3) entry with zero probability. Lemma 3: Let us define, for an arbitrarily small > 0, the following constant: where p, q and the various α and β constants have been introduced in (36) -- (38). We have that: (cid:107)h(yn)(cid:107) ≤ αh γn(cid:107)y0(cid:107) + ασ γi(cid:107)xn−i(cid:107) + κ, for some κ > 0. i=1 June 24, 2019 DRAFT  γ (cid:44) ασαgαh(cid:107)A(cid:107) + , ασαgβh(cid:107)A(cid:107), ασαhβg(cid:107)A(cid:107), if p > 0 and q > 0 if p = 1 and q = 0 if p = 0 and q = 1 (cid:32) n(cid:88) (cid:33) (63) (64) Proof: Using (36) -- (38) we have the following chain of inequalities: (cid:107)σ (g(y) (cid:12) Ah(y) + x)(cid:107) ≤ ασ(cid:107)g(y) (cid:12) Ah(y) + x(cid:107) + βσ ≤ ασ((cid:107)g(y) (cid:12) Ah(y)(cid:107) + (cid:107)x(cid:107)) + βσ ≤ ασ(cid:107)A(cid:107)(cid:107)Dg(y)(cid:107)(cid:107)h(y))(cid:107) + ασ(cid:107)x(cid:107) + βσ ≤ ασ(cid:107)A(cid:107) (αg(cid:107)y(cid:107)p + βg) (αh(cid:107)y(cid:107)q + βh) + ασ(cid:107)x(cid:107) + βσ ≤ ασαgαh(cid:107)A(cid:107)(cid:107)y(cid:107)(p+q) + ασ(cid:107)x(cid:107) + βσ + ασ(cid:107)A(cid:107)βgβh + ασαgβh(cid:107)A(cid:107)(cid:107)y(cid:107)p + ασαhβg(cid:107)A(cid:107)(cid:107)y(cid:107)q. 27 (65) Now, consider first the case where both p > 0 and q > 0. In this case, since p + q = 1, we have that both p and q are strictly less than 1, implying that, for an arbitrarily small , and for a certain constant c we can write: ασαgβh(cid:107)A(cid:107)(cid:107)y(cid:107)p ≤ c + ασαhβg(cid:107)A(cid:107)(cid:107)y(cid:107)p ≤ c + (cid:107)y(cid:107), (cid:107)y(cid:107). 2 2 Accordingly, from (65) we get: (cid:107)σ (g(y) (cid:12) Ah(y) + x)(cid:107) ≤ γ(cid:107)y(cid:107) + ασ(cid:107)x(cid:107) + β, (66) (67) where we have collected all the constant terms into the factor β, and we have used the first definition of γ in (63). Next consider the case q = 0. This implies that h(y) is bounded by a constant, and we accordingly have αh = 0. Therefore, Eq. (65) becomes: (cid:107)σ (g(y) (cid:12) Ah(y) + x)(cid:107) ≤ γ(cid:107)y(cid:107) + ασ(cid:107)x(cid:107) + β, (68) where we have collected the constant terms into β, and where we have used the second definition of γ in (63). The proof for the case p = 0 follows similarly. Lemma 4: If the constant γ in (63) is strictly smaller than 1, then the sequence of random variables {(cid:107)h(yn)(cid:107)2} is uniformly integrable. Proof: First, we observe that: un (cid:44) (cid:107)h(yn)(cid:107)2 = ϕ(y0, x1, x2, . . . , xn), June 24, 2019 (69) DRAFT for a deterministic function ϕ. In view of Lemma 3, we can write: (cid:34) (cid:32) ϕ(y0, x1, x2, . . . , xn) ≤ γn(cid:107)y0(cid:107) + ασ αh n(cid:88) i=1 γi(cid:107)xn−i(cid:107) Now, since the xi's are i.i.d., the function: u(cid:48) n = ϕ(y0, xn, xn−2, . . . , x1) (cid:33) (cid:35)2 + κ . 28 (70) (71) that is obtained by applying the function ϕ to a reversed sequence xn, xn−1, . . . , x1 has the same distribution of un. Applying (71) to (70) we can write: (cid:34) (cid:32) u(cid:48) n = ϕ(y0, xn, xn−1, . . . , x1) ≤ γn(cid:107)y0(cid:107) + ασ n(cid:88) αh γi(cid:107)xi(cid:107) i=1 (cid:33) (cid:35)2 ∞(cid:88) + κ . (72) Now, in view of Kolmogorov two-series theorem [61], it makes sense to introduce the random variable: (73) which has the first two moments bounded since γ < 1 and E[(cid:107)x(cid:107)2] < ∞. Therefore, from (72) we have: i=1 ξ (cid:44) γi(cid:107)xi(cid:107), This clearly shows that: n ≤ [αh((cid:107)y0(cid:107) + ασ ξ) + κ]2 u(cid:48) n ≤ u(cid:48), u(cid:48) (74) (75) where u(cid:48) is an integrable random variable that does not depend on n. This implies that u(cid:48) n are uniformly integrable. Since uniform integrability is a property of the distribution, also the original un are uniformly integrable, and the claim of the lemma follows. Proof of Proposition 1: We need to show that the following expectations are well-defined: F0(n) = E[h(yn)h(yn)(cid:62)], E[ω(yn)h(yn)(cid:62)]. (76) That F0(n) is well-posed follows directly from Lemma 4. For what concerns E[ω(yn)h(yn)(cid:62)], we know that (cid:107)ω(yn)(cid:107)2 is integrable in view of Lemma 2, while (cid:107)h(yn)(cid:107)2 is integrable in view of Lemma 4. Thus, the claim follows since the product of two L2 functions is L1. In a more explicit form, if we take the (i, j)-th entry of the matrix ω(yn)h(yn)(cid:62), we get: and: wi(yi(n))hj(yj(n)), E[wi(yi(n))hj(yj(n))] ≤(cid:113)E[w2 i (yi(n))] E[h2 j (yj(n))] by simple application of the Cauchy-Schwartz inequality. June 24, 2019 (77) (78) DRAFT APPENDIX B PROOF OF PROPOSITION 2 Proof of Proposition 2: We will prove the claim by contradiction. Assume F0(n) not invertible. In view of (42), this corresponds to saying that h(yn)(cid:62)v = 0 almost surely, i.e.: 29 where: P[h(yn) ∈ V⊥] = 1, V⊥ = {z : z(cid:62)v = 0}. On the other hand, we have that (µLeb is the Lebesgue measure in RN ): µLeb(V⊥) = 0 ⇓ µLeb(h−1(V⊥)) = 0 ⇓ P[yn ∈ h−1(V⊥)] = 0, (79) (80) (81) where the first equality holds true because V⊥ is a lower-dimensional subspace of RN ; the intermediate implication comes from Assumption 3; whereas the last implication comes from the fact that yn is absolutely continuous with respect to the Lebesgue measure, as we will now show. Indeed, from (12) we can write: yn = σ (g(yn−1) (cid:12) Ah(yn−1) + xn) . (82) Now, the argument of the function σ is absolutely continuous because it is the sum of two independent random variables, with one of these being absolutely continuous. Since the mapping σ is a diffeomor- phism, it preserves absolute continuity. Indeed, letting y = σ(z), with z being absolutely continuous with a density fz, we have that (D denotes the Jacobian): P[y ∈ A] = P[σ(z) ∈ A] = P[z ∈ σ−1(A)] (cid:90) (cid:90) A = = σ−1(A) fz(ζ)dζ fz(σ−1(y))(cid:12)(cid:12)det(Dσ−1(y))(cid:12)(cid:12) dy, (83) and absolute continuity of y = σ(z) follows by absolute continuity of z. We conclude that yn is absolutely continuous. But we have the equality: P[yn ∈ h−1(V⊥)] = P[h(yn) ∈ V⊥], which violates condition (79), yielding a contradiction. June 24, 2019 (84) DRAFT 30 APPENDIX C PROOF OF PROPOSITION 3 Proof: In order to prove ergodicity, it is convenient to use the additive noise model representation in (16), which we report here for convenience: zn+1 = G(zn) + xn+1, where we have used the definitions, zn = σ−1(yn), (cid:101)g(y) = g(σ(y)), (cid:101)h(y) = h(σ(y)), and, for z ∈ RN : G(z) =(cid:101)g(z) (cid:12) A(cid:101)h(z). (85) (86) (87) In view of Example 7.4.6 in [58], if F is continuous, if the noise xn is absolutely continuous with respect to the Lebesgue measure with almost everywhere positive density with finite mean, and if there exist positive constants β and γ < 1 such that4: (88) then the chain is w-geometrically ergodic with weight function ω(z) = 1 +z [58]. Since by assumption continuity of F and the properties of the noise are fulfilled, it remains to verify that (88) holds true in E[(cid:107)G(z) + x(cid:107)] ≤ γ(cid:107)z(cid:107) + β, our setting. Reasoning as done in (65), we get: (cid:107)G(z) + x(cid:107) ≤ (cid:107)Dg(σ(z))Ah(σ(z))(cid:107) + (cid:107)x(cid:107) ≤ (cid:107)A(cid:107)(cid:107)Dg(σ(z))(cid:107)(cid:107)h(σ(z))(cid:107) + (cid:107)x(cid:107) ≤ (cid:107)A(cid:107) (αh(cid:107)σ(z)(cid:107)p + βh) (αg(cid:107)σ(z)(cid:107)q + βg) + (cid:107)x(cid:107) ≤ (cid:107)A(cid:107)αhαg(cid:107)σ(z)(cid:107) + (cid:107)A(cid:107)βhβg + (cid:107)x(cid:107) + (cid:107)A(cid:107)αhβg(cid:107)σ(z)(cid:107)p + (cid:107)A(cid:107)αgβh(cid:107)σ(z)(cid:107)q. (89) We shall consider the case p > 0, q > 0. The proof for the remaining cases is similar. Reasoning along the same lines as in (66), we can conclude that, for an arbitrarily small > 0: (cid:107)G(z) + x(cid:107) ≤ ((cid:107)A(cid:107)αhαg + )(cid:107)σ(z)(cid:107) + c + (cid:107)x(cid:107), (90) where c is a certain positive constant. On the other hand, in view of (36) -- (38) we can further write: E[(cid:107)G(z) + x(cid:107)] ≤ γ(cid:107)z(cid:107) + β, 4Actually, condition (88) rephrases the condition reported in [58] in a way that is more convenient in our setting. June 24, 2019 (91) DRAFT where γ = (cid:107)A(cid:107)αhαgασ + (cid:48), and β is a proper constant. Therefore, we have proved that (88) is verified, which implies w-ergodicity of zn in light of Example 7.4.6 in [58]. Since zn = σ−1(yn), we have in fact proved w-ergodicity for yn. Since we now have w-geometric ergodicity, we also have convergence in total variation to the unique 31 invariant measure πy, i.e., (cid:107)νP n − πy(cid:107)TV n→∞−→ 0 (92) for any initial distribution ν, where P is the transition kernel characterizing the Markov process {yn}, and (cid:107) · (cid:107)TV is the total variation norm [58]. In other words, the unique invariant measure πy is a global attractor of the dynamical system νP n in the space of probability measures on RN . Now, the aforementioned convergence in total variation will imply the strong law in (48) via application of the ergodic theorem [55], [57], if we prove that the following moment (the notation Eπy [·] means that the expectation of the random variable under brackets is computed under the measure πy): (cid:104) h(y)h(y)(cid:62)(cid:105) F0 (cid:44) Eπy (93) is well-defined. To this end, let us first observe that the constant κ appearing in (45) is strictly smaller than 1 by assumption, which implies that, for a sufficiently small , it is possible to choose a constant γ < 1 in Lemma 4. Therefore, we have that the sequence of random variables h(yn)2 is uniformly integrable. Now, since: i) yn converges in total variation (and, hence, in distribution); ii) h is a continuous mapping; iii) the sequence h(yn)2 is uniformly integrable, we can conclude that the expectation in (93) is well- defined, and that the first convergence of expectations in (47) holds true [60]. We switch to the analysis of the one-lag matrix function. First, in view of (24), we have: F1(n) = AF0(n) → AF0, (94) since we have proved that F0(n) → F0. Second, we want to show that (49) holds true. By applying the definitions in (20) and (21) to (27), we get: F1(yn+1, yn) = AF0(yn)I[yn /∈ Z] + e(xn+1, yn), where we have defined the quantity: e(xn+1, yn) = xn+1 (cid:12) ω(yn)h(yn)(cid:62)I[yn /∈ Z]. In light of (31), from (95) we can also write: (cid:98)F1(n) = A (cid:80)n−1 k=0 F0(yk)I[yk /∈ Z] n (cid:80)n−1 k=0 e(xk+1, yk) n . + June 24, 2019 (95) (96) (97) DRAFT 32 Since uniform integrability is not impaired by the presence of the indicator, and since the probability that yk /∈ Z is equal to one, from the previous analysis it is clear that: a.s.−→ AF0. (cid:80)n−1 k=0 F0(yk)I[yk /∈ Z] (98) n Therefore, we need to establish that the second term on the RHS in (97) converges to zero almost surely. To this aim, let us consider the joint process vn ∈ R2N : vn (cid:44) (yn, xn+1). (99) Since yn and xn+1 are statistically independent, and since the vectors xn+1 are i.i.d. across time (with a certain measure πx), it is immediately seen that vn admits a unique invariant measure, given by the product measure πv = πx × πy, between the invariant measure of yn and the (stationary) measure of xn+1. Moreover, for any n we obvisouly have that πvn = πx × πyn, where πyn = νP n is the distribution of yn, for a certain initial distribution ν. Using then (92), we can conclude that: (cid:107)πvn − πv(cid:107)TV n→∞−→ 0. Therefore, we can apply the ergodic theorem [55], [57] to vn. In particular, we can observe that: n−1(cid:88) k=0 1 n e(yk, xk+1) a.s.−→ Eπx×πy [e(x, y)], (100) (101) provided that the latter expectation, computed under the invariant distribution, exists. However, from (96) we can write: E[e(xn+1, yn)] = E[diag(xn+1)]E[ω(yn)h(yn)(cid:62)], (102) since integrability of the pertinent random variables has been already established in Lemma 1. More- over, since we have also established that ω(yn) and h(yn) are L2-integrable, with integrals that are bounded by a constant independent of n, joint application of Skorohod's representation theorem and by Fatou's lemma [60] imply that also the expectation Eπy [ω(y)h(y)(cid:62)] (i.e., computed under the invariant distribution πy) exists. But since the invariant distribution πv has the product form πx× πy, we can write: Eπx×πy [e(x, y)] = Eπx[diag(x)] Eπy [ω(y)h(y)(cid:62)] = 0, (103) (cid:124) (cid:123)(cid:122) =0 (cid:125) which applied in (101) yields the desired claim. In order to conclude the proof of the proposition, we have to prove invertibility of F0. To this aim, let us consider an initial state y0 distributed according to the invariant measure πy. In view of (12), we have: y1 = σ (g(y0) (cid:12) Ah(y0) + x1) , (104) DRAFT June 24, 2019 33 and we observe that the state y1 will be still distributed according to πy due to invariance, implying that we can write F0 = Eπy . (105) (cid:104) h(y)h(y)(cid:62)(cid:105) = E(cid:104) h(y1)h(y1)(cid:62)(cid:105) Therefore, the fact that F0 is invertible follows by applying to the random variable in (104) the reasoning used to prove Proposition 2. REFERENCES [1] A. Barrat, M. Barth´elemy, and A. Vespignani, Dynamical Processes on Complex Networks. Cambridge University Press, 2012. [2] 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. [3] C. J. Honey, O. Sporns, L. Cammoun, X. Gigandet, J. P. Thiran, R. Meuli, and P. Hagmann, "Predicting human resting-state functional connectivity from structural connectivity," Proceedings of the National Academy of Sciences, vol. 106, no. 6, pp. 2035 -- 2040, Feb. 2009. [4] S. Wang, E. D. Herzog, I. Z. Kiss, W. J. Schwartz, G. Bloch, M. Sebek, D. Granados-Fuentes, L. Wang, and J.-S. Li, "Inferring dynamic topology for decoding spatiotemporal structures in complex heterogeneous networks," Proceedings of the National Academy of Sciences, vol. 115, no. 37, pp. 9300 -- 9305, Sep. 2018. [5] F. Alderisio, G. Fiore, and M. di Bernardo, "Reconstructing the structure of directed and weighted networks of nonlinear oscillators," Phys. Rev. E, vol. 95, no. 4, pp. 042302-1 -- 042302-6, Apr. 2017. [6] L. J. S. Allen, "Some discrete-time SI, SIR, and SIS epidemic models," Mathematical Biosciences, vol. 124, no. 1, pp. 83 -- 105, Nov. 1994. [7] M. J. Keeling and K. T. D. Eames, "Networks and epidemic models," Journal of the Royal Society Interface, vol. 2, pp. 295 -- 307, Jun. 2005. [8] X. Wan, J. Liu, K.-W. Cheung, and T. Tong, "Inferring epidemic network topology from surveillance data," PloS one, vol. 9, no. 6, pp. e100661-1 -- e100661-9, Jun. 2014. [9] A. Santos, J. M. F. Moura, and J. M. F. Xavier, "Sufficient condition for survival of the fittest in a bi-virus epidemics," in Proc. Asilomar Conference on Signals, Systems, and Computers, Pacific Grove, CA, USA, Nov. 2015, pp. 1323 -- 1327. [10] A. Santos, J. M. F. Moura, and J. Xavier, "Bi-virus SIS epidemics over networks: Qualitative analysis," IEEE Trans. Netw. Sci. Eng., vol. 2, no. 1, pp. 17 -- 29, Jan. 2015. [11] A. Korobeinikov and P. K. Maini, "Non-linear incidence and stability of infectious disease models," Mathematical Medicine and Biology, vol. 22, no. 2, pp. 113 -- 128, Jun. 2005. [12] A. Korobeinikov, "Global properties of infectious disease models with nonlinear incidence," Bulletin of Mathematical Biology, vol. 69, no. 6, pp. 1871 -- 1886, Aug. 2007. [13] P. Van Mieghem, J. Omic, and R. Kooij, "Virus spread in networks," IEEE/ACM Trans. Netw., vol. 17, no. 1, pp. 1 -- 14, Feb. 2009. [14] J. R. Norris and R. W. R. Darling, "Differential equation approximations for Markov chains," Probability Surveys, vol. 5, pp. 37 -- 79, 2008. [15] N. Antunes, C. Fricker, P. Robert, and D. Tibbi, "Stochastic networks with multiple stable points," The Annals of Applied Probability, vol. 36, no. 1, pp. 255 -- 278, Jan. 2008. June 24, 2019 DRAFT 34 [16] A. Santos and J. M. F. Moura, "Emergent behavior in large scale networks," in Proc. IEEE Conference on Decision and Control and European Control Conference (CDC-ECC), Orlando, FL, USA, Dec. 2011, pp. 4485 -- 4490. [17] A. Santos, J. M. F. Moura, and J. Xavier, "Diffusion and topology: Large densely connected bipartite networks," in Proc. IEEE Conference on Decision and Control (CDC), Maui, HI, USA, Dec. 2012, pp. 2738 -- 2743. [18] 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. [19] A. H. Sayed, "Adaptation, Learning, and Optimization over Networks," Found. Trends Mach. Learn., vol. 7, no. 4-5, pp. 311 -- 801, 2014. [20] V. Matta and A. H. Sayed, "Estimation and detection over adaptive networks," in Cooperative and Graph Signal Processing, P. Djuric and C. Richard, Eds. Elsevier, 2018, pp. 69 -- 106. [21] V. Matta, P. Braca, S. Marano, and A. H. Sayed, "Diffusion-based adaptive distributed detection: Steady-state performance in the slow adaptation regime," IEEE Trans. Inf. Theory, vol. 62, no. 8, pp. 4710 -- 4732, Aug. 2016. [22] V. Matta, P. Braca, S. Marano, and A. H. Sayed, "Distributed detection over adaptive networks: Refined asymptotics and the role of connectivity," IEEE Trans. Signal Inf. Process. Netw., vol. 2, no. 4, pp. 442 -- 460, Dec. 2016. [23] S. Kar and J. M. F. Moura, "Convergence rate analysis of distributed gossip (linear parameter) estimation: Fundamental limits and tradeoffs," IEEE J. Sel. Topics Signal Process., vol. 5, no. 4, pp. 674 -- 690, Aug. 2011. [24] B. Pasdeloup, V. Gripon, G. Mercier, D. Pastor, and M. G. Rabbat, "Characterization and inference of graph diffusion processes from observations of stationary signals," IEEE Trans. Signal Inf. Process. Netw., vol. 4, no. 3, pp. 481 -- 496, Sep. 2018. [25] J. Mei and J. M. F. Moura, "Signal processing on graphs: Causal modeling of unstructured data," IEEE Trans. Signal Process., vol. 65, no. 8, pp. 2077 -- 2092, Apr. 2017. [26] D. Materassi and M. V. Salapaka, "On the problem of reconstructing an unknown topology via locality properties of the Wiener filter," IEEE Trans. Autom. Control, vol. 57, no. 7, pp. 1765 -- 1777, Jul. 2012. [27] C. J. Quinn, N. Kiyavash, and T. P. Coleman, "Directed information graphs," IEEE Trans. Inf. Theory, vol. 61, no. 12, pp. 6887 -- 6909, Dec. 2015. [28] A. Anandkumar, V. Y. F. Tan, F. Huang, and A. S. Willsky, "High-dimensional structure estimation in Ising models: Local separation criterion," The Annals of Statistics, vol. 40, no. 3, pp. 1346 -- 1375, Jun. 2012. [29] A. Anandkumar, V. Y. F. Tan, F. Huang, and A. S. Willsky, "High-dimensional Gaussian graphical model selection: Walk summability and local separation criterion," Journal of Machine Learning Research, vol. 13, pp. 2293 -- 2337, Jan. 2012. [30] G. Bresler, "Efficiently learning Ising models on arbitrary graphs," in Proc. ACM Symposium on Theory of Computing (STOC), Portland, OR, USA, Jun. 2015, pp. 771 -- 782. [31] A. Anandkumar and R. Valluvan, "Learning loopy graphical models with latent variables: Efficient methods and guarantees," The Annals of Statistics, vol. 41, no. 2, pp. 401 -- 435, Apr. 2013. [32] V. Chandrasekaran, P. A. Parrilo, and A. S. Willsky, "Latent variable graphical model selection via convex optimization," The Annals of Statistics, vol. 40, no. 4, pp. 1935 -- 1967, Aug. 2012. [33] G. Bresler, F. Koehler, A. Moitra, and E. Mossel, "Learning restricted Boltzmann machines via influence maximization," in Proc. ACM Symposium on Theory of Computing (STOC), Phoenix, AZ, USA, Jun. 2019. [34] D. Materassi and M. V. Salapaka, "Network reconstruction of dynamical polytrees with unobserved nodes," in Proc. IEEE Conference on Decision and Control (CDC), Maui, HI, USA, Dec. 2012, pp. 4629 -- 4634. [35] J. Etesami, N. Kiyavash, and T. Coleman, "Learning minimal latent directed information polytrees," Neural Computation, vol. 28, no. 9, pp. 1723 -- 1768, Aug. 2016. June 24, 2019 DRAFT 35 [36] P. Geiger, K. Zhang, B. Scholkopf, M. Gong, and D. Janzing, "Causal inference by identification of vector autoregressive processes with hidden components," in Proc. International Conference on Machine Learning (ICML), vol. 37, Lille, France, Jul. 2015, pp. 1917 -- 1925. [37] D. Materassi and M. V. Salapaka, "Identification of network components in presence of unobserved nodes," in Proc. IEEE Conference on Decision and Control (CDC), Osaka, Japan, Dec. 2015, pp. 1563 -- 1568. [38] V. Matta and A. H. Sayed, "Tomography of adaptive multi-agent networks under limited observation," in Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Calgary, Canada, Apr. 2018, pp. 6638 -- 6642. [39] 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. [40] A. Santos, V. Matta, and A. H. Sayed, "Divide-and-conquer tomography for large-scale networks," in Proc. IEEE Data Science Workshop (DSW), Lausanne, Switzerland, Jun. 2018, pp. 170 -- 174. [41] A. Santos, V. Matta, and A. H. Sayed, "Consistent tomography over diffusion networks under the low-observability regime," in Proc. IEEE International Symposium on Information Theory (ISIT), Vail, CO, USA, Jun. 2018, pp. 1839 -- 1843. [42] V. Matta, A. Santos, and A. H. Sayed, "Tomography of large adaptive networks under the dense latent regime," in Proc. Asilomar Conference on Signals, Systems, and Computers, Pacific Grove, CA, USA, Oct. 2018, pp. 2144 -- 2148. [43] V. Matta, A. Santos, and A. H. Sayed, "Graph learning with partial observations: Role of degree concentration," in Proc. IEEE International Symposium on Information Theory (ISIT), Paris, France, Jul. 2019, pp. 1 -- 5. [44] E. S. C. Ching and H. C. Tam, "Reconstructing links in directed networks from noisy dynamics," Physical Review E, vol. 95, no. 1, pp. 010301-1 -- 010301-5, Jan. 2017. [45] A. Mauroy and J. Goncalves, "Linear identification of nonlinear systems: A lifting technique based on the Koopman operator," in Proc. IEEE Conference on Decision and Control (CDC), Las Vegas, NV, USA, Dec. 2016, pp. 6500 -- 6505. [46] M. Nitzan, J. Casadiego, and M. Timme, "Revealing physical interaction networks from statistics of collective dynamics," Science Advances, vol. 3, no. 2, 2017. [47] Y. Shen, B. Baingana, and G. B. Giannakis, "Kernel-based structural equation models for topology identification of directed networks," IEEE Trans. Signal Process., vol. 65, no. 10, pp. 2503 -- 2516, May 2017. [48] G. B. Giannakis, Y. Shen, and G. V. Karanikolas, "Topology identification and learning over graphs: Accounting for nonlinearities and dynamics," Proceedings of the IEEE, vol. 106, no. 5, pp. 787 -- 807, May 2018. [49] J. Etesami and N. Kiyavash, "Measuring causal relationships in dynamical systems through recovery of functional dependencies," IEEE Trans. Signal Inf. Process. Netw., vol. 3, no. 4, pp. 650 -- 659, Dec. 2017. [50] F. S. Borges, E. L. Lameu, K. C. Iarosz, P. R. Protachevicz, I. L. Caldas, R. L. Viana, E. E. N. Macau, A. M. Batista, and M. S. Baptista, "Inference of topology and the nature of synapses, and the flow of information in neuronal networks," Phys. Rev. E, vol. 97, no. 2, pp. 022303-1 -- 022303-7, Feb. 2018. [51] F. Giardina, E. O. Romero-Severson, J. Albert, T. Britton, and T. Leitner, "Inference of transmission network structure from HIV phylogenetic trees," PLOS Computational Biology, vol. 13, no. 1, pp. 1 -- 22 , Jan. 2017. [52] O. Kallenberg, Foundations of Modern Probability. [53] A. Moneta, N. Chlass, D. Entner, and P. Hoyer, "Causal search in structural vector autoregressive models," in Proc. Neural Springer, 2002. Information Processing Systems (NIPS), Vancouver, Canada, Dec. 2009, pp. 95 -- 118. [54] P.-Y. Lai, "Reconstructing network topology and coupling strengths in directed networks of discrete-time dynamics," Physical Review E, vol. 95, no. 2, pp. 022311-1 -- 022311-13, Feb. 2017. [55] S. Meyn and R. L. Tweedie, Markov Chains and Stochastic Stability. Cambridge University Press, 2009. June 24, 2019 DRAFT 36 [56] A. Lasota and M. C. Mackey, Chaos, Fractals, and Noise. Springer, 1994. [57] O. Hern´andez-Lerma and J. B. Lasserre, Markov Chains and Invariant Probabilities. Birkhauser, 2003. [58] O. Hern´andez-Lerma and J. B. Lasserre, Further Topics on Discrete-Time Markov Control Processes. Springer, 1999. [59] B. Bollob´as, Random Graphs. Cambridge University Press, 2001. [60] A. DasGupta, Asymptotic Theory of Statistics and Probability. Springer, 2008. [61] W. Feller, An Introduction to Probability and Its Applications, vol. 2, Wiley, NY, 1971. June 24, 2019 DRAFT
1907.12648	1	1907	2019-07-21T22:41:33	Multi-Agent Path Finding with Capacity Constraints	[ "cs.MA", "cs.AI" ]	In multi-agent path finding (MAPF) the task is to navigate agents from their starting positions to given individual goals. The problem takes place in an undirected graph whose vertices represent positions and edges define the topology. Agents can move to neighbor vertices across edges. In the standard MAPF, space occupation by agents is modeled by a capacity constraint that permits at most one agent per vertex. We suggest an extension of MAPF in this paper that permits more than one agent per vertex. Propositional satisfiability (SAT) models for these extensions of MAPF are studied. We focus on modeling capacity constraints in SAT-based formulations of MAPF and evaluation of performance of these models. We extend two existing SAT-based formulations with vertex capacity constraints: MDD-SAT and SMT-CBS where the former is an approach that builds the model in an eager way while the latter relies on lazy construction of the model.	cs.MA	cs	Multi-Agent Path Finding with Capacity Constraints Pavel Surynek1[0000−0001−7200−0542], T. K. Satish Kumar2, and Sven Koenig3 1 Faculty of Information Technology, Czech Technical University in Prague Th´akurova 9, 160 00 Praha 6, Czechia [email protected] http://users.fit.cvut.cz/surynek 2 University of Southern California, Henry Salvatori Computer Science Center 941 Bloom Walk, Los Angeles, USA [email protected] 3 University of Southern California, Information Sciences Institute 4676 Admiralty Way, Marina del Rey, USA [email protected] Abstract. In multi-agent path finding (MAPF) the task is to navigate agents from their starting positions to given individual goals. The problem takes place in an undirected graph whose vertices represent positions and edges define the topol- ogy. Agents can move to neighbor vertices across edges. In the standard MAPF, space occupation by agents is modeled by a capacity constraint that permits at most one agent per vertex. We suggest an extension of MAPF in this paper that permits more than one agent per vertex. Propositional satisfiability (SAT) mod- els for these extensions of MAPF are studied. We focus on modeling capacity constraints in SAT-based formulations of MAPF and evaluation of performance of these models. We extend two existing SAT-based formulations with vertex ca- pacity constraints: MDD-SAT and SMT-CBS where the former is an approach that builds the model in an eager way while the latter relies on lazy construction of the model. Keywords: multi agent path finding, propositional satisfiability (SAT), capacity constraints, cardinality constraints 1 Introduction In multi-agent path finding (MAPF) [9,18,19,21,24,28,33] the task is to navigate agents from given starting positions to given individual goals. The standard version of the problem takes place in undirected graph G = (V, E) where agents from set A = {a1, a2, ..., ak} are placed in vertices with at most one agent per vertex. The initial configuration of agents in vertices of the graph can be written as α0 : A → V and similarly the goal configuration as α+ : A → V . The task of navigating agents can be expressed as a task of transforming the initial configuration of agents α0 : A → V into the goal configuration α+ : A → V . Movements of agents are instantaneous and are possible across edges into neighbor vertices assuming no other agent is entering the same target vertex. This formulation permits agents to enter vertices being simultaneously vacated by other agents. Trivial 2 P. Surynek et al. case when a pair of agents swaps their positions across an edge is forbidden in the standard formulation. We note that different versions of MAPF exist where for example agents always move into vacant vertices [29]. We usually denote the configuration of agents at discrete time step t as αt : A → V . Non-conflicting movements transform configuration αt instantaneously into next configuration αt+1. We do not consider what happens between t and t + 1 in this discrete abstraction. Multiple agents can move at a time hence the MAPF problem is inherently parallel. In order to reflect various aspects of real-life applications variants of MAPF have been introduced such as those considering kinematic constraints [8], large agents [11], or deadlines [14] - see [13] for more variants. Particularly in this work we are dealing with an extension of MAPF that generalizes the constraint of having at most one agent per vertex. There are many situations where we need to model nodes that could hold more than agent at a time. Such situations include various graph-based evacuation models where for example nodes correspond to rooms in evacuated buildings [10] which naturally can hold more than one agent. Various spatial projections could also lead to having multiple agents per vertex such as upper projection of agents representing aerial drones where a single node corresponds to x,y-coordinate that could hold multiple agents at different z-coordinates [12]. Generally the need to consider nodes capable of containing multiple agents appears in modeling of multi-agent motion planning task at higher levels of granularity. 1.1 Contributions The contribution of this paper consists in showing how to generalize existing proposi- tional satisfiability (SAT) [4] models of MAPF for finding optimal plans with general capacity constraints that bound the number of agents in vertices. Two existing SAT- based models are generalized: MDD-SAT [32] that builds the propositional model in an eager way and SMT-CBS [30,31] that builds the model in a lazy way inspired by satisfiability modulo theories (SMT) [16]. The eager style of building the propositional model means that all constraints are posted into the model in advance. Such model is complete that is, it is solvable (sat- isfiable) if and only if the instance being modeled is solvable. In contrast to this, the lazy style does not add all constraints at once and works with incomplete models. The incomplete model preserve only one-sided implication w.r.t. solvability: if the instance being modeled is solvable then the incomplete model is solvable (satisfiable). The SMT-CBS algorithm iteratively refines the incomplete model towards the com- plete one by eliminating conflicts. That is, a candidate solution is extracted from the satisfied incomplete model. The candidate is checked for conflicts - whether any of the MAPF rules is violated - for example if a collision between agents occurred. If there are no conflicts, we are finished as the candidate is a valid solution of the input MAPF instance. If a conflict is detected, then a constraint that eliminates this particular conflict is added to the incomplete model resulting in a new model and the process is repeated. That is, a new candidate solution is extracted from the new model etc. Eventually the process may end up with a complete model after eliminating all possible conflicts. How- ever, we hope that the process finishes before constructing a complete model and we solve the instance with less effort. Multi-Agent Path Finding with Capacity Constraints 3 In the presented generalization with capacity constraints we need to distinguish between the eager and lazy variant. The capacity constraint concerning given vertex v bounding the number of agents that can simultaneously occupy v by some integer constant say 2 can be literally translated into the requirement that no 3 distinct agents can occupy v at the same time. Such a constraint can be directly posted in the eager variant: we either forbid all possible triples of agents in v or post the corresponding cardinality constraint [3,23]. The situation is different in the lazy variant. To preserve the nature of the lazy ap- proach we cannot post the capacity bound entirely as conceptually at the low level as we are informed only about a particular MAPF rule violation, say for example agents a1, a5 and a8 occurred simultaneously in v which is forbidden in given MAPF. The information that there is a capacity constraint on v bounding the number of agents in v by 2 may even not be accessible at the low level. Hence we can forbid simultaneous occurrence of only the given triple of agents, a1, a5 and a8 in this case. The paper is organized as follows. We first introduce the standard multi-agent path finding problem formally including commonly used objectives. Then we introduce two major existing SAT-based encodings. On top of this, we show how to extend these en- codings with vertex capacities. Finally we evaluate extended models on standard bench- marks including open grids and large game maps. 2 Formal Definition of MAPF and Vertex Capacities The Multi-agent path finding (MAPF) problem [24,18] consists of an undirected graph G = (V, E) and a set of agents A = {a1, a2, ..., ak} such that A ≤ V . Each agent is placed in a vertex so that at most one agent resides in each vertex. The placement of agents is denoted α : A → V . Next we are given initial configuration of agents α0 and goal configuration α+. At each time step an agent can either move to an adjacent vertex or wait in its current vertex. The task is to find a sequence of move/wait actions for each agent ai, moving it from α0(ai) to α+(ai) such that agents do not conflict, i.e., do not occupy the same location at the same time nor cross the same edge in opposite directions simultaneously. The following definition formalizes the commonly used movement rule in MAPF. Definition 1. Valid movement in MAPF. Configuration α(cid:48) results from α if and only if the following conditions hold: (i) α(a) = α(cid:48)(a) or {α(a), α(cid:48)(a)} ∈ E for all a ∈ A (agents wait or move along (ii) for all a ∈ A it holds α(a) (cid:54)= α(cid:48)(a) ⇒ ¬(∃b ∈ A)(α(b) = α(cid:48)(a) ∧ α(cid:48)(b) = α(a)) (iii) and for all a, a(cid:48) ∈ A it holds that a (cid:54)= a(cid:48) ⇒ α(cid:48)(a) (cid:54)= α(cid:48)(a(cid:48)) (no two agents share a (no two agents crosses an edge in opposite directions); edges); vertex in the next configuration). Solving the MAPF instance is to find a sequence of configurations [α0, α1, ..., αµ] such that αi+1 results using valid movements from αi for i = 1, 2, ..., µ − 1, and αµ = α+. 4 P. Surynek et al. A version of MAPF with vertex capacities generalizes the above definition by adding capacity function c : V → Z+ that assigns each vertex a positive integer capacity. The interpretation is that a vertex v can hold up to the specified number of agents c(v) at any time-step. The definition of the valid movement will change only in point (iii) where instead of permitting at most one agent per vertex we allow any number of agents not exceeding the capacity of the vertex: Definition 2. Vertex capacities in MAPF. (iii') for all v ∈ V it holds that a α(cid:48)(a) = v ≤ c(v) (the number of agents in each vertex does not exceed the capacity in the next configuration). Using generalized vertex capacities relaxes the problem in fact as illustrated in Fig- ure 1. Intuitively, capacities greater than one induce additional parking place in the environment which we hypothetise to make the problem easier to solve. Fig. 1: Illustration of the standard MAPF (c = 1) and MAPF with generalized vertex capacity (uniform capacity c = 2 us used). With c = 2 two agents a2 and a3 can both enter vertex D. In contrast to this, a3 must wait in vertex F in the standard MAPF. 2.1 Common Objectives in MAPF We address here optimal MAPF solving hence we need to introduce objective functions more formally. In case of makespan [29] we just need to minimize µ in the aforemen- tioned solution sequence. For introducing the sum-of-costs objective [7,26,21,19] we need the following notation: Definition 3. Sum-of-costs objective is the summation, over all agents, of the number i=1 ξ(path(ai)), where ξ(path(ai)) is an individual path cost of agent ai connecting α0(ai) calculated as the number of edge traversals and wait actions. 4 4 The notation path(ai) refers to path in the form of a sequence of vertices and edges connecting α0(ai) and α+(ai) while ξ assigns the cost to a given path. of time steps required to reach the goal vertex. Denoted ξ, where ξ =(cid:80)k α+ α0 A a1 a2 a3 a2 a1 a3 B C D E F A B C C C D D D A A B B E E E F F F Capacity c = 2 Capacity c = 1 a1 a2 a3 a3 a2 a1 C D A B E F Capacity c = 1 a1 a2 a3 A B C D E F Capacity c = 2 a3 a2 a1 Multi-Agent Path Finding with Capacity Constraints 5 Observe that in the sum-of-costs we accumulate the cost of wait actions for agents not yet reaching their goal vertices. For the sake of brevity we focus here on the sum- of-costs, but we note that all new concepts can be introduced for different cumulative objectives like the makespan.5 We note that finding a solution that is optimal (minimal) with respect to the sum-of- costs objective is NP-hard [17]. The same result holds for the variant with capacities as it is a straight generalization of the standard MAPF version. 3 Related Work Let us now recall existing SAT-based optimal MAPF solvers. We here focus on aspects important for introducing capacities. We recall MDD-SAT the sum-of-costs optimal solver based on eager SAT encoding [32] and SMT-CBS [31], the most recent SAT- based, or more precisely SMT-based, algorithm using lazy encoding. 3.1 SAT-based Approach The idea behind the SAT-based approach is to construct a propositional formula F(ξ) such that it is satisfiable if and only if a solution of a given MAPF of sum-of-costs ξ exists [29]. Moreover, the approach is constructive; that is, F(ξ) exactly reflects the MAPF instance and if satisfiable, solution of MAPF can be reconstructed from satisfy- ing assignment of the formula. We say F(ξ) to be a complete propositional model of MAPF. Definition 4. (complete propositional model). Propositional formula F(ξ) is a com- plete propositional model of MAPF Σ if the following condition holds: F(ξ) is satisfiable ⇔ Σ has a solution of sum-of-costs ξ. Being able to construct such formula F one can obtain optimal MAPF solution by checking satisfiability of F(0), F(1), F(2),... until the first satisfiable F(ξ) is met. This is possible due to monotonicity of MAPF solvability with respect to increasing values of common cumulative objectives like the sum-of-costs. In practice it is however impractical to start at 0; lower bound estimation is used instead - sum of lengths of shortest paths can be used in the case of sum-of-costs. The framework of SAT-based solving is shown in pseudo-code in Algorithm 1. 3.2 Details of the MDD-SAT Encoding Construction of F(ξ) as used in the MDD-SAT solver relies on time expansion of un- derlying graph G. Having ξ, the basic variant of time expansion determines the maxi- mum number of time steps µ (makespan) such that every possible solution of the given MAPF with the sum-of-costs less than or equal to ξ fits within µ timesteps. Given ξ we 5 Dealing with objectives is out of scope of this paper. We refer the reader to [32] for more detailed discussion. 6 P. Surynek et al. can calculate µ as maxk path connecting α0(ai) and α+(ai); ξ0 = (cid:80)k i=1{ξ0(ai)} + ξ − ξ0 where ξ0(a1) is the length of the shortest i=1 ξ0(ai). The detailed justification of 0(ai) to αµ this equation is given in [32]. Time expansion itself makes copies of vertices V for each timestep t = 0, 1, 2, ..., µ. That is, we have vertices vt for each v ∈ V and time step t. Edges from G are con- verted to directed edges interconnecting timesteps in the time expansion. Directed edges (ut, vt+1) are introduced for t = 1, 2, ..., µ − 1 whenever there is {u, v} ∈ E. Wait ac- tions are modeled by introducing edges (ut, tt+1). A directed path in the time expansion corresponds to trajectory of an agent in time. Hence the modeling task now consists in construction of a formula in which satisfying assignments correspond to directed paths from α0 Assume that we have time expansion T EGi = (Vi, Ei) for agent ai. Propositional variable X t v(ai) is that it is TRUE if and only if agent ai resides in v at time step t. Similarly we intro- duce Eu, vt(ai) for every directed edge (ut, vt+1) in Ei. Analogously the meaning of u,v(ai) is that is TRUE if and only if agent ai traverses edge {u, v} between time E t steps t and t + 1. Constraints are added so that truth assignment are restricted to those that correspond to valid solutions of a given MAPF. Added constraints together ensure that F(ξ) is a complete propositional model for given MAPF. v(aj) is introduced for every vertex vt in Vi. The semantics of X t +(ai) in the time expansion. We here illustrate the model by showing few representative constraints. We omit here constraints that concern objective function. For the detailed list of constraints we again refer the reader to [32]. Collisions among agents are eliminated by the following constraint for every v ∈ V and timestep t expressed on top of X t v(ai) variables: X t v(ai) ≤ 1 (1) (cid:88) ai∈A vt∈Vi One efficient way is to introduce ¬X t There are various ways how to translate the constraint using propositional clauses. v(aj) for all possible pairs of ai and aj. Next, there is a constraint stating that if agent ai appears in vertex u at time step t then it has to leave through exactly one edge (ut, vt+1). This can be established by following constraints: v(ai)∨¬X t u(ai) ⇒ (cid:95) X t E t u,v(ai), (cid:88) (ut,vt+1)∈Ei E t u,v(ai) ≤ 1 vt+1 (ut,vt+1)∈Ei (2) (3) Similarly, the target vertex of any movement except wait action must be empty. This is ensured by the following constraint for every (ut, vt+1) ∈ Ei: E t u,v(ai) ⇒ ¬X t v(aj) (4) aj∈A aj(cid:54)=ai∧vt∈Vj (cid:94) Multi-Agent Path Finding with Capacity Constraints 7 Algorithm 1: Framework of SAT-based MAPF solving 1 SAT-Based (G = (V, E), A, α0, α+) 2 paths ← {shortest path from α0(ai) to α+(ai)i = 1, 2, ..., k} ξ ←(cid:80)k i=1 ξ(N.paths(ai)) while TRUE do 3 4 5 6 7 8 9 10 F(ξ) ← encode(ξ, G, A, α0, α+) assignment ← consult-SAT-Solver(F(ξ)) if assignment (cid:54)= UNSAT then paths ← extract-Solution(assignment) return paths ξ ← ξ + 1 Other constraints ensure that truth assignments to variables per individual agents form paths. That is if agent ai enters an edge it must leave the edge at the next time step. E t u,v(ai) ⇒ X t v(ai) ∧ X t+1 v (ai) (5) A common measure how to reduce the number of decision variables derived from the time expansion is the use of multi-value decision diagrams (MDDs) [21]. The basic observation that holds for MAPF is that an agent can reach vertices in the distance d (distance of a vertex is measured as the length of the shortest path) from the current position of the agent no earlier than in the d-th time step. Analogical observation can be made with respect to the distance from the goal position. Above observations can be utilized when making the time expansion of G. For a given agent, we do not need to consider all vertices at time step t but only those that are reachable in t timesteps from the initial position and that ensure that the goal can be reached in the remaining µ − t timesteps. 3.3 Resolving Conflicts Lazily in SMT-CBS SMT-CBS is inspired by search-based algorithm CBS [22,20] that uses the idea of re- solving conflicts lazily; that is, a solution of MAPF instance is not searched against the complete set of movement constraints that forbids collisions between agents but with respect to initially empty set of collision forbidding constraints that gradually grows as new conflicts appear. SMT-CBS follows the high-level framework of CBS but rephrases the process into propositional satisfiability in a similar way as done in formula satisfia- bility testing in satisfiability modulo theory paradigm [16,15,5]. The high-level of CBS searches a constraint tree (CT) using a priority queue in breadth first manner. CT is a binary tree where each node N contains a set of collision avoidance constraints N.constraints - a set of triples (ai, v, t) forbidding occurrence of agent ai in vertex v at time step t, a solution N.paths - a set of k paths for individual agents, and the total cost N.ξ of the current solution. 8 P. Surynek et al. The low-level process in CBS associated with node N searches paths for individual agents with respect to set of constraints N.constraints. For a given agent ai, this is a standard single source shortest path search from α0(ai) to α+(ai) that avoids a set of vertices {v ∈ V (ai, v, t) ∈ N.constraints} whenever working at time step t. For details see [19]. CBS stores nodes of CT into priority queue OPEN sorted according to the ascending costs of solutions. At each step CBS takes node N with the lowest cost from OPEN and checks if N.paths represent paths that are valid with respect to MAPF movements rules - that is, N.paths are checked for collisions. If there is no collision, the algorithms returns valid MAPF solution N.paths. Otherwise the search branches by creating a new pair of nodes in CT - successors of N. Assume that a collision occurred between agents ai and aj in vertex v at time step t. This collision can be avoided if either agent ai or agent aj does not reside in v at timestep t. These two options correspond to new successor nodes of N - N1 and N2 that inherit the set of conflicts from N as follows: N1.conflicts = N.conflicts ∪ {(ai, v, t)} and N2.conflicts = N.conflicts ∪ {(aj, v, t)}. N1.paths and N1.paths inherit paths from N.paths except those for agents ai and aj respectively. Paths for ai and aj are recalculated with respect to extended sets of conflicts N1.conflicts and N2.conflicts respectively and new costs for both agents N1.ξ and N2.ξ are determined. After this, N1 and N2 are inserted into the priority queue OPEN. SMT-CBS compresses CT into a single branch in which the propositional model taken from MDD-SAT is iteratively refined. The high-level branching from CBS is deferred to the low level of SAT solving. In the MDD-SAT encoding collision avoidance constraints are omitted initially, only constraints ensuring that assignments form valid paths interconnecting starting positions with goals are be preserved. This will result in an incomplete propositional model denoted H(ξ). The important component of SMT- CBS is a paths validation procedure that reports back the set of conflicts found in the current solution that are used for making model refinements. SMT-CBS is shown in pseudo-code as Algorithm 2. The algorithm is divided into two procedures: SMT-CBS representing the main loop and SMT-CBS-Fixed solving the input MAPF for fixed cost ξ. The major difference from the standard CBS is that there is no branching at the high-level. The high-level SMT-CBS roughly correspond to the main loop of MDD-SAT. The set of conflicts is it- eratively collected during the entire execution of the algorithm. Procedure encode from MDD-SAT is replaced with encode-Basic that produces encoding that ignores specific movement rules (collisions between agents) but in contrast to encode it encodes col- lected conflicts into H(ξ). The conflict resolution in the standard CBS implemented as high-level branching is here represented by refinement of H(ξ) with disjunction (line 20). The presented SMT-CBS can eventually build the same formula as MDD-SAT but this is done lazily in SMT-CBS. Multi-Agent Path Finding with Capacity Constraints 9 Algorithm 2: SMT-CBS algorithm for solving MAPF 1 SMT-CBS (Σ = (G = (V, E), A, α0, α+)) 2 3 conf licts ← ∅ paths ← {path∗(ai) a shortest path from α0(ai) to α+(ai)i = 1, 2, ..., k} ξ ←(cid:80)k i=1 ξ(paths(ai)) while TRUE do (paths, conf licts) ← SMT-CBS-Fixed(conf licts, ξ, Σ) if paths (cid:54)= UNSAT then 4 5 6 7 8 9 19 20 21 22 return paths ξ ← ξ + 1 H(ξ) ← encode-Basic(conf licts, ξ, Σ) while TRUE do 10 SMT-CBS-Fixed(conf licts, ξ, Σ) 11 12 13 14 15 16 17 18 assignment ← consult-SAT-Solver(H(ξ)) if assignment (cid:54)= U N SAT then paths ← extract-Solution(assignment) collisions ← validate(paths) if collisions = ∅ then return (paths, conf licts) for each (ai, aj, v, t) ∈ collisions do v (ai) ∨ ¬X t H(ξ) ← H(ξ) ∪ {¬X t conf licts ← conf licts ∪ {[(ai, v, t), (aj, v, t)]} v (aj)} return (UNSAT,conf licts) 4 Handling Capacity Constraints in MAPF To adapt the SAT-based approach for MAPF with capacities we need minor modifica- tions only in both MDD-SAT and SMT-CBS. However in each algorithm the integra- tion of capacity constraints in profoundly different. While in MDD-SAT we integrate capacity constraints eagerly in the line with the original design of the algorithm (that is, the constraint in introduced as a whole), in SMT-CBS we integrate capacity constraint lazily which means part by part as new conflicts appear. 4.1 Details of the Encoding with Capacities We need only a small modification of the MDD-SAT encoding to handle vertex capac- ities. We need to replace constraint 1 with the following constraint that is again posted for every vertex v and time step t:(cid:88) X t v(ai) ≤ c(v) (6) ai∈A vt∈Vi Unlike in the standard MAPF we need here a more sophisticated translation of the constraint to propositional clauses. Using the approach of forbidding individual 10 P. Surynek et al. c(v) + 1-tuples can be highly inefficient especially in cases when c(v) is large. There- fore we use cardinality constraints encodings commonly used in SAT [3,25,23]. Gen- erally the cardinality constraint over set of propositional variables {X1,X2, ...,Xn} permits at most a specified number of variables from the set to be TRUE , denoted ≤k{X1,X2, ...,Xn} means that at most k variables from the set can be TRUE . In our case of MAPF with capacities we need to introduce following cardinality constraints for every vertex v and time step t. The practical implementation of cardinaliy constraints is done through encoding adder circuits inside the formula [23]. ≤c(v){X t v(ai) ai ∈ A ∧ vt ∈ Vi} (7) 4.2 Capacities in SMT-CBS Capacities in SMT-CBS are resolved lazily as well. That is, the capacity constraint is not posted entirely as a cardinality constraint but instead individual sets of agents that violate the capacity are forbidden one by one as they appear. That is for example if a generalized conflict occurs with agents ai1, ai2, ..., aim in vertex v (in other words if m > c(v)) we post a conflict elimination clause concerning the colliding set of agents: ¬X t v(ai1) ∨ ¬X t Hence in the SMT-CBS algorithm we modify only lines 20 and 21 that handle gen- eralized vertex conflicts. Also we need to modify the validate procedure called at line 15 to reflect generalized vertex capacities. v(ai2) ∨ ... ∨ ¬X t v(aim ). 5 Experimental Evaluation To evaluate the performance of capacity handling in context of SAT-based algorithms we performed an extensive evaluation on both standard synthetic benchmarks [6,21] and large maps from games [27]. 5.1 Setup of Experiments and Benchmarks We took the existing implementations of MDD-SAT and SMT-CBS written in C++. Both implementations are built on top of the Glucose 4 SAT solver [1,2]. In the imple- mentations we modified the capacity constraint from the original at-most-one to gener- alized variants as mentioned above. All experiments were run on a Ryzen 7 CPU 3.0 Ghz under Kubuntu linux 16 with 16GB RAM. The timeout in all experiments was set to 500 seconds. Presented are only results finished within this time limit. The second part of experimental evaluation took place on large 4-connected maps taken from Dragon Age [19,27]. We took three structurally different maps focusing on various aspects such as narrow corridors, large almost isolated rooms, or topologically complex open space. In contrast to small instances, these were only sparsely populated with agents. Initial and goal configuration were generated at random again. Up to 80 agents were used in these instances and uniform capacities of 1, 2, and 3. On large maps we measured the runtime. Multi-Agent Path Finding with Capacity Constraints 11 Fig. 2: Sorted runtimes and the number of clauses on the 8 × 8 grid. MDD-SAT and SMT-CBS are compared. 5.2 Results on Small Grids Results obtained for small open grids are presented in Figures 2 and 3. We can see that in comparison with the standard MAPF capacities bring significant reduction of difficulty of instances. This difference can be seen in both MDD-SAT and SMT-CBS. The starkest performance difference is between c = 1 and c = 2. The least performance difference is between c = 3 and c = 4. The similar picture can be seen in for the number of clauses. 5.3 Results on Large Maps Results for large game maps are shown in Figures 4 and 5. A different picture can be seen here. Adding capacities does not cause any significant simplification except the brc202d map which consists of narrow corridors. The interpretation is that adding extra parking place through capacities may lead to simplification only when it is not available normally. Otherwise generalized capacity constraints lead to harder instances. 6 Discussion and Conclusion We introduced multi-agent path finding problem with vertex capacity constraints. We modified two existing state-of-the-art SAT-based optimal MAPF solvers to reflect vertex 0,0010,010,11101001000080160240320400480560Runtime (seconds) Instance Runtime Grid 8×8 MDD-SAT c = 1c = 2c = 3c = 40,0010,010,11101001000080160240320400480560Runtime (seconds) Instance Runtime Grid 8×8 MDD-SAT c = 1c = 2c = 3c = 40,0010,010,11101001000080160240320400480560Runtime (seconds) Instance Runtime Grid 8×8 SMT-CBS c = 1c = 2c = 3c = 41162564096655361048576080160240320400480560Number of clauses Instance Clauses Grid 8×8 SMT-CBS c = 1c = 2c = 3c = 412 P. Surynek et al. Fig. 3: Sorted runtimes and the number of clauses on the 16× 16 grid. MDD-SAT and SMT-CBS are compared. capacities, the MDD-SAT solver using the eager encoding and the SMT-CBS solver using the lazy encoding. In both solvers we observed that adding an extra room by increasing the capacity of vertices dramatically reduces the difficulty of instances. However adding further ca- pacity does has less significant effect. In large maps using higher capacities even lead to performance degradation which we attribute to more complex constraints. In the future work we would like to apply the MAPF formulation with capacities in the real-life multi-robot problem. References 1. Audemard, G., Lagniez, J., Simon, L.: Improving glucose for incremental SAT solving with assumptions: Application to MUS extraction. In: SAT. pp. 309 -- 317 (2013) 2. Audemard, G., Simon, L.: Predicting learnt clauses quality in modern SAT solvers. In: IJCAI. pp. 399 -- 404 (2009) 3. Bailleux, O., Boufkhad, Y.: Efficient CNF encoding of boolean cardinality constraints. In: CP. pp. 108 -- 122 (2003) 4. Biere, A., Biere, A., Heule, M., van Maaren, H., Walsh, T.: Handbook of Satisfiability: Vol- ume 185 Frontiers in Artificial Intelligence and Applications. IOS Press (2009) 5. Bofill, M., Palah´ı, M., Suy, J., Villaret, M.: Solving constraint satisfaction problems with SAT modulo theories. Constraints 17(3), 273 -- 303 (2012) 0,0010,010,1110100100003206409601280160019202240Runtime (seconds) Instance Runtime Grid 16×16 MDD-SAT c = 1c = 2c = 3c = 411625640966553610485761677721603206409601280160019202240Number of clauses Instance Clauses Grid 16×16 MDD-SAT c = 1c = 2c = 3c = 40,0010,010,1110100100003206409601280160019202240Runtime (seconds) Instance Runtime Grid 16×16 SMT-CBS c = 1c = 2c = 3c = 411625640966553610485761677721603206409601280160019202240Number of clauses Instance Clauses Grid 16×16 SMT-CBS c = 1c = 2c = 3c = 4Multi-Agent Path Finding with Capacity Constraints 13 Fig. 4: Sorted runtimes of MDD-SAT on ost003d, brc202d, and den520d maps. Fig. 5: Sorted runtimes of MDD-SAT on ost003d, brc202d, and den520d maps. 6. Boyarski, E., Felner, A., Stern, R., Sharon, G., Tolpin, D., Betzalel, O., Shimony, S.: ICBS: improved conflict-based search algorithm for multi-agent pathfinding. In: IJCAI. pp. 740 -- 746 (2015) 7. Dresner, K., Stone, P.: A multiagent approach to autonomous intersection management. JAIR 31, 591 -- 656 (2008) 8. Honig, W., Kumar, T.K.S., Cohen, L., Ma, H., Xu, H., Ayanian, N., Koenig, S.: Summary: Multi-agent path finding with kinematic constraints. In: Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence, IJCAI 2017. pp. 4869 -- 4873 (2017) 9. Kornhauser, D., Miller, G.L., Spirakis, P.G.: Coordinating pebble motion on graphs, the di- ameter of permutation groups, and applications. In: FOCS, 1984. pp. 241 -- 250 (1984) 10. Kumar, K., Romanski, J., Hentenryck, P.V.: Optimizing infrastructure enhancements for evacuation planning. In: Proceedings of the Thirtieth AAAI Conference on Artificial In- telligence. pp. 3864 -- 3870. AAAI Press (2016) 11. Li, J., Surynek, P., Felner, A., Ma, H., Koenig, S.: Multi-agent path finding for large agents. In: AAAI. AAAI Press (2019) 12. Liu, S., Mohta, K., Atanasov, N., Kumar, V.: Towards search-based motion planning for micro aerial vehicles. CoRR abs/1810.03071 (2018), http://arxiv.org/abs/1810.03071 13. Ma, H., Koenig, S., Ayanian, N., Cohen, L., Honig, W., Kumar, T.K.S., Uras, T., Xu, H., Tovey, C.A., Sharon, G.: Overview: Generalizations of multi-agent path finding to real-world scenarios. CoRR abs/1702.05515 (2017), http://arxiv.org/abs/1702.05515 14. Ma, H., Wagner, G., Felner, A., Li, J., Kumar, T.K.S., Koenig, S.: Multi-agent path find- ing with deadlines. In: Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI 2018, July 13-19, 2018, Stockholm, Sweden. pp. 417 -- 423 (2018) 0,11101001000080160240320Runtime (seconds) Instance Ost003d MDD-SAT c = 1c = 2c = 30,11101001000080160240Runtime (seconds) Instance Brc202d MDD-SAT c = 1c = 2c = 30,111010010000120240360480600Runtime (seconds) Instance Den520d MDD-SAT c = 1c = 2c = 3 0,0010,010,11101001000080160240320400480Runtime (seconds) Instance Ost003d SMT-CBS c = 1c = 2c = 30,010,11101001000080160240320400Runtime (seconds) Instance Brc202d SMT-CBS c = 1c = 2c = 30,010,111010010000160320480640Runtime (seconds) Instance Den520d SMT-CBS c = 1c = 2c = 314 P. Surynek et al. 15. Nieuwenhuis, R.: SAT modulo theories: Getting the best of SAT and global constraint fil- tering. In: Principles and Practice of Constraint Programming - CP 2010 - 16th Interna- tional Conference, CP 2010, St. Andrews, Scotland, UK, September 6-10, 2010. Proceed- ings. pp. 1 -- 2 (2010) 16. Nieuwenhuis, R., Oliveras, A., Tinelli, C.: Solving SAT and SAT modulo theories: From an abstract davis -- putnam -- logemann -- loveland procedure to dpll(T). J. ACM 53(6), 937 -- 977 (2006) 17. Ratner, D., Warmuth, M.K.: Finding a shortest solution for the N x N extension of the 15- puzzle is intractable. In: AAAI. pp. 168 -- 172 (1986) 18. Ryan, M.R.K.: Exploiting subgraph structure in multi-robot path planning. J. Artif. Intell. Res. (JAIR) 31, 497 -- 542 (2008) 19. Sharon, G., Stern, R., Felner, A., Sturtevant, N.: Conflict-based search for optimal multi- agent pathfinding. Artif. Intell. 219, 40 -- 66 (2015) 20. Sharon, G., Stern, R., Felner, A., Sturtevant, N.R.: Conflict-based search for optimal multi- agent pathfinding. Artif. Intell. 219, 40 -- 66 (2015) 21. Sharon, G., Stern, R., Goldenberg, M., Felner, A.: The increasing cost tree search for optimal multi-agent pathfinding. Artif. Intell. 195, 470 -- 495 (2013) 22. Sharon, G., Stern, R., Felner, A., Sturtevant, N.R.: Conflict-based search for optimal multi- agent path finding. In: AAAI (2012) 23. Silva, J., Lynce, I.: Towards robust CNF encodings of cardinality constraints. In: CP. pp. 483 -- 497 (2007) 24. Silver, D.: Cooperative pathfinding. In: AIIDE. pp. 117 -- 122 (2005) 25. Sinz, C.: Towards an optimal CNF encoding of boolean cardinality constraints. In: CP (2005) 26. Standley, T.: Finding optimal solutions to cooperative pathfinding problems. In: AAAI. pp. 173 -- 178 (2010) 27. Sturtevant, N.R.: Benchmarks for grid-based pathfinding. Computational Intelligence and AI in Games 4(2), 144 -- 148 (2012) 28. Surynek, P.: A novel approach to path planning for multiple robots in bi-connected graphs. In: ICRA 2009. pp. 3613 -- 3619 (2009) 29. Surynek, P.: Time-expanded graph-based propositional encodings for makespan-optimal solving of cooperative path finding problems. Ann. Math. Artif. Intell. 81(3-4), 329 -- 375 (2017) 30. Surynek, P.: Lazy modeling of variants of token swapping problem and multi-agent path find- ing through combination of satisfiability modulo theories and conflict-based search. CoRR abs/1809.05959 (2018), http://arxiv.org/abs/1809.05959 31. Surynek, P.: Unifying search-based and compilation-based approaches to multi-agent path finding through satisfiability modulo theories. In: Proceedings of the 28th International Joint Conference on Artificial Intelligence, IJCAI 2019. in press (2019) 32. Surynek, P., Felner, A., Stern, R., Boyarski, E.: Efficient SAT approach to multi-agent path finding under the sum of costs objective. In: ECAI. pp. 810 -- 818 (2016) 33. Wang, K., Botea, A.: MAPP: a scalable multi-agent path planning algorithm with tractability and completeness guarantees. JAIR 42, 55 -- 90 (2011)
1310.4753	2	1310	2013-10-18T01:58:01	Society Functions Best with an Intermediate Level of Creativity	[ "cs.MA", "q-bio.NC" ]	In a society, a proportion of the individuals can benefit from creativity without being creative themselves by copying the creators. This paper uses an agent-based model of cultural evolution to investigate how society is affected by different levels of individual creativity. We performed a time series analysis of the mean fitness of ideas across the artificial society varying both the percentage of creators, C, and how creative they are, p using two discounting methods. Both analyses revealed a valley in the adaptive landscape, indicating a tradeoff between C and p. The results suggest that excess creativity at the individual level can be detrimental at the level of the society because creators invest in unproven ideas at the expense of propagating proven ideas.	cs.MA	cs	Society Functions Best with an Intermediate Level of Creativity Liane Gabora ([email protected]) University of British Columbia (Okanagan campus) Department of Psychology, Arts Building, 3333 University Way Kelowna BC, V1V 1V7, CANADA Hadi Firouzi ([email protected]) University of British Columbia (Okanagan campus) Department of Engineering, EME Building, 3333 University Way Kelowna BC, V1V 1V7, CANADA 3 1 0 2 t c O 8 1 ] A M . s c [ 2 v 3 5 7 4 . 0 1 3 1 : v i X r a Reference: Gabora, L., & Firouzi, H. (2012). Society func- tions best with an intermediate level of creativity. Proceed- ings of the Annual Meeting of the Cognitive Science Society (pp. 1578-1583). August 1-4, Sapporo Japan. Houston TX: Cognitive Science Society. Abstract In a society, a proportion of the individuals can benefit from creativity without being creative themselves by copying the creators. This paper uses an agent-based model of cultural evolution to investigate how society is affected by different levels of individual creativity. We performed a time series analysis of the mean fitness of ideas across the artificial society varying both the percentage of creators, C, and how creative they are, p using two discounting methods. Both analyses revealed a valley in the adaptive landscape, indicating a tradeoff between C and p. The results suggest that excess creativity at the individual level can be detrimental at the level of the society because creators invest in unproven ideas at the expense of propagating proven ideas. Keywords: adaptive landscape; agent-based model; creativity; cultural evolution; discounting; EVOC; imitation; individual differences; time series analysis Introduction Our capacity for self-expression, problem solving, and mak- ing aesthetically pleasing artifacts, all stem from our creative abilities. Psychologists have almost universally converged on the definition of creativity proposed by Guilford over sixty years ago at his annual address to the American Psycholog- ical Association (Moran, 2011). Guilford defined creativity in terms of two criteria: originality or novelty, and appro- priateness or adaptiveness, i.e., relevance to the task at hand. Individuals vary from not particularly creative to highly cre- ative. Not only are humans individually creative, but we build on each other's ideas such that over centuries, art, science, and technology, as well as customs and folk knowledge, can be said to evolve (Cavalli-Sforza & Feldman, 1981; Gabora, 1996; Mesoudi, Whiten, & Laland, 2006; Whiten, Hinde, La- land, & Stringer, 2011). Creativity has long been associated with personal fulfillment (May, 1975; Rogers, 1959), self- actualization (Maslow, 1959), and maintaining a competitive edge in the marketplace, and it is often assumed that more creativity is necessarily better. However, there are significant drawbacks to creativity (Cropley, Cropley, Kaufman, & Runco, 2010; Ludwig, 1995). Generating creative ideas is time consuming, and a creative solution to one problem often generates other problems, or has unexpected negative side effects that may only become apparent after much effort has been invested. Creative peo- ple often reinvent the wheel, and may be more likely to bend rules, break laws, and provoke social unrest (Sternberg & Lubart, 1995; Sulloway, 1996). They are more prone to af- fective disorders such as depression and bipolar disorder, and have a higher incidence of schizophrenic tendencies, than other segments of the population (Andreason, 1987; Flaherty, 2005; Goodwin & Jamison, 1990). Given these negative aspects of creativity, it is perhaps just as well that, in a group of interacting individuals, not all of them need be particularly creative for the benefits of creativ- ity to be felt throughout a social group. The rest can reap the rewards of the creator's ideas by simply copying, using, or admiring them. Few of us know how to build a computer, or write a symphony, but they are nonetheless ours to use and enjoy. Clearly if everyone relied on the strategy of imi- tating others, the generation of cultural novelty would grind to a halt. This raises the following questions: what is the ideal ratio of creators to imitators, and how creative should the "creative types" be? The Model We investigated this using an agent-based model of cultural evolution referred to as "EVOlution of Culture", abbrevi- ated EVOC. To our knowledge, EVOC is the only computa- tional model that enables one to create an artificial world with agents of varying levels of creativity and observe the effect of varying creativity on mean fitness and diversity of ideas in the artificial society. It uses neural network based agents that (1) invent new ideas, (2) imitate actions implemented by neighbors, (3) evaluate ideas, and (4) implement successful ideas as actions. EVOC is an elaboration of Meme and Vari- ations, or MAV (Gabora, 1995), the earliest computer pro- gram to model culture as an evolutionary process in its own right, as opposed to modelling the interplay of cultural and biological evolution1. The approach was inspired by genetic 1The approach can thus be contrasted with computer models of how individual learning affects biological evolution (Best, 1999, 2006; Higgs, 1992; Hinton & Nowlan, 1992; Hutchins & Hazel- hurst, 1991). algorithm (Holland, 1975), or GA. The GA is a search tech- nique that finds solutions to complex problems by generating a population of candidate solutions through processes akin to mutation and recombination, selecting the best, and repeating until a satisfactory solution is found. The goal behind MAV, and also behind EVOC, was to distil the underlying logic of not biological evolution but cultural evolution, i.e., the pro- cess by which ideas adapt and build on one another in the minds of interacting individuals. Agents do not evolve in a biological sense -- they neither die nor have offspring -- but do in a cultural sense, by generating and sharing ideas for ac- tions. In cultural evolution, the generation of novelty takes place through invention instead of through mutation and re- combination as in biological evolution, and the differential replication of novelty takes place through imitation, instead of through reproduction with inheritance as in biological evo- lution. EVOC has been used to address such questions as how does the presence of leaders or barriers to the diffusion of ideas affect cultural evolution. We now summarize briefly the architecture of EVOC in sufficient detail to explain our results; for further details on the model, we refer the reader to previous publications (Gabora, 1995, 2008a, 2008b; Gabora & Leijnen, 2009; Lei- jnen & Gabora, 2009; Gabora & Saberi, 2011). Agents Agents consist of (1) a neural network, which encodes ideas for actions and detects trends in what constitutes a fit action, (2) a 'perceptual system', which carries out the evaluation and imitation of neighbours' actions, and (3) a body, consisting of six body parts which implement actions. The neural network is composed of six input nodes and six corresponding out- put nodes that represent concepts of body parts (LEFT ARM, RIGHT ARM, LEFT LEG, RIGHT LEG, HEAD, and HIPS), and seven hidden nodes that represent more abstract concepts (LEFT, RIGHT, ARM, LEG, SYMMETRY, OPPOSITE, and MOVEMENT). Input nodes and output nodes are connected to hidden nodes of which they are instances (e.g., RIGHT ARM is connected to RIGHT.) Each body part can occupy one of three possible positions: a neutral or default positions, and two other positions, which are referred to as active po- sitions. Activation of any input node activates the MOVE- MENT hidden node. Same-direction activation of symmetri- cal input nodes (e.g., positive activation -- which represents up- ward motion -- of both arms) activates the SYMMETRY node. In the experiments reported here the OPPOSITE hidden node was not used. Invention An idea for a new action is a pattern consisting of six elements that dictate the placement of the six body parts. Agents gener- ate new actions by modifying their initial action or an action that has been invented previously or acquired through imita- tion. During invention, the pattern of activation on the output nodes is fed back to the input nodes, and invention is biased according to the activations of the SYMMETRY and MOVE- (Were this not the case there would MENT hidden nodes. be no benefit to using a neural network.) To invent a new idea, for each node of the idea currently represented on the input layer of the neural network, the agent makes a proba- bilistic decision as to whether the position of that body part will change, and if it does, the direction of change is stochas- tically biased according to the learning rate. If the new idea has a higher fitness than the currently implemented idea, the agent learns and implements the action specified by that idea. Imitation The process of finding a neighbour to imitate works through a form of lazy (non-greedy) search. The imitating agent ran- domly scans its neighbours, and adopts the first action that is fitter than the action it is currently implementing. If it does not find a neighbour that is executing a fitter action than its own current action, it continues to execute the current action. Evaluation Following Holland (1975), we refer to the success of an ac- tion in the artificial world as its fitness, with the caveat that unlike its usage in biology, here the term is unrelated to num- ber of offspring (or ideas derived from a given idea). Fitness of an action is determined using a predefined equation, Eq. 1, that ascribes a range of fitness values from 0 to 10 to the 729 possible actions. (Six body parts that can be in three possi- ble positions gives a total of 729.) The fitness function used in these experiments rewards activity of all body parts except for the head, and symmetrical limb movement. Total body movement, m, is calculated by adding the number of active body parts, i.e., body parts not in the neutral position. The fitness F of an action is calculated as follows: Fnc = m + 1.5(sa + st ) + 2(1− mh) (1) sa = 1 if arms are moving symmetrically; 0 otherwise st = 1 if legs are moving symmetrically; 0 otherwise mh = 1 if head is stationary; 0 otherwise Note that actions have a cultural version of what in biology is referred to as epistasis, wherein what is optimal with respect to one component depends on what is done with respect to another. Epistasis occurs because what is optimal for the left arm depends on what the right arm is doing, and vice versa, and same for the legs. Learning Invention makes use of the ability to detect, learn, and re- spond adaptively to trends. Since no action acquired through imitation or invention is implemented unless it is fitter than the current action, new actions provide valuable information about what constitutes an effective idea. Knowledge acquired through the evaluation of actions is translated into educated guesses about what constitutes a successful action by updat- ing the learning rate. For example, an agent may learn that more overall movement tends to be either beneficial (as with the fitness function used here) or detrimental, or that sym- metrical movement tends to be either beneficial (as with the fitness function used here) or detrimental, and bias the gener- ation of new actions accordingly. A Typical Run Fitness of actions starts out low because agents are initially immobile. They are all implementing the same action, with all body parts in the neutral position; thus action diversity is at a minimum. Soon some agent invents an action that has a higher fitness than immobility, and this action gets imitated, so fitness increases. Fitness increases further as other ideas get invented, assessed, implemented as actions, and spread through imitation. The diversity of actions increases due to the proliferation of new ideas, and then decreases as agents hone in on the fittest actions. In the version of the model used here, fitness values hit a ceiling and converge2. Thus, over successive rounds of invention and imitation, the agents' actions improve. EVOC thereby models how "descent with modification" occurs in a purely cultural context. Experiments To carry out our investigation of how varying the level of creativity of individuals affects the fitness of ideas in soci- ety as a whole, these experiments used a default artificial world: a toroidal lattice with 1024 nodes, each occupied by a single, stationary agent, and a von Neumann neighbor- hood structure (agents only interacted with their four adjacent neighbors). Creators and imitators were randomly dispersed.3 Runs lasted 100 iterations. In an earlier version of EVOC, in which the ratio of in- venting and imitating was always the same for all agents, we found that the society as a whole did best when the ratio of creating to imitating was approximately 2:1 (Gabora, 1995). To incorporate individual differences in degree of creativity, we constructed a version of EVOC that enables us to distin- imitators, that only obtain new guish two types of agents: actions by imitating neighbors, and creators, that obtain new actions by either inventing one or by imitating a neighbor. Imitators never invent at all; they simply copy the creators' successful inventions. Thus all new actions are generated by creators. We also made it possible to vary the probability that creators create versus imitate; each agent can be a pure imi- tator, a pure creator, or something in between. Whereas any given agent is either a creator or an imitator throughout the entire run, the proportion of creators innovating or imitating in a given iteration fluctuates stochastically. The proportion of creators relative to imitators in the society is referred to as C. The creativity of the creators -- that is, the probability that a creator invents a new action instead of imitating a neighbor -- is referred to as p. If a creator decides to invent on a partic- 2This is not the case for another version of the model (Gabora & Saberi, 2011). 3In other experiments (Leijnen & Gabora, 2009) we investigated the results of clustering creators. ular iteration, the probability of changing the position of any body part involved in an action is 1/6.4 The society consists of three subgroups: • C× p× N creators attempting to innovate • C× (1− p)× N creators attempting to imitate • (1−C)× N imitators attempting to imitate where the number of agents, N is 1024. In previous investigations we measured, for different val- ues of C and p, the diversity of ideas over the course of a run. We found that the cultural diversity, i.e., the number of different ideas implemented by one or more agent(s), was positively correlated with both the proportion of creators to imitators, and with how creative the creators were. We also obtained suggestive evidence that when creators are relatively uncreative, the mean fitness of ideas increases as a function of the percentage of creators in the society, but when cre- ators are highly creative, the society appears to be better off with fewer creators (Leijnen & Gabora, 2009). However, those simulations were performed with small societies (100 agents), and since action fitness was obtained at only one time slice (after 50 iterations) for all ratios of creators to inventors, these results did not reflect the dynamics of the time series. Given a set of series of accumulating value over time, it is unclear which series is most representative. The series cannot be unambiguously ordered unless for each pair of series one strictly dominates the other, and that is not the case here; the curves representing mean fitness at different values of {C, p} increase monotonically but they often cross and re-cross as time progresses. Thus here we present a more extensive in- vestigation of the relationship between creativity and society as a whole that employs a sophisticated solution to the time series problem. Analysis We used time series discounting which associates a "present value" with any future benefit such that the present value of any given benefit diminishes as a function of elapsed time un- til the benefit is realized (McDonald & Siegel, 1986). The standard approach in financial settings is exponential dis- counting. Given a series of benefits bt, the Net Present Value (NPV) is defined as: NPV (b) = N ∑ t=1 rt−1bt with 0 < r ≤ 1 (2) The discount rate r is normally set as r = ( 100+i 100 )−1 where i is the interest rate (in percentage) for the unit period that an investor can obtain from a safe investment. This basic idea was adapted to analyze the benefit ac- crued by attaining fit actions for different values of C and 4This gave on average a probability of one change per newly invented action, which previous experiments (Gabora, 1995) showed to be optimal. p in EVOC. The first discounting method used was Time-to- Threshold (TTT) discounting. Since all fitness trajectories were monotonically increasing, those that reached a reason- ably high threshold τ sooner should be valued higher. We measured how many iterations (time to threshold) it took for fitness to reach τ. For these runs, τ = 9 was used as a measure of optimal fitness to allow for a realistic averaging over time. Whereas imitators need creators, creators should ignore others if they could do better on their own (p = 1). In other words, the fitness prospects of creators working alone can be viewed in a manner analogous to the interest yield of treasury bonds in investment decisions. This logic suggests another kind of modification of the standard discounting method. The second adaptation to the basic notion of discounting we refer to as Present Innovation Value (PIV) discounting. Let FC,p be the mean action fitness at period t for parameter setting {C, p}. Note that F1,1 is the fitness expectation with no in- teraction amongst agents. We define the PIV for any fitness curve as: t t PIV (FC,p) = −N + FC,p F1,1 N ∑ t=1 (3) Therefore, PIV (F1,1) = 0; creators are indifferent to work- ing alone or in a community with imitation. Results All results are averages across 100 runs. The 3D graph and contour plot for the log10 TTT discounting analysis of the time series for different C, p settings are shown in Figures 1 and 2 respectively. Note that by definition a low TTT value corresponds to high mean fitness of actions across the society. The TTT method clearly demonstrates a valley in the adaptive landscape. The line running along the bottom of the valley in Figure 2 indicates, for any given value of p the optimal value for C, and vice versa. When p = 1 the optimal values of C = 0.38. When C = 1 the optimal values of p is 0.19. The global optimum is at approximately {C, p} = {0.4,1.0}. The 3D graph and contour plot for the PIV discounting analysis of the time series for different C, p settings are shown in Figures 3 and 4 respectively. The pattern is very similar to that obtained with the log10 TTT discounting analysis. Thus both log10 TTT and PIV analyses of the time se- ries showed that, although some creativity is essential to get the fitness of cultural novelty increasing over time, more cre- ativity is not necessarily better. For optimal mean fitness of agents actions across the society there is a tradeoff between C, the proportion of creators in the artificial society, and p, how creative these creators are. Discussion and Future Directions investigation yielded results This the widespread assumption that creativity is necessarily desir- able. The model is highly idealized, and caution must be taken in extrapolating to human societies. The PIV results assume that creators avoid input from neighbors if doing so that contradict Figure 1: 3D graph of the log10 Time-to-Threshold (TTT) landscape of the average mean fitness for different values of C and p, with τ = 9. The valley in the fitness landscape indicates that the optimal values of C and p for the society as a whole are less than their maximum values for most C, p settings. would maximize the fitness of their actions. In reality, cre- ative individuals may not behave so rationally. However, the PIV results were corroborated by the TTT results, indicating that the basic pattern does not depend on the assumption of economic rationality. EVOC agents are too rudimentary to suffer the affective penalties of creativity but the model incorporates another im- portant drawback to creativity mentioned in the introduction: an iteration spent inventing is an iteration not spent imitat- ing. Creative agents, absorbed in their creative process, ef- fectively rupture the fabric of the artificial society; they act as insulators that impede the diffusion of proven solutions. Imi- tators, in contrast, serve as a cultural memory that ensures the preservation of successful ideas. This suggests that the reason people are not more creative than they are is not just because it is difficult to be creative; there is a cost to society as well. Our results suggest that families, organizations, or soci- eties may self-organize to achieve levels of both imitation and creativity that are intermediate in order to achieve a balance between continuity and change. The results suggest that im- itation is neither just the greatest compliment, nor a form of free-riding, but a valuable social mechanism that serves inno- vators and imitators alike. Without invention there is nothing to imitate, but invention is considerably more effective in con- junction with imitation. Limitation of this work include that the fitness function was static throughout a run, and agents had only one action to op- timize. In real life, there are many tasks, and a division of labor such that each agent specializes in a few tasks, and imi- tates other agents to carry out other tasks. Another limitation is that EVOC currently does not allow an agent to imitate only Figure 2: Top-view contour plot of the log10 Time-to- Threshold (TTT) landscape of the average mean fitness for different values of C and p, with τ = 9. The line, obtained by visually extrapolating over minimum values C and p, in- dicates the set of optima. certain features of an idea while retaining features the idea it is currently implementing. Creative change can break up co- adapted partial solutions. Recall that actions have a cultural version of what in biology is referred to as epistasis, wherein what is optimal with respect to one component depends on what is done with respect to another. Once both components have been optimized in a mutually beneficial way (for exam- ple, the arms are moving symmetrically), excess creativity can cause co-adapted partial solutions to break down. In fu- ture studies we will investigate the effects of using a dynamic fitness function, and enabling partial imitation. We will also compare our findings to real world data. If it is the case that social groups can be too creative for their own good, then expensive and widely used programs to enhance creativity through methods such as brainstorm- ing may be counterproductive. The results of these exper- iments help make sense of findings that creativity is often suppressed in the classroom and in society at large, and that creative individuals often experience discrimination, or worse (Craft, 2005; Cropley & Cropley, 2005; Scott, 1999; Tor- rance, 1963b, 1963a). (It is well-known that Einstein's disser- tation was rejected by the Techniche Hochschule in Vienna; he wrote his papers on relativity while working at a patent office.) On the other hand, once the merits of ones' creative efforts become known, this individual's creativity is generally supported or even idolized. In future work we plan to inves- tigate the hypothesis that the social practice of discouraging creativity until the creative individual has proven him- or her- self serves as a form of social self-regulation ensuring that Figure 3: 3D graph of the Present Innovation Value (PIV) landscape of the average mean fitness for different values of C and p. Since the x axis has been inverted to aid visibility of the adaptive landscape, the valley again indicates that the optimal values of C and p for the society as a whole are less than their maximum values for most C, p settings. creative efforts are not squandered. Specifically, we will use EVOC to test the hypothesis that if individuals who generate creative outputs of low fitness fitness are exposed to social pressures that discourage creativity, and individuals who gen- erate creative outputs of high fitness fitness are encouraged to be creative, the society may self-organize such that it achieves a balance of creative and uncreative individuals (such as the C, p values indicated by the red line in our experiments). Acknowledgments This work was supported by grants to the first author from the Social Sciences and Humanities Research Council of Canada, the Natural Sciences and Engineering Research Council of Canada, and the Flemish Fund for Scientific Research, Bel- gium. We thank Tiha von Ghyczy for help with the analysis. References Andreason, N. C. (1987). Creativity and mental illness. prevalence rates in writers and their first degree relatives. American Journal ofPsychiatry, 144, 1288 -- 1292. Best, M. (1999). How culture can guide evolution: An in- quiry into gene/meme enhancement and opposition. Adap- tive Behavior, 7, 289 -- 293. Best, M. (2006). Adaptive value within natural language discourse. Interaction Studies, 7, 1 -- 15. Cavalli-Sforza, L. L., & Feldman, M. W. (1981). Cul- tural transmission and evolution: A quantitative approach. Princeton, NJ: Princeton University Press. Craft, A. R. (2005). Creativity in schools: Tensions and dilemmas. London: Routledge. Gabora, L., & Saberi, M. (2011). How did human creativity arise? an agent-based model of the origin of cumulative open-ended cultural evolution. In Proceedings of the acm conference on cognition and creativity (pp. 299 -- 306). New York: ACM Press. Goodwin, F., & Jamison, K. (1990). Manic-depressive ill- ness. New York: Oxford University Press. Higgs, P. G. (1992). The mimetic transition: a simulation study of the evolution of learning by imitation. Proceedings of the Royal Society B - Biological Sciences, 267, 1355 -- 1361. Hinton, G. E., & Nowlan, S. J. (1992). How learning can guide evolution. Complex Systems, 267, 495 -- 502. Holland, J. (1975). Adaptation in natural and artificial sys- tems. Ann Arbor: University of Michigan Press. Hutchins, E., & Hazelhurst, B. (1991). Learning in the cul- tural process. In C. Langton, J. Taylor, D. Farmer, & S. Ras- mussen (Eds.), Artificial life ii. Redwood City: Addison- Wesley. Leijnen, S., & Gabora, L. (2009). The tradeoff between de- gree of creativity and number of creators in a computational model of society. In B. Cooper & V. Danos (Eds.), Proceed- ings of developments in computational models from nature (pp. 108 -- 119). Rhodes, Greece: ICALP. Ludwig, A. M. (1995). The price of greatness. New York: Cliffs, NJ: Prentice-Hall. Whiten, A., Hinde, R. A., Laland, K. N., & Stringer, C. B. (2011). Culture evolves. Philosophical Transactions of the Royal Society B, 366, 938 -- 948. Guilford Press. Maslow, A. H. (1959). Creativity in self-actualizing people. In H. . Brothers (Ed.), Creativity and its cultivation. New York: McGraw-Hill. May, R. (1975). The courage to create. New York: Bantam. McDonald, R., & Siegel, D. R. (1986). The value of waiting to invest. Quarterly Journal of Economics, 101, 707 -- 728. (2006). Toward a unified science of cultural evolution. Behavioral & Brain Sciences, 29, 329 -- 347. Mesoudi, A., Whiten, A., & Laland, K. Moran, S. (2011). The roles of creativity in society. In J. K. . R. Sternbergn (Ed.), Cambridge handbook of cre- ativity (pp. 74 -- 90). Cambridge UK: Cambridge University Press. Rogers, C. (1959). Toward a theory of creativity. In H. Ander- son (Ed.), Creativity and its cultivation. New York: Harper & Row. Scott, C. L. (1999). Teachers biases toward creative children. Creativity Research Journal, 12, 321337. Sternberg, R. J., & Lubart, T. I. (1995). Defying the crowd: Cultivating creativity in a culture of conformity. New York: Free Press. Sulloway, F. (1996). Born to rebel. New York: Pantheon. Torrance, E. P. (1963a). Education and the creative potential. Minneapolis: University of Minnesota Press. Torrance, E. P. (1963b). Guiding creative talent. Englewood Figure 4: Top-view contour plot of the Present Innovation Value (PIV) landscape of average mean fitness for different values of C and p. The line, obtained by visually extrap- olating over maximum values C and p, indicates the set of optima. Cropley, D. H., & Cropley, A. J. (2005). Engineering cre- ativity: A systems concept of functional creativity. In J. C. Kaufman & J. Baer (Eds.), Faces of the muse: How people think, work and act creatively in diverse domains. Hillsdale, NJ: Lawrence Erlbaum. Cropley, D. H., Cropley, A. J., Kaufman, J. C., & Runco, M. (2010). Creativity in schools: Tensions and dilemmas. Cambridge: Cambridge University Press. Flaherty, A. W. (2005). Frontotemporal and dopaminergic control of idea generation and creative drive. Journal of Computational Neurology, 493, 147 -- 153. Gabora, L. (1995). Meme and variations: A computational model of cultural evolution. In L. Nadel & D. Stein (Eds.), 1993 lectures in complex systems. Reading MA: Addison- Wesley. Gabora, L. (1996). A day in the life of a meme. Philosophica, 57, 901 -- 938. Gabora, L. (2008a). Evoc: A computational model of cultural evolution. In Proceedings of the 30th annual meeting of the cognitive science society (pp. 18 -- 25). New York: Sheridan Publishing. Gabora, L. (2008b). Modelling cultural dynamics. In Pro- ceedings of the association for the advancement of artificial intelligence (aaai) (pp. 18 -- 25). Menlo Park, CA: AAAI Press. Gabora, L., & Leijnen, S. (2009). How creative should cre- ators be to optimize the evolution of ideas? a computational model. Electronic Proceedings in Theoretical Computer Science, 9, 108 -- 119.
1901.04836	1	1901	2018-12-24T01:27:48	Inferring Causality in Agent-Based Simulations - Literature Review	[ "cs.MA", "stat.AP" ]	Complex systems have interested researchers across a broad range of fields for many years and as computing has become more accesible and feasible, it is now possible to simulate aspects of these systems. A major point of research is how emergent behaviour arises and the underlying causes of it. This paper aims to discuss and compare different methods of identifying causal links between agents in such systems in order to gain further understanding of the structure.	cs.MA	cs	Inferring Causality in Agent-Based Simulations - Literature Review George Hassan-Coring Department of Computer Science University College London December 2018 Abstract Complex systems have interested researchers across a broad range of fields for many years and as computing has become more accesible and feasible, it is now possible to simulate aspects of these systems. A major point of research is how emergent behaviour arises and the underlying causes of it. This paper aims to discuss and compare different methods of identifying causal links between agents in such systems in order to gain further understanding of the structure. 1 Introduction The statistical analysis of complex systems has interested academics for the last sixty years and this is not only due to the difficulty of the problem but also the wide range of possible applications of it. Complex systems can vary between biological ecosys- tems [4 -- 6, 9, 12] to social media interactions [19, 28], financial markets [2, 17, 25] to weather prediction [32]. This range of possibilities for complex systems means that researchers seek to find some common ground that link these systems which can help in understanding them as a whole and how these systems develop over time. Computer simulations are an essential tool [8, 10, 16] in this investigation as they allow repeated analysis in great detail. This literature review first provides basic definitions and then presents an overview of relevant work in this area. 2 Definitions A complex system is "one in which there are multiple interactions between many different components" [29]. The time series of output values from these components are values measured at each discrete unit of time. The time series are influenced by the interactions between components and we are interested in discovering causal links between these time series. An example of this could be the price of an individual com- pany's stock over a time period and how it is linked to interest rates. A related property is the stationarity of a time series where "the stochastic mechanism generating the sequence is not changing" [24] and so the parameters like the mean or variance do not change over time. This property allows us to properly use statistical tests and tech- niques on the time series, making it very important. 1 Causality, as mentioned in this paper, will generally refer to Granger Causal- ity [20] and not True Causality [7] unless stated otherwise. True Causality is the underlying relationship between causally linked variables and discovering it is the ulti- mate goal of investigating causality. However, problems arise from things like random noise and error in measurements leading attempts to find true causal links astray so alternative methods are needed in practice. Granger Causality is a statistically testable measure of causality applied to time series in a system. To determine whether one variable X 'causes' another variable Y , a hypothesis test is used to compare whether a prediction model using the time series of X up to that moment as well as the time series of Y prior to that moment is better at predicting the next value of Y than an alternative prediction model using just the time series of Y . A related term is Spurious Causality; this is "incorrect causality, in a multivariate system, due to common drivers or indirect interactions." [7] Finding these spurious links is important because it allows us to get a better picture of the whole system without these unnecessary links polluting it. When discussing complex systems, a major point of inquiry is Emergent Be- haviour, defined as "A property is emergent if and only if it is present in a macrostate and it is not present in the microstate." [31] These emergent properties are so inter- esting because they are built incrementally by the simple definitions and interactions of multiple agents across an entire complex system and are the subject of a great deal of research. [8, 10 -- 15] A simple example could be an upper limit for a value in a time series that is not explicitly defined but which emerges over the time span of the sys- tem. Understanding the causal links of a system is the first step to investigating and understanding the emergent properties. 3 Literature Review The first paper that truly initiated research into the analysis of complex systems was Weiner's paper in 1956 [32]. In this piece of work, the context was meteorology and the prediction of weather systems. He discusses what makes the prediction of systems like this so difficult like the error of measurements and how these errors can propagate in deeply connected systems such that our predictions are affected. His work began to introduce statistical methods into the field so that it could be formally reasoned about and inspired others to build on his work. The impressive thing about this paper is that his comments on common errors that arise are just as relevant to this paper today as they were sixty years ago. Inspired by Wiener's previous work, Granger went on to publish his hugely impor- 2 tant paper in 1969 [20]. In this paper, he outlined the method which can be used to determine a statistical cause and when this method can be used, giving rise to the term Granger Causality as defined above. Differing to Wiener, Granger discussed these methods to be used in the field of econometrics, showing that even this early just how diverse the applications were. This paper is a true cornerstone of the whole subject, being cited over 20000 times according to JSTOR and his methodology is favoured today in fields like econometrics and neuroscience. The simplicity of the method allows it to be easily applied and is one of the reasons it has been chosen for my own work concerning the Interdyne simulator. In Granger's 1969 paper [20], he begins by explaining how a stationary time series can be "decomposed into unrelated components associated with a particular frequency", where stationary time series refers to a time series whose statistical properties like mean and variance stay constant over time. The variance of this time series is the sum of the variance of these individual component time series. Granger also mentions some interesting situations that can occur like spurious causality in the trivariate case (mentioned previously) and instantaneous causality [20]. Instantaneous causality describes a causal relationship between variables X and Y for example, where X is best predicted at a moment in time, t, using the previous values of X before point t as well as the values of Y before and including the value at time t. The opposite of this would be simple causality where X depends only on X and Y values previously to time t. A fascinating point that Granger makes about these possibilities is that a relationship may appear to be instantaneous causality but this may be due to the frequency of the measurements of the data rather than the actual properties of the variables. For example, instantaneous causality is detected when using data measured every 2 seconds but in fact there is a time lag of one second between the variables that is missed due to the infrequent sampling. Granger aims to investigate causal feedback in systems, something which he feels was not pursued prior to his paper so he sets out some formal definitions for causality and feedback which provide the basis of the Granger test. • Causality: if σ2(X ¯X, ¯Y ) < σ2(X ¯X) then Y → X • Feedback: X ↔ Y His definition of causality allows him to model possible distributions for variable X and evaluate them using the variance in order to check which one provides the best estimate and thus, whether Y has a causal relationship with X. This is the essence of the Granger test for causality but the test statistic can be altered from variance of the model to different values. Granger himself acknowledged the limitation of using the 3 variance to judge the models and suggested further investigation. He also mentioned that this theory only applies to stationary time series as non-stationary series are much harder to create tests for. Examples of bivariate and trivariate tests are presented in the paper. An outline of the process of determining Granger Causality between two variables, X and Y , would be first to generate two possible models for the variable Y at a time t: Yt+1 = t X i=1 αiYt−i + ǫt Yt+1 = t X i=1 αiYt−i + t X j=1 βjXt−j + ǫt (1) (2) These models provide the null and alternative hypothesis respectively for our hy- pothesis test. Typically a Wald Test is used to determine whether the null hypothesis is rejected or not as it is used to determine whether exploratory variables are significant with respect to a test variable. According to Aggresti [1], Wald tests are generally more widely applicable compared to alternatives like the Likelihood Ratio Test or Lagrange Multiplier Test as Wald tests require less knowledge about the distribution of the time series. If the Wald test finds variable X to have statistical significance in predicting variable Y , we reject the null hypothesis (1) and accept the alternative hypothesis (2). Thus suggesting that X "Granger causes" Y . Further papers in 1980 [21] and 1988 [22] further discuss the limitations and criti- cisms of Granger's methods which range from philosophical considerations of whether it's appropriate to use the word "cause" in relation to these variables to whether defin- ing instantaneous causality is necessary. Granger et al [23] give a good summary of his methods, and uses Granger causality tests extended with further econometric mea- surements to analyse causality between Asian stock markets and exchange rates. The authors use advanced statistical techniques along with observations about the context of the situation to address the non-stationary property of the variables in this time period. An alternative way of investigating causality between time series is proposed by Holland [26]. Inspired by the work of Rubin [30], Holland was interested in "measuring the effects of causes rather than the causes of effects" and used the idea of experiments to consider causality. To test whether a variable X causes a variable Z, let Z be the response value; this is the variable that is measured in order to determine if the other variables have a causal influence on it. X and Y are the experiment and control values respectively and are the two possible causes of Z. He proposed looking at the difference 4 in the response variable after being exposed to the experiment value and the control value to determine causal inference. Holland splits the response variable Z in two to represent the two potential responses from each of the causes: ZX for the response caused by X and ZY for the response caused by Y . This differs from Granger's ap- proach in that it always considers causation to be relative, either Z is caused by the experiment value, X, or it is caused by control variable, Y , and there are no other causes outside of the tester's control. This approach has its merits when conducting a carefully controlled experiment, for example: a medical experiment, but becomes harder in a less controllable situation. There is also a chance of X and Y themselves being causally linked such that Holland's method would provide no relevant results as ZX and ZY would have similar values. In the context of the Interdyne simulator with many different agents and rules, this method might not produce as many valuable results as Granger's methods because a large number of variables can be present and there could be spurious links present, resulting in many pointless tests using Holland's method. Care needs to be taken with the testing method to ensure that complex cases such as spurious causality are considered as they can easily arise within the system and negatively affect the results of the tests. Granger's simple testing method has also been extended by others in order to be usefully applied in other fields. For example, Geweke [18] presented a method to de- compose the causal relationships at certain frequencies, similar to what was suggested by Granger [20]. This method of decomposition was of great interest to neuroscientists with regards to signals in the brain over time and means that Granger testing can be applied to this field too [27]. Granger's theories are still an active area of research 50 years later [3, 7] and this shows the quality of his work and the developments of others who have built upon it. 4 Conclusion In conclusion, Granger causality tests appear to be the best choice for trying to identify causal links between time series as they are hugely popular across a range of disciplines concerning the study of complex systems and are relatively easy to implement. Having investigated some of the alternate ideas to Granger causality, they seem to be more con- cerned with problems regarding the specific definition of causality and are less practical when implemented. After consideration of the related literature, I believe that Granger causality is the most appropriate choice to evaluate causal relationships in a complex system like the Interdyne simulator as it is more adaptable and easier to implement than the alternatives. 5 References [1] Agresti A. Categorical Data Analysis. John Wiley and Sons, 1990. [2] P. W. Anderson. The economy as an evolving complex system. CRC Press, 2018. [3] T. Aste and T. Di Matteo. Sparse causality network retrieval from short time series. Complexity, 2017. [4] K. Bentley and C. Clack. The artificial cytoskeleton for lifetime adaptation of morphology. SODANS workshop proceedings of the Ninth International Conference on the Simulation and Synthesis of Living Systems, pages 13 -- 16, 2004. [5] K. Bentley and C. Clack. Morphological plasticity: Environmentally driven mor- phogenesis. VIIIth European Conference on Artificial Life, pages 118 -- 127, 2005. [6] K. Bentley, C. Clack, and E. J. Cox. Diatom colony formation: A computational study predicts a single mechanism can produce both linkage and separation valves due to an environmental switch. Journal of Phycology, 48(3), 2012. [7] L. Carlos-Sandberg. An investigation of spurious causality in trivariate granger analysis. Journal of Time Series Analysis, In Preparation. [8] C. C. Chen, S. Nagl, and C. Clack. A calculus for multi-level emergent behaviours in component-based systems and simulations. Proceedings of Emergent Properties in Natural and Artificial Complex Systems, pages 35 -- 51, 2007. [9] C. C. Chen, S. Nagl, and C. Clack. Context sensitivity in individual-based mod- eling. BMC Systems Biology, page 44, 2007. [10] C. C. Chen, S. Nagl, and C. Clack. Specifying, detecting and analysing emer- gent behaviours in multi-level agent-based simulations. Proceedings of the Summer Computer Simulation Conference, pages 969 -- 976, 2007. [11] C. C. Chen, S. Nagl, and C. Clack. A method for validating and discovering associations between multi-level emergent behaviours in agent-based simulations. 2nd KES International Symposium on Agent and Multi-Agent Systems, 2008. [12] C. C. Chen, S. Nagl, and C. Clack. Multi-level behaviours in agent-based sim- ulation: colonic crypt cell populations. 7th International conference on Complex Systems, 2008. [13] C. C. Chen, S. Nagl, and C. Clack. Complexity and emergence in engineering systems. Complex Systems in Knowledge-Based Environments: Theory, Models and Applications, 2009. 6 [14] C. C. Chen, S. Nagl, and C. Clack. A formalism for multi-level emergent behaviours in designed component-based systems and agent-based simulations. From System Complexity to Emergent Properties, 2009. [15] C. C. Chen, S. Nagl, and C. Clack. Identifying multi-level emergent behaviours in agent-directed simulations using complex event type specifications. Simulation, 86:41 -- 51, 2010. [16] C. Clack. Bioscience computing and the role of computational simulation in biology and medicine. Intelligent Algorithms in Ambient and Biomedical Computing, 7:3 -- 19, 2006. [17] C. Clack and T. Chiotis. Nonlinearity linkage detection for financial time series analysis. Proceedings of the Genetic and Evolutionary Computation Conference, pages 1179 -- 1186, 2007. [18] J Geweke. Measurement of linear dependence and feedback between multiple time series. Journal of the American Statistical Association, 77(378):304 -- 313, 1982. [19] P. A. Grabowicz, J. J. Ramasco, E. Moro, J. M. Pujol, and V. M. Eguiluz. Social features of online networks: The strength of intermediary ties in online social media. PLOS ONE, 2012. [20] C. W. J. Granger. Investigating causal relations by econometric models and cross- spectral methods. Econometrica, 37(3):424 -- 438, 1969. [21] C. W. J. Granger. Testing for causality a personal viewpoint. Journal of Economic Dynamics and Control, 2:329 -- 352, 1980. [22] C. W. J. Granger. Some recent developments in a concept of causality. Journal of Econometrics, 39:199 -- 211, 1988. [23] C. W. J. Granger, B. N. Huang, and C. W. Yang. A bivariate causality between stock prices and exchange rates: evidence from recent asian u. The Quarterly Review of Economics and Finance, 40:337 -- 354, 2000. [24] E. J. Hannan. Time Series and Statistics. Palgrave Macmillan UK, 1990. [25] C. Hiemstra and J. D. Jones. Testing for linear and nonlinear granger causality in the stock price-volume relation. Journal of Finance, 49(5):1639 -- 1664, 1994. [26] P. W. Holland. Statistics and causal inference. Journal of the American Statistical Association, 81(396):945 -- 960, 1986. [27] M. Kaminski, M. Ding, W. A. Truccolo, and S. L. Bressler. Evaluating causal relations in neural systems: Granger causality, directed transfer function and sta- tistical assessment of significance. Biological Cybernetics, 85:145 -- 157, 2001. 7 [28] D. Lazer, A. Pentland, L. Adamic, S. Aral, A. L. Barabsi, D. Brewer, N. Christakis, N. Contractor, J. Fowler, M. Gutmann, T. Jebara, G. King, M. Macy, D. Roy, and M. Van Alstyne. Life in the network: the coming age of computational social science. Science, 2009. [29] D. Rind. Complexity and climate. Science, 284:105 -- 107, 1999. [30] D. B. Rubin. Estimating causal effects of treatments in randomized and nonran- domized studies. Journal of Educational Psychology, 66(5):688 -- 701, 1974. [31] A. Ryan. Emergence is coupled to scope, not level. Complex Systems Engineering, 13(2):67 -- 77, 2007. [32] N. Wiener. Nonlinear prediction and dynamics. Proc. Third Berkeley Symp. on Math. Statist. and Prob., 3:247 -- 252, 1956. 8
1702.03037	1	1702	2017-02-10T01:48:40	Multi-agent Reinforcement Learning in Sequential Social Dilemmas	[ "cs.MA", "cs.AI", "cs.GT", "cs.LG" ]	Matrix games like Prisoner's Dilemma have guided research on social dilemmas for decades. However, they necessarily treat the choice to cooperate or defect as an atomic action. In real-world social dilemmas these choices are temporally extended. Cooperativeness is a property that applies to policies, not elementary actions. We introduce sequential social dilemmas that share the mixed incentive structure of matrix game social dilemmas but also require agents to learn policies that implement their strategic intentions. We analyze the dynamics of policies learned by multiple self-interested independent learning agents, each using its own deep Q-network, on two Markov games we introduce here: 1. a fruit Gathering game and 2. a Wolfpack hunting game. We characterize how learned behavior in each domain changes as a function of environmental factors including resource abundance. Our experiments show how conflict can emerge from competition over shared resources and shed light on how the sequential nature of real world social dilemmas affects cooperation.	cs.MA	cs	Multi-agent Reinforcement Learning in Sequential Social Dilemmas Joel Z. Leibo1 DeepMind, London, UK [email protected] Vinicius Zambaldi1 DeepMind, London, UK [email protected] Marc Lanctot DeepMind, London, UK [email protected] Janusz Marecki DeepMind, London, UK [email protected] Thore Graepel DeepMind, London, UK [email protected] 7 1 0 2 b e F 0 1 ] A M . s c [ 1 v 7 3 0 3 0 . 2 0 7 1 : v i X r a ABSTRACT Matrix games like Prisoner's Dilemma have guided research on social dilemmas for decades. However, they necessarily treat the choice to cooperate or defect as an atomic action. In real-world social dilemmas these choices are temporally extended. Cooperativeness is a property that applies to poli- cies, not elementary actions. We introduce sequential social dilemmas that share the mixed incentive structure of matrix game social dilemmas but also require agents to learn poli- cies that implement their strategic intentions. We analyze the dynamics of policies learned by multiple self-interested independent learning agents, each using its own deep Q- network, on two Markov games we introduce here: 1. a fruit Gathering game and 2. a Wolfpack hunting game. We char- acterize how learned behavior in each domain changes as a function of environmental factors including resource abun- dance. Our experiments show how conflict can emerge from competition over shared resources and shed light on how the sequential nature of real world social dilemmas affects cooperation. CCS Concepts •Computing methodologies → Multi-agent reinforce- ment learning; Agent / discrete models; Stochastic games; Keywords Social dilemmas, cooperation, Markov games, agent-based social simulation, non-cooperative games 1. INTRODUCTION Social dilemmas expose tensions between collective and individual rationality [1]. Cooperation makes possible bet- ter outcomes for all than any could obtain on their own. However, the lure of free riding and other such parasitic strategies implies a tragedy of the commons that threatens the stability of any cooperative venture [2]. 1These authors contributed equally. Appears in: Proceedings of the 16th International Confer- ence on Autonomous Agents and Multiagent Systems (AA- MAS 2017), S. Das, E. Durfee, K. Larson, M. Winikoff (eds.), May 8–12, 2017, Sao Paulo, Brazil. Copyright © 2017, International Foundation for Autonomous Agents and Multiagent Systems (www.ifaamas.org). All rights reserved. The theory of repeated general-sum matrix games pro- vides a framework for understanding social dilemmas. Fig. 1 shows payoff matrices for three canonical examples: Pris- oner's Dilemma, Chicken, and Stag Hunt. The two actions are interpreted as cooperate and defect respectively. The four possible outcomes of each stage game are R (reward of mutual cooperation), P (punishment arising from mu- tual defection), S (sucker outcome obtained by the player who cooperates with a defecting partner), and T (tempta- tion outcome achieved by defecting against a cooperator). A matrix game is a social dilemma when its four payoffs satisfy the following social dilemma inequalities (this formu- lation from [3]): 1. R > P defection. Mutual cooperation is preferred to mutual (1) 2. R > S exploited by a defector. Mutual cooperation is preferred to being (2) 3. 2R > T + S This ensures that mutual cooperation is preferred to an equal probability of unilateral cooper- ation and defection. (3) 4. either greed : T > R Exploiting a cooperator is preferred over mutual cooperation or over being exploited. fear : P > S Mutual defection is preferred (4) Matrix Game Social Dilemmas (MGSD) have been fruit- fully employed as models for a wide variety of phenomena in theoretical social science and biology. For example, there is a large and interesting literature concerned with mecha- nisms through which the socially preferred outcome of mu- tual cooperation can be stabilized, e.g., direct reciprocity [4, 5, 6, 7], indirect reciprocity [8], norm enforcement [9, 10], simple reinforcement learning variants [3], multiagent rein- forcement learning [11, 12, 13, 14, 15], spatial structure [16], emotions [17], and social network effects [18, 19]. However, the MGSD formalism ignores several aspects of real world social dilemmas which may be of critical impor- tance. 1. Real world social dilemmas are temporally extended. 2. Cooperation and defection are labels that apply to poli- cies implementing strategic decisions. 3. Cooperativeness may be a graded quantity. C D C R, R S, T D T, S P, P Chicken C D C 3, 3 4, 1 D 1, 4 0, 0 Stag Hunt C D C 4, 4 3, 0 D 0, 3 1, 1 Prisoners C D C 3, 3 4, 0 D 0, 4 1, 1 Figure 1: Canonical matrix game social dilemmas. Left: Outcome variables R, P , S, and T are mapped to cells of the game matrix. Right: The three canonical matrix game social dilemmas. By convention, a cell of X, Y represents a utility of X to the row player and Y to the column player. In Chicken, agents may defect out of greed. In Stag Hunt, agents may defect out of fear of a non-cooperative partner. In Prisoner's Dilemma, agents are motivated to defect out of both greed and fear simultaneously. 4. Decisions to cooperate or defect occur only quasi-si- multaneously since some information about what player 2 is starting to do can inform player 1's decision and vice versa. 5. Decisions must be made despite only having partial information about the state of the world and the ac- tivities of the other players. We propose a Sequential Social Dilemma (SSD) model to better capture the above points while, critically, maintaining the mixed motivation structure of MGSDs. That is, analo- gous inequalities to (1) – (4) determine when a temporally- extended Markov game is an SSD. To demonstrate the importance of capturing sequential structure in social dilemma modeling, we present empiri- cal game-theoretic analyses [20, 21] of SSDs to identify the empirical payoff matrices summarizing the outcomes that would arise if cooperate and defect policies were selected as one-shot decisions. The empirical payoff matrices are them- selves valid matrix games. Our main result is that both of the SSDs we considered, Gathering and Wolfpack, have empirical payoff matrices that are Prisoner's Dilemma (PD). This means that if one were to adhere strictly to the MGSD- modeling paradigm, PD models should be proposed for both situations. Thus any conclusions reached from simulating them would necessarily be quite similar in both cases (and to other studies of iterated PD). However, when viewed as SSDs, the formal equivalence of Gathering and Wolfpack disappears. They are clearly different games. In fact, there are simple experimental manipulations that, when applied to Gathering and Wolfpack, yield opposite predictions con- cerning the emergence and stability of cooperation. More specifically, we describe a factor that promotes the emergence of cooperation in Gathering while discouraging its emergence in Wolfpack, and vice versa. The straight- forward implication is that, for modeling real-world social dilemmas with SSDs, the choice of whether to use a Gathering- like or Wolfpack-like model is critical. And the differences between the two cannot be captured by MGSD modeling. Along the way to these results, the present paper also makes a small methodological contribution. Owing to the greater complexity arising from their sequential structure, it is more computationally demanding to find equilibria of SSD models than it is for MGSD models. Thus the standard evo- lution and learning approaches to simulating MGSDs cannot be applied to SSDs. Instead, more sophisticated multiagent reinforcement learning methods must be used (e.g [22, 23, 24]). In this paper we describe how deep Q-networks (e.g [25]) may be applied to this problem of finding equilibria of SSDs. Figure 2: Venn diagram showing the relationship between Markov games, repeated matrix games, MGSDs, and SSDs. A repeated matrix game is an MGSD when it satisfies the social dilemma inequal- ities (eqs. 1 – 4). A Markov game with S > 1 is an SSD when it can be mapped by empirical game- theoretic analysis (EGTA) to an MGSD. Many SSDs may map to the same MGSD. 2. DEFINITIONS AND NOTATION We model sequential social dilemmas as general-sum Markov (simultaneous move) games with each agent having only a partial observation onto their local environment. Agents must learn an appropriate policy while coexisting with one another. A policy is considered to implement cooperation or defection by properties of the realizations it generates. A Markov game is an SSD if and only if it contains outcomes arising from cooperation and defection policies that satisfy the same inequalities used to define MGSDs (eqs. 1 – 4). This definition is stated more formally in sections 2.1 and 2.2 below. 2.1 Markov Games A two-player partially observable Markov game M is de- fined by a set of states S and an observation function O : S×{1, 2} → Rd specifying each player's d-dimensional view, along with two sets of actions allowable from any state A1 and A2, one for each player, a transition function T : S × A1 × A2 → ∆(S), where ∆(S) denotes the set of discrete probability distributions over S, and a reward function for each player: ri : S × A1 × A2 → R for player i. Let Oi = {oi s ∈ S, oi = O(s, i)} be the observation space of player i. To choose actions, each player uses policy πi : Oi → ∆(Ai). For temporal discount factor γ ∈ [0, 1] we can define the i (s0) to player i when the joint policy long-term payoff V (cid:126)π (cid:126)π = (π1, π2) is followed starting from state s0 ∈ S. (cid:34) ∞(cid:88) (cid:35) i (s0) = E(cid:126)at∼(cid:126)π(O(st)),st+1∼T (st,(cid:126)at) V (cid:126)π γtri(st, (cid:126)at) . (5) t=0 Matrix games are the special case of two-player perfectly observable (Oi(s) = s) Markov games obtained when S = 1. MGSDs also specify A1 = A2 = {C, D}, where C and D are called (atomic) cooperate and defect respectively. The outcomes R(s), P (s), S(s), T (s) that determine when a matrix game is a social dilemma are defined as follows. 1 R(s) := V πC ,πC P (s) := V πD ,πD S(s) := V πC ,πD T (s) := V πD ,πC 1 1 1 2 (s) = V πC ,πC (s) = V πD ,πD (s) = V πD ,πC (s) = V πC ,πD 2 2 2 (s), (s), (s), (s), (6) (7) (8) (9) where πC and πD are cooperative and defecting policies as described next. Note that a matrix game is a social dilemma when R, P, S, T satisfy the inequalities (1) – (4). 2.2 Definition of Sequential Social Dilemma This definition is based on a formalization of empirical Definition: game-theoretic analysis [20, 21]. We define the outcomes (R, P, S, T ) := (R(s0), P (s0), S(s0), T (s0)) induced by ini- tial state s0, and two policies πC , πD, through their long- term expected payoff (5) and the definitions (6) – (9). We refer to the game matrix with R, P , S, T organized as in Fig. 1-left. as an empirical payoff matrix following the ter- minology of [21]. A sequential social dilemma is a tuple (M, ΠC , ΠD) where ΠC and ΠD are disjoint sets of poli- cies that are said to implement cooperation and defection respectively. M is a Markov game with state space S. Let the empirical payoff matrix (R(s), P (s), S(s), T (s)) be in- duced by policies (πC ∈ ΠC , πD ∈ ΠD) via eqs. (5) – (9). A Markov game is an SSD when there exist states s ∈ S for which the induced empirical payoff matrix satisfies the social dilemma inequalities (1) – (4). Remark: There is no guarantee that ΠC(cid:83) ΠD = Π, the set of all legal policies. This reflects the fact that, in practice for sequential behavior, cooperativeness is usually a graded property. Thus we are forced to define ΠC and ΠD by thresh- olding a continuous social behavior metric. For example, to construct an SSD for which a policy's level of aggressiveness α : Π → R is the relevant social behavior metric, we pick threshold values αc and αd so that α(π) < αc ⇐⇒ π ∈ ΠC and α(π) > αd ⇐⇒ π ∈ ΠD. 3. LEARNING ALGORITHMS Most previous work on finding policies for Markov games takes the prescriptive view of multiagent learning [26]: that is, it attempts to answer "what should each agent do?" Sev- eral algorithms and analyses have been developed for the two-player zero-sum case [22, 27, 28, 29, 30]. The general- sum case is significantly more challenging [31], and algo- rithms either have strong assumptions or need to either track several different potential equilibria per agent [32, 33], model other players to simplify the problem [34], or must find a Figure 3: Left: Gathering. In this frame the blue player is directing its beam at the apple respawn location. The red player is approaching the apples from the south. Right: Wolfpack. The size of the agent's view relative to the size of the map is il- lustrated. If an agent is inside the blue diamond- shaped region around the prey when a capture occurs-when one agent touches the prey-both it and its partner receive a reward of rteam. cyclic strategy composed of several policies obtained through multiple state space sweeps [35]. Researchers have also stud- ied the emergence of multi-agent coordination in the decen- tralized, partially observable MDP framework [36, 37, 38]. However, that approach relies on knowledge of the underly- ing Markov model, an unrealistic assumption for modeling real-world social dilemmas. In contrast, we take a descriptive view, and aim to answer "what social effects emerge when each agent uses a partic- ular learning rule?" The purpose here then is to study and characterize the resulting learning dynamics, as in e.g., [13, 15], rather than on designing new learning algorithms. It is well-known that the resulting "local decision process" could be non-Markovian from each agent's perspective [39]. This is a feature, not a bug in descriptive work since it is a prop- erty of the real environment that the model captures. We use deep reinforcement learning as the basis for each agent in part because of its recent success with solving com- plex problems [25, 40]. Also, temporal difference predictions have been observed in the brain [41] and this class of rein- forcement learning algorithm is seen as a candidate theory of animal habit-learning [42]. 3.1 Deep Multiagent Reinforcement Learning Modern deep reinforcement learning methods take the perspective of an agent that must learn to maximize its cu- mulative long-term reward through trial-and-error interac- tions with its environment [43, 44]. In the multi-agent setting, the i-th agent stores a function Qi : Oi × Ai → R represented by a deep Q-network (DQN). See [25] for details in the single agent case. In our case the true state s is observed differently by each player, as oi = O(s, i). However for consistency of notation, we use a shorthand: Qi(s, a) = Qi(O(s, i), a). During learning, to encourage exploration we parameter- ize the i-th agent's policy by (cid:26) argmaxa∈Ai U(Ai) πi(s) = Qi(s, a) with probability 1 − with probability where U(Ai) denotes a sample from the uniform distribution over Ai. Each agent updates its policy given a stored batch1 of experienced transitions {(s, a, ri, s(cid:48))t : t = 1, . . . T} such that Qi(s, a) ← Qi(s, a) + α ) − Qi(s, a) (cid:48) Qi(s (cid:21) (cid:20) ri + γ max a(cid:48)∈Ai (cid:48) , a This is a "growing batch" approach to reinforcement learn- ing in the sense of [45]. However, it does not grow in an un- bounded fashion. Rather, old data is discarded so the batch can be constantly refreshed with new data reflecting more recent transitions. We compared batch sizes of 1e5 (our default) and 1e6 in our experiments (see Sect. 5.3). The network representing the function Q is trained through gra- dient descent on the mean squared Bellman residual with the expectation taken over transitions uniformly sampled from the batch (see [25]). Since the batch is constantly refreshed, the Q-network may adapt to the changing data distribution arising from the effects of learning on π1 and π2. In order to make learning in SSDs tractable, we make the extra assumption that each individual agent's learning depends only on the other agent's learning via the (slowly) changing distribution of experience it generates. That is, the two learning agents are "independent" of one another and each regard the other as part of the environment. From the perspective of player one, the learning of player two shows up as a non-stationary environment. The independence as- sumption can be seen as a particular kind of bounded ratio- nality: agents do no recursive reasoning about one another's learning. In principle, this restriction could be dropped through the use of planning-based reinforcement learning methods like those of [24]. 4. SIMULATION METHODS Both games studied here were implemented in a 2D grid- world game engine. The state st and the joint action of all players (cid:126)a determines the state at the next time-step st+1. Observations O(s, i) ∈ R3×16×21 (RGB) of the true state st depended on the player's current position and orientation. The observation window extended 15 grid squares ahead and 10 grid squares from side to side (see Fig. 3B). Ac- tions a ∈ R8 were agent-centered: step forward, step back- ward, step left, step right, rotate left, rotate right, use beam and stand still. Each player appears blue in its own local view, light-blue in its teammates view and red in its oppo- nent's view. Each episode lasted for 1, 000 steps. Default neural networks had two hidden layers with 32 units, in- terleaved with rectified linear layers which projected to the output layer which had 8 units, one for each action. During training, players implemented epsilon-greedy policies, with epsilon decaying linearly over time (from 1.0 to 0.1). The default per-time-step discount rate was 0.99. 5. RESULTS In this section, we describe three experiments: one for each game (Gathering and Wolfpack), and a third experi- ment investigating parameters that influence the emergence of cooperation versus defection. 5.1 Experiment 1: Gathering The goal of the Gathering game is to collect apples, repre- sented by green pixels (see Fig. 3A). When a player collects Figure 4: Social outcomes are influenced by envi- ronment parameters. Top: Gathering. Shown is the beam-use rate (aggressiveness) as a function of re-spawn time of apples Napple (abundance) and re- spawn time of agents Ntagged (conflict-cost). These results show that agents learn aggresssive policies in environments that combine a scarcity of resources with the possibility of costly action. Less aggressive policies emerge from learning in relatively abundant environments with less possibility for costly action. Bottom: Wolfpack. Shown is two minus the aver- age number of wolves per capture as a function of the capture radius and group capture benefit (rteam). Again as expected, greater group benefit and larger capture radius lead to an increase in wolves per cap- ture, indicating a higher degree of cooperation. an apple it receives a reward of 1 and the apple is temporar- ily removed from the map. The apple respawns after Napple frames. Players can direct a beam in a straight line along their current orientation. A player hit by the beam twice is "tagged" and removed from the game for Ntagged frames. No rewards are delivered to either player for tagging. The only potential motivation for tagging is competition over the apples. Refer to the Gathering gameplay video2 for demon- stration. Intuitively, a defecting policy in this game is one that is aggressive-i.e., involving frequent attempts to tag rival players to remove them from the game. Such a policy is mo- tivated by the opportunity to take all the apples for oneself that arises after eliminating the other player. By contrast, a cooperative policy is one that does not seek to tag the other player. This suggests the use of a social behavior met- 1The batch is sometimes called a "replay buffer" e.g. [25]. 2https://goo.gl/2xczLc ric (section 2.2) that measures a policy's tendency to use the beam action as the basis for its classification as defec- tion or cooperation. To this end, we counted the number of beam actions during a time horizon and normalized it by the amount of time in which both agents were playing (not removed from the game). By manipulating the rate at which apples respawn after being collected, Napple, we could control the abundance of apples in the environment. Similarly, by manipulating the number of timesteps for which a tagged agent is removed from the game, Ntagged, we could control the cost of po- tential conflict. We wanted to test whether conflict would emerge from learning in environments where apples were scarce. We considered the effect of abundance (Napple) and conflict-cost (Ntagged) on the level of aggressiveness (beam- use rate) that emerges from learning. Fig. 4A shows the beam-use rate that evolved after training for 40 million steps as a function of abundance (Napple) and conflict-cost (Ntagged). Supplementary video 3 shows how such emergent conflict evolves over the course of learning. In this case, differences in beam-use rate (proxy for the tendency to defect) learned in the different environments emerge quite early in training and mostly persist throughout. When learning does change beam-use rate, it is almost always to increase it. We noted that the policies learned in environments with low abundance or high conflict-cost were highly aggressive while the policies learned with high abundance or low conflict- cost were less aggressive. That is, the Gathering game pre- dicts that conflict may emerge from competition for scarce resources, but is less likely to emerge when resources are plentiful. 1 , πD 2 , πD 1 ) and (πC To further characterize the mixed motivation structure of the Gathering game, we carried out the empirical game- theoretic analysis suggested by the definition of section 2.2. We chose the set of policies ΠC that were trained in the high abundance / low conflict-cost environments (low aggression policies) and ΠD as policies trained in the low abundance and high conflict-cost environments (high aggression poli- cies), and used these to compute empirical payoff matrices as follows. Two pairs of policies (πC 2 ) are sampled from ΠC and ΠD and matched against each other in the Gathering game for one episode. The resulting re- wards are assigned to individual cells of a matrix game, in which πC i corresponds the cooperative action for player i, and πD j , the defective action for player j. This process is repeated until convergence of the cell values, and generates estimates of R, P, S, and T for the game corresponding to each abundance / conflict-cost (Napple, Ntagged) level tested. See Figure 5 for an illustration of this workflow. Fig. 6A summarizes the types of empirical games that were found given our parameter spectrum. Most cases where the social dilemma inequalities (1) – (4) held, i.e., the strategic sce- nario was a social dilemma, turned out to be a prisoner's dilemma. The greed motivation reflects the temptation to take out a rival and collect all the apples oneself. The fear motivation reflected the danger of being taken out oneself by a defecting rival. P is preferred to S in the Gathering game because mutual defection typically leads to both play- ers alternating tagging one another, so each gets some time alone to collect apples. Whereas the agent receiving the out- come S does not try to tag its rival and thus never gets this chance. 5.2 Experiment 2: Wolfpack The Wolfpack game requires two players (wolves) to chase a third player (the prey). When either wolf touches the prey, all wolves within the capture radius (see Fig. 3B) receive a reward. The reward received by the capturing wolves is pro- portional to the number of wolves in the capture radius. The idea is that a lone wolf can capture the prey, but is at risk of losing the carcass to scavengers. However, when the two wolves capture the prey together, they can better protect the carcass from scavengers and hence receive a higher re- ward. A lone-wolf capture provides a reward of rlone and a capture involving both wolves is worth rteam. Refer to the Wolfpack gameplay video4 for demonstration. The wolves learn to catch the prey over the course of train- ing. Fig. 4B shows the effect on the average number of wolves per capture obtained from training in environments with varying levels of group capture bonus rteam/rlone and capture radius. Supplementary video 5 shows how this de- pendency evolves over learning time. Like in the Gathering game, these results show that environment parameters in- fluence how cooperative the learned policies will be. It is interesting that two different cooperative policies emerged from these experiments. On the one hand, the wolves could cooperate by first finding one another and then moving to- gether to hunt the prey, while on the other hand, a wolf could first find the prey and then wait for the other wolf to arrive before capturing it. Analogous to our analysis of the Gathering game, we choose ΠC and ΠD for Wolfpack to be the sets of policies learned in the high radius / group bonus and low radius /group bonus environments respectively. The procedure for esti- mating R, P, S, and T was the same as in section 5.1. Fig. 6B summarizes these results. Interestingly, it turns out that all three classic MGSDs, chicken, stag hunt, and prisoner's dilemma can be found in the empirical payoff matrices of Wolfpack. 5.3 Experiment 3: Agent parameters influenc- ing the emergence of defection So far we have described how properties of the environ- ment influence emergent social outcomes. Next we consider the impact of manipulating properties of the agents. Psy- chological research attempting to elucidate the motivational factors underlying human cooperation is relevant here. In particular, Social Psychology has advanced various hypothe- ses concerning psychological variables that may influence co- operation and give rise to the observed individual differences in human cooperative behavior in laboratory-based social dilemmas [2]. These factors include consideration-of-future- consequences [46], trust [47], affect (interestingly, it is nega- tive emotions that turn out to promote cooperation [48]), and a personality variable called social value orientation characterized by other-regarding-preferences. The latter has been studied in a similar Markov game social dilemma setup to our SSD setting by [49]. Obviously the relatively simple DQN learning agents we consider here do not have internal variables that directly cor- respond to the factors identified by Social Psychology. Nor 3https://goo.gl/w2VqlQ 4https://goo.gl/AgXtTn 5https://goo.gl/vcB8mU Figure 5: Workflow to obtain empirical payoff matrices from Markov games. Agents are trained under different environmental conditions, e.g., with high or low abundance (Gathering case) or team capture bonus (Wolfpack case) resulting in agents classified as cooperators (πC ∈ ΠC ) or defectors (πD ∈ ΠD). Empirical game payoffs are estimated by sampling (π1, π2) from ΠC × ΠC , ΠC × ΠD, ΠD × ΠC , and ΠD × ΠD. By repeatedly playing out the resulting games between the sampled π1 and π2, and averaging the results, it is possible to estimate the payoffs for each cell of the matrix. should they be expected to capture the full range of human individual differences in laboratory social dilemmas. Never- theless, it is interesting to consider just how far one can go down this road of modeling Social Psychology hypotheses using such simple learning agents6. Recall also that DQN is in the class of reinforcement learning algorithms that is generally considered to be the leading candidate theory of animal habit-learning [50, 42]. Thus, the interpretation of our model is that it only addresses whatever part of coop- erative behavior arises "by habit" as opposed to conscious deliberation. Experimental manipulations of DQN parameters yield con- sistent and interpretable effects on emergent social behavior. Each plot in Fig. 7 shows the relevant social behavior metric, conflict for Gathering and lone-wolf behavior for Wolfpack, as a function of an environment parameter: Napple, Ntagged (Gathering) and rteam/rlone (Wolfpack). The figure shows that in both games, agents with greater discount parame- ter (less time discounting) more readily defect than agents that discount the future more steeply. For Gathering this likely occurs because the defection policy of tagging the other player to temporarily remove them from the game only provides a delayed reward in the form of the increased op- portunity to collect apples without interference. However, when abundance is very high, even the agents with higher discount factors do not learn to defect. In such paradisia- cal settings, the apples respawn so quickly that an individ- ual agent cannot collect them quickly enough. As a con- sequence, there is no motivation to defect regardless of the temporal discount rate. Manipulating the size of the stored- and-constantly-refreshed batch of experience used to train each DQN agent has the opposite effect on the emergence of defection. Larger batch size translates into more experience with the other agent's policy. For Gathering, this means that avoiding being tagged becomes easier. Evasive action benefits more from extra experience than the ability to tar- get the other agent. For Wolfpack, larger batch size allows 6The contrasting approach that seeks to build more struc- ture into the reinforcement learning agents to enable more interpretable experimental manipulations is also interesting and complementary e.g., [24]. greater opportunity to learn to coordinate to jointly catch the prey. Possibly the most interesting effect on behavior comes from the number of hidden units in the neural network be- hind the agents, which may be interpreted as their cogni- tive capacity. Curves for tendency to defect are shown in the right column of Fig. 7, comparing two different network sizes. For Gathering, an increase in network size leads to an increase in the agent's tendency to defect, whereas for Wolfpack the opposite is true: Greater network size leads to less defection. This can be explained as follows. In Gathering, defec- tion behavior is more complex and requires a larger network size to learn than cooperative behavior. This is the case because defection requires the difficult task of targeting the opposing agent with the beam whereas peacefully collecting apples is almost independent of the opposing agent's behav- ior. In Wolfpack, cooperation behavior is more complex and requires a larger network size because the agents need to co- ordinate their hunting behaviors to collect the team reward whereas the lone-wolf behavior does not require coordina- tion with the other agent and hence requires less network capacity. Note that the qualitative difference in effects for network size supports our argument that the richer framework of SSDs is needed to capture important aspects of real social dilemmas. This rather striking difference between Gathering and Wolfpack is invisible to the purely matrix game based MGSD-modeling. It only emerges when the different com- plexities of cooperative or defecting behaviors, and hence the difficulty of the corresponding learning problems is modeled in a sequential setup such as an SSD. 6. DISCUSSION In the Wolfpack game, learning a defecting lone-wolf pol- icy is easier than learning a cooperative pack-hunting pol- icy. This is because the former does not require actions to be conditioned on the presence of a partner within the capture radius. In the Gathering game the situation is re- versed. Cooperative policies are easier to learn since they need only be concerned with apples and may not depend Figure 6: Summary of matrix games discovered within Gathering (Left) and Wolfpack (Right) through extracting empirical payoff matrices. The games are classified by social dilemma type indicated by color and quandrant. With the x-axis representing fear = P − S and the y-axis representing greed = T − R, the lower right quadrant contains Stag Hunt type games (green), the top left quadrant Chicken type games (blue), and the top right quadrant Prisoner's Dilemma type games (red). Non-SSD type games, which either violate social dilemma condition (1) or do not exhibit fear or greed are shown as well. on the rival player's actions. However, optimally efficient cooperative policies may still require such coordination to prevent situations where both players simultaneously move on the same apple. Cooperation and defection demand dif- fering levels of coordination for the two games. Wolfpack's cooperative policy requires greater coordination than its de- fecting policy. Gathering's defection policy requires greater coordination (to successfully aim at the rival player). Both the Gathering and Wolfpack games contain embed- ded MGSDs with prisoner's dilemma-type payoffs. The MGSD model thus regards them as structurally identical. Yet, viewed as SSDs, they make rather different predictions. This suggests a new dimension on which to investigate classic questions concerning the evolution of cooperation. For any to-be-modeled phenomenon, the question now arises: which SSD is a better description of the game being played? If Gathering is a better model, then we would expect coop- eration to be the easier-to-learn "default" policy, probably requiring less coordination. For situations where Wolfpack is the better model, defection is the easier-to-learn "default" behavior and cooperation is the harder-to-learn policy re- quiring greater coordination. These modeling choices are somewhat orthogonal to the issue of assigning values to the various possible outcomes (the only degree of freedom in MGSD-modeling), yet they make a large difference to the results. SSD models address similar research questions as MGSD models, e.g. the evolution of cooperation. However, SSD models are more realistic since they capture the sequential structure of real-world social dilemmas. Of course, in mod- eling, greater verisimilitude is not automatically virtuous. When choosing between two models of a given phenomenon, Occam's razor demands we prefer the simpler one. If SSDs were just more realistic models that led to the same conclu- sions as MGSDs then they would not be especially useful. This however, is not the case. We argue the implication of the results presented here is that standard evolutionary and learning-based approaches to modeling the trial and error process through which societies converge on equilibria of so- cial dilemmas are unable to address the following important learning related phenomena. 1. Learning which strategic decision to make, abstractly, whether to cooperate or defect, often occurs simulta- neously with learning how to efficiently implement said decision. 2. It may be difficult to learn how to implement an effec- tive cooperation policy with a partner bent on defection- or vice versa. 3. Implementing effective cooperation or defection may involve solving coordination subproblems, but there is no guarantee this would occur, or that cooperation and defection would rely on coordination to the same ex- tent. In some strategic situations, cooperation may require coordination, e.g., standing aside to allow a partner's passage through a narrow corridor while in others defection may require coordination e.g. block- ing a rival from passing. 4. Some strategic situations may allow for multiple dif- ferent implementations of cooperation, and each may require coordination to a greater or lesser extent. The same goes for multiple implementations of defection. Figure 7: Factors influencing the emergence of defecting policies. Top row: Gathering. Shown are plots of average beam-use rate (aggressiveness) as a function of Napple (scarcity) Bottom row: Wolfpack. Shown are plots of (two minus) average-wolves-per-capture (Lone-wolf capture rate) as a function of rteam (Group Benefit). For both Gathering and Wolfpack we vary the following factors: temporal discount (left), batch size (centre), and network size (right). Note that the effects of discount factor and batch size on the tendency to defect point in the same direction for Gathering and Wolfpack, network size has the opposite effect (see text for discussion.) 5. The complexity of learning how to implement effec- tive cooperation and defection policies may not be equal. One or the other might be significantly easier to learn-solely due to implementation complexity- in a manner that cannot be accounted for by adjusting outcome values in an MGSD model. Our general method of tracking social behavior metrics in addition to reward while manipulating parameters of the learning environment is widely applicable. One could use these techniques to simulate the effects of external inter- ventions on social equilibria in cases where the sequential structure of cooperation and defection are important. No- tice that several of the examples in Schelling's seminal book Micromotives and Macrobehavior [51] can be seen as tempo- rally extended social dilemmas for which policies have been learned over the course of repeated interaction, including the famous opening example of lecture hall seating behav- ior. It is also possible to define SSDs that model the extrac- tion of renewable vs non-renewable resources and track the sustainability of the emergent social behaviors while taking into account the varying difficulties of learning sustainable (cooperating) vs. non-sustainable (defecting) policies. Ef- fects stemming from the need to learn implementations for strategic decisions may be especially important for informed policy-making concerning such real-world social dilemmas. Acknowledgments The authors would like to thank Chrisantha Fernando, Toby Ord, and Peter Sunehag for fruitful discussions in the lead- up to this work, and Charles Beattie, Denis Teplyashin, and Stig Petersen for software engineering support. REFERENCES [1] Anatol Rapoport. Prisoner's dilemma–recollections and observations. In Game Theory as a Theory of a Conflict Resolution, pages 17–34. Springer, 1974. [2] Paul AM Van Lange, Jeff Joireman, Craig D Parks, and Eric Van Dijk. The psychology of social dilemmas: A review. Organizational Behavior and Human Decision Processes, 120(2):125–141, 2013. [3] Michael W Macy and Andreas Flache. Learning dynamics in social dilemmas. Proceedings of the National Academy of Sciences, 99(suppl 3):7229–7236, 2002. [4] Robert L. Trivers. The evolution of reciprocal altruism. Quarterly Review of Biology, pages 35–57, 1971. [5] Robert Axelrod. The Evolution of Cooperation. Basic Books, 1984. [6] Martin A Nowak and Karl Sigmund. Tit for tat in heterogeneous populations. Nature, 355(6357):250–253, 1992. [7] Martin Nowak, Karl Sigmund, et al. A strategy of win-stay, lose-shift that outperforms tit-for-tat in the prisoner's dilemma game. Nature, 364(6432):56–58, 1993. [8] Martin A Nowak and Karl Sigmund. Evolution of indirect reciprocity by image scoring. Nature, 393(6685):573–577, 1998. [9] Robert Axelrod. An evolutionary approach to norms. American political science review, 80(04):1095–1111, 1986. Scarcity0.040.060.080.100.12Aggressiveness+Discount: 0.99Discount: 0.995Scarcity0.040.050.060.070.080.090.100.110.12_Batch size: 1e+04Batch size: 1e+05Scarcity0.040.060.080.10+Network size: 16Network size: 64Group benefit0.540.560.580.600.620.640.660.680.70Lone-wolf capture rate+Discount: 0.99Discount: 0.995Group benefit0.500.550.600.650.700.75_Batch size: 1e+04Batch size: 1e+05Group benefit0.400.450.500.550.600.650.70_Network size: 16Network size: 64[10] Samhar Mahmoud, Simon Miles, and Michael Luck. Cooperation emergence under resource-constrained peer punishment. In Proceedings of the 2016 International Conference on Autonomous Agents & Multiagent Systems, pages 900–908. International Foundation for Autonomous Agents and Multiagent Systems, 2016. [11] T.W. Sandholm and R.H. Crites. Multiagent reinforcement learning in the iterated prisoner's dilemma. Biosystems, 37(1–2):147–166, 1996. and joint intentions in social interaction. In Proceedings of the 38th Annual Conference of the Cognitive Science Society, 2016. [25] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski, S. Petersen, C. Beattie, A. Sadik, I. Antonoglou, H. King, D. Kumaran, D. Wierstra, S. Legg, and D. Hassabis. Human-level control through deep reinforcement learning. Nature, 518(7540):529–533, 2015. [12] Enrique Munoz de Cote, Alessandro Lazaric, and [26] Y. Shoham, R. Powers, and T. Grenager. If Marcello Restelli. Learning to cooperate in multi-agent social dilemmas. In Proceedings of the Fifth International Joint Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2006. [13] M. Wunder, M. Littman, and M. Babes. Classes of multiagent Q-learning dynamics with greedy exploration. In Proceedings of the 27th International Conference on Machine Learning, 2010. [14] Erik Zawadzki, Asher Lipson, and Kevin Leyton-Brown. Empirically evaluating multiagent learning algorithms. CoRR, abs/1401.8074, 2014. multi-agent learning is the answer, what is the question? Artificial Intelligence, 171(7):365–377, 2007. [27] M. G. Lagoudakis and R. Parr. Value function approximation in zero-sum Markov games. In Proceedings of the 18th Conference on Uncertainty in Artificial Intelligence (UAI), pages 283–292, 2002. [28] J. P´erolat, B. Scherrer, B. Piot, and O. Pietquin. Approximate dynamic programming for two-player zero-sum Markov games. In Proceedings of the International Conference on Machine Learning (ICML), 2015. [15] Daan Bloembergen, Karl Tuyls, Daniel Hennes, and [29] J. P´erolat, B. Piot, M. Geist, B. Scherrer, and Michael Kaisers. Evolutionary dynamics of multi-agent learning: A survey. Journal of Artificial Intelligence Research, 53:659–697, 2015. [16] Martin A Nowak and Robert M May. Evolutionary games and spatial chaos. Nature, 359(6398):826–829, 1992. [17] Chao Yu, Minjie Zhang, Fenghui Ren, and Guozhen Tan. Emotional multiagent reinforcement learning in spatial social dilemmas. IEEE Transactions on Neural Networks and Learning Systems, 26(12):3083–3096, 2015. [18] Hisashi Ohtsuki, Christoph Hauert, Erez Lieberman, and Martin A Nowak. A simple rule for the evolution of cooperation on graphs and social networks. Nature, 441(7092):502–505, 2006. [19] Francisco C Santos and Jorge M Pacheco. A new route to the evolution of cooperation. Journal of Evolutionary Biology, 19(3):726–733, 2006. [20] William E Walsh, Rajarshi Das, Gerald Tesauro, and Jeffrey O Kephart. Analyzing complex strategic interactions in multi-agent systems. In AAAI-02 Workshop on Game-Theoretic and Decision-Theoretic Agents, pages 109–118, 2002. [21] Michael Wellman. Methods for empirical game-theoretic analysis (extended abstract). In Proceedings of the National Conference on Artificial Intelligence (AAAI), pages 1552–1555, 2006. [22] M. L. Littman. Markov games as a framework for multi-agent reinforcement learning. In Proceedings of the 11th International Conference on Machine Learning (ICML), pages 157–163, 1994. [23] Ann Now´e, Peter Vrancx, and Yann-Michael De Hauwere. Game theory and multiagent reinforcement learning. In Marco Wiering and Martijn van Otterlo, editors, Reinforcement Learning: State-of-the-Art, chapter 14. Springer, 2012. O. Pietquin. Softened approximate policy iteration for Markov games. In Proceedings of the International Conference on Machine Learning (ICML), 2016. [30] Branislav Bosansk´y, Viliam Lis´y, Marc Lanctot, Jir´ı Cerm´ak, and Mark H.M. Winands. Algorithms for computing strategies in two-player simultaneous move games. Artificial Intelligence, 237:1–40, 2016. [31] M. Zinkevich, A. Greenwald, and M. Littman. Cyclic equilibria in Markov games. In Neural Information Processing Systems, 2006. [32] J. Hu and M. P. Wellman. Multiagent reinforcement learning: Theoretical framework and an algorithm. In Proceedings of the 15th International Conference on Machine Learning (ICML), pages 242–250, 1998. [33] A. Greenwald and K. Hall. Correlated-Q learning. In Proceedings of the 20th International Conference on Machine Learning (ICML), pages 242–249, 2003. [34] Michael Littman. Friend-or-foe Q-learning in general-sum games. In Proceedings of the Eighteenth International Conference on Machine Learning, pages 322–328, 2001. [35] J. P´erolat, B. Piot, B. Scherrer, and O. Pietquin. On the use of non-stationary strategies for solving two-player zero-sum Markov games. In Proceedings of the 19th International Conference on Artificial Intelligence and Statistics, 2016. [36] Piotr J Gmytrasiewicz and Prashant Doshi. A framework for sequential planning in multi-agent settings. Journal of Artificial Intelligence Research, 24:49–79, 2005. [37] Pradeep Varakantham, Jun-young Kwak, Matthew E Taylor, Janusz Marecki, Paul Scerri, and Milind Tambe. Exploiting coordination locales in distributed POMDPs via social model shaping. In Proceedings of the 19th International Conference on Automated Planning and Scheduling, ICAPS, 2009. [24] Max Kleiman-Weiner, M K Ho, J L Austerweil, [38] Raphen Becker, Shlomo Zilberstein, Victor Lesser, and Michael L Littman, and Josh B Tenenbaum. Coordinate to cooperate or compete: abstract goals Claudia V Goldman. Solving transition independent decentralized Markov decision processes. Journal of Artificial Intelligence Research, 22:423–455, 2004. [39] Guillaume J. Laurent, Laetitia Matignon, and N. Le Fort-Piat. The world of independent learners is not Markovian. Int. J. Know.-Based Intell. Eng. Syst., 15(1):55–64, 2011. [40] David Silver, Aja Huang, Chris J. Maddison, Arthur Guez, Laurent Sifre, George van den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy Lillicrap, Madeleine Leach, Koray Kavukcuoglu, Thore Graepel, and Demis Hassabis. Mastering the game of Go with deep neural networks and tree search. Nature, 529:484–489, 2016. [41] W. Schultz, P. Dayan, and P.R. Montague. A neural substrate of prediction and reward. Science, 275(5306):1593–1599, 1997. [42] Y. Niv. Reinforcement learning in the brain. The Journal of Mathematical Psychology, 53(3):139–154, 2009. [43] Richard S. Sutton and Andrew G. Barto. Introduction to Reinforcement Learning. MIT Press, 1998. [44] Michael L Littman. Reinforcement learning improves behaviour from evaluative feedback. Nature, 521(7553):445–451, 2015. [45] Sascha Lange, Thomas Gabel, and Martin Riedmiller. Batch reinforcement learning. In Reinforcement learning, pages 45–73. Springer, 2012. [46] Katherine V Kortenkamp and Colleen F Moore. Time, uncertainty, and individual differences in decisions to cooperate in resource dilemmas. Personality and Social Psychology Bulletin, 32(5):603–615, 2006. [47] Craig D Parks and Lorne G Hulbert. High and low trusters' responses to fear in a payoff matrix. Journal of Conflict Resolution, 39(4):718–730, 1995. [48] Hui Bing Tan and Joseph P Forgas. When happiness makes us selfish, but sadness makes us fair: Affective influences on interpersonal strategies in the dictator game. Journal of Experimental Social Psychology, 46(3):571–576, 2010. [49] Joseph L. Austerweil, Stephen Brawner, Amy Greenwald, Elizabeth Hilliard, Mark Ho, Michael L. Littman, James MacGlashan, and Carl Trimbach. How other-regarding preferences can promote cooperation in non-zero-sum grid games. In Proceedings of the AAAI Symposium on Challenges and Opportunities in Multiagent Learning for the Real World, 2016. [50] Nathaniel D Daw, Yael Niv, and Peter Dayan. Uncertainty-based competition between prefrontal and dorsolateral striatal systems for behavioral control. Nature neuroscience, 8(12):1704–1711, 2005. [51] Thomas C. Schelling. Micromotives and macrobehavior. WW Norton & Company, 1978 Rev. 2006.
1607.00695	1	1607	2016-07-03T22:44:57	Can we reach Pareto optimal outcomes using bottom-up approaches?	[ "cs.MA", "cs.AI", "cs.GT" ]	Traditionally, researchers in decision making have focused on attempting to reach Pareto Optimality using horizontal approaches, where optimality is calculated taking into account every participant at the same time. Sometimes, this may prove to be a difficult task (e.g., conflict, mistrust, no information sharing, etc.). In this paper, we explore the possibility of achieving Pareto Optimal outcomes in a group by using a bottom-up approach: discovering Pareto optimal outcomes by interacting in subgroups. We analytically show that Pareto optimal outcomes in a subgroup are also Pareto optimal in a supergroup of those agents in the case of strict, transitive, and complete preferences. Then, we empirically analyze the prospective usability and practicality of bottom-up approaches in a variety of decision making domains.	cs.MA	cs	Can we reach Pareto optimal outcomes using bottom-up approaches? Victor Sanchez-Anguix1 and Reyhan Aydogan2,4 and Tim Baarslag3 and Catholijn M. Jonker4 1 Coventry University, Coventry, United Kingdom, [email protected] 2 Ozyegin University, Istanbul,Turkey, [email protected] 3 Centrum Wiskunde & Informatica, Amsterdam, Netherlands, [email protected] 4 Technical University of Delft, Delft, Netherlands, [email protected] Abstract. Traditionally, researchers in decision making have focused on attempting to reach Pareto Optimality using horizontal approaches, where optimality is calculated taking into account every participant at the same time. Sometimes, this may prove to be a difficult task (e.g., con- flict, mistrust, no information sharing, etc.). In this paper, we explore the possibility of achieving Pareto Optimal outcomes in a group by using a bottom-up approach: discovering Pareto optimal outcomes by interact- ing in subgroups. We analytically show that Pareto optimal outcomes in a subgroup are also Pareto optimal in a supergroup of those agents in the case of strict , transitive, and complete preferences. Then, we em- pirically analyze the prospective usability and practicality of bottom-up approaches in a variety of decision making domains. Keywords: pareto optimality, agreement technologies, group decision making, multi-agent systems, artificial intelligence 1 Introduction Group decision making has been studied within different disciplines with aim of reaching a mutually acceptable outcome. One of the desired properties of that outcome is Pareto optimality. However, reaching Pareto optimal agreements is not straightforward in practice. In open and dynamic environments, decision makers may not know each other's preferences completely. It may even be the case that it becomes more complicated to find Pareto optimal solutions when the number of participants increases, as the number of interactions required to achieve an optimal deal for the group may increase due to internal conflicts or lack of trust. A number of works in the field focus on finding a global Pareto optimal solution by involving all agents at the same time [11,16,10,29], which may lead to complicated interactions and lengthy decision making processes. However, we believe that, in many situations, agents can benefit from taking a bottom-up approach: calculating Pareto optimal outcomes in subgroups. In other words, we pursue the question of whether or not it is possible to estimate some Pareto optimal outcomes without knowing or predicting the preferences of all agents. In essence, solving the Pareto optimal set problem in a smaller group may be less complicated than in larger groups (e.g., less privacy concerns, less interactions needed, more willingness to cooperate, etc.) and it may provide a relatively important ratio of the final Pareto Optimal outcomes. Such kind of property can be used in some complex group decision making scenarios. Imagine that a group of agents is negotiating in unison with an unknown opponent [22,27,25]. If the agents can find the Pareto optimal outcomes within the team, they may use these outcomes in their bidding strategy to reach a Pareto optimal agreement with their opponent. In this paper we explore bottom-up strategies. For that, first we prove that any Pareto optimal outcome in a subgroup is also Pareto optimal in a larger group that contains the subgroup, as long as agents' preferences are strict linear order. Second, we empirically simulate how bottom-up approaches may perform in realistic scenarios. More specifically, we show that we can obtain a reasonable ratio of the Pareto optimal outcomes within a group of agents by only finding the Pareto optimal outcomes within the subgroup of these agents. The remainder of this paper is organized as follows: first we present a proof of how Pareto optimal solutions in subgroups are also Pareto optimal in larger groups when agents have strict, transitive, and complete preferences. Section 3 discusses some of the implications of the proof, and how it can be applied to solve a wide variety of problems in multi-agent systems. In Section 4, we empir- ically validate the theory in practice and analyze empirically compare the ratio of Pareto optimal outcomes within subgroups to the Pareto optimal outcomes within the entire group in a wide variety of real domains. After discussing the related work, we finally conclude the paper with future lines of work. 2 Pareto optimality in subgroups In this section we prove that any Pareto optimal outcome in a subgroup of agents is also Pareto optimal in any group of agents containing the subgroup. First, we provide some of the necessary definitions and introduce some notation. Let A = {a1, ..., an} be a set of agents where k is the index of agent ak and A(cid:48) = {a1, ..., am} be a superset of A, A ⊂ A(cid:48) where m > n. O is the set of all possible solutions in a given domain, and o ∈ O represents a possible solution in the domain. We assume that (cid:23)i represents agent's ai preference relation over outcomes in O. If o (cid:23)i o(cid:48) then agent ai likes o at least as well as o(cid:48), we write o (cid:31)i o(cid:48) to denote a strict preference for o and o = o(cid:48) to denote indifference. We assume that the agents' preference relations are strict, transitive and complete. An outcome o∗ is Pareto optimal with respect to A and O, denoted by po(o∗,A,O) iff (cid:64)o ∈ O ∃j ≤ n o (cid:23)i o∗ ∧ o (cid:31)j o∗. n(cid:94) i=1 We denote the set of all Pareto optimal outcomes over A by O∗ A = {o∗ ∈ O po(o∗,A,O)} . A ⊂ O∗ A(cid:48). Theorem 1. Given a set of outcomes O. For all two sets of agents A and A(cid:48), if A ⊂ A(cid:48), then O∗ Proof. Let us assume by reductio ad absurdum that A ⊂ A(cid:48), but O∗ This means there exists an o∗ ∈ O∗ of Pareto optimal outcomes, we have A (cid:54)⊂ O∗ A(cid:48). A(cid:48). Expanding the definition A such that o∗ /∈ O∗ o∗ /∈ {o ∈ O (cid:64)o(cid:48) ∈ O ∃k ≤ m, o(cid:48) (cid:23)i o ∧ o(cid:48) (cid:31)k o}. m(cid:94) i=1 m(cid:86) This means there must exist an o ∈ O and a k ≤ m such that We consider two scenarios: either ak ∈ A or ak /∈ A. -- If ak ∈ A then o is an outcome that dominates o∗ over A, which is not o (cid:23)i o∗∧o (cid:31)k o∗. i=1 possible as o∗ is Pareto optimal over A. n(cid:86) -- Otherwise, k > n, so we have o (cid:23)i o∗. In that case, as o∗ is Pareto optimal over A, the condition is only true if all of the agents in A are indifferent between o and o∗. As preferences are strict, that cannot be true either. i=1 Since both sides lead to a contradiction, we have proven the theorem. At this point the reader may be wondering how the theorem above behaves in a scenario where agents' preferences are not strict. As we will discuss later, the likeliness of such as scenario is small, but the conclusion of the theorem above may in fact not hold in that case. Basically, an outcome that is Pareto optimal in a subgroup A may not be Pareto optimal in the group A(cid:48) when all of the agents in A are indifferent between such outcome and another Pareto optimal outcome. Then, one of the two outcomes may not be Pareto optimal with A(cid:48) when one of the agents in the group is not indifferent between those outcomes. Nevertheless, as we shall outline in Section 4.2, such situations are rare in practice, as all of the agents need to be indifferent between outcomes. This becomes increasingly unlikely as the group size grows and thus, for large enough groups, we can consider that the theorem is true for practically any scenario. 3 Prospective applications In Section 2 we have demonstrated that an outcome that is Pareto optimal in a subgroup of agents will also remain Pareto optimal in a larger group5. It should be highlighted that we are not depicting achieving Pareto optimality as a simple task. However, there is value in computing Pareto optimality in smaller groups as long as we are able to use those solutions in more challenging scenarios: 5 For strict, transitive, and complete preferences -- Negotiation teams: In this scenario, a group of individuals negotiate as a party with opponent(s) to achieve a deal [22,27,26,25,23]. In that case, finding the outcomes that are Pareto optimal within the team may play in favor of the team as (i) if the team sticks to these outcomes while negotiating with opponents, it can ensure efficiency in the final outcome, (ii) the set of calculated deals may be reused in multiple negotiations with different opponents as they remain Pareto optimal, and (iii) finding Pareto optimal outcomes once may reduce the time spent in negotiation threads as the team exactly knows which outcomes are more beneficial for team members. On top of that, one can also assume that team members may be more willing to share information with teammates, which may make easier the search for Pareto optimal outcomes inside the team. -- Multi-party negotiations: Some participants in a multi-party negotia- tion [29,6,3,8,10,11] may decide to collude and bias the agreement with their preferences. For that, the subgroup of agents may calculate Pareto optimal outcomes within the subgroup, and decide on the Pareto optimal outcomes that they plan to use in the upcoming multi-party negotiation. This way, there may be higher probabilities for the negotiation to finish with an out- come that satisfies the subgroups' interests and that is efficient. Another possible application for this proof in multi-party settings is precisely the idea of looking for Pareto optimal agreements within subgroups of agents. For instance, agents with high degrees of trust may decide to share some information that facilitates the search of Pareto optimal outcomes within the subgroup. Then, once outcomes are found in subgroups, these may be shared among all of the agents, and the whole group may need to decide on the most appropriate Pareto optimal outcome. -- Decision making in open environments: Open multi-agent systems [2,24,12] have the particularity of being systems where agents enter and leave the system dynamically. In such environments, decision making tasks may suffer from the same characteristic and agents may enter and leave decision making tasks as needed, resulting in a real time problem. For those situa- tions, agents in a decision making task may benefit from a continuous search for Pareto optimal outcomes. As new agents join the task, those Pareto opti- mal outcomes calculated should be kept as they will remain Pareto optimal in the new group. When agents leave, remaining group members can get rid of some outcomes that have become dominated in the new setting. As the reader may have noticed, the range of applications where this ap- proach could be applied is varied. We are not claiming that those are the sole applications for this approach, and there may be others in domains like social choice, group recommendations, and so forth. 4 Experimental study Section 2 shows theoretically that Pareto optimal outcomes within a group of agents having complete, transitive and strict preferences are still Pareto optimal when the group size increases with incoming agents. Even when preferences are non strict, we expect for the theorem to hold in most of the cases. In this section, we empirically analyze the prospective performance and applicability of bottom- up approaches. For that purpose, we selected a variety of domains: -- Sushi domain: 5000 preference profiles over 10 types of sushi [15]. -- AGH course selection: 153 students' preferences over 6 courses offered by AGU University of Science and Technology in 2004 [28]. -- Book crossing domain: The original dataset contains preferences of 278,858 users that produced 1,149,780 ratings over 271,379 books [30]. In order to calculate Pareto optimality, we require preferences to be complete on at least a subset of items. We kept 7 users that had rated 23 books in common. -- Movielens domain: The original dataset contains 138,000 users that pro- vided ratings over 27,000 movies [19]. As we require complete preferences, we picked 10 preference profiles that had rated a total of 298 movies in common. -- Holiday domain: A multi-party negotiation domain available in Genius [18]. In this scenario, participants need to decide on the details of a holiday trip. In total, 9 preference profiles over 1024 possible outcomes are avail- able. These preferences have been elicited from TU Delft computer science students, but not with serious plans for a joint holiday in mind. -- Symposium domain: Another multi-party negotiation domain that is avail- able in Genius [18] concerning the organization of a conference. There are 9 preference profiles over 2305 possible outcomes. These preferences have been elicited from faculty members in computer science of TU Delft experienced in organizing conferences, but not having a specific conference in mind. -- Party domain: Another multi-party negotiation domain, where agents de- cide on the details of a party [18]. We elicited preferences from students in a Master level AI course. Students were asked to input their real preferences via Genius based on their tastes for hosting parties. In total, we elicited 24 real preference profiles over 3072 outcomes. From a global perspective, the sushi, agh, and book domain are small attend- ing to the number of outcomes. These domains correspond to decision making domains where outcomes are non customizable objects (e.g., a movie, a book, a course, etc.). The data in the Movielens domain is less sparse and we were able to find 10 users that had rated 298 outcomes in common. This is again a domain where outcomes are non customizable, but the size of the domain is one order of magnitude larger than that of the small domains. The three remaining multi-party negotiation domains (i.e., holiday, symposium, and party domain) represent scenarios where the final outcome can be customized via the negotiable issues. As a result, the number of possible outcomes is larger. We consider these domains and the Movielens domain as the large domains in our study. 4.1 Validation and Performance Analysis Our performance metric is the ratio of the Pareto optimal outcomes within a subgroup with a size of {2, ..., n-1} to the Pareto optimal outcomes within the n-sized group. If the ratio remains low even for large subgroups, then this means that the performance of our theoretical finding may be of little value in practice, as only a small ratio of the final Pareto outcomes may be achievable. However, if the ratio is large, then it may indicate that bottom-up approaches may be valuable. Additionally, common sense indicates that, the larger the subgroup, the higher the ratio of final Pareto optimal outcomes that may be obtained in the subgroup. However, one question that arises is the actual speed by which the ratio of final Pareto outcomes increases, and whether or not subgroups may be able to calculate a respectable ratio of the final Pareto optimal outcomes. For testing the practical performance of our bottom-up approaches, we ran- domly generated groups of size n based on the preference profiles available for each domain. For each randomly generated group, we built all possible sub- groups with varying sizes {2, .., n − 1} and estimated the Pareto optimal set in each (sub)group. More specifically, for each domain we tested a maximum of 1000 groups6 of size n = {5, 7, 9}. Fig. 1. Average ratio of the final Pareto optimal obtained in subgroups of different size 6 The total number is min(1000,(cid:0)m (cid:1)), where m is the total number of available pref- erence profiles and n is the size of the group n The results of this experiment can be observed in Figure 1. As expected, the results show that the larger the subgroup is, the larger the average ratio of the final Pareto Optimal set that we get. The increase is clearly continuous for all of the domains and group sizes. When we look at the results for groups of 7 and 9 members we observe a non-linear increase with the size of the subgroup. This non-linear increase is not as evident in the case of 5 members' groups, as in that case we only have 3 data points7. One should highlight that for n − 1 agents in the subgroup, n being the total number of agents in the group, the average ratio of the Pareto optimal set obtained in the subgroup is always over 50% of the final set, being close to 80% in some cases (e.g, smaller domains, larger groups). This is a good result, especially for negotiation team scenarios [22,27,25], where the team could calculate the Pareto set inside the team and use those outcomes in the negotiation with an opponent. This is a clear case where a subgroup of size n− 1 can be formed (i.e., all of the team members) and, according to the experimental results, obtain a notably high ratio of final Pareto optimal outcomes. Consequently, they can propose Pareto optimal bids without knowing their opponent's preferences. The result is also notable for smaller subgroups. For instance, in groups of size 5, we are able to obtain between an average of 68% of the final Pareto set for small domains and 32% for the larger domains with just about half of the group members (i.e., 3). In the case of groups of size 7, we get 68% of the final Pareto set for small domains and 28% for larger domains with just about half of the group members (i.e., 4). Similarly, for groups of size 9 we are able to obtain an average of 76% of the final Pareto set in small domains, and 30% in large domains with just about half of group members (i.e., 5). The trends in the graphics and the results mentioned above may also suggest that larger domains may result in lower ratios of the final Pareto optimal set achievable by subgroups. Nevertheless, as we have been able to observe above, the results can still be considered as positive. Although the current results are promising, we would like to test a wider range of domains and domain sizes to strengthen the results of this study. 4.2 Applicability Analysis There are still other aspects that we need to analyze to determine the applicabil- ity of bottom-up approaches in real situations. Even though considerable ratios of the final Pareto optimal set are obtainable within subgroups, this may be useless in practice if the total number of Pareto optimal outcomes is very close to all possible outcomes. In those cases, there would be no point in calculating Pareto optimal outcomes in subgroups, as almost any outcome would be Pareto optimal. Therefore, we are interested in checking that the set of final Pareto op- timal outcomes does not dramatically approach the total number of outcomes. In [20], O'Neill studied how Pareto optimality was affected by the number of agents participating in a decision making process. To put it simply, the author 7 Even non-linear functions may look like linear when the number of points is reduced proved that the number of Pareto optimal outcomes grows exponentially with the number of agents, with the assumption that all preference profiles are equally probable. Additionally, he proposed a formula to estimate the number of out- comes that are expected to be Pareto optimal based on the size of the domain m, and the number of agents in the group n: E(Km,n) = − m(cid:80) (−1)i(cid:0)m (cid:1) 1 in−1 . i i=1 He also stated that the size of the domain had an effect on the number of outcomes that were Pareto optimal: larger outcome spaces tend to slow down the exponential growth of the Pareto optimal set, although the growth is still exponential. Of course, for drawing such a conclusion, the author had to assume that all preference profiles were equally probable. We argue that, in practice, all preference profiles are not equally probable as in some domains not all of the outcomes may be equally feasible (e.g., high prices in a team of buyers, popular choices in movies, popular choices in travel destinations, etc.). Hence, we argue that the exponential growth may not be as fast as in the theoretical case, and bottom-up approaches may be applicable to more scenarios. In order to examine this theoretical finding in practice, we calculated the ratio of the Pareto optimal outcomes to the total number of outcomes for each domain and group size. Figure 2 shows the average ratio of outcomes that are Pareto optimal for different groups sizes and domains. In these graphs, blue dots represent the average ratios calculated in real scenarios while green dots denote the theoretical estimation provided by [20] for domains of the same size. In addition to this, for each data point we provide the total number of cases8 that were considered for calculating the average. Numbers in red represent less than 30 samples and such averages should be ignored. As it can be observed in Figure 2, the growth in the number of outcomes that are Pareto optimal is usually slower in real domains than in the theoretical esti- mation. Being more specific, we observe that only the symposium domain shows a similar behavior to that of the theoretical case. The rest of the domains deviate from the theoretical behavior sooner or later, showing a slower saturation. We can observe that this difference is specially acute in the Movielens, Book, Sushi, and Agh domain, which are the ones whose preferences have been rigorously elicited from real users (except for the party domain). This may reinforce our initial intuition, that, in real domains, the exponential growth on the number of Pareto optimal outcomes may not be as drastic as in the theoretical case. In other domains like the party and the holiday domain, the difference is less acute but still existent. In fact, if one analyzes the proposed domains one by one, it is possible to realize that there are general preferential trends. This is clear in domains like Movielens or the Book domain, where we know that some movies and some books tend to be more popular than others. For instance, The Shawshank Redemption is one of the most popular movies of all times, and it has been able to obtain average ratings of 9.3 over 10 stars in sites like IMDB.9, where it has been voted 8 Again, the total number is min(1000,(cid:0)m (cid:1)) n 9 http://www.imdb.com. Visited on 16th November 2015 Fig. 2. Average ratio of the final Pareto optimal set obtained with subgroups of differ- ent size by more than 1 million users. Similarly, we can find books like Harry Potter and the Deathly Hallows that have received an average rating of 4.59 over 5 stars with more than 1 million ratings on sites like GoodReads 10. Finding users that did not like these items has low odds, and as a consequence we can state that not all preference profiles are equally probable. Not only there are general trends in users preferences, but many times we find that there are clusters of users with similar preferences. For instance, in the book domain, we can expect that users that have liked The Lord of the Rings will also like other fantasy themed books like Song of Ice and Fire. This is the type of patterns exploited by recommender systems, and suggests that the number of likely preference profiles is even smaller. With respect to the other small domains (e.g., AGH, Sushi), we analyzed the preferences of users. In fact, for analyzing the preferences of users on items we performed a Borda count with all of the preference profiles. We could observe that, in the Sushi domain, there are also some popular choices the toro (a total score of 39445) and some choices that are usually the least liked by users like the kappa-maki (a total score of 14928). In the case of the AGH domain, we could also observe that one of the courses (e.g., course 3) was the most preferred one with a score of 731, whereas the least preferred score had almost half the score. 10 http://www.goodreads.com/. Visited on 16th November 2015 This means that in these domains, preferences are not equally distributed and one should not expect such an exponential growth as in the theoretical case. With respect to negotiation domains, we elicited real preferences from the Party domain, whereas we used the preference profiles provided by Genius in the Holiday and Symposium domain. Interestingly, we could observe that real users in the Party domain tend to consider the type of food, the type of drinks, and the music as the most important attributes. Even in some specific attributes, we could find that there were popular choices like for instance Beer only for drinks, and Finger-food and Chips and Nuts for food choices. With respect to the rest of negotiation domains, it has to be considered that they were not strictly and rigor- ously elicited like in the case of the party domain. Users were not contextualized in a specific scenario and their preferences were just elicited from their previous experiences in similar scenarios. In the case of the Holiday domain, we were able to observe some patterns like users considering the duration and the activities as the most important attributes. The users usually preferred longer durations to shorter durations, and we observed a slight positive inclination towards His- torical Places and Restaurants. Even in the rest of less important attributes we were able to find some patterns like the fact that most users preferred Mi- ami and Amsterdam as destinations. These patterns again show that not all preference profiles are equally likely, and that is reflected in the fact that the experimental growth depicted for Figure 2 is slower than the theoretical growth. On the other hand, the Symposium domain is closer to the theoretical expec- tation. This may be explainable due to the fact that the Symposium domain preferences were not elicited thinking on an specific symposium. In contrast to the Holiday domain, which did not follow a rigorous preference elicitation pro- cess either, in the Symposium domain it is harder to relate to the scenario, as it includes totally fictional speakers (e.g. Mr. Talkolot), whereas in the holiday domain one always can think about his/her own preferences on a trip. This may explain why the increase in the ratio of Pareto optimal outcomes is similar to the theoretical case where preference profiles are equally probable. It should be highlighted that in many negotiation domains, preferences are made different to test the performance of negotiation algorithms in conflicting scenarios. The fact that, as we have shown, not all preference profiles are equally likely makes bottom-up approaches more applicable to real life scenarios than the results depicted in theory[20]. However, it should be noted that, even though the growth is slower, the graphics still suggest an increase with the size of the group and eventually the proof may not be applicable for domains involving a large number of agents. These results raise an interesting trade-off that should be analyzed in the future: the relation between the performance of bottom-up approaches, which increases with the subgroup size, and its applicability, which decreases with the group size, as nearly all outcomes may be Pareto optimal. There is another additional issue to be studied concerning the applicability of bottom-up approaches. As the reader may have guessed, the aforementioned domains do not guarantee strict preferences. Therefore, some Pareto optimal outcomes calculated in subgroups may not be Pareto optimal in the whole group Group size 5 7 9 3 4 Subgroup size 7 5 2 7% 4% 1% - - 5% 2% 1% 0.7% 0.3% - 4% 2% 1% 0.7% 0.4% 0.2% 0.07% 8 - - 6 - Table 1. Average % of false positives calculated in a subgroup (we call these false positives). In order to study this, we measured the ratio of false positives in the previous experimental setting. The results are summarized in Table 1. As it can be observed, the percentage of false positives remains low for every possible scenario, and it tends to decrease with the size of the subgroup. This matches our initial intuition, and shows that the proof presented in this paper practically holds in every situation. Hence, this result supports the applicability of bottom-up approaches in practice. 5 Related work As far as these authors are concerned, most of the studies have dedicated their efforts on reaching Pareto optimal solutions using a horizontal approach that involves interactions with all group members. In [16], the authors propose a gen- eral framework for bilateral negotiations where agents are able to reach near Pareto optimal outcomes by decomposing the negotiation process into iso-utility curves, from where outcomes are proposed based on the similarity to the last offer proposed by the opponent. Later, the authors extend their findings to a multilateral and multi-issue environment where convergence is guaranteed [29]. Ehtamo et al. [11] propose a centralized mechanism for achieving Pareto optimal outcomes based on real valued linear additive utility functions and information sharing. Amador et al. [1] propose a task allocation method for agents with temporal constraints that is capable of providing envy free and Pareto opti- mal solutions under specific conditions. Other works like [21] have extended the concept of Pareto optimality to argumentation frameworks. The authors study different agent attitudes, how they relate to the problem of efficiency in abstract argumentation dialogues, and define several situations and scenarios that lead to Pareto optimal arguments. Recently, Hara et al. [10] proposed a mediated mechanism based on genetic algorithms that is capable of achieving near Pareto optimal outcomes for multi-party negotiations where agents preferences present non-linear relationships and change over time. However, none of these works employ bottom-up approaches, which may prove more useful in some scenarios. Another field related to our study is that of multi-objective optimization. Pareto optimality is a well-known efficiency measure in multi-objective opti- mization [13,17,14]. Similarly to our multiagent decision setting, researchers in centralized multi-objective optimization have noticed the exponential increase on the number of Pareto optimal outcomes with the number of objective functions [7,5]. Due to this unfortunate property of Pareto optimality, some researchers have offered practical alternatives to the selection of Pareto optimal outcomes. Di Pierro et al. define the concept of k optimality for deciding over Pareto opti- mal outcomes. Basically, a non-dominated outcome is defined as k-optimal when that outcome is non-dominated over every possible combination of k objectives. Thus, it results in a stronger concept of optimality that may help to choose a so- lution over a set of Pareto optimal outcomes. We want to highlight the practical usability of k-optimality on future decision making mechanisms for agents and how it complements our current findings. First, based on our proof, a subgroups of agents may calculate Pareto optimal outcomes on subgroups and communi- cate them to the rest of subgroups. Then, a mechanism may be devised to allow agents to select a k-optimal outcome over calculated Pareto optimal outcomes. Finally, economic and theoretical studies are also a source of related work. As introduced in the text, [20] analyzed how the number of Pareto optimal outcomes exponentially increases with the number of agents by assuming that all preference profiles are equally probable. In our present study, we have, among other contributions, shown how real domains in practice behave with regards to Pareto optimality. More specifically, we have shown that, despite the increase in the number of Pareto optimal outcomes with the number of agents, the growth speed is not as quick as portrayed by [20]. This is, as far as we know, our closest work in the study of the underlying properties of Pareto optimality. Of course, there have been other successful studies on Pareto optimality for specific domains and problems like characterizing fairness, or studying the relationship between monotonic solutions and Pareto optimality [4,9], but their focus of study has not been on the exploration of bottom-up approaches for reaching Pareto optimality. 6 Conclusion In this paper, we have explored the applicability and performance of bottom-up approaches for reaching Pareto optimal outcomes in groups. Our analysis shows that Pareto optimal outcomes in a group remain optimal when increasing the number of agents in the group in many practical scenarios. This has implications for bottom-up approaches, as Pareto optimal outcomes may be calculated in subgroups first, and then be used in scenarios involving the whole group. We performed experimental analysis on preferences elicited from users in real-life scenarios and validated that this principle can be applied to a wide range of domains. Our results on performance and applicability indicate that we are able to calculate a significant ratio of the final Pareto optimal frontier within subgroups. Conversely, we analyzed the applicability of our approach by considering how the ratio of Pareto outcomes increases with the size of the group. Our findings highlight that this increase is not as abrupt as expected in theoretical studies, as not all preference profiles are equally likely in many real-life domains. Still, the increase of the ratio of final Pareto optimal outcomes points to a necessary trade off in practice, which we plan to analyze in the future. Additionally, as a future work, we plan to design novel negotiation approaches for intra-team negotiations that benefit from our findings. In particular, we plan to design a negotiation strategy for negotiation teams, which first calculate the Pareto optimal solutions within the team using our approach, and then target that set of Pareto optimal proposals when negotiating with the opponent. References 1. S. Amador, S. Okamoto, and R. Zivan. Dynamic multi-agent task allocation with spatial and temporal constraints. In Proceedings of the 2014 international confer- ence on Autonomous agents and multi-agent systems, pages 1495 -- 1496. Interna- tional Foundation for Autonomous Agents and Multiagent Systems, 2014. 2. E. Argente, V. Botti, C. Carrascosa, A. Giret, V. Julian, and M. Rebollo. An abstract architecture for virtual organizations: The thomas approach. Knowledge and Information Systems, 29(2):379 -- 403, 2011. 3. R. Aydogan, K. V. Hindriks, and C. M. Jonker. Multilateral mediated negotiation In Novel Insights in Agent-based Complex Automated protocols with feedback. Negotiation, pages 43 -- 59. Springer, 2014. 4. A. Bogomolnaia and H. Moulin. Size versus fairness in the assignment problem. Games and Economic Behavior, 90:119 -- 127, 2015. 5. D. W. Corne and J. D. Knowles. Techniques for highly multiobjective optimisation: Some nondominated points are better than others. In Proceedings of the 9th Annual Conference on Genetic and Evolutionary Computation, GECCO '07, pages 773 -- 780, New York, NY, USA, 2007. ACM. 6. D. de Jonge and C. Sierra. \ hbox {NB}{3}: a multilateral negotiation algorithm for large, non-linear agreement spaces with limited time. Autonomous Agents and Multi-Agent Systems, 29(5):896 -- 942, 2015. 7. F. di Pierro. Many-objective evolutionary algorithms and applications to water resources engineering. PhD thesis, University of Exeter, 2006. 8. S. Esparcia, V. Sanchez-Anguix, and R. Aydogan. A negotiation approach for energy-aware room allocation systems. In Highlights on Practical Applications of Agents and Multi-Agent Systems, pages 280 -- 291. Springer, 2013. 9. J. Garc´ıa-Segarra and M. Gin´es-Vilar. The impossibility of paretian monotonic solutions: A strengthening of roths result. Operations Research Letters, 43(5):476 -- 478, 2015. 10. K. Hara and T. Ito. A mediation mechanism for automated negotiating agents whose utility changes over time. In Twenty-Seventh AAAI Conference on Artificial Intelligence, 2013. 11. P. Heiskanen, H. Ehtamo, and R. P. Hamalainen. Constraint proposal method for computing pareto solutions in multi-party negotiations. European Journal of Operational Research, 133(1):44 -- 61, 2001. 12. C. Hewitt. Open information systems semantics for distributed artificial intelli- gence. Artificial Intelligence, 47(1-3):79 -- 106, 1991. 13. J. Horn, N. Nafpliotis, and D. E. Goldberg. A niched pareto genetic algorithm for multiobjective optimization. In Evolutionary Computation, 1994. IEEE World Congress on Computational Intelligence., Proceedings of the First IEEE Conference on, pages 82 -- 87. Ieee, 1994. 14. X.-B. Hu, M. Wang, and E. Di Paolo. Calculating complete and exact pareto front for multiobjective optimization: a new deterministic approach for discrete problems. IEEE Transactions on Cybernetics, 43(3):1088 -- 1101, 2013. 15. T. Kamishima. Nantonac collaborative filtering: recommendation based on order responses. In Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining, pages 583 -- 588. ACM, 2003. 16. G. Lai, C. Li, and K. Sycara. Efficient multi-attribute negotiation with incomplete information. Group Decision and Negotiation, 15(5):511 -- 528, 2006. 17. H. Li and Q. Zhang. Multiobjective optimization problems with complicated pareto sets, moea/d and nsga-ii. Evolutionary Computation, IEEE Transactions on, 13(2):284 -- 302, 2009. 18. R. Lin, S. Kraus, T. Baarslag, D. Tykhonov, K. Hindriks, and C. M. Jonker. Genius: An integrated environment for supporting the design of generic automated negotiators. Computational Intelligence, 30(1):48 -- 70, 2014. 19. B. N. Miller, I. Albert, S. K. Lam, J. A. Konstan, and J. Riedl. Movielens un- plugged: experiences with an occasionally connected recommender system. In Pro- ceedings of the 8th international conference on Intelligent user interfaces, pages 263 -- 266. ACM, 2003. 20. B. O'Neill. The number of outcomes in the pareto-optimal set of discrete bargaining games. Mathematics of Operations Research, pages 571 -- 578, 1981. 21. I. Rahwan and K. Larson. Pareto optimality in abstract argumentation. In AAAI, pages 150 -- 155, 2008. 22. V. Sanchez-Anguix, R. Aydogan, V. Julian, and C. Jonker. Unanimously accept- able agreements for negotiation teams in unpredictable domains. Electronic Com- merce Research and Applications, 13(4):243 -- 265, 2014. 23. V. Sanchez-Anguix, T. Dai, Z. Semnani-Azad, K. Sycara, and V. Botti. Modeling power distance and individualism/collectivism in negotiation team dynamics. In 45 Hawaii International Conference on System Sciences (HICSS-45), pages 628 -- 637, 2012. 24. V. Sanchez-Anguix, A. Espinosa, L. Hernandez, and A. Garcia-Fornes. Mamsy: A management tool for multi-agent systems. In 7th International conference on practical applications of agents and multi-agent systems (PAAMS 2009), pages 130 -- 139. Springer, 2009. 25. V. Sanchez-Anguix, V. Julian, V. Botti, and A. Garcia-Fornes. Reaching Unani- mous Agreements within Agent-Based Negotiation Teams with Linear and Mono- tonic Utility Functions. IEEE Transactions on Systems, Man and Cybernetics - Part B, 42(3):778 -- 792, 2012. 26. V. Sanchez-Anguix, V. Julian, V. Botti, and A. Garcia-Fornes. Studying the Im- pact of Negotiation Environments on Negotiation Teams' Performance. Informa- tion Sciences, 219:17 -- 40, 2013. 27. V. Sanchez-Anguix, V. Julian, V. Botti, and A. Garc´ıa-Fornes. Tasks for agent- based negotiation teams: Analysis, review, and challenges. Engineering Applica- tions of Artificial Intelligence, 26(10):2480 -- 2494, 2013. 28. P. Skowron, P. Faliszewski, and A. Slinko. Achieving fully proportional represen- tation is easy in practice. In Proceedings of the 2013 international conference on Autonomous agents and multi-agent systems, pages 399 -- 406. International Foun- dation for Autonomous Agents and Multiagent Systems, 2013. 29. C. N. D. T. Zhenh, Ronghuo and K. Sycara. Automated multilateral negotiation on multiple issues with private information. INFORMS Journal on Computing, page In Press, 2015. 30. C.-N. Ziegler, S. M. McNee, J. A. Konstan, and G. Lausen. Improving recommen- dation lists through topic diversification. In Proceedings of the 14th international conference on World Wide Web, pages 22 -- 32. ACM, 2005.
1006.1450	1	1006	2010-06-08T04:16:45	Separating Agent-Functioning and Inter-Agent Coordination by Activated Modules: The DECOMAS Architecture	[ "cs.MA" ]	The embedding of self-organizing inter-agent processes in distributed software applications enables the decentralized coordination system elements, solely based on concerted, localized interactions. The separation and encapsulation of the activities that are conceptually related to the coordination, is a crucial concern for systematic development practices in order to prepare the reuse and systematic integration of coordination processes in software systems. Here, we discuss a programming model that is based on the externalization of processes prescriptions and their embedding in Multi-Agent Systems (MAS). One fundamental design concern for a corresponding execution middleware is the minimal-invasive augmentation of the activities that affect coordination. This design challenge is approached by the activation of agent modules. Modules are converted to software elements that reason about and modify their host agent. We discuss and formalize this extension within the context of a generic coordination architecture and exemplify the proposed programming model with the decentralized management of (web) service infrastructures.	cs.MA	cs	Separating Agent-Functioning and Inter-Agent Coordination by Activated Modules: The DECOMAS Architecture Jan Sudeikat∗and Wolfgang Renz Multimedia Systems Laboratory (MMLab), Faculty of Engineering and Computer Science, Hamburg University of Applied Sciences, Berliner Tor 7, 20099 Hamburg, Germany {jan.sudeikatwolfgang.renz}@.haw-hamburg.de The embedding of self-organizing inter-agent processes in distributed software applications enables the decentralized coordination system elements, solely based on concerted, localized interactions. The separation and encapsulation of the activities that are conceptually related to the coordination, is a crucial concern for systematic development practices in order to prepare the reuse and systematic integration of coordination processes in software systems. Here, we discuss a programming model that is based on the externalization of processes prescriptions and their embedding in Multi-Agent Systems (MAS). One fundamental design concern for a corresponding execution middleware is the minimal-invasive augmentation of the activities that affect coordination. This design challenge is approached by the activation of agent modules. Modules are converted to software elements that reason about and modify their host agent. We discuss and formalize this extension within the context of a generic coordination architecture and exemplify the proposed programming model with the decentralized management of (web) service infrastructures. 1 Introduction Self-Organization describes adaptive processes among system elements, as found in physical, biological, and social systems, that establish and maintain structures [19]. The utilization of self-organization prin- ciples is an alternative approach for the construction of self-adaptive, distributed software systems [23]. This approach is attractive, as it allows to embed adaptive properties in the interplay of system entities. Consequently, centralized responsibilities are avoided that may imply bottle necks and single points of failure. Self-organizing system phenomena are governed by feedback loops, i.e. circular interdependen- cies among system elements (e.g. discussed in [1]). Unlike the control loops in self-managing software systems [5], the loops are decentralized, i.e. distributed among system elements. In the research project "Selbstorganisation durch Dezentrale Koordination in Verteilten Systemen"1 (Sodeko VS), the utilization of self-organizing inter-agent processes as reusable design elements is stud- ied [29]. Distributed feedbacks, as structures of mutual influences among system elements, are elevated to discrete design elements. These structures are used to define inter-agent processes and a corresponding programming model can be used to integrate these processes in agent-based software systems. A founda- tional building block is a middleware layer that provides an execution context for the process enactment and integration [29] (see Section 3). ware agents, i.e. Multiagent Systems (MAS). Developers can add decentralized coordination, i.e. A key design criterion is that adaptive features can be supplemented to functioning sets of soft- the ∗Jan Sudeikat is doctoral candidate at the Distributed Systems and Information Systems (VSIS) group, Department of Informatics, Faculty of Mathematics, Informatics and Natural Sciences, University of Hamburg, Vogt -- Kolln -- Str. 30, 22527 Hamburg, Germany, [email protected] 1Self-Organisation by Decentralized Coordination in Distributed Systems Tom Van Cutsem and Mark S. Miller (Eds.): Decentralized Coordination of Distributed Processes 2010 EPTCS 27, 2010, pp. 17 -- 31, doi:10.4204/EPTCS.27.3 c(cid:13) J. Sudeikat & W. Renz This work is licensed under the Creative Commons Attribution License. 18 Separating Agent-Functioning and Inter-Agent Coordination by Activated Modules self-organization of system aspects, to working systems. A prerequisite is the conceptual and practical separation of the activities that are conceptually related to the inter-agent coordination. These activities concern the participation in a coordinating process and define a supplement, which influences the core functionality of the agents. In this paper, we present an approach for this separation that is based on extending agent-oriented implementation modules. This allows the minimal-intrusive encapsulation and automation of inter-agent coordination. In addition, the discussed enhancement is attractive for MAS developers as it allows to modularize crosscutting concerns in MAS (see Section 2.2). This paper is structured as follows. In the following section, related work is outlined. In Section 3, a programming model for self-organization is outlined. The technological foundation for the encapsulation and automation of coordination activities is the activation of agent modules that is discussed in Section 4. The utilization of the programming model is exemplified in Section 5 before we conclude and give prospects for future work (see Section 6). 2 Related Work Agent technology provides tools and concepts for the construction of autonomous software elements and is a prominent grounding for the development of self-organizing applications [25]. Natural self- organizing systems are composed of autonomous system elements [19], e.g. particles and cells, and the coaction of these elements can be metaphorically resembled with autonomous software agents. 2.1 Integrating Coordination Mechanisms The construction of self-organization, i.e. an adaptive, coordinating process that structures the config- urations of system elements, is based on two foundational types of implementation mechanisms [31]. First, generic interaction-level mechanisms have been proposed that allow to establish information flows between system elements (e.g. reviewed in [8]). Among others, these mechanisms support the stochastic dissemination of information and the attenuation of outdated data. Secondly, adaptation-level mecha- nisms control the participation in interactions, the processing of the exchanged information, and affect the conclusive adjustments within software agents that result from the exchanged information. The encapsulation of interaction-level mechanisms is typically approached by dedicated communi- cation infrastructures and languages [12]. These are means to decouple software components but the coordination logic, e.g. when to interact and how to (locally) respond to interactions is blended in the control flow the system elements. In previous works, three foundational approaches have been followed to separate this control of the coordination from the control of the element functioning. First, special- ized agent architectures (e.g. [27]) have been proposed that outsource coordination-related activities to specific modules. Secondly, the separation can be enforced by the execution infrastructure that is given by the utilized programming-language or middleware. An example is the outsourcing of coordination by using aspect-orientation [24]. Finally, approaches use networked elements to control the localized adjustments [26]. In this paper, a coordination middleware layer is presented that extends the modularization-based ap- proaches (see above) to enable the separation and integration of coordination logic in generic agent archi- tectures. Preparing the integration in established, general-purpose agent-architectures allows to reuse the existing constructive knowledge and tool support that concerns these agent models, e.g. methodology- specific agent design techniques. The direct integration, by reusing agent-modularization concepts avoids the communication overhead of externalized approaches. J. Sudeikat & W. Renz 19 2.2 Crosscutting Concerns in Agent-Orientation Modularization enforces the decomposition of software systems into functional clusters with minimum overlapping functionality. These so-called core concerns are typically separated into different com- ponents or modules [17], complex software systems often comprise additional crosscutting concerns, so-called aspects [15], which are to be referenced from various modules. Prime examples are amongst others failure recovery, monitoring and logging. While these functionalities can be clustered in mod- ules, these will be referenced throughout the agent model. Thus the information when to invoke the functionality is spread and references are scattered. In this respect, the embedding of coordination is regarded as another crosscutting concern. When the logic how to adjust and interact, as to participate in a collaborative process, is encapsulated in specific module, the contained activities have to be frequently referenced in the agent model and these references with be scattered as well. Consequently, it is desirable to contains the functionality, as well as the context if it's invocation in a single agent element. The notion of crosscutting concerns for agent modularization has to date found minor attention. Numerous MAS infrastructures are build with object-oriented programming languages, thus Aspect- oriented Programming (AOP) [15] is one approach to embed crosscutting functionalities with additional programming language tool sets. In [16], it is exemplified how AOP techniques can be used to modularize object-oriented agent models by encapsulating mobility related API calls that are available in the JADE2 agent platform and Garcia et al. [9], examined how AOP frameworks facilitate the realization of object- oriented agents. Examples of aspects in software agents are interaction, mobility and learning [10]. 2.3 Agent Modularization Using the Example of BDI Agents Agent platforms [2] provide distributed middlewares for the construction of MAS. One prominent archi- tectural model is the Belief Desire Intention (BDI) architecture [20] that allows to express both longterm goal-directed objectives as well as reactivity. Following this architectural style, agents are structured as sets of Beliefs, Goals, and Plans. Beliefs contains the local knowledge of the agent about itself and the environment. Goals represent the objectives and plans are the executable means of agents. BDI-specific reasoning engines control the agent execution. The currently active goals are deliberated and means-end reasoning is used to select plans for the achievement of goals. Modularity in terms of functional indepen- dent clusters has been introduced to BDI agents by the Capability concept [6, 4]. Capabilities describe clusters of BDI concepts, i. e. Beliefs, Goals and Plans in a name-space. These enable the recursive inclusion of other capabilities (sub-capability). The interplay with a surrounding agent/capability (super- capability) is controlled by scoping rules. The visibility of the comprised elements is specified as well as the visibility of relevant events, generated outside of the capability. In [22], a goal-centric modularization scheme, so-called Goal-Oriented Modularity, has been pro- posed. Modules encapsulate the information how sets of related goals can be satisfied. This enables a higher degree of encapsulation of behaviors.A behavior-based stance towards agent modularization is given in [7] where role concepts encapsulate sets of beliefs, goals, plans and reasoning rules. In addition, the Enactment and deactment of roles at run-time is prepared. Modularization by policy-based intentions is proposed in [13]. Developers can explicitly declare in which context a module is to be activated. Due to the widespread, recognition, and practicability of the BDI agent model, the prototype real- ization of the here discussed coordination middleware (see Section 3) is based on this agent type. In this paper, the utilized enhancements to modularize crosscutting concerns are discussed in general (see Section 4) and are then detailed for this particular agent model (see Section 4.1). 2http://jade.tilab.com/ 20 Separating Agent-Functioning and Inter-Agent Coordination by Activated Modules 3 The DECOMAS Architecture Within the SodekoVS research project [29], a programming model for distributed feedbacks among sys- tem elements is revised. A key objective of this framework is that the ability to self-organize is provided as an optional tool that development teams can integrate in their applications when needed. The aim is to enable the supplementation of self-organizing properties when the need for decentralized coordination of system elements is revealed. The foundational elements are a declarative configuration language [33] and an architectural model for the supplementation of externally prescribed inter-agent processes. Here, the configuration of self-organizing processes is not discussed. Details on the configuration language can be found in [33] and a graphical representation of the process description is exemplified in the Sections 5.1 and 5.2 to illustrate the intended application dynamics. This integration architecture follows a layered structure that is illustrated in Figure 1. The Application Layer contains the application functionality. Within this layer, agents act as providers of application- dependent functionalities. An underlying Coordination Layer controls the enactment of coordinating processes among (sub-)sets of agents. Coordination Media are conceptual entities that contain interaction mechanisms [8]. Inside these media, these are realized by the utilization of communication infrastruc- tures (e.g. see [12]). The details of the interaction mechanisms are hidden by a generic publish/subscribe interface. The utilization of these media is shielded from the agent internals by intermediate Coor- dination Endpoints. These are associated to an agent and control the participation in a coordination process. Endpoints are enabled to observe and modify the execution of associated agents. The rationale is that Endpoints interact, via Media, in place of the associated agent and decide the local adjustments. Therefore, coordination-relevant activities, including adaptation-level mechanisms, are encapsulated (see Section 2.1). The operation of Endpoints are declaratively configured (see above). These declarations indicate which changes in the agent-internal configuration are significant for their participation in an inter-agent process. These events are then propagated via Media and processed by perceiving Endpoints. If these perceptions indicate the need for adjustments, these are made by triggering agent-internal be- haviors. Agent models are often capable to show concurrent conduct of behaviors and their scheduling is realized in the agent execution environment (e.g. see Section 4.1.1). Agents can be associated to more then one Endpoint, so the enactment of different processes is separated as well. A prototypical implementation of this architecture is reported in [32] and it's utilization is exemplified in [34]. Figure 1: The SodekoVS-Architecture to the embedding of decentralized coordination in MAS [34]. J. Sudeikat & W. Renz 21 4 Activating Agent Modules: Enabling Contributive Processing Agent-oriented software development is supported by comprehensive development environments. These provide execution middleware and programming languages for the utilization of agent-based implemen- tation concepts [2]. It is desirable that agent developers can utilize these concepts throughout the whole development cycle, also when expressing cross-cutting concerns (see Section 2.2). Conventional mod- ules cluster functional concerns. These are typically used inside agents by explicitly referencing con- tained elements, e.g. dispatching (sub-)goals that are contained modules [4]. The aim of the proposed extension is to automate these references. Both the functionality and the information when it is to be invoked are contained in modules. These modules extend conventional agent modules, as modules are equipped with the ability of observe and modify the agent execution. These enhanced modules operate as autonomous actors that react to changes in the immediate context, i.e. the state of their host agent or super module. We name these modules co-efficient, since they register for contributive processing on certain agent reasoning events. The presented module concepts allows to compose agents as sets of independent actors. Besides the structuring of agent models, these modules facilitate the embedding of crosscutting mechanisms, like logging, failure recovery etc., to be automatically triggered. A example is the encapsulation of the monitoring of agent-behaviors in [30]. A module observes the reasoning of the host agent and decides the recording, when the course of action is significantly adjusted. Figure 2 illustrates the conceptual model of a co-efficient agent module. Abstracting from specific agent architectures, we assume that the execution of an Agent Model is managed by a Reasoner, e.g. using reactive or deliberative mechanisms. The reasoning component processes Agent Reasoning Events that characterize the agent execution and reference agent elements which are modified. Examples are changes in agent-intern data structures (knowledge) or the execution of plans. A module concept allows to structure agents by containing sets of agent elements (e.g. [6]). Co-efficient modules extend these with two additional components. First, an Observation / Adjustment Component allows to observe and modify agent execution by registering for and dispatching reasoning events. This component makes use of platform-specific interfaces (Observation / Inducement). Secondly, developers specify an Event Mapping that describes which events are subject of observation as well as which events are to be dispatched to the reasoning mechanism. Figure 2: Conceptual model of co-efficient agent modules. Figure 3 outlines the operating principle of a coefficient module. On agent start-up the module is registering (Observation interface) for the observation of a set of reasoning events in the surrounding agent (1). Subsequently, it is notified about these events (2) and the event mapping is interpreted to 22 Separating Agent-Functioning and Inter-Agent Coordination by Activated Modules infer which events to dispatch via the Inducement interface. These events reference agent elements (e.g. goals, beliefs) inside the module (3), or in the surrounding agent (4). The available reasoning events can be classified according to their effects on the agent execution. Events denote modifications of the agent state, the agent behavior, the agent model, or describe communicative activities. The concrete realization of these event categories depend on the utilized agent platform and architecture (e.g. see section 4.2). Figure 3: Execution of a Co -- Efficient Capability. 4.1 Formalization of Coefficiency in Operational Semantics In [20] the BDI architecture has been defined by an abstract interpreter and a corresponding proof theory. This seminal work bridges the gap between the theory and practice of BDI agents and led to the logic -- based Agentspeak(L) programming language [21]. Based on this language and further formalizations, the Operational Semantics for BDI agents have been given in [35] to support implementation and verifi- cation of BDI-based MAS. Here, we adopt these semantics to formally express the impact of co-efficient modules. Within BDI agents these are Coefficient Capabilities (CECs) (see Section 2.3)). 4.1.1 Operational Semantics for BDI Agents Operational Semantics are an established formalism to describe the semantics of programming lan- guages. The operation of programs is expressed by a transition relation between program configurations [18]. The complete specification of the Operational Semantics of BDI reasoning is given in [35]. In this section, key definitions are summarized that are used to specify the effects of crosscutting concerns on the execution of BDI agents (cf. Section 4.1.2). An agent configuration is defined as a a tuple < ag,C,M,T,s > [35]. ag is the agent program, given by a set of beliefs and plans. C is a Circumstance that resembles the execution context of an individual agent. M is a tuple that characterizes the agent communication. T is a tuple that provides temporary information that is used by the reasoner and s is the current step in the reasoning cycle. In the following, these elements are detailed. The Circumstance C is defined as a tuple < I,E,A >, where the element I is a set of intentions {i,i(cid:48), . . .}. An intention i is a stack of partially instantiated plans. E is a set of events {(te,i), (te(cid:48),i(cid:48)), . . .}. These are denoted as pairs (te,i) of a triggering event (te) and a related intention (i). When events result from the processing of other events, e.g. from the Agent Execution PlatformAgentCo-Efficient Capability Agent Element (Goal/Plan) Behavior Observation / Adjustment Component:1 : register agent observation2 : notification of reasoning event3 : induce event (intern)4 : induce event (extern) Agent Reasoning Component Agent-Internal Reasoning21243::J. Sudeikat & W. Renz 23 achievement of (sub)goals, these follow-up events are associated to the currently active intention. An empty intention is denoted by (cid:62). In this respect, intentions represent different courses of actions. Their concurrent execution is controlled by the agent reasoner [35]. A is a set of actions that is available to the agent to modify the environment. The asynchronous communication of agents is characterized by the tuple M. The communicative abilities of agents are not influenced by CECs, therefore the management of communications is not discussed. Details can be found in [35]. Temporary information is kept in the tuple T . The elements < R,Ap,ι,ρ,ε > provide volatile data to the reasoner. The set R is the set of plans that are relevant for the current event. These plans are capable to handle the event. Ap denotes the set of plans that are not only relevant but can be activated. The elements ι,ε,ρ refer to the current intention, event and applicable plan that are considered during one reasoning cycle. Finally, the current step in the reasoning cycle is given by the element s. Altogether the reasoning cycle is composed of nine steps s ∈ {ProcMsg,SelEv,RelPl,ApplPl,SelAppl,AddIm,SelInt,ExecInt, ClrInt}. These steps are: processing incoming messages, selecting an event to be handed, computing the relevant plans, computing the applicable plans, adding means, i.e. plans, to an intention, selecting an intention, executing an intention, clearing an intention [35]. CECs affect the selection of the handled events. Therefore, we summarize here the semantics of the original Event Selection rule (SelEv). This rule picks an event and marks it for further processing. BDI agents employ reactive planning they handle the events in E. Events are added to E by transition rules or elements of the general architecture outside the agent interpreter, e. g. belief updates. The rule SelEv1 refers to the selection function SE that selects events from the set E. Selected events are removed from E and added to ε for further processing. If no event is to be handled, the rule SelEv2 skips directly to the intention execution that is initialized by the selection of an intention (SelInt) [35]: SelEv1 SE(CE) = (cid:104)te,i(cid:105) (cid:104)ag,C,M,T,SelEv(cid:105) −→ (cid:104)ag,C(cid:48),M,T (cid:48),RelPl(cid:105) E = CE\{(cid:104)te,i(cid:105)} where : C(cid:48) ε = (cid:104)te,i(cid:105) T (cid:48) SelEv2 SE(CE) = {} (cid:104)ag,C,M,T,SelEv(cid:105) −→ (cid:104)ag,C,M,T,SelInt(cid:105) (1) (2) In order to handle selected events, the relevant and applicable plans are calculated and one applicable plan is selected. A group of rules is responsible to execute this applicable plan. Additional rules control the execution of different intentions. External events, i. e. events that are perceived and not generated by previous plan executions trigger the creation of a novel intention. These stacks of partially instantiated plans are added, removed, and selected for execution by dedicated transition rules [35]. 4.1.2 The Operational Semantics for Co-Efficient Capabilities Co -- efficient modules register themselves for agent observation (see Section 4). These modules are no- tified when an event in a subset of reasoning events occurs. Upon these occurrences additional BDI reasoning events are dispatched in the surrounding agent. Realizations of these modules contain the con- figuration of (1) the events that are to be added by certain observations and (2) the execution context that permits the addition of the event. This configuration can be described as a set (K) of tuples (cid:104)tes,ted,λ ,κ(cid:105): 24 Separating Agent-Functioning and Inter-Agent Coordination by Activated Modules • the element tes is a triggering event that is to be observed by the capability. The set of observed events is given by the set S (tes ∈ S), which is a subset of the available reasoning events (S ∈ E). • the element ted is the corresponding event that is to be added to the agent reasoner when tes is observed. The set of actuated events (D) is also a subset of the available reasoning events (ted ∈ D,D ∈ E). • the element λ denotes the logical location of event additions. Events (ted) can be placed in the currently active intention (i) or in a new intention ((cid:62);λ ∈ {i,(cid:62)}). • the optional element κ is a boolean expression that defines when the event ted is applicable to be added to the agent configuration. This expression takes into account the current agent state (ag,C,M,T ) and denotes the context that permits the introduction of the additional event (ted). The expressiveness of these statements depends on the utilized agent platform. Three functions are introduced, which access this configuration, to simplify the semantics of agent state modifications. An auxiliary mapping function m(x) is assumed that maps triggering events to corresponding events (m : S −→ D). According to the configurations in K this function returns the event(s) that are to be introduced (m(tes) = ted). The function l : S,D −→ λ returns the target intention (λ ) for an event mapping (S,D), i.e. two corresponding events (tes,ted). λ can have two values and indicates either that the event is to be added to the current intention (λc) or that the event is to be included in a new intention (λn). In the following, we also assume the availability of a platform specific function eval(tes,ted) that extracts the agent execution context, looks-up the corresponding condition statement (κ) and evaluates the statement, i.e. returns a boolean (eval : S,D −→ {true, f alse}). The functions m and l are prescribed by the agent programmer. The given mapping defines the set of observed events and their counterparts, which are to be induced. For each of the latter events, its is also specified whether it is to be placed in the current intention or in a new one. New intentions initially contain only the triggering event and this placement initializes another concurrent course of actionfor the agent. Implementations of CECs are aware of the mapping function and dispatch the corresponding event in the surrounding agent. Afterwards the agent execution continues unaltered. Therefore, only the event selection rule SelEv is supplemented with another rule (SelEvCEC) that defines the the contribution of a CEC when events are selected that are in the set of observed events S. If te /∈ S, the unaltered selection rules (SelEv1,2) is used (cf. Section 4.1.1). The inserting of the event (ted ∈ D) that corresponds the observed event m(te) into the agent circumstance (C(cid:48) E) is guarded by the annotated condition (κ) that is evaluated for the currently handled event mapping (eval(te,m(te))). Events are inserted into the intention that is indicated by the function l(tes,ted). The original event (te), which was selected for processing (SE(CE) = (cid:104)te,i(cid:105)), is added to the temporal structure (T (cid:48) ε ) and subsequently processed by the reasoning cycle. This rule does not enforce immediate event processing. The additional event (ted) is added to the events that will be subsequently processed by the agent but the operation of the interpreter decides how agent execution proceeds. SelEvCEC SE(CE) = (cid:104)te,i(cid:105) (te ∈ S) (cid:104)ag,C,M,T,SelEv(cid:105) −→ (cid:104)ag,C(cid:48),M,T (cid:48),RelPl(cid:105) where: C(cid:48) E = (cid:26) CE ∪{(cid:104)m(te),l(te,m(te)(cid:105)}\{(cid:104)te,i(cid:105)} i f eval(te,m(te)) = true CE\{(cid:104)te,i(cid:105)} otherwise (3) ε = (cid:104)te,i(cid:105) T (cid:48) J. Sudeikat & W. Renz 25 4.2 Prototype Realization The described mechanism has been implemented using the Jadex reasoning engine3, that provides an execution environment for BDI-style agents on top of arbitrary distributed systems middleware [3]. The Jadex platform allows implementations of a certain interface (jadex.run-time.ISystemEventListener), to be registered at individual agents. The Jadex Introspector and Tracer development tools, accompanying the Jadex distribution, use this mechanism to observe agent reasoning. Implementations of this interface can be registered for a specified set of reasoning events (defined in jadex.runtime.Systemevent) and an API allows to modify the BDI facilities of the surrounding agent via an API. 4.3 Structuring Agents with Active, Context-Aware Modules In terms of comprehensibility and reusability, it is advisable to cluster distinguishable functionalities in modules. Capabilities contain libraries of BDI elements that can be referenced to reuse functionality. Besides, the goal-oriented and event-based processing of the reactive planning mechanism, can as well be exploited to handle functionalities that would be commonly understood as crosscutting concerns. For example, message based communication are encapsulated in [10] in an interaction aspect while the modularization of roles in a negotiation protocol is the prime example for BDI-based capabilities [6] (cf. section 2.3). The proposed extension to the capability concept allows to automate the references to elements that are contained in capabilities. This facilitates information-hiding, as both the agent elements and the information when these are to be activated/modified are contained in a single entity. In addition, this approach facilitates the provision of additional, quasi contributive, processing of BDI reasoning events. The original handling of events leads to a series of subsequent events. For example the achievement of a goal may involve the successive achievement of subgoals. Besides this sequential execution path, which is controlled by the agent reasoner, the described mechanism allows to declare additional activities should be activated as well. These are caused by the activation of reasoning events (see equation 3). The authority of the reasoning mechanism, e.g. a BDI interpreter, is untouched and the contributive events integrate in the currently active or a in a concurrently executed intention. An example usage is the embedding of monitoring routines that decide the significance of agent state changes and communicate these to remote observers [30] as well as the embedding of assertion validations [28]. The here discussed activation of agent modules reveals another perspective on the modularization of agents. Besides functional clusters (see Section 2.3), agents can be structured as composites of context- aware actors that decide locally when to provide the contained functionalities. Despite the outlined benefits, the presented mechanism reduces the traceability of the agent actions. it makes it more difficult to trace the causes of specific agent actions. Since the additional events are handed over to the agent reasoner for subsequent processing, arbitrary agent events can be selected for inclusion. Subsequently, all events are similarly processed. Therefore, coefficient activities can be arbitrarily selected. However, the performance of the agent functionality is reduced by the coefficient inclusion of exhaustive processings. 4.4 Construction of Coordination Endpoints A requirement for the realization of the Decomas architecture is the ability to associate Endpoints with agents and enable these endpoints to observe, reason about, and influence the agent execution (see Sec- tion 3). The discussed coefficiency mechanism allows to construct Coordination Endpoints as agent mod- ules. Conceptually, this approach is attractive since the agent coordination is separated is represented as 3http://jadex.informatik.uni-hamburg.de 26 Separating Agent-Functioning and Inter-Agent Coordination by Activated Modules an orthogonal aspect. The uncommunicative overhead, imposed by remote endpoints is avoided. Using coefficient agent modules, a generic Endpoint model has been conceived that abstracts from the utilized agent platform. This structure is illustrated in Figure 4. A Coordination Endpoint is composed of three (sub-) mod- ules. First, a Communication module contains the abilities to publish and perceive events that are sig- nificant for the inter-agent coordination. These modules exchange specific data elements (Coordination Information) and realize the Medium-specific information propagation. Endpoints contain also two in- terpreter elements. The Coordination Information Interpreter is responsible to process the perceived information and decides, based on a declarative configuration of the enacted coordination process [33], the appropriate modifications of the host agent. This is a conventional agent module except that it affects the injection of modifications in the surrounding agent (see Section 4.1.2). The Agent State Interpreter is a coefficient module that registers for the observation of the surrounding host agent. The events of interest are given by a declarative model of the desired inter-agent coordination [33] that also contains the conditions and constraints that control the publication of agent-internal events and data. Using coef- ficiency, the observation and affection of publications is realized as an autonomous background process inside agents. Figure 4: Excerpt from [32] that illustrates the structure of Coordination Endpoints. 5 Shoaling Glassfishes: Decentralized (Web) Service Management Distributed software systems imply an administrative overhead that originates in the manual maintenance of computational infrastructure, e.g. the provision of server installation and server deployments in Ser- vice Oriented Architectures. The reduction of this overhead in dynamic environments, where request loads and resource usages fluctuate, is a prominent research topic [14]. Often it is necessary to augment existing middlewares with adaptive mechanisms [11]. Here, we outline the development of a decentralized, agent-based management framework for the maintenance of distributed software infrastructures. Figure 5 illustrates the conceptual architecture. Ser- vices are deployed on application server instances (App. Server). These servers reside on different Server-Clusters, i.e. are distributed in different computing / data centers. Service endpoints and applica- tion servers are managed by remote software agents (Agent-based Management). These agents monitor and manipulate the configurations of services and servers. The management by agents is realized with J. Sudeikat & W. Renz 27 the SUN Appserver Management EXtensions4 (AMX), a generic API to control J2EE Application Server configurations. Managers of application servers are bound to specific installations, while service end- points are free to reallocate to different servers and change their service offer. Agents are equipped with plans for the deployment and undeployment of services. Unknown of service configurations can be fetched from remotely accessible repositories. Service consumers (cf. figure 5, top-left) invoke web services. The dynamics of service deployments and server utilizations are hidden by Service Brokers. These maintain local registries of the physical locations of service deployments. Therefore, clients can retrieve the addresses of the current deployments from the static locations of the Brokers. In addition, the Brokers load-balance the service utilizations. Figure 5: Decentralized, agent-based service management framework. This prototype architecture has been realized with the Jadex agent framework (cf. Section 4.2). It prepares the agent-based management of service infrastructures as agents are capable to administrate services and servers, i.e. to deploy / undeploy services. The server management has been tested with the freely available Glassfish5 application server. The actual coordination logic is supplemented with the systemic programming model that is discussed in Section 3. This management concerns two aspects of the dynamic deployment of (web) services. First, the allocation of physical servers is balanced to maintain averaged utilization levels. Servers are associated with preferential workload-levels. Based on the communication of available capacities, services are moved to ensure that all servers are in their preferred operational condition. Secondly, the adjustment of static service deployments is automated to meet dynamically changing service workloads. The deployments of highly-demanded services are reinforced and less-demanded services are reduced. 4https://glassfish.dev.java.net/javaee5/amx/index.html 5Version 2 (ur2-b04), https://glassfish.dev.java.net/ Service Users (Workflows):......Client i=0Client i=mClient i=NServer-Cluster (Domain)ServiceServiceApp.ServerServiceServiceApp.ServerApp.Server......Agent-based Management:ServiceRepositoryService ManagerServer ManagerService ManagerService ManagerService EndpointManagerServer ManagerServer ManagerInteractionModels...Coordination MediumServiceServiceBrokerServiceBrokerServiceBroker......Coordination MediumInteractionService Conf. Provision AgentAgentManagement by AgentCoordination MediumRequest Service Address Register Service AdresseService (Re-)Deployment28 Separating Agent-Functioning and Inter-Agent Coordination by Activated Modules 5.1 Server Utilization Management The aim of the utilization management is to maintain a preferential number of service deployments on servers. The rational is, that servers are dimensioned for a specific utilization. Therefore, the maintenance of several underutilized servers is a waste of resources, e.g. energy [14], when the services could be handled by a single server that is properly utilized. The decentralized management is based on the publication of capacities by underutilized servers and services are attracted to servers that are almost well utilized. The dynamics of the adaptive placement of services is illustrated in Figure 6 (I). The variable Un- derloaded describes the number of servers that are below their intended utilization. This utilization is maintained by a balancing feedback loop (-). A Coordination Endpoint determines whether the server is underutilized or not (a logical condition, defined on the set of agent beliefs). The Endpoints within underloaded servers publish the availability of capacities (β ). These publications propagate in a Coor- dination Medium to the Coordination Endpoints in Service Endpoint agents (Moveable). These decide whether to change the service deployment, i.e. move to another server, or not. Movements affect the system-wide rate of (re-)deploying agents (Server Change) and consequently influences the number of (re-)deployed services. Since these deployment actions were caused by the availability of resources, the number of underloaded servers decreases. Figure 6 (II) shows simulation results for the described coordinating process. Two application servers are initialized with slightly different workloads. Servers are configured to host up to 5 services. The availability of server capacity gradually spreads in the system and services are (re-)deployed. Figure 6 (A) denotes the movement of services that is carried out be their un- and (re-)deployment. Figure 6: Server utilization management: The dynamics of the management (I) and a simulation snapshot (II). The simulation results relate the deployments per server (II.a) to all service registrations (II.b). 5.2 Balancing Service Deployments The adaptive reinforcement of service deployments, balanced with fluctuating request loads, is based on the publications of demand changes by Broker agents. Using the coordination architecture, the mea- surement and interpretation of demand changes is encapsulated in Endpoint modules. These publish significant changes and upon the reception of these information, the Endpoints of Service agents decide whether to adjust the local deployment or not. I:II:II.a:II.b:J. Sudeikat & W. Renz 29 The dynamics of this adaptive process is illustrated in Figure 7 (I) Requesters increase, with a fluc- tuating rate, the number of Service Requests that the system has to work off. The amount of requests causally influences the measured Service Demand. The Allocated variable denotes the number of agents that offer services. Service Brokers are supplemented with the ability to publish (α) significant changes in request workloads (Service Reinforcement). The publications affect that agents consider an adjust- ment of their current service offer (Changing Service Allocation). The additional behaviors, required to participate in this process, are implemented in Coordination Endpoints. Within Service Brokers, these decide the significance of workload changes and in Service Endpoints, these decide whether to change deployments or not. Figure 7 (II) shows simulation results for 10 virtual application servers (domains) that run on a Glassfish installation. Each of them is configured to host a maximum of 5 web service implementations. An additional constraint is that every service type is only deployed once on every domain. Initialized with an arbitrary configuration of service deployments, the system is exposed to a sudden workload of service type 1. The publication of this demand change enforces that Service Endpoints locally adjust and switch their deployments. Details on the integration of this coordination model and the declaration of agent-internal data to be communicated, including code fragments, can be found in [34, 33]. Figure 7: J2EE (Web-)Service management: Balancing Service Workload. 6 Conclusions In this paper, we presented an architecture for the integration of decentralized, self-organizing coordina- tion in MAS. This architecture provides a middleware layer that contains coordination mechanisms and automates their invocation. A key design concern is that the architecture can be integrated in established agent architectures and that coordination can be supplemented to functional MAS. This allows that self- organization can be supplemented to conventionally developed MAS. The accomplishment of this design objective requires that invocations can be supplemented without affecting the structure of the original, i.e. not self-organizing, MAS. This has been approached with the concept of activated agent modules. Their generic structure and operating principle is discussed. Their implementation is outlined and formalized for a particular agent architecture. The utilization of the coordination framework is exemplified for the management of service oriented computing infrastructures. Future work comprises the programming language techniques that are used to control the execution of Media and Endpoint elements. Ongoing work concerns the revision of a declarative configuration approach [33] to integrate the prescriptions of self-organizing processes in MAS development frameworks. I:II:30 Separating Agent-Functioning and Inter-Agent Coordination by Activated Modules Acknowledgment We cordially thank Rafael H. Bordini6 from the Institute of Informatics at the Federal University of Rio Grande do Sul for discussing and reviewing the formalization of the endpoint operation principle that is discussed in Section 4.1. We also thank the Distributed Systems and Information Systems (VSIS) group at Hamburg University, particularly W. Lamersdorf, L. Braubach, A. Pokahr, and A. Vilenica for helpful discussion. The SodekoVS-project is funded by the Deutsche Forschungsgemeinschaft (DFG)7. References [1] E. Bonabeau, M. Dorigo & G. Theraulaz (1999): Swarm Intelligence: From Natural to Artificial Systems. Santa Fe Institute Studies on the Sciences of Complexity. Oxford University Press. [2] Rafael Bordini, Lars Braubach, Mehdi Dastani, Amal El Fallah Seghrouchni, Jorge Gomez-Sanz, Joao Leite, Gregory O'Hare, Alexander Pokahr & Alessandro Ricci (2006): A Survey of Programming Languages and Platforms for Multi-Agent Systems. In: Informatica 30, pp. 33 -- 44. [3] L. Braubach, A. Pokahr & W. Lamersdorf (2005): Jadex: A BDI Agent System Combining Middleware and Reasoning. In: Software Agent-Based Applications, Platforms and Development Kits, Birkhauser. [4] Lars Braubach, Alexander Pokahr & Winfried Lamersdorf (2005): Extending the Capability Concept for Flexible BDI Agent Modularization. In: Proc. of PROMAS-2005. [5] Yuriy Brun, Giovanna Marzo Serugendo, Cristina Gacek, Holger Giese, Holger Kienle, Marin Litoiu, Hausi Muller, Mauro Pezz`e & Mary Shaw (2009): Software Engineering for Self-Adaptive Systems, chapter Engi- neering Self-Adaptive Systems through Feedback Loops, pp. 48 -- 70. Springer-Verlag, Berlin, Heidelberg. [6] Paolo Busetta, Nicholas Howden, Ralph Ronnquist & Andrew Hodgson (2000): Structuring BDI Agents in Functional Clusters. In: ATAL '99, Springer, pp. 277 -- 289. [7] Mehdi Dastani, M. Birna van Riemsdijk, Joris Hulstijn, Frank Dignum & John-Jules Ch. Meyer (2005): Enacting and Deacting Roles in Agent Programming. In: Lecture Notes in Computer Science, 3382. [8] T. DeWolf & T. Holvoet (2007): Decentralised Coordination Mechanisms as Design Patterns for Self- Organising Emergent Systems. In: Engineering Self-Organising Systems, 4335/2007, pp. 28 -- 49. [9] Alessandro Garcia, Uir Kulesza & Carlos Lucena (2005): Aspectizing Multi-agent Systems: From Architec- ture to Implementation. In: Lecture Notes in Computer Science, 3390, pp. 121 -- 143. [10] Alessandro F. Garcia, Carlos J. P. de Lucena & Donald D. Cowan (2004): Agents in object-oriented software engineering. Softw. Pract. Exper. 34(5), pp. 489 -- 521. [11] D. Garlan, S.-W. Cheng, A.-C. Huang, B. Schmerl & P. Steenkiste (2004): Rainbow: architecture-based self-adaptation with reusable infrastructure. Computer 37(10), pp. 46 -- 54. [12] David Gelernter & Nicholas Carriero (1992): Coordination languages and their significance. Commun. ACM 35(2), pp. 97 -- 107. [13] Koen Hindriks (2008): Modules as Policy-Based Intentions: Modular Agent Programming in GOAL. In: Programming Multi-Agent Systems, 4908/2008, Springer Berlin / Heidelberg. [14] Rajarshi Das Jeffrey, O. Kephart, Charles Lefurgy, Gerald Tesauro, David W. Levine & Hoi Chan (2008): Autonomic Multi-Agent management of Power and Performance in Data Centers. In: Proc. of the 7th Int. Conf. on Autonomous Agents and Multiagent Systems (AAMAS 2008), pp. 107 -- 114. [15] Gregor Kiczales, John Lamping, Anurag Menhdhekar, Chris Maeda, Cristina Lopes, Jean-Marc Loingtier & John Irwin (1997): Aspect-Oriented Programming. In: Proc. of ECOOP, Springer. [email protected] 7http://www.dfg.de J. Sudeikat & W. Renz 31 [16] Cidiane Lobato, Alessandro Garcia, Alexander Romanovsky, Cludio Sant'Anna, Uir Kulesza & Carlos Lu- cena (2004): Mobility as an Aspect: The AspectM Framework. In: Proceedings of the 1st Brazilian Workshop on Aspect-Oriented Software Development WASP04. [17] D. L. Parnas (1972): On the criteria to be used in decomposing systems into modules. Commun. ACM 15(12), pp. 1053 -- 1058. [18] Gordon D. Plotkin (2004): The origins of structural operational semantics. Journal of Logic and Algebraic Programming 60-61, pp. 3 -- 15. Available at http://dx.doi.org/10.1016/j.jlap.2004.03.009. [19] Mikhail Prokopenko (2008): Advances in Applied Self -- organizing Systems, chapter Design vs. Self -- organization, pp. 3 -- 17. Springer London. [20] A. S. Rao & M. P. Georgeff (1995): BDI-agents: from theory to practice. In: Proceedings of the First Int. Conference on Multiagent Systems. Available at citeseer.ist.psu.edu/rao95bdi.html. [21] Anand S. Rao (1996): AgentSpeak(L): BDI agents speak out in a logical computable language. In: Proceed- ings of the 7th European workshop on Modelling autonomous agents in a multi-agent world, pp. 42 -- 55. [22] M. Birna van Riemsdijk, Mehdi Dastani, John-Jules Meyer & Frank de Boer (2006): Goal-Oriented Mod- ularity in Agent Programming. In: AAMAS '06: Proceedings of the fifth international joint conference on Autonomous agents and multiagent systems, ACM Press. [23] Mazeiar Salehie & Ladan Tahvildari (2009): Self-adaptive software: Landscape and research challenges. ACM Trans. Auton. Adapt. Syst. 4(2), pp. 1 -- 42. [24] Linda M. Seiter, Daniel W. Palmer & Marc Kirschenbaum (2006): An aspect-oriented approach for modeling self-organizing emergent structures. In: SELMAS '06: Proceedings of the 2006 international workshop on Software engineering for large-scale multi-agent systems, ACM Press, New York, NY, USA, pp. 59 -- 66. [25] G. D. M. Serugendo, M. P. Gleizes & A. Karageorgos (2006): Self-Organisation and Emergence in MAS: An Overview. In: Informatica, 30, pp. 45 -- 54. [26] G. Di Marzo Serugendo & J. Fitzgerald (2009): Designing and Controlling Trustworthy Self-Organising Systems. Perada Magazine . [27] Amit Shabtay, Zinovi Rabinovich & Jeffrey S. Rosenschein (2006): Behaviosites: a novel paradigm for affecting Distributed Behavior. In: Proceedings of ESOA'06, pp. 23 -- 39. [28] Jan Sudeikat, Lars Braubach, Alexander Pokahr, Winfried Lamersdorf & Wolfgang Renz (2006): Validation of BDI Agents. In: Programming Multi-Agent Systems (ProMAS 2006), number 4411 in LNAI, pp. 185 -- 200. [29] Jan Sudeikat, Lars Braubach, Alexander Pokahr, Wolfgang Renz & Winfried Lamersdorf (2009): Systemati- cally Engineering SelfOrganizing Systems: The SodekoVS Approach. EASST 17. ISSN 1863-2122. [30] Jan Sudeikat & Wolfgang Renz (2007): Monitoring Group Behavior in Goal -- Directed Agents using Co -- Efficient Plan Observation. In: Agent-Oriented Software Engineering VII, 4405/2007, pp. 174 -- 189. [31] Jan Sudeikat & Wolfgang Renz (2008): Applications of Complex Adaptive Systems, chapter Building Com- plex Adaptive Systems: On Engineering Self -- Organizing Multi -- Agent Systems, pp. 229 -- 256. IGI Global. [32] Jan Sudeikat & Wolfgang Renz (2009): DeCoMAS: An Architecture for Supplementing MAS with Systemic Models of Decentralized Agent Coordination. In: Proc. of the 2009 IEEE/WIC/ACM Int. Conf. on Intelligent Agent Technology, IEEE Computer Society Press, pp. 104 -- 107. [33] Jan Sudeikat & Wolfgang Renz (2009): MASDynamics: Toward Systemic Modeling of Decentralized Agent Coordination. In: K. David & K. Geihs, editors: Kommunikation in Verteilten Systemen, pp. 79 -- 90. [34] Jan Sudeikat & Wolfgang Renz (2009): Programming Adaptivity by Complementing Agent Function with Agent Coordination: A Systemic Programming Model and Development Methodology Integration. Commu- nications of SIWN 7, pp. 91 -- 102. ISSN 1757-4439. [35] Renata Vieira, ´Alvaro Moreira, Michael Wooldridge & Rafael H. Bordini (2007): On the formal semantics of speech-act based communication in an agent-oriented programming language. J. Artif. Int. Res. 29(1), pp. 221 -- 267.
1909.02964	1	1909	2019-09-06T15:09:31	Multi-Objective Multi-Agent Decision Making: A Utility-based Analysis and Survey	[ "cs.MA", "cs.AI", "cs.LG" ]	The majority of multi-agent system (MAS) implementations aim to optimise agents' policies with respect to a single objective, despite the fact that many real-world problem domains are inherently multi-objective in nature. Multi-objective multi-agent systems (MOMAS) explicitly consider the possible trade-offs between conflicting objective functions. We argue that, in MOMAS, such compromises should be analysed on the basis of the utility that these compromises have for the users of a system. As is standard in multi-objective optimisation, we model the user utility using utility functions that map value or return vectors to scalar values. This approach naturally leads to two different optimisation criteria: expected scalarised returns (ESR) and scalarised expected returns (SER). We develop a new taxonomy which classifies multi-objective multi-agent decision making settings, on the basis of the reward structures, and which and how utility functions are applied. This allows us to offer a structured view of the field, to clearly delineate the current state-of-the-art in multi-objective multi-agent decision making approaches and to identify promising directions for future research. Starting from the execution phase, in which the selected policies are applied and the utility for the users is attained, we analyse which solution concepts apply to the different settings in our taxonomy. Furthermore, we define and discuss these solution concepts under both ESR and SER optimisation criteria. We conclude with a summary of our main findings and a discussion of many promising future research directions in multi-objective multi-agent systems.	cs.MA	cs	Multi-Objective Multi-Agent Decision Making: A Utility-based Analysis and Survey Roxana Radulescu · Patrick Mannion · Diederik M. Roijers · Ann Now´e Abstract The majority of multi-agent system (MAS) implementations aim to optimise agents' policies with respect to a single objective, despite the fact that many real-world problem domains are inherently multi-objective in nature. Multi- objective multi-agent systems (MOMAS) explicitly consider the possible trade-offs between conflicting objective functions. We argue that, in MOMAS, such compro- mises should be analysed on the basis of the utility that these compromises have for the users of a system. As is standard in multi-objective optimisation, we model the user utility using utility functions that map value or return vectors to scalar val- ues. This approach naturally leads to two different optimisation criteria: expected scalarised returns (ESR) and scalarised expected returns (SER). We develop a new taxonomy which classifies multi-objective multi-agent decision making settings, on the basis of the reward structures, and which and how utility functions are ap- plied. This allows us to offer a structured view of the field, to clearly delineate the current state-of-the-art in multi-objective multi-agent decision making approaches and to identify promising directions for future research. Starting from the execu- Roxana Radulescu Artificial Intelligence Lab Vrije Universiteit Brussel Belgium E-mail: [email protected] Patrick Mannion School of Computer Science National University of Ireland Galway Ireland E-mail: [email protected] Diederik M. Roijers Computational Intelligence Vrije Universiteit Amsterdam The Netherlands E-mail: [email protected] Ann Now´e Artificial Intelligence Lab Vrije Universiteit Brussel Belgium E-mail: [email protected] 9 1 0 2 p e S 6 ] A M . s c [ 1 v 4 6 9 2 0 . 9 0 9 1 : v i X r a 2 Roxana Radulescu et al. tion phase, in which the selected policies are applied and the utility for the users is attained, we analyse which solution concepts apply to the different settings in our taxonomy. Furthermore, we define and discuss these solution concepts under both ESR and SER optimisation criteria. We conclude with a summary of our main findings and a discussion of many promising future research directions in multi-objective multi-agent systems. Keywords Multi-agent systems · Multi-objective decision making · Multi- objective optimisation criteria · Solution concepts · Reinforcement learning 1 Introduction A multi-agent system (MAS) features multiple agents deployed into a common environment. This is an inherently distributed paradigm, which benefits from scal- ability (agents can be added as required) and fault tolerance (the failure of any one agent does not imply failure of the whole system). The agents within a MAS may act cooperatively, competitively, or may exhibit a mixture of these behaviours [121, 127]. The majority of MAS implementations aim to optimise agent's policies with respect to a single objective, despite the fact that many real world problems are in- herently multi-objective in nature. Single-objective approaches seek to find a single policy to a problem, whereas in reality a system may have multiple possibly con- flicting objectives. Multi-objective optimisation (MOO) [21] approaches consider these possibly conflicting objectives explicitly. In multi-objective multi-agent sys- tems (MOMAS) the reward signal for each agent is a vector, where each component represents the performance on a different objective. By taking a multi-objective perspective on decision making problems, complex trade-offs can be managed; e.g., when selecting energy sources for electricity generation, there is an inherent trade-off between using cheap sources of energy which damage the environment, versus using renewable energy sources which are more expensive but better for the environment. Such trade-offs appear in a wide range of domains such as ur- ban transportation, aviation, management of natural resources and robotics; these are all domains where multi-objective multi-agent approaches could confer huge benefits. Compromises between competing objectives should be made on the basis of the utility that these compromises have for the users. In other words, if we can define a utility function that maps the vector value of a compromise solution to a scalar utility -- called a utility or scalarisation function -- then we can derive what to optimise [86], and how to measure the quality of solutions [132]. In some rare cases, we might even be able to apply the utility function a priori, and try to solve the decision problem as a single-objective problem. However, as is known from single- agent multi-objective decision making [85], it is often impossible, undesirable, or unfeasible to perform such a priori scalarisation. For example, if the utility function is non-linear, this typically renders a priori scalarisation and subsequent single- objective solution methods intractable. Moreover, while trying to find compromise policies, i.e., while the agents are planning or learning, the utility function is often unknown or uncertain. In such cases, it is often desirable to construct a so-called coverage set, a set of solutions that has at least one optimal policy with respect to every possible utility function that a user might have. Multi-Objective Multi-Agent Decision Making 3 The utility-based approach naturally leads to two different optimisation cri- teria for agents in a MOMAS: expected scalarised returns (ESR) and scalarised expected returns (SER). In the former, the users derive their utility from single roll-outs of the policy, while in the latter, the utility is derived from the expected outcomes, i.e., the mean over multiple roll-outs. To date, the differences between the SER and ESR approaches have received little attention in multi-agent settings, despite having received some attention in single-agent settings (see e.g. [84, 85]). Consequently, the implications of choosing either ESR or SER as the optimisation criterion for a MOMAS are currently not well-understood. In single-agent or fully cooperative multi-agent settings [86], it is typically as- sumed that there is one, possibly unknown, utility function that determines the possibly unknown preferences of the users, and that the users are interested in the utility of the expected vector-valued returns (i.e., the SER optimality criterion). In the execution phase all agents will ultimately pursue the best utility with respect to this single function. Therefore, the coverage set can be derived from everything that is known about this utility function, and the types of policies allowed. For example, for deterministic stationary policies and possibly non-linear utility func- tions, a coverage set is a so-called Pareto front of deterministic stationary policies. That is the set of policies that are not Pareto-dominated, i.e., for which there is no other deterministic stationary policy that has a better or equal value for all objectives and is better in at least one objective. A Pareto-undominated policy is also called Pareto-optimal. Another well-known coverage set is a convex coverage set (CCS), which is a coverage set with respect to all possible linear utility func- tions. Incidentally, a Pareto-coverage set for stochastic policies can be constructed from a CCS of deterministic stationary policies [118, 85]. We discuss related work on single-objective and fully cooperative multi-agent systems in Section 2.2. In multi-agent settings however, the situation can become much more complex. While in fully cooperative multi-agent systems, the agents are assumed to all receive the same team rewards, the individual reward vectors received by agents may be different in general multi-agent settings. We review various models with different assumptions regarding the reward functions, as well as observability and statefulness, in Section 3. Then, we consider settings where individual agents value objectives according to their own preferences, i.e., where each agent can have their own utility function. This has important consequences for what constitutes a solution set. We build a taxonomy of what constitutes a solution for a multi- objective multi-agent decision problem based on modelling assumptions, utility functions, and optimisation criteria, by analysing what happens at execution time in Section 4. We note that many of the different settings we identify in Section 4 are under-explored in the current literature, and offer examples of decision problems that would merit further investigation for each part of our taxonomy. Using our taxonomy, we review the literature on multi-objective multi-agent decision problems in terms of optimal solution sets (Section 5), and solution meth- ods (Section 6). Finally, we discuss what we consider the key open problems in this new and exciting field (Section 7). 4 Roxana Radulescu et al. Player Y A B Player X (xA,A, yA,A) A B (xB,A, yB,A) (xA,B, yA,B) (xB,B, yB,B) Table 1 General format of a Single-Objective Normal Form Game. A and B represent different actions which are available to the agents, and scalar payoffs x and y are given to each agent X and Y respectively depending on which combination of actions was selected. Player Y A B Player X A (xA,A, yA,A) B (xB,A, yB,A) (xA,B, yA,B) (xB,B, yB,B) Table 2 General format of a Multi-Objective Normal Form Game. A and B represent different actions which are available to the agents, and vector payoffs x and y are given to each agent X and Y respectively depending on which combination of actions was selected. 1.1 Motivating Example As a motivating example for adopting a multi-objective perspective on multi-agent decision making, we introduce a Multi-Objective Normal Form Game (MONFG) which is called the Commuting MONFG. By contrast to the usual Single-Objective Normal Form Game (SONFG) format, which is common in the literature, in a MONFG the agents receive payoffs in vector rather than scalar format after se- lecting their actions. This difference is illustrated in Tables 1 and 2. In the Commuting MONFG, two agents wish to commute from a common ori- gin to the same destination. There are two transportation options available: travel by taxi or travel by train. If both agents choose the taxi option, they may split the cost equally between them. If they both choose to travel by train, they must each purchase their own ticket individually. If one chooses to travel by taxi and the other chooses to travel by train, they must also pay their own fares individually. A train ticket is cheaper than a taxi fare (even when agents share a taxi ride); however, the taxi journey takes less time than the train journey. The expected travel time and cost for each mode of transport is listed in Table 3. The individual or local vector payoffs for each agent are shown in Table 4. Note that the values in the matrix are negative as this is a minimisation problem for both objectives (commuters in general do not want to spend any additional time or money on their commute above what is necessary). From an utility-based perspective, each commuter will try to balance these con- flicting objectives such that his/her derived utility is maximised. Each commuter can of course have a different utility function depending on how each objective is valued. Furthermore, depending on when each commuter evaluates the commute (e.g., on a monthly basis or after each trip), the two different optimisation criteria come into play: ESR or SER. For example, for some commuters travelling costs should be maintained within a certain budget every month, while still being on time at least 75% of the time. This requires the use of the scalarised expected returns for the month. For other commuters the time component might be crucial and they cannot be late on any given day, thus imposing the expected scalarised return criterion. Multi-Objective Multi-Agent Decision Making 5 cost 20 5 time 10 30 taxi train Table 3 Cost and travel time for different modes of transport in the Commuting MONFG. Player X taxi ([−10, −10], [−10, −10]) ([−5, −30], [−20, −10]) train ([−20, −10], [−5, −30]) ([−5, −30], [−5, −30]) taxi train Player Y Table 4 Individual (local) payoff matrix for the Commuting MONFG 2 Background Before we address the specifics of multi-objective multi-agent systems, and how to define optimal solutions for the Commuting MONFG in the previous section, we first introduce relevant background work on multi-agent decision theory, multi- objective decision making, optimisation criteria and utility functions which is nec- essary to understand the material covered later in this article. 2.1 Multi-Agent Decision Theory Multi-agent systems appeared as a natural paradigm for modelling numerous real- world problems (e.g., health-care [42], smart grid management [69, 49], traffic [37], and Internet of Things [18]) as they lend themselves perfectly to the idea of large distributed systems. They combine several disciplines ranging from artificial in- telligence, software engineering, economics to social sciences [127]. We are mostly interested here in autonomous intelligent systems, where multiple agents are de- ployed in the same environment and are faced with a series of decision making problems. Multi-agent decision making problems can often be modelled as a stochastic (or Markov) game (SG) [102, 121]. A stochastic game can be formally defined as a tuple: M = (S,A, T,R), with n ≥ 2 agents, where: -- S is the system state space -- A = A1 × ··· × An is the set of joint actions, Ai is the action set of agent i -- T : S × A × S → [0, 1] is a probabilistic transition function -- R = R1×···×Rn are the environment reward functions, where Ri : S×A×S → R is the reward function of agent i At every timestep, the environment emits a joint state s = (cid:104)s1, . . . , sn(cid:105), where s ∈ S. Depending on the system, the agents can fully observe this state, or can only observe a local view si. Notice that the reward received by an agent depends on the joint action taken by all the agents in the environment and not just on her own. However, the SG model is not the most general model. The stochastic game model can be further generalised to a partially observable stochastic game (POSG) 6 Roxana Radulescu et al. [38, 124] to include the possibility that the agents do not have full access to the environment state. In this case, each agent receives an observation and maintains a set of beliefs over possible states. Because the issue of partial observability is orthogonal to the existence of multiple objectives, but does make the model sig- nificantly more complex, we will restrict ourselves to fully observable models in this article. Note however that the solution concepts presented here generalise to partially observable environments as well. The behaviour of an agent is defined by its policy πi : S × Ai → [0, 1], meaning that given a state, actions are selected according to a certain probability dis- tribution. In the discounted infinite-horizon case, optimising πi is equivalent to maximising the expected discounted long-term reward: (cid:34) ∞(cid:88) (cid:35) V πi = E γtri,t π, µ0 (1) t=0 where π is the joint policy of the agents acting in the environment, µ0 is the distribution over initial states s0, γ is the discount factor and ri,t = Ri(st, at, st+1) is the reward obtained by agent i at timestep t, for the joint action at ∈ A, at state st ∈ S and transitioning to the next state st+1 ∈ S. Learning in multi-agent systems is considered a vital component, as envi- ronments are often characterised by high complexity and stochasticity, meaning that optimal behaviours are often impossible to achieve using pre-programmed approaches [4]. However, we note that transitioning from single- to multi-agent learning is not straightforward. Building an intelligent distributed system is a notoriously difficult problem as it involves dealing with non-stationarity, limited resource sharing, and often requires coordination or overcoming conflicting goals [101]. As a learning paradigm, we mainly consider reinforcement learning (RL) [110], but we will also discuss approaches from related fields such as game theory, planning or negotiation. Multi-agent decision making is a multifaceted problem that can be explored through the lens of many fields and from different perspectives (e.g., system versus agent point of view). But perhaps the most important distinction we observe in multi-agent learning problems concerns the definition of the reward function. Whether agents in a MAS will learn to act cooperatively or competitively depends directly on the reward function definitions. The literature usually distinguishes between three different settings: -- cooperative, where the reward function is the same for all agents: R1 = ··· = Rn = R. Examples of this setting include domains where all agents work together to optimise the performance of a larger system, such as urban traffic control [61], air traffic control [20] and electricity generator scheduling [64]. -- competitive, where any win for one agent implies a loss for another. Some competitive settings are zero-sum. Examples of this setting include fully com- petitive games such as Backgammon [116], Go [107] and Starcraft 2 [120]. -- mixed, where no restriction is imposed on the reward function definitions. Mixed games may incorporate elements of both cooperation and competition. Examples of this setting include games with opposing teams of agents, such as RoboCup soccer [50]. This classification of multi-agent decision making problems is also reflected in the taxonomy we define in Section 4. Multi-Objective Multi-Agent Decision Making 7 Discrete, tabular representations are the simplest way for agents to store in- formation which they have learned (e.g., policies, environment models, or action values in the case of model-free RL agents). When information is stored discretely, each additional feature tracked in the state leads to an exponential growth in the number of state-action pair values that must be stored [110]. This problem is commonly referred to in the literature as the "curse of dimensionality", a term originally coined by Bellman [11]. While this rarely occurs in simple environments, it may lead to an intractable learning task in complex real-world domains due to memory and/or computational constraints. Learning over a large state -- action space is possible, but may take an unacceptably long time to learn useful policies. Function approximation methods, such as artificial neural networks (ANNs), may be employed to represent policies, environment models or action values. These methods allows one to handle higher dimensional inputs, as well as allowing gener- alisation between similar observations and actions. Agents using function approxi- mators can also potentially deal with continuous observation and/or action spaces. Recent advances in single and multi-agent RL make use of deep ANNs as function approximators; this emerging paradigm is known as deep reinforcement learning (DRL). For further information on recent deep MARL methods, the interested reader is referred to recent survey articles [40, 75]. Another dimension to characterise multi-agent systems is represented by the degree of decentralisation. The planning/learning phase and the execution phase may be either (partially) centralised or fully decentralised. The paradigm of cen- tralised training with decentralised execution represents a middle ground between fully centralised and decentralised settings often used in cooperative or mixed set- tings [30, 31, 55, 97, 108]. The aim here is to enrich and aid the training/learning phase with extra information shared between the agents, however during the policy execution phase, the agents act in a fully decentralised manner. 2.2 Multi-Objective Decision-Making Single-objective decision making requires the existence of a single scalar reward function that agents can observe. The goal for the agents is then to find a policy that maximises the expected sum of these scalar rewards. However, most real- world problems do not adhere to this requirement. Specifically, there are typically multiple objectives that agents should care about. For example, consider the cost and time objectives of our transportation example in Section 1.1. When there are multiple objectives, one might in some special cases still be able to use single-objective methods by using a priori scalarisation. Specifically, if there exists a function that maps every possible outcome, i.e., a vector with policy values in each objective, which captures the exact preferences that the user(s) might have with respect to all these possible outcomes of the decision problem; this function is known a priori; and can be applied to the decision problem in such a way that the resulting single-objective problem remains tractable; then single- objective methods could still be used to solve the problem [85]. However, often such a priori scalarisation is either impossible, infeasible or undesirable. Roijers et al. [85] identify three use-case scenarios for multi-objective decision making, shown in Figure 1, in which this is indeed the case. 8 Roxana Radulescu et al. Fig. 1 Use-case scenarios for multi-objective decision making [85] In the unknown weights scenario, or more precisely the unknown utility function scenario (Figure 1a), a priori scalarsation is undesirable, as the utility that the user is able to get from the alternatives is too uncertain, or even unknown at the moment when planning or learning must occur. For example, when the objectives correspond to things that can be purchased or sold at an open market, but due to the complexity of the planning problem the prices can change significantly before planning or learning is complete. In such cases, it is desirable to compute a coverage set in order to be able to respond as quickly as possible whenever the available information about the market prices is updated. In the decision support scenario (Figure 1b), a priori scalarisation is infeasible or impossible, as a utility function that corresponds to the preferences of the user is never known explicitly. For example, consider a decision on the medical treat- ment of a serious illness. This decision problem has objectives such as maximising the probability of being cured and minimising the side effects. However, it is very difficult for a patient to articulate an exact utility function that makes all hypo- thetical trade-offs between these objectives a priori. In such cases it is therefore highly preferable to create a set containing the available possibly optimal alterna- tives, and present this set to the user. The decision support scenario thus proceeds almost identically to the unknown weights scenario. The only difference is that in the selection phase, the user selects a policy from the coverage set according to her arbitrary preferences, rather than explicit scalarization according to given weights. Finally, in the known weights scenario or known utility function scenario, a priori scalarisation would in principle be possible, as an exact utility function is available before planning or learning. However, even if this is indeed the case, it can still be undesirable to do so, because performing a priori scalarisation can lead to an intractable problem [85]. Key to all of these use cases is the notion of user utility. Indeed, we argue that for any (multi-objective) decision making problem, the agent should always aim to maximise the user's utility. Specifically, following the work of Roijers et al. [85], we take the utility-based approach to multi-objective decision making. In short, this algorithmMOMDPsolution setscalarizationweightsalgorithmsolution setuser selectionsingle solutionplanning or learning phaseexecution phasealgorithmMOMDP+weights(a)(b)(c)MOMDPsingle solutionexecution phaseselection phasesingle solutionexecution phaseselection phaseplanning or learning phaseplanning or learning phaseMulti-Objective Multi-Agent Decision Making 9 means that the ultimate goal is to maximise user utility, and that what constitutes a solution to a multi-objective decision problem should be derived from what is known about the user utility. User utility is characterised by a utility function u that maps vector-valued (expected) returns to a scalar value. We first discuss one optimisation criterion, SER, which will be discussed in the next section to give an impression of how an optimal solution set in multi-objective problems can be constructed. In SER, it is the value vector V πi , i.e., the expected vector-valued return of policy πi, that is projected to a scalar value: V πi = u(V πi ) (2) In order to derive the optimal solution set -- which is what a planning or learning algorithm should output -- one should start at the back of the use-case scenario, i.e., the execution phase, and work back, through the selection phase, until a specification of the optimal output of the algorithm is reached. As shown in Figure 2, in single-agent multi-objective decision making, the execution phase is straight- forward. The agent uses its policy π to interact with the environment, which leads Fig. 2 The execution phase for single-agent multi-objective decision making. to a value vector Vπ. Under SER (see next section) the utility function u is applied to Vπ. This means that in the selection phase, the policy that maximises u(Vπ), must be available, which brings us to the selection phase. In a known weights scenario, this is trivial, as u is known, so let us focus on the decision support and unknown weights scenarios. In both these scenarios, u is (implicitly or explicitly) applied to a set of alternative value vectors, leading to the maximising policy from a set of alternatives to be chosen.1 Because in the unknown weights and decision support scenarios, u is at least partially unknown when the agent needs to plan or learn, the planning or learning algorithm should output a set of alternative policies, that for every possible u that a user might have (subject to what is already known at the beginning of the planning or learning phase), contains at least one optimal policy. This is called a coverage set [85]. In multi-agent settings, the execution phase is much less straight-forward. In fact, there are different settings, all with their own uses, that lead to very different schemas for the execution phase. Therefore, after discussing the various multi- agent multi-objective decision-theoretic problem settings in Section 3, we conduct a thorough analysis of the execution phase for multi-agent multi-objective decision making in Section 4. Before going into different choices with respect to the execu- tion phase in multi-agent settings however, we must first discuss another, perhaps even more fundamental choice: when to apply the utility function. 1 Note that we are assuming here that there is a small discrete set of alternatives, and that this maximisation can explicitly be computed in reasonable time. If this is not the case, for example if their set of alternatives is continuous, the user can be assisted in selecting a good policy using specific algorithms designed for the selection phase [133]. However, in such cases optimality can typically not be guaranteed. environmentVu(V) π10 Roxana Radulescu et al. 2.3 Multi-Objective Optimisation Criteria In the previous section, we have seen an example of how to construct an optimal solution for a given problem setting. From a more general point of view, when agents consider multiple conflicting objectives, they should balance these in such a way that the user utility derived from the outcome of a decision problem is max- imised. This is known as the utility-based approach [85]. Following this approach, we assume that there exists a utility function that maps a vector with a value for each of the d objectives to a scalar utility: u : Rd → R. We recall that the value function vector is defined similarly to Equation 1, as the expected discounted long-term reward: (cid:34) ∞(cid:88) (cid:35) Vπ = E γtrt π, µ0 t=0 where µ0 is the distribution over initial states, π is the agent's policy, γ is the discount factor and rt is the reward vector received for each of the objectives at timestep t. When deciding what to optimise in a multi-objective decision making problem, we thus need to apply this utility function to the vector-valued outcomes of the decision problem in some way. There are two choices for how to do this [85, 86]. Computing the expected value of the payoffs of a policy first and then applying the utility function, leads to the scalarised expected returns (SER) optimisation criterion, i.e., (cid:32) (cid:34) ∞(cid:88) t=0 V π u = u E γtrt π, µ0 (cid:34) (cid:32) ∞(cid:88) (cid:33) u = E V π u γtrt π, µ0 (cid:35)(cid:33) (cid:35) (3) (4) u is the return derived by the agent from the vector Vπ. SER is employed where V π in most of the multi-objective planning and reinforcement learning literature. Al- ternatively, the utility function can be applied before computing the expectation, leading to the expected scalarised returns (ESR) optimisation criterion [84], i.e., t=0 Which of these criteria should be considered best depends on how we are interested in evaluating the outcome of a policy. SER is the correct criterion if we want to execute a policy multiple times, and it is the average return over multiple executions that determines the agent's utility. ESR is the correct formulation if the return of a single policy execution is what is important to the agent. 2.4 Utility Functions Linear combinations are a widely used canonical example of a utility function: u(r) = wdrd (5) (cid:88) d∈D Multi-Objective Multi-Agent Decision Making 11 where D is the set of objectives, w is a weight vector2, wd ∈ [0, 1] is the weight for objective d and rd is the component for objective d of some reward vector r. Interestingly, for such linear utility functions, there is no difference between SER and ESR, as the inner product with the weight vector distributes over the expectation. Non-linear, discontinuous utility functions may arise in the case where it is important for an agent to achieve a minimum payoff on one of the objectives; such a utility function may look like the following: (cid:40) u(r) = rtd 0 if rd ≥ td otherwise (6) where rd represents the component of r for objective d, td is the required threshold value for d, and rtd is the reward for reaching the threshold value on d. In general we are interested in the class of all monotonically increasing (possibly non-linear) utility functions. Definition 1 A scalarization function f is monotonically increasing if: (∀o, V π o ≥ V π(cid:48) o ) ⇒ u(Vπ) > u(Vπ(cid:48) ). (7) This means that if for all objectives, the value for that objective under policy π is greater than or equal than the value for that same objective under policy π(cid:48), then policy π yields equal or higher utility than policy π(cid:48). This is a rather minimal assumption to make, as it translates to: we always want more of each objective. Non-linear utility functions may yield different optimal policies under SER and ESR. Utility functions may not always be known a priori and/or may not be easy to define depending on the setting. For example, in the decision support scenario [85] it may not be possible for users to specify utility functions directly; instead users may be asked to provide their preferences by scoring or ranking different possible outcomes. After the preference elicitation process is complete, users' responses may then be used to model their utility functions [133]. 3 Modelling Multi-Objective Multi-Agent Settings In this section, we discuss how multi-objective problems with multiple agents can be modelled. We discuss the multi-objective stochastic game, and partially observ- able stochastic game models, as the most general models, and then show which additional assumptions can be made to arrive at more restricted models. 3.1 The Multi-Objective Stochastic Game Model As a framework for defining multi-objective multi-agent decision making settings we will use the Multi-Objective Stochastic Game (MOSG). We formally define a MOSG as a tuple M = (S,A, T,R), with n ≥ 2 agents and d ≥ 2 objectives, where: 2 A vector whose coordinates are all non-negative and sum up to 1. 12 Roxana Radulescu et al. -- S state space -- A = A1 × ··· × An set of joint actions, Ai is the action set of agent i -- T : S × A × S → [0, 1] probabilistic transition function -- R = R1 × ··· × Rn reward functions, Ri : S × A × S → Rd is the vectorial reward function of agent i for each of the d objectives Furthermore, the same as in the stochastic game case, the MOSG can be ex- tended to also incorporate partial observability. We can thus also define a multi- objective partially observable stochastic game (MOPOSG), where agents do not have access anymore to the full state of the environment. In this situation, agents receive observations from the environment and have to maintain beliefs over possi- ble states. While, for the scope of this work, it is sufficient to consider the MOSG model, we will build our categorisations having the MOPOSG model as the most general case. An agent behaves according to a policy πi : S×Ai → [0, 1], meaning that given a state, actions are selected according to a certain probability distribution. Opti- mising πi is equivalent to maximising the expected discounted long-term reward: (cid:34) ∞(cid:88) (cid:35) V πi = E γtRi(st, at, st+1) π, µ0 (8) t=0 where π is the joint policy of the agents acting in the environment, µ0 is the distribution over initial states s0, γ is the discount factor and Ri(st, at, st+1) is the vectorial reward obtained by agent i for the joint action at ∈ A, at state st ∈ S. We also note that it is also possible to extend this framework to include the case in which the discount factor is different for each agent i by replacing γ with γi. The value function is also vectorial, V πi ∈ Rd. We consider that each agents also has an individual utility function ui to project V πi to a scalar value, as described in Section 2.4. Starting from this model, we will further develop a taxonomy in Section 4 focusing on the utility and reward axis. Furthermore, we show how the approaches found in the literature can be mapped to this model by limiting various dimensions such as states, observability, individual rewards, or utilities. 3.2 Special Case Models The MOPOSG model is general enough to encompass a wide range of multi- objective multi-agent decision making settings; consequently, many prior decision making models may be viewed as special cases of a MOPOSG. By restricting certain degrees of freedom in the MOPOSG model, one can derive many commonly used decision making models from the single-agent and multi-agent literature, as well as the single-objective and multi-objective literature; e.g. by setting the number of agents n = 1 and the number of objectives d = 1 in a MOPOSG, one may obtain a traditional POMDP. Fig. 3 outlines the relationship between the MOPOSG model and many other common multi-objective multi-agent decision making models. Table 5 sumarises other common decision making models, and outlines which degrees of freedom of the MOPOSG model must be restricted to derive each other model. We hope that readers will be able to use this as a reference, Multi-Objective Multi-Agent Decision Making 13 Fig. 3 The MOPOSG is a general model, which encompasses many other common de- cision making models. The abbreviations in the Venn diagram stand for: multi-objective partially observable stochastic game (MOPOSG), multi-objective stochastic game (MOSG), multi-objective decentralised partially observable Markov Decision Process (MODec-POMDP), multi-objective Bayesian game (MOBG), multi-objective multi-agent Markov Decision Pro- cess (MOMMDP), multi-objective normal form game (MONFG), multi-objective collaborative Bayesian game (MOCBG), and multi-objective coordination graph (MOCoG) so they can easily identify ways in which problem settings and algorithms from the single-objective literature could be extended/reanalysed from a multi-objective perspective. Furthermore, it should be possible to easily spot methods developed specifically for multi-objective models which could be applied to the corresponding single-objective model. 4 The Execution Phase As mentioned in Section 3, multi-objective multi-agent models are typically named according to assumptions about observability, whether the problem is sequential, and the structure of the reward function. These are indeed important distinctions. However, following the utility-based approach [85], this information is not sufficient to determine what constitutes a solution for such a problem. Specifically, we should aim to optimise the utility of the user(s). In single-agent multi-objective problems, we can typically assume that at execution time we aim to optimise the utility of a single user with a single utility function3. The shape of the utility function, 3 Or multiple users whose utility functions can be aggregated in an overall aggregated utility function MOSGMOCBG MOPOSGfully observablecooperativeMOCoGMODec-POMDPMOMMDPstatelessMONFGMOBG14 multi-agent multi-objective single-agent multi-agent single-objective single-agent Model MOPOSG MOSG MODec-POMDP MOMMDP MOCoG MOBG MOCBG MONFG MOMG MOPOMDP MOMDP MO Multi-armed bandit POSG SG Dec-POMDP MMDP CoG BG CBG NFG MG POMDP MDP Multi-armed bandit Roxana Radulescu et al. d n S observability full full full full full full full full full full full full full full 1 1 1 1 1 1 1 1 1 1 1 1 2 2 1 1 1 2 2 1 1 1 1 1 1 1 1 1 1 1 1 1 1 Table 5 Summary of which degrees of freedom must be restricted to derive common decision making models from the MOPOSG model. Here d is the number of objectives, n is the number of agents and S is the size of the system state space. Blank cells indicate no restriction, whereas numeric values indicate a required parameter setting. (See Figure 3 for a list of the abbreviations.) in conjunction with the allowed policy space, can be used to derive the optimal solution set that a multi-objective decision-theoretic algorithm should produce. In multi-agent settings, the situation is more complex than in single-agent settings. Particularly, each individual agent can represent one or more distinct users. In other words, the utility function may vary per agent. For example, assume we have a group of friends deciding where to go on holiday, who outsource the decision making to a group of agents (one agent per friend). The objectives they agree on are minimising costs, minimising the distance from the hotel to the beach, maximising the expected number of hours of sun, and maximising the number of museums and other points of cultural interest within a 20km radius. After a decision is reached, every friend will get the same (expected) returns vector. However, each friend may have a different utility for each possible vector -- in fact this is the entire reason that this decision problem may be hard. Furthermore, it depends on which perspective we take, as the algorithm designers. In the example, we have taken the perspective of the individual users, but we could also take the perspective of an external observer that wants the outcome to be fair (for whatever definition of fair), i.e., wants to optimise some form of social welfare. We propose a taxonomy based on the reward as well as the utility functions. We distinguish between two types of reward functions: a team reward, in which each agent receives the same value or return vector for executing the policy, and individual rewards in which each agent receives a different value/return vector. Furthermore, we make a distinction in three types of utility -- more or less orthog- Multi-Objective Multi-Agent Decision Making 15 onally to the types of rewards -- i.e., team utility, which is what happens when all the agents serve the same interest, e.g., when they all work for a single company or are on the same football team; social choice utility, when we are interested in optimising the overall social welfare across all agents; and individual utility, which is what happens if each agent serves a different agenda and just tries to optimise for that. This results in the taxonomy provided in Figure 4. We further note that the utility functions may be applied according to the ESR or SER criteria for every setting. Fig. 4 Taxonomy of multi-objective multi-agent decision making settings. We note that in the taxonomy, the team reward and team utility setting could be translated to a single-agent setting, by flattening out the multi-agent aspect. Specifically, we could define a single agent that would control the actions of all other agents, i.e., one agent choosing its actions from the entire joint action space. As such, the solution concepts from the single-agent multi-objective literature apply [85]. However, the problem can still be significantly harder than a single- agent problem, due to the size of the joint action space, as we discuss in Section 4.1.1. Furthermore, we note that the individual rewards with a team utility setting is not realistic; even if the utility function of all the individual agents would be the same (i.e., the agents have the same opinion about what is important), that would still lead to different individual utilities due to different input (expected) return vectors. Hence, even when the utility functions are identical, we treat these as individual utilities. In the remainder of this section, we discuss each of the remaining settings in our taxonomy in more detail. In Section 5 we discuss the various solution concepts that apply to these settings (see Figure 4). Multi-objectivemulti-agentdecision makingTeamTeamSocial choiceRewardIndividualIndividualSocial choiceIndividualUtility16 4.1 Team Reward Roxana Radulescu et al. First, we consider the top row of Figure 4, team reward. In this setting each agent receives the same reward vectors, R1 = ··· = Rn = R. As a result, the (expected) return vector is the same for each agent when a given joint policy is executed. This is for example the case in multi-objective multi-agent Markov decision processes (MOMMDPs) [83, Section 5.2.1], as we discussed in the previous section. At first glance, this may appear to be a fully cooperative setting. However, this depends on how much the individual agents value their (expected) cumu- lative reward vectors, i.e., on the utility function of each agent. We distinguish between three cases: team utility, individual utility, and social choice with respect to individual utilities. 4.1.1 Team Reward Team Utility Perhaps due to its relative simplicity, the most common case by far in the multi- objective multi-agent decision-theoretic planning and reinforcement learning liter- ature is the team reward with team utility setting, i.e., all the agents together aim to strive for a single maximum utility, under SER, ∗ V = max π u(V π), ∗ V = max π E[u(ρ)π, µ0], or under ESR: where ρ =(cid:80)∞ t=0 γtrt. u (including its parameterisation) may or may not be known to the agents. This is a truly fully cooperative setting. Therefore, the optimal solution sets, i.e., coverage sets, can be derived from the same information as in single-agent multi-objective settings (see Section 2), and the same types of solution methods apply. Fig. 5 The execution phase for the Team Reward Team Utility setting. This figure depicts the SER optimality criterion, where the expected values (i.e., the average over many executions of the policies) will be input to u. Under ESR the input to u would be ρ, i.e., the returns for an individual roll-out. Even though techniques similar to single-agent multi-objective settings can be used to solve multi-agent multi-objective settings, multi-agent multi-objective problems are much more complex than their single-agent counterparts. Specifi- cally, the number of possible joint actions increases exponentially in the number of agents, leading to a much larger policy space. In turn, in cases where the utility function is unknown during planning or learning this leads to much larger coverage sets. To keep multi-objective multi-agent planning and reinforcement learning tractable in these settings, it is key to exploit so called loose couplings [36, 51], i.e., each environmentVu(V)πijkMulti-Objective Multi-Agent Decision Making 17 agent's actions only directly affect a subset of the other agents. Loose couplings can be expressed using a factorised reward function. Such a factorised reward function can be visually represented as a graphical model known as a coordi- nation graph in the multi-agent literature. The single-shot setting -- the multi- objective coordination graph (MO-CoG) -- is one of the most well-studied models in the multi-objective multi-agent literature [22, 24, 23, 65, 66, 90, 87, 86, 94, 93, 125, etc.]. Exploiting loose couplings also plays an important role in sequential multi- objective multi-agent settings [99, 83]. We discuss the solution concepts for this setting in Section 5.2. 4.1.2 Team Reward Individual Utility When a group of agents receives a single shared reward vector, that does not mean that all agents value that reward equally. For example, imagine that you are playing a massive multiplayer online role playing game (MMORPG), and you set out on a quest with teammates. You will play multiple quests with the same team, so you are interested in the expected returns rather than the returns of a single quest (SER). The expected value of doing a quest in terms of experience points, currency and gear is the same for each member of the team, but for different players each of these objectives may be more or less important. Therefore, even when the team gets team rewards for all quests, the members of the team may prefer different quests. This is because mathematically, each agent tries to optimise its own utility via the team value of a joint policy: under SER, or, V π i = ui(V π), i = E[ui(ρ)π, µ0], V π under ESR, where π is the joint team policy. This leads to the execution phase depicted in Figure 6. Fig. 6 The execution phase for the Team Reward Individual Utility setting. This figure depicts the SER optimality criterion, where the expected values (i.e., the average over many executions of the policies) will be input to u. Under ESR the input to u would be ρ, i.e., the returns for an individual roll-out. The existence of individual utilities immediately poses a problem for the agents. Each agent can only control a small part of the joint policy, i.e., its own actions, and a lack of coordination may lead to a very bad policy for all agents. In other words, an agent cannot simply maximise its utility by changing its own policy without taking the policies, and policy changes, of the other agents into account. environmentVui(V)uj(V)uk(V)πijk18 Roxana Radulescu et al. Therefore, in the selection phase -- immediately preceding the execution phase -- it is vitally important to coordinate, and agree on a joint policy. There are two main ways to go about this. Firstly, let us view the game- theoretic perspective, in which we aim to find a joint policy that is in some sense stable, i.e., agents do not have an incentive to deviate from the joint policy. Stable solutions come in many different levels of strictness [19], from core stability, to Nash equilibria, to individual stability. Particularly challenging in this respect is how to figure out what the individual preferences are. When agents do not or cannot divulge their individual utility functions a priori, for example because it would be hard or even impossible to specify this utility function exactly, algorithms that aim to find stable outcomes must learn about the individual utility functions of the agents to learn whether a joint policy is stable or not [43]. Finding a stable joint policy in the planning or learning phase may seem to mitigate the need for an extensive selection phase; as no agent will have an in- centive to deviate from it, deviations should not happen. There are however two problems that could still arise. Firstly, if there are multiple possible stable solu- tions, the agents still need to agree on which of these to pick. Secondly, in repeated interaction settings, an agent could be spiteful, i.e., aim to be as disruptive to the elected stable solution as possible, in order to strengthen its hand the next time a stable solution must be selected. Secondly, there is the negotiation perspective, i.e., agents will try to ham- mer out a deal on which policy they will jointly execute. This has the advantage that even non-stable solutions -- that may offer better utility for all agents than the stable ones -- could be selected, as long as the agents are obligated to follow through. For example, the Automated Negotiating Agents Competition (ANAC) [46] considers three-agent negotiations in which agents negotiate about possible alternative outcomes. When each alternative is associated with its own vector cor- responding to different objectives, the agents will know that some outcomes are Pareto-dominated, and should therefore be excluded from consideration, but for the solutions that are in the Pareto coverage set, different agents may have dif- ferent preferences. In general, the outcome of such a negotiation should thus be a "deal" between the agents about which alternative joint policy from the coverage set to execute. Finally, we note that there is a special case of the team rewards and individual utilities, in which the number of objectives is equal to the number of agents, and the utility function of each agent would just be the value of the objective corresponding to that agent. This special case may seem identical to the single- objective multi-agent case with individual rewards, but there is in fact a significant difference. Specifically, it reflects the situation in which the agents can care about the rewards of the other agents, and can make (a priori or a posteriori) agreements on which division of rewards is admissible. In other words, it can be used to model various degrees of altruism. At a very minimum, the agents could all exclude Pareto-dominated solutions, leading to the situation in which agents always prefer to help the other agents to increase their rewards, as long as it does not cost them anything. A bit more drastically, the agents could agree to exclude a joint policy from consideration if another policy exists in which the total sum of the values for each objective/agent is at least the same, but more fairly distributed over the agents. This leads to the solution concept of Lorenz optimality [34], which we will discuss in Section 5.2.4. Multi-Objective Multi-Agent Decision Making 19 4.1.3 Team Reward and Social Welfare with Respect to Individual Utilities In the individual rewards setting, it is hard to predict, let alone optimise for, utility. This is because the agents have different agendas, leading to complex behavioural dynamics, in which agents react to each other's behaviours. This process may not converge to stable solutions. Furthermore, the individual utility functions may not be common knowledge. Fig. 7 The execution phase for the Team Reward and Social Welfare with Respect to Individ- ual Utilities setting. Please note that the social welfare can depend both on the utilities of the agents and the value/return vector. This figure depicts the SER optimality criterion, where the expected values (i.e., the average over many executions of the policies) will be input to u. Under ESR the input to u would be ρ, i.e., the returns for an individual roll-out. A different perspective on this problem is to take a step back from the self- interested agents and optimising for their individual utilities, and instead look at what would be a desirable outcome. For example, we can focus on what would be socially favourable by the agents. Once we have decided on what would be desirable, we can define social welfare as a social choice function, corresponding to the desirability of each outcome, and construct a system of payments that will make the joint policy converge to the desired outcome. This is known as mechanism design [121, Ch. 6]. It is important to note that the social welfare function can depend both on the value or return vector, as well as the individual utilities of the agents, as illustrated in Fig. 7. For example, in traffic, social welfare may depend on the pollution levels, as well as fairness between different vehicles in terms of their total expected time that they have to wait for traffic lights. In mechanism design, the challenge is to formulate the system of payments in such a way that the agents will be non-manipulable, i.e., do not have an incentive to lie about their preferences. If this succeeds, the agents will report their preferences truthfully, and from a planning perspective, the decision-problem becomes fully cooperative, i.e., aiming to collectively optimise the social choice function. For multi-objective decision problems, the social welfare perspective can for example be used by governments to control the parameters of tenders, to balance the different objectives for projects. For example, in a traffic network maintenance planning setting [100], the balancing of traffic delays and costs can be made a posteriori, by computing a convex coverage set for a cooperative multi-objective multi-agent MDP [89], because a non-manipulable mechanism exists for every dif- ferent weighting of the objectives. While mechanism design methods are very pow- erful, they do pose challenges. Specifically, they typically require (near-)optimal policies to be guaranteed, and they require agents to articulate their preferences exactly, in order for the mechanism to be non-manipulable. The first condition environmentVui(V)uj(V)uk(V)πijkW( ui(V), uj(V), uk(V) )20 Roxana Radulescu et al. poses restrictions on the type of planning methods than can be used; which is par- ticularly important in highly complex sequential decision problems. The second condition poses restrictions on the way the utility functions can be accessed. We discuss the implications of this in Section 5. 4.2 Individual Rewards Up until now, we have considered situations in which all agents have the same vector input to the utlity function, V under SER and ρ under ESR, but may have separate individual utilities, ui(V) or ui(ρ), with respect to this vector. We now consider situations in which the rewards, and therefore (expected) return vectors, are different for the individual agents. First, we note that we consider only two settings for individual rewards: in- dividual utilities and social choice. This is because when individual rewards are received, even if the utility functions for all agents are the same, the resulting utilities are still individual, and the interest of the agents may still be opposed. We observe that individual reward settings may seem similar to the team re- ward but individual utilities settings, regarding the fact that ultimately the joint policies will be selected on the utilities of the individual agents. However, whether the value (or return) vectors are identical or not, can have a profound impact on how complex it is to solve the decision problem. Specifically, a joint policy can often be excluded from consideration if all agents agree that executing a different policy would be better for all agents.4 When the rewards are shared, all agents will share the same joint policy outcomes and will thus always agree on whether a policy is Pareto-dominated or not. When the rewards are individual however, a joint policy can be the only Pareto-optimal policy (in terms of value or return vectors) for one agent, while it is dominated for another. In other words, settings with individual rewards are considerably more difficult to solve. 4.2.1 Individual Reward Individual Utility First let us consider the completely self-interested setting of agents with individual rewards and utilities. This results in the execution phase depicted in Figure 8. Fig. 8 The execution phase for the Individual Reward Individual Utility setting. This figure depicts the SER optimality criterion, where the expected values for each agent (i.e., the average over many executions of the policies) will be input to u. Under ESR the input to u would be ρi, i.e., the returns for an individual roll-out for an individual agent. 4 Note that this is not a sufficient condition in multi-agent settings though, as there may be equilibria that are Pareto-dominated. environmentViVjVkui(Vi )uj(Vj )uk(Vk )πijkMulti-Objective Multi-Agent Decision Making 21 For example [28] study cooperative games, in which coalitions of agents are formed that can obtain rewards in different objectives, and then divide the value of these objectives amongst themselves, leading to individual rewards. Subsequently, they consider what information regarding the utility functions of the agents is available, and whether stable coalitions can be found given this information. Because the individual rewards and individual utilities setting is highly com- plex, it is vitally important to exploit all available information there is with regards to the utility functions of the agents. For example, consider the situation in which all individual agents have the same utility function [28, 114], but it is not a priori clear what this utility function is, or the utility function is not fixed. For exam- ple, consider the case in which the objectives correspond to resources that can be sold on an open market. Because these prices can vary (possibly rapidly) over time, the agents will need to adjust their policies according to the latest possible price information. A multi-objective multi-agent model with individual rewards and individual utilities, may then help to predict how the agents will respond to changing prices. In general, the individual utility functions may be different for each agent, and various degrees of knowledge may exist about their shape or properties. In such settings, it may be hard to produce a sufficiently compact set of possibly viable joint policies to choose from or negotiate with. In this case, we suspect that interactive approaches [43], in which more information about the utility functions is actively pursued by querying the agents while planning or learning to limit the set of viable alternatives, will play an important role in future research. 4.2.2 Individual Reward and Social Choice with Respect to Individual Utilities Finally, let us consider the individual rewards and utilities, from the perspective of social choice. This leads to the situation in Figure 9, in which agents obtain individual value or return vectors, value these according to individual utilities, which are then weighed up, together with the individual value or return vectors, through a social welfare function. Fig. 9 The execution phase for the Individual Reward and Social Welfare with Respect to In- dividual Utilities setting. Please note that the social welfare can depend both on the utilities of the agents and the value/return vectors for each agent. This figure depicts the SER optimality criterion, where the expected values for each agent (i.e., the average over many executions of the policies) will be input to u. Under ESR the input to u would be ρi, i.e., the returns for an individual roll-out for an individual agent. As in the team reward setting, it is important to note that the social welfare function can depend both on the individual value or return vectors, as well as the individual utilities of the agents. For example, in auctions [78], social welfare may environmentViVjVkui(Vi )uj(Vj )uk(Vk )πijkW( ui(Vi ), uj(Vj ), uk(Vk ) )22 Roxana Radulescu et al. Utility Team Social choice Individual m a e T Coverage sets Mechanism design d r a w e lR a u d v d n I i i Mechanism design Coverage sets (+ Negotiation) Equilibria and stability concepts Equilibria and stability concepts Coverage Sets as best responses Fig. 10 Taxonomy mapping to solution concepts for multi-objective multi-agent system depend on attributes of the winning bid(s), as well as a fair outcome in terms of payments to the individual agents, that together the costs the agents need to make to execute their bids if chosen, typically determine the individual utilities. As in the team reward but individual utilities case, we aim to find a mechanism, i.e., a social welfare function, that forces agents to be truthful about their utility functions, such that the joint policy can be optimised with respect to a notion of social welfare. An interesting -- but to our knowledge unexplored -- aspect would be to investigate, in the case when individual reward vectors are common knowledge, but the preferences are (partially) unknown, whether such mechanisms could still be established, possibly through active querying to obtain information about the individual utility functions. 5 Solution Concepts In this section we introduce some of the main solution concepts which are featured in MAS and multi-objective optimisation research, as well as explaining how they relate to the scenarios described in our taxonomy above. In the context of MAS, it is difficult to identify what constitutes an optimal behaviour, as the agents' strategies are interrelated, each decision depending on the choices of the others. For this reason, we usually try to determine interesting groups of outcomes (i.e., solution concepts), to determine when the system can reach some form of equilibrium. Fig. 10 provides an overview of which of these solution concepts are relevant to each of the five settings in our multi-agent decision making taxonomy. Multi-Objective Multi-Agent Decision Making 23 5.1 Policies We introduce a few preliminary definitions regarding types of behaviour agents can learn, depending on the action selection strategy given a certain state or on whether or not time plays any role in the policy definition. A deterministic (or pure) policy is one where the same action a is always selected for a given state s (i.e., P r(as) = 1). A stochastic policy is one where actions in a given state are selected according to a probability distribution (i.e., P r(as) ∈ [0, 1],∀a ∈ A). The output of a stationary policy depends only on state, not on time. The output of a non-stationary policy may vary with both state and time. A deterministic game is one where the transition function is deterministic, i.e., the game always transitions to the same next state, for a given system state and joint action. While in single-objective decision problems it is often sufficient to take only deterministic stationary policies into account, it is known that in multi- objective decision problems stochastic or non-stationary policies can lead to better utility [85, 123] both under SER and ESR [84]. A mixture policy [104, 118] is a stochastic combination of deterministic policies (referred to as base policies). This technique has been used in single agent multi- objective settings to combine two or more deterministic Pareto optimal policies to satisfy a user's preferences. Mixing happens inter-episode only, rather than intra- episode. Vamplew et al. [118] note that switching between base policies during an episode will likely result in erratic and sub-optimal behaviour. Therefore, one of the available base policies is selected probabilistically at the beginning of each episode and followed for the entire episode duration. The aim is to determine mixture probabilities, which on average, after a large number of runs, will yield the desired long-term average return on each objective. As noted in Section 4.1.1, the team reward team utility setting is similar enough to single-agent multi-objective settings such that methods developed for one may be easily applied to the other; mixture policies are one such method which could feasibly be used in the team reward team utility setting. 5.2 Coverage Sets The optimal solution in single-agent multi-objective decision making is called a coverage set (CS) [85, 86]. A coverage set contains at least one optimal policy for each possible utility function, u(Vπ), i.e., if a set C is a coverage set then, under SER, ∀u ∈ U : max π∈Π u(Vπ) = max π∈C u(Vπ), and under ESR, ∀u ∈ U : max π∈Π E[u(ρ)π, µ0] = max π∈C E[u(ρ)π, µ0], returns, i.e., ρ = (cid:80)∞ where Π is the space of all possible (and allowed) policies, ρ are the vector-valued t=0 γtrt and U is the set of all possible utility functions. Furthermore, coverage sets do not contain dominated policies, π ∈ C → ∃u ∈ U : u(Vπ) = max π(cid:48)∈C u(Vπ(cid:48) ), 24 under SER, and, Roxana Radulescu et al. π ∈ C → ∃u ∈ U : E[u(ρ)π, µ0] = max π(cid:48)∈C E[u(ρ)π (cid:48) , µ0], under ESR, i.e., a coverage set should only contain policies that are optimal for some utility function u. Finally, algorithms should aim to construct coverage sets that are as small as possible, but as coverage sets are not unique, so constructing a minimally sized one is far from trivial. 5.2.1 Motivations for Coverage Sets in Multi-Agent Settings In single-agent settings, coverage sets need to be constructed with respect to any possible utility function allowed by the problem specification. However, due to the single-agent nature, it can be assumed that ultimately, in the execution phase, there will be one true utility function that governs user utility. Multi-agent settings are more complex; the different agents can represent different interests, and may be optimising for different utility functions. Nonetheless, there are many multi-agent settings for which coverage sets are the appropriate solution concept. The first and most straightforward motivation is the team reward and team utility setting described in Section 4.1.1. This is a fully cooperative setting; all rewards and the utility derived from that is shared between all agents. Therefore, there is only one true utility function in the execution phase, and the motivation for coverage sets being the right solution concept is the same as for single-agent multi- objective decision making. For example, this is the case when there are multiple agents belonging to the same team or organisation are tackling a problem together, e.g., a soccer team or different agents belonging to the same company. However, team utility is not strictly necessary for coverage sets to be useful. In a team reward but individual utility setting, coverage sets could be used if all agents will agree (preferably contractually) that they will always execute a policy that is potentially optimal. In this case, a coverage set can be computed as the input to a negotiation [45, 46] between the agents of which policy to execute. Note that this strategy of computing a coverage set and then negotiating does not trivially apply to individual reward settings. In the case of individual rewards, a joint policy can be optimal for one agent, while it can be strictly dominated for another. Generalising the concept of a coverage set to individual reward settings is an open question that would merit investigation. Furthermore, in an individual utility setting, a coverage set can also be a set of possible best responses to the behaviours of the other agents. Of course, one needs a different coverage set per combination of possible behaviours for all the individual other agents. This may quickly become infeasible if the set of possible policies of the other agents becomes large. However, if one can model the opponents using a small set of possible behaviours this may be a viable approach. Finally, there is a uniquely individual rewards setting coverage set, for the special case that each objective corresponds to one agent, and the objectives of other agents are seen as secondary objectives. In other words, this is the case where agents are at least partially altruistic. This concept is called a Lorenz optimal set, which we discuss in Section 5.2.4. Multi-Objective Multi-Agent Decision Making 25 5.2.2 Convex Coverage Sets A convex coverage set is the optimal solution set when it can be assumed that the utility functions of all agents are linear. This is a salient case in the multi- objective decision making literature, and for example applies in the case where each objective corresponds to a resource that can be bought or sold on an open market. Specifically the utility functions are assumed to be the inner product between a vector of weights w and the value vector of the joint policy Vπ, i.e., ui(Vπ) = wi · Vπ. (9) Please note that for this type of utility function, there is no difference between SER and ESR, as E[w · ρπ, µ0] = w · E[ρπ, µ0] = w · Vπ. In the case of linear utility functions, the undominated set -- the convex hull -- is defined as follows: Definition 2 The convex hull (CH) is the subset of the set of all admissible joint policies Π for which there exists a w for which the linearly scalarised value is maximal: CH(Π) = {π : π ∈ Π ∧ ∃w∀(π (cid:48) ∈ Π) w · Vπ ≥ w · Vπ(cid:48)}. (10) Please note that in this definition, we assume a team reward setting, such that Vπ is a single vector. One problem with the CH is that it can be undesirably large; and in the case of stochastic policies, often infinitely large. However, in such cases a convex coverage set (CCS) can often been defined that is much more compact5: Definition 3 A set CCS(Π) is a convex coverage set (CCS) if it is a subset of CH(Π) and if, for every w, it contains a policy whose linearly scalarised value is maximal: CCS(Π) ⊆ CH(Π) ∧ (∀w)(∃π) (cid:48) ∈ Π) w · Vπ ≥ w · Vπ(cid:48)(cid:17) (cid:16) π ∈ CCS(Π) ∧ ∀(π . While in the case of individual utility, the actual wi can differ per agent, the CCS contains at least one optimal policy for every wi, and therefore forms a suitable starting point for finding possible compromises based on the assumption that all agents have a linear utility function. For example, a strategy could be to try to estimate each wi and take the average weights vector across all agents to select the default compromise. Of course, agents may also want to negotiate [45, 46] in order to get a better deal than such a default compromise. (11) 5.2.3 Pareto Coverage Sets For monotonically increasing but non-linear utility functions, the undominated and coverage sets become significantly larger than for linear utility functions. To be able to define a coverage set for this setting under SER we must first define the concept of Pareto dominance: 5 For details on why this is so, please refer to [85]. 26 Roxana Radulescu et al. Definition 4 A joint policy π Pareto-dominates another joint policy π(cid:48) when its value is at least as high in all objectives and strictly higher in at least one objective: Vπ (cid:31)P Vπ(cid:48) ⇔ ∀i, V π i ≥ V π(cid:48) i ∧ ∃i, V π i > V π(cid:48) i . (12) Looking at this definition, it is clear that no Pareto-dominated policy can ever have a higher utility under a monotonically increasing utility function: Vπ (cid:31)P Vπ(cid:48) → ui(Vπ) ≥ ui(Vπ(cid:48) ). As long as being monotonically increasing is the only assumption we can make about the utility function, this is in fact the only thing that can be said of the rel- ative preferences across all possible utility functions. Therefore, we use the concept of Pareto dominance to define the undominated set for monotonically increasing utility functions, the Pareto front: Definition 5 The Pareto front is the set of all joint policies that are not Pareto dominated by any other joint policy in the set of all admissible joint policies Π: P F (Π) = {π : π ∈ Π ∧ ¬∃(π (cid:48) ∈ Π), Vπ(cid:48) (cid:31)P Vπ}. (13) A Pareto coverage set (PCS) of minimal size can be constructed by selecting only one policy of the policies with identical value vectors from the P F (Π): Definition 6 A set P CS(Π) is a Pareto coverage set if it is a subset of P F (Π) and if, for every policy π(cid:48) ∈ Π, it contains at least one policy that either dominates π(cid:48) or has equal value to π(cid:48): P CS(Π)⊆ P F (Π) ∧ ∀(π π∈ P CS(Π) ∧ (Vπ (cid:31)P Vπ(cid:48) ∨ Vπ = Vπ(cid:48) (cid:48)∈ Π)(∃π) (cid:16) (cid:17) ) . (14) Negotiating a good compromise from a set of alternatives with different values for all objectives is a typical setting for negotiation [46]. Note that in multi-objective settings, agents and/or users are often incapable of specifying their utilities numer- ically [133]. However, recently there has been research in automated negotiation focusing on preference uncertainty [117], i.e., uncertainty about the individual util- ity functions, and eliciting preferences [9], making realistic negotiation with the PCS of a multi-objective decision problem as input, possible. Under ESR the situation becomes significantly more complex, i.e., in general, the undominated set is defined as: Definition 7 The undominated set of policies (U) under possibly non-linear mono- tonically increasing u, under ESR, is the subset of the set of all admissible joint policies Π for which there exists a u for which the scalarised value is maximal: U (Π) = {π : π ∈ Π ∧ ∃(u∈U) ∀(π (cid:48)∈ Π) E[u(ρ)π, µ0] ≥ E[u(ρ)π (cid:48) , µ0]}. (15) However, this is very hard to determine without further information about u. To our knowledge no research has yet been done into constructing (approximate) undominated or coverage sets under ESR. In fact, the available research in single- agent MORL typically assumes that u is known [84]. Multi-Objective Multi-Agent Decision Making 27 5.2.4 Lorenz Optimal Sets A uniquely multi-agent coverage set is the Lorenz optimal set [77]. Underlying this solution concept is the assumption that each objective corresponds to the interest of each individual agent. Furthermore, it is assumed that the interests of the other agents are an objective for every agent. In other words, the agents are at least in part altruistic. Finally, it is assumed that "more equal" solutions - we will define this exactly below -- are better if sum of utilities does not increase. This final assumption corresponds to a (rather minimal) concept of fairness. The use-case for Lorenz optimal sets is: all agents agree that fair solutions are better, hence a set of possibly fair solutions will be computed, after which the agents will negotiate which solution from this set to select. It is thus vital that the agents can rely on the selected solution being followed and that no individual agent will enrich itself to the detriment of the group. This can either be enforced contractually, or simply by the notion that the group of agents will have to rely on each other in the future, and that agents that do not follow the convention will no longer be allowed to participate in other decision problems in which the same agents will need to cooperate. The underlying idea of the Lorenz notion of fairness, is the so-called Robin- Hood transfer: if in a vector objective i has a higher value, vi than objective j, vj, then transferring part of the difference, i.e., setting the value of vi to vi − δ and vj to vj + δ, for 0 < δ ≤ vi − vj, yields a fairer, and therefore better value vector. More formally, this can be captured in the concept of Lorenz domination. To test whether a vector Vπ Lorenz dominates a vector Vπ(cid:48) , both vectors are first projected to their corresponding Lorenz vectors: Definition 8 The Lorenz vector L(Vπ) of a vector Vπ is defined as: v(1), v(1) + v(2), ... , v(i) , where, v(1) ≤ v(2) ≤ ... ≤ v(N ), correspond to the values in the vector Vπ sorted in increasing order. i=0 NB: this definition is under SER. To our knowledge no research has been done with regards to Lorenz optimality under ESR. Definition 9 A vector Vπ Lorenz dominates ((cid:31)L) a vector Vπ(cid:48) when: Vπ (cid:31)L Vπ(cid:48) ⇔ L(Vπ) (cid:31)P L(Vπ(cid:48) ), i.e., when the Lorenz vector of Vπ Pareto dominates the Lorenz vector of Vπ(cid:48) . Definition 10 The Lorenz Optimal Set is the set of all joint policies that are not Lorenz dominated by any other joint policy in the set of all admissible joint policies Π: LOS(Π) = {π : π ∈ Π ∧ ¬∃(π (cid:48) ∈ Π), Vπ(cid:48) (cid:31)L Vπ}. (16) A Lorenz coverage set (LCS) of minimal size can be constructed by selecting only one policy of the policies with identical value vectors from the LOS(Π), similar to constructing a PCS from a PF. (cid:32) (cid:33) N(cid:88) 28 5.3 Nash Equilibria Roxana Radulescu et al. When multiple self-interested agents learn and act together in the same environ- ment, it is generally not possible for all agents to receive the maximum possible reward. Therefore, MAS are often designed to converge to a Nash equilibrium [105] (NE). This notion of equilibrium was first introduced by Nash [72], and is one of the most important concepts used to analyse MAS [127]. Consider a multi-agent system with n agents, where π = (π1, . . . , πi, . . . , πn) is their joint policy, with πi representing the stochastic policy of agent i. We also define π−i = (π1, . . . , πi−1, πi+1, . . . , πn) to be a joint policy without the policy of agent i. We can thus write π = (πi, π−i). Definition 11 A joint policy πN E leads to a Nash equilibrium if for each agent i ∈ {1, ..., n} and for any alternative policy πi: ≥ V ,πN E−i ) (17) (πi,πN E−i ) i (πN E V i i Whenever the above inequality holds true for all possible policies and for all agents in a MAS, a Nash equilibrium exists. In other words, a Nash equilibrium occurs whenever any individual agent cannot improve its own return by changing its behaviour, assuming that all other agents in the MAS continue to behave in the same way. In cooperative MAS (i.e., the team reward scenario), coordinating agents' actions to achieve the highest possible system welfare is already a difficult prob- lem. While it is possible for multiple individual learners in a cooperative MAS to converge to a point of equilibrium, whether they will converge to an optimal joint policy (one which maximises the system welfare) depends on the specific learning algorithm and reward scheme used. 5.3.1 Nash Equilibrium in Multi-Objective Multi-agent Settings In multi-objective decision making, each agent is trying to optimise his return along a set of objectives. Each agent needs to also make compromises between competing objectives on the basis of his/her utility function. As a motivation for why NE is an appropriate solution concept in MOMAS, we look at the team reward individual utility (section 4.1.2) and individual reward individual utility (section 4.2.1) scenarios. In both these cases, the utility derived by each agent from the received reward is different, regardless if this reward is the same or not for all the agents. These constitute the most difficult scenarios in our taxonomy. Furthermore, one should also consider which optimisation criteria are best to use, based on what each agent is looking to optimise. Depending on whether an agent cares about average performance over a number of policy executions, or just the performance of a single policy execution [84], we can define the concept of Nash Equilibrium from the perspective of the two multi-objective optimisation criteria defined in Section 2.3: ESR and SER. To simplify our notation let us denote the discounted sum of rewards received γtri,t. We can then re-write the expected value of a joint by agent i by: ρi = policy π, given the distribution µ0 over initial states as: V π = E [ρi π, µ0]. t=0 ∞(cid:80) Multi-Objective Multi-Agent Decision Making 29 Definition 12 (Nash equilibrium for Expected Scalarised Returns) A joint policy πN E leads to a Nash equilibrium under the Expected Scalarised Returns cri- terion if for each agent i ∈ {1, ..., n} and for any alternative policy πi: ui(ρi) (πN E i , πN E−i ), µ0 ui(ρi) (πi, πN E−i ), µ0 (18) (cid:105) ≥ E(cid:104) E(cid:104) (cid:105) (cid:105)(cid:17) i.e., πN E is a Nash equilibrium under ESR if no agent can increase the expected utility of her returns by deviating unilaterally from πN E. Definition 13 (Nash equilibrium for Scalarised Expected Returns) A joint policy πN E leads to a Nash equilibrium under SER if for each agent i ∈ {1, ..., n} and for any alternative policy πi: (cid:16)E(cid:104) ui ρi (πN E i , πN E−i ), µ0 (cid:105)(cid:17) ≥ ui (cid:16)E(cid:104) ρi (πi, πN E−i ), µ0 (19) i.e. πN E is a Nash equilibrium under SER if no agent can increase the utility of her expected returns by deviating unilaterally from πN E. Under non-linear utility functions, it has been shown that the choice of optimi- sation criterion can alter the set of Nash equilibria [98]. Furthermore, under SER, even in a multi-objective normal form game, NE need not exist. 5.4 -approximate Nash Equilibria An -approximate Nash equilibrium [76] occurs when an individual agent cannot increase its return by more than an additive > 0 by deviating from its policy, assuming that all other agents continue to behave in the same way. In other words, an agent will not care to switch his policy, if the obtained gain is too small. Definition 14 A joint policy πN E leads to a -Nash equilibrium if for each agent i ∈ {1, ..., n} and for any alternative policy πi: (πN E V i i ,πN E−i ) ≥ V (πi,πN E−i ) i − (20) -Nash equilibria can be envisioned as regions surrounding any Nash equilib- rium. All the definitions for NE under SER and ESR can also be adapted for the case of -Nash equilibria by subtracting from the right side of each inequality. 5.5 Correlated Equilibria A correlated equilibrium (CE) is a game theoretic solution concept proposed by Aumann [7] in order to capture correlation options available to the agents when some form of communication can be established prior to the action selection phase. Another way to think about this concept is to envision an external device sampling from a given distribution and providing each agent with a private signal (e.g., a recommended action) at each time-step. Given this private signal, each agent can then independently make a decision on how to act next. For this work we will consider that the signals take the form of action recommendations. 30 Roxana Radulescu et al. While previously discussed policies define state-based action probabilities in- dependently for each agent, a correlated policy σ represents a probability distribu- tion over the joint-action space A of all the agents in the system (i.e., P r(as) ∈ [0, 1],∀a ∈ A). Thus, correlated policies introduce explicit dependencies between the agents' behaviours. Let us define πσ = (πσ n) as the joint policy of the agents when following the recommendation provided according to a correlated policy σ. Definition 15 A correlated policy σCE is a correlated equilibrium if for each agent i ∈ {1, ..., n} with its corresponding policy under σCE, πσCE , and for any alterna- tive policy πi: 1 , . . . , πσ i (πσCE V i i ,πσCE −i ) ≥ V (πi,πσCE i −i ) (21) Thus, a correlated equilibrium ensures that no player can gain additional return by deviating from the suggestions, given that the other players follow them as well. 5.5.1 Correlated Equilibria in Multi-Objective Multi-agent Settings Similarly to the Nash equilibria case, solution concepts such as correlated equilibria can be used in scenarios in which each agent derives a different utility from the received reward vector, i.e., the team reward individual utility (section 4.1.2) and individual reward individual utility (section 4.2.1) settings. Furthermore, we can also define correlated equilibria from the perspective of the two possible optimisation criteria: ESR and SER, when considering multi-objective multi-agent decision making problems. We will again denote the value of a joint policy π as V π = E [ρi π, µ0], where ρi = ∞(cid:80) γtri,t. t=0 Definition 16 (Correlated equilibrium for Expected Scalarised Returns) A correlated policy σCE is a correlated equilibrium under ESR if for any agent i ∈ {1, ..., n} with its corresponding policy under σCE, πσCE , and for any alterna- tive policy πi: i ui(ρi) (πσCE i , πσCE −i ), µ0 ui(ρi) (πi, πσCE −i ), µ0 (22) (cid:105) ≥ E(cid:104) E(cid:104) (cid:105) (cid:105)(cid:17) i.e. σCE is a correlated equilibrium under ESR if no agent can increase the ex- pected utility of her returns by deviating unilaterally from the action recommen- dations in σCE. Definition 17 (Correlated equilibrium for Scalarised Expected Returns) A correlated policy σCE is a correlated equilibrium under SER if for any agent i ∈ {1, ..., n} with its corresponding policy under σCE, πσCE , and for any alterna- tive policy πi: i ρi (πσCE i , πσCE −i ), µ0 ρi (πi, πσCE −i ), µ0 (23) (cid:105)(cid:17) ≥ ui (cid:16)E(cid:104) (cid:16)E(cid:104) ui i.e. σCE is a correlated equilibrium under SER if no agent can increase the utility of her expected returns by deviating unilaterally from the given action recommen- dations in σCE. 6 6 When considering a CE-based approach, an agent is able to calculate his expected return given one correlation signal, but also an expected return given all the possible signals. This Multi-Objective Multi-Agent Decision Making 31 It has also been shown that the set of correlated equilibria will be altered, depending on the optimisation criteria used [98]. Furthermore, under SER, with non-linear utility functions, in multi-objective normal form games, CE need not exist when taking the expectation over all the possible correlation signals. 5.6 Coalition Formation and Stability Concepts A different perspective on multi-agent decisions is that taken by cooperative game theory [19]. Cooperative game theory studies settings where binding agreements among agents are possible. A central problem is therefore that of coalition forma- tion, i.e., finding (sub)groups of agents that are willing to make such a binding agreement with each other. In the models in cooperative game theory, the utility for each agent is directly derived from the coalition the agents end up in, however, one can imagine that under the hood, the coalition works together cooperatively (based on their binding agreement) in a sequential decision problem that results in this utility. We further note, that the word cooperative does not imply team util- ity; typically, the agents will have their own utility functions. Hence, the solution concepts from cooperative game theory apply to the individual utility settings. To illustrate the solution concepts for multi-objective cooperative game theory, we use the multi-criteria coalition formation game (MC2FG) [43, 113, 114]. Such a game consists of a set of agents, N , each with their own utility function ui(q), and a quality/reward function q(S) that maps each possible subset, i.e., coalition, of the agents S ∈ N to a value or quality vector, that each agent in that coalition will receive. That is, we are in an individual utility setting. Definition 18 A multi-criteria coalition formation game (MC2FG) is a triple (N, q,U) where N a finite set of agents, q : 2N → Rd is a vector-valued reward function that represents the quality q(S) of a subset, i.e. coalition, of agents S ⊆ N , and ui ∈ U are the utility functions for each agent i ∈ N . The MC2FG is a useful model to study for multi-objective multi-agent systems. Specifically, if in a multi-agent system with multiple objectives, the agents need to form coalitions to cooperate to gain a value vector, the most straightforward case is a MC2FG, i.e., given the coalition the value vector can exactly be predicted independently of the other coalitions, but agents can have different preferences between possible value vectors. Therefore, MC2FGs form a minimal model to study the feasibility of contract negotiations between agents in multi-objective multi-agent decision making. The goal in an MC2FG is to find a partition, ψ, of agents into coalitions that are stable. That is, the coalitions will not break apart. For this notion of stability, there are multiple possible versions, from strong to weak: core stability, Nash stability, and individual stability. We denote the coalition (subset of agents) which agent i is in according to ψ as ψ(i). A partition ψ is individually rational if no agent strictly prefers staying alone to their own coalitions, i.e. ∀i : ui(q(ψ(i)) ≥ ui(q({i})). allows one to define two variants for CE under SER: the single-signal CE (when agents have multiple interactions under the same given signal) and multi-signal CE (when agents receive a new signal after every interaction) [98]. For this work we define the more general case of multi-signal CE. 32 Roxana Radulescu et al. Definition 19 A coalition S ⊆ N is said to block a partition ψ if every agent strictly prefers S to ψ(i), i.e., ∀(i ∈ S) : ui(q(ψ(i)) < ui(q(S)). Definition 20 A partition ψ of N is core stable (CR) if no (non-empty) coalition S ⊆ N blocks ψ. Beside CR, there are two key stability concepts that represent immunity to deviations by individual players. An agent i, wants to deviate from ψ(i) to another coalition in ψ, S, if it prefers S ∪ {i} to ψ(i), i.e., ui(q(ψ(i)) < ui(q(S ∪ {i})). A player j ∈ S would accept such a deviation if it prefers S ∪ {i} to S, i.e., uj(q(S) ≤ uj(q(S ∪ {i})). Definition 21 A deviation of i from ψ(i) to S is an NS-deviation if i wants to deviate from ψ(i) to S. Definition 22 A deviation of i from ψ(i) to S is an IS-deviation if it is an NS- deviation and all players in S accept it. Definition 23 A partition ψ is Nash stable (NS) if there are no NS-deviations for any agent i, from its coalition ψ(i) to any other coalition S ∈ ψ or to ∅. Definition 24 A partition ψ is Individually stable (IS) if there are no IS-deviations for any agent i, from its coalition ψ(i) to any other coalition S ∈ ψ or to ∅. Every single-criterion coalition formation game has at least one partition that is core stable and individually stable [43]. However, this is not necessarily so in the multi-objective case. This is because in a single-objective coalition formation game, the utility of a coalition is the same for each agent, i.e., the scalar quality/reward of the coalition. However, in the multi-objective case, all agents that are in a coalition S receive the same reward vector q(S), but they may value these vectors differently. In fact, Igarashi and Roijers [43] show that MC2FGs do not necessarily have core, Nash, nor individually stable partitions by counter-example resulting in the following Theorem: Theorem 1 For any positive integer n and for any 0 < ε < 1 there exists an MC2FG (N, q,{wi : i ∈ N}), where wi is the weights vector for the linear utility function of agent i, which admits neither a core nor individually stable partition, where the number of players N = n, the number of criteria m = 2, and wi,k − wi,k ≤ ε for any i, j ∈ N and either objective (k). So, this theorem implies that even when the number of objectives is smaller than the number of agents, and the difference between the utility functions (even if they are linear) is arbitrarily small, stable partitions do not need to exist. This has important consequences for multi-objective multi-agent systems in general, as MC2FGs are such a minimal model of finding cooperative subsets of agents that could contractually agree on a value vector. Because no stable solutions need to exist, such contract negotiations could go on forever (agents repeatedly switching between coalitions before signing the contract), if all agents just aim to optimise their individual utilities. We believe this means that a thorough investigation of (the compatibility of) negotiation techniques for various multi-objective multi- agent decision problems on the basis of coverage sets, under different optimali- sation criteria (i.e., ESR versus SER) is required. Furthermore, the fact that the Multi-Objective Multi-Agent Decision Making 33 stability of coalitions cannot be guaranteed could have a strong impact on future interactive approaches 7 as well. While the prospects of such interactive approaches seem good, as [43] have shown that individually stable coalitions can often be found interactively under linear utility functions in MC2FGs, it is not clear what will happen for non-linear utility functions under SER or ESR, or in learning settings where the estimated value vectors of different joint policies of changing coalitions, may change. 5.7 Social Welfare and Mechanism Design In this section, we have so far taken the position of the individual agents. How- ever, we can also take a system perspective, i.e., we can look at what the socially desirable outcomes of a multi-agent decision problem would be. In Section 4.1.3 and 4.2.2, we have looked at the execution phase of such settings and defined the social welfare function, i.e., a function that should be maximised if we want to find socially desirable outcomes. In game theory, the field of mechanism design takes the system's perspective for multi-agent decision problems: taking an original decision problem where the agents have individual reward functions that are unknown to the other agents and the "owner" of the game, as well as a social welfare function as input, the aim is to design a system of additional payments that would a) force the agents to be truthful about their individual utilities, and b) leads to solutions that are (approximately) optimal under the social welfare function. In single-objective multi-agent decision problems, the individual utilities of the agents are simply the individual (expected) (cumulative) rewards that the agents receive. The agents can be assumed to know these rewards, and act accordingly. This is for example the case in public tenders, where different companies know their own costs and profit margins of their possible proposals, but do not broadcast this information to others. In multi-objective settings, the situation is more complex, as the individually received rewards determine the individual utilities via individual private utility functions. These utility functions can have different properties. In general, it might even be very hard, or even impossible to articulate these functions, so being "truthful" about their utilities might be infeasible from the get-go. Nevertheless, it is possible to design mechanisms for some multi-objective multi-agent problems if the individual utilities can be articulated. First, we observe that if the utility functions are linear, the inner product with weights distributes over all expectations. Hence, it is possible to even design mechanisms that are agnostic about the weights, compute a convex coverage set (see Section 5.2.2) of possibly socially desirable outcomes, and choose the weights a posteriori. This enables the designer/owner of the decision problem to make an informed decision about which weights to use. For example, in a public tender for traffic maintenance by Scharpff et al. [100, 89], the objectives of costs and traffic hindrance should both be minimised. Because of mechanism design, all agents need to be truthful; what- ever weights (and resulting penalties) are put on traffic hindrance, it is in the best interest of the agents to be truthful about their costs, making it possible for 7 Interactive approaches intertwine preference elicitation and learning about the decision problem [92, 91]. 34 Roxana Radulescu et al. the owner of the game to assume that given the mechanism, all agents will be fully cooperative, solve the problem as an MOMMDP, and choose the weights a posteriori. For specific cases of non-linear utility functions, it is also possible to devise mechanisms. For example, Grandoni et al. [35] assume individual utility functions with a primary objective that should be maximised, and other objectives that need to achieve at least a threshold value. The utility is the value of the first objective in the case that all thresholds are met, but negative infinity if the thresholds are not met. They show that for such cases, effective mechanisms can be designed, and solutions can be found within a reasonable amount of time. An interesting and different approach to social welfare is taken by Mouaddib et al. [71], who cast a decentralised sequential multi-agent problem with individual (scalar) reward functions as a multi-objective problem. Specifically, besides its main objective an agent will model its positive impact on the group as well as the nuisance it causes to other group members as separate objectives. Even though this work provides no strong guarantees, the authors show empirically that these additional objectives in combination with a social welfare function can lead to good emergent group behaviour in very hard -- decentralised partially observable multi-agent -- decision problems. 5.8 Other Solution Concepts The concepts discussed so far do not form in any way an exhaustive list for what constitutes a solution in a MOMAS. We briefly present below a few other possible solution concepts that have been discussed in the literature. An early discussion on how to extend equilibria concepts from single-objective games to multi-objective settings can be found in [103], where the concepts of weak and strong equilibria are proposed as extensions of NE. These concepts are defined using vector domination and thus are called Pareto-Nash Equilibria [13, 122, 56]. Continuing the game theoretic perspective, [48] extends the concept of evolutionary stability for multi-objective games. Cyclic equilibria [29, 68, 131] have been proposed as a solution concept for games where no stationary equilibrium exists. A cyclic equilibrium is a non-stationary joint policy where agents have no incentive to deviate unilaterally [131]. Cyclic equilibria cycle repeatedly through a set of stationary policies. Similar to -NE, an -correlated cyclic equilibrium is defined as a situation where no agent can improve its value by more than at any stage by deviating unilaterally [131]. 6 Algorithmic Approaches and Applications In the section, we survey related work on algorithmic approaches to MOMAS, as well as applications of multi-objective multi-agent decision making. The works we survey are organised into three broad categories: those which aim to derive coverage sets are discussed in Section 6.1, those which aim to apply stability and equilibria concepts are discussed in Section 6.2 and finally methods which employ mechanism design are discussed in Section 6.3. Within these categories, the works Multi-Objective Multi-Agent Decision Making 35 surveyed are further classified according to our taxonomy on the basis of the reward and utility functions used. 6.1 Coverage Sets 6.1.1 Team Reward - Team Utility (TRTU) Multi-objective coordination graphs (MOCoGs)8 are one of the most studied mod- els for cooperative multi-objective multi-agent systems, and in particular for team reward team utility. One reason for this is that it is the simplest model to express and exploit loose couplings, i.e., the fact that in multi-agent systems the rewards can often be factorised into a sum over small components, i.e., local reward func- tions, that depend on small (but possibly overlapping) subsets of agents. However, finding suitable joint actions in a MOCoG is also key to finding, e.g., coverage sets for MOMMDPs (as is also the case for single-objective CoGs and MMDPs [51, 52]). In MOCoGs, a lot of research focuses on finding (approximate) Pareto coverage sets (PCSs), using various algorithmic approaches. These approaches often extend single-objective methods by adapting the inner workings of such methods to be able to handle sets of value vectors rather than single scalar values. Examples of such methods are multi-objective bucket elimination (MOBE) [94, 93], also known as multi-objective variable elimination (MOVE, which is the more common in the planning and reinforcement learning communities), which solves a series of local sub-problems to eliminate all agents from a MOCoG in sequence, by finding local coverage sets as best responses to neighbouring agents. Other such methods include multi-objective Russian doll search [95], multi-objective (AND/OR) branch-and- bound tree search [65, 66, 96] using mini-bucket heurtistics [94, 65], Pareto local search [44], and multi-objective max-sum [22]. Many of these papers note that PCSs can grow very large very quickly, making finding exact PCSs infeasible. Therefore computing bounded approximations [66] can be necessary. On the other hand, Roijers et al. [88, 89, 90] compare the computational and memory complexity of computing convex and Pareto coverage sets (CCSs and PCSs). They observe that the size of PCSs typically grows much faster with the number of agents in a MOCoG than the size of CCSs, and that often CCSs suffice, e.g., in the case that mixture policies are allowed (Section 5.1). It can therefore be highly preferable to focus on finding CCSs, especially in problems with many agents. They propose several methods to do so with different computation-memory trade-offs. Specifically for finding CCSs, they propose and compare inner loop methods to outer loop methods, i.e., methods that construct CCSs by iteratively solving scalarised instances of the multi-objective decision problem. Outer loop methods are more memory-efficient, and significantly faster for smaller numbers of objectives, while inner loop methods are faster for many objectives. Finally, they note that ε-approximate CCS can efficiently be computed using outer loop 8 Because the MOCoG model is a flexible multi-objective graphical model that can be used for many types of problems, and has been used by many research communities, the MOCoG is known under many different names. These include: multi-objective weighted constraint sat- isfaction problems (MO-WCSPs) [94] and Multi-objective Constraint Optimisation Problems (MOCOPs) [65, 67]. 36 Roxana Radulescu et al. methods. Anytime approximations to CCSs can also be effectively constructed using an outer loop method that employs variational (inference) methods [87]. Wilson et al. [125] consider methods to compute coverage sets for MOCoGs when more information about the possible utility function(s) is available. They assume that along with the standard notion of the shape of the utility function9, they are provided with a set of preferences that users expressed a priori. They integrate these preferences into AND/OR branch-and-bound, and show that this can often lead to much more efficient computation than would be required to compute a PCS. To achieve this, they pose additional constraints on the utility function, that are only fulfilled by linear utility functions. Therefore, doing the same for arbitrarily shaped utility functions is still an open problem. Multi-objective multi-agent MDPs (MOMMDPs) -- in the literature often re- ferred to as cooperative multi-objective stochastic games (cooperative MOSGs) -- are another frequently used model for cooperative multi-objective multi-agent decision making problems. Some recent works have sought to derive coverage sets in MOMMDPs using reinforcement learning or evolutionary algorithms (e.g. [129, 128, 63, 60, 59, 58]). As in single-objective MMDPs, learning joint policies which co- ordinate agents' actions to get the desired outcome(s) in MOMMDPs is a difficult problem. When individual agents learn using the system reward and the same util- ity function (i.e TRTU), it is difficult for any one agent to evaluate how its actions affected the system utility, due to the effect of the other agents in the system. This is referred to as the credit assignment problem in the single-objective MAS literature. Reward shaping is one solution which has been proposed to address this problem, where the reward which is usually received from the environment is augmented with an additional shaping term, with the goal of providing a more informative reward signal to the agents in a MAS. Specific forms of reward shaping which have been applied to cooperative MOMAS include difference rewards (D) [126] and potential-based reward shaping (P BRS) [73]10. Yliniemi [129] and Yliniemi and Tumer [128] present the first work that con- siders the use of reward shaping in a cooperative MOMARL setting. Their work compares the effectiveness of the difference reward with that of two typical MARL reward structures: local rewards (L) and global rewards (G). Experimental work in a multi-objective congestion problem, and a multi-objective robot coordination problem confirms that D can improve MOMARL performance when compared to L or G, both in terms of learning speed and the quality of the set of non-dominated solutions found. [128] also demonstrates that D can be used effectively with multi- objective evolutionary algorithms, in a series of experiments where it is applied to shape the fitness function of the Elitist Non-dominated Sorting Genetic Algorithm- II (NSGA-II). [64] evaluate the effect of D in an electricity generator scheduling problem. Mannion et al. [60] provide a theoretical analysis of P BRS in single- and multi-agent MORL settings, demonstrating that the Pareto relation between (joint) policies is preserved when P BRS is used. Mannion et al. [62] also provide 9 Or, in their original paper, the equivalent concept of domination. 10 Although individualised reward shaping implies that each agent receives a different reward, we have classified these works under the TRTU setting as all agents use the same utility function and the individual shaped rewards are still aligned with the global (system) rewards. Reward shaping might also be useful in combination some of the other settings in our taxonomy and solution concepts discussed in Section 5, although only the TRTU setting with coverage sets has been explored to date. Multi-Objective Multi-Agent Decision Making 37 a theoretical analysis of D in MOMMDPs, demonstrating that the relative values of actions (and therefore the Pareto relation between actions) is preserved when D is applied in MOMMDPs. A comprehensive analysis of the effects of both D and P BRS when generating coverage sets via learning in cooperative MOMARL settings is presented in [59] and [58]. On the application side, Ahmad et al. [2] consider the problem of multi-core processing, or, more specifically, energy aware task allocation for optimising energy use versus performance. They employ a cooperative game theory perspective and transform the problem into a max-max-min one to generate solutions for differ- ent energy-time trade-offs. Bone and Dragi´cevi´c [12] are interested in the setting of natural resource management, where each agent represents a forest harvesting company. Agents need to learn how to harvest wood in order to maximise eco- nomic profit and minimise ecological impact. They use a step utility function to transform the problem into a single-objective one and learn an optimal policy as independent reinforcement learners. Babbar-Sebens and Mukhopadhyay [10] con- sider a water resource management system modelled as a multi-objective game, where the players use a simple reinforcement learning or genetic algorithm ap- proach in order to find a set of solutions corresponding to various linear utility functions. Focusing on traffic optimisation, Houli et al. [41] develop the multi-RL algorithm, a multi-agent reinforcement learning approach which selects an optimi- sation objective depending on the real-time traffic conditions. 6.1.2 Team Reward - Individual Utility (TRIU) Aoki et al. [6] model a multi-stage flow systems as a MAS, where each agent (i.e., service centre) represents a different objective. They use a distributed reinforce- ment learning framework and propose a bi-directional decision making mechanism to address the multi-objective nature of the problem. 6.1.3 Individual Reward - Individual Utility (IRIU) Investigating multi-objective games, Avigad et al. [8] propose an evolutionary search algorithm for finding the Pareto set of strategies for a player, given each pos- sible strategy of an opponent. Also looking at competitive multi-objective games, Eisenstadt et al. [26] proposes a co-evolutionary algorithm for finding solutions simultaneously for both players, under worst-case assumptions, given that the oponent's preferences are unknown. Brys et al. [16] apply MOMARL to a traffic signal control problem, where each intersection in a 2 × 2 grid is controlled by an individual agent. Their work demonstrats that rewarding agents with a linear scalarised combination of delay and throughput can improve delay times when compared to agents rewarded using delay alone. However, their approach uses local rewards (i.e. each agent is rewarded based on conditions at its assigned intersection only, and hence this work is classi- fied as IRIU), and does not make any attempt to explicitly encourage coordination between the agents. Dusparic and Cahill [25] propose the Distributed W-Learning algorithm, an RL-based approach for multi-policy optimisation in collaborative multi-agent sys- tems, such as urban traffic. Each agent represents a traffic light at an intersection and implements a set of local and remote policies (i.e., involving its neighbours). 38 Roxana Radulescu et al. Even though the agents here have possibly conflicting goals and receive an indi- vidual reward, at every time-step they exchange information with their neighbours regarding their current states and rewards, allowing them to develop, if necessary, a cooperative behaviour. Van Moffaert et al. [119] apply MORL to a multi-objective multi-agent smart camera problem. They develop an adaptive weight algorithm (AWA) which is used to choose the weighting between the two system objectives when linear scalarisa- tion is applied to individual reward vectors for each camera agent in the system. The AWA algorithm is found to improve learning speed, obtaining solutions with a better spread in the objective space, when compared with other weight selection methods that were tested. 6.2 Equilibria and Stability methods 6.2.1 Team Reward - Individual Utility (TRIU) Lee [54] takes a game theory perspective on the reservoir watershed management domain and develops for this purpose a multi-objective game theory model. The goal in this case is to balance between maximising economic gain and minimising environmental impact. Each player represents a different objective and multiple bargaining rounds are used in order to arrive at a Nash equilibrium. 6.2.2 Individual Reward - Individual Utility (IRIU) Qu et al. [79] examine multi-objective non-zero sum game, where each agent has a set of weights for each objectives (i.e., linear utility function where the exact weights are not known). They propose an approach for finding a robust weighted Nash Equilibrium. Taylor et al. [115] propose Parallel Transfer Learning (PTL) as a mechanism to accelerate learning, by sharing experience among agents. PTL is tested on a multi-objective multi-agent smart grid problem, and is found to improve learning speed and final performance when compared to agents learning without PTL. Madani and Lund [57] consider a Monte Carlo Game Theory approach for stochastic multi-criteria decision making settings, such as water resource manage- ment. They propose the use of Monte Carlo simulations to map the problem to deterministic settings transformed then into strategic games, solved using non- cooperative stability methods. 6.3 Social Welfare and Mechanism Design On the line of reward engineering in multi-objective congestion problems, Ramos et al. [80] consider the route choice traffic problem and develop a reward signal based on the marginal cost tolling mechanism. This allows one to reach a system optimum performance, even when agents have heterogeneous preferences with re- spect to travelling time and monetary costs. To model the agent preferences they Multi-Objective Multi-Agent Decision Making 39 use a linear utility function that attributes different weights to the considered ob- jectives. They show that if all the agents use their proposed learning approach, i.e., Generalised Toll-based Q-learning, they will converge to a system optimum. Mouaddib et al. [71] develop a multi-objective multi-agent planning framework in the form of a regret-based algorithm to improve the resulting social behaviour for the considered vector-valued Dec-MDP. This framework assumes a true ob- jective -- the regular Dec-MDP objective -- and then adds two extra objectives, i.e., the positive and negative impact an agent has on the team. This leads to so- cial behaviour, and therefore better functioning teams. Such addition of artificial objectives to attain better policies can be seen as a form of multi-objectivisation [15]. Multi-objectivisation to improve team behaviour through social welfare is an interesting research direction that we believe needs to be further explored. Grandoni et al. [35] assume individual utility functions with a primary objec- tive that should be maximised, and other objectives that need to achieve at least a threshold value. This type of utility function is similar to constrained MDPs [5] in the single-agent literature, and can be seen as a special case of multi-objective MDPs, i.e., one where the utility has this shape. They show that truthful mecha- nisms exist that allow finding solutions in reasonable time. Pla et al. [78] study auctions in which agents make bids that lead to a value- vector in different objectives. For such auctions the social welfare must be opti- mised via a mechanism that determines the payments w.r.t. the bid. These pay- ments, together with the costs of the bids, constitute the individual utilities. They show that the social welfare function must obey three properties for a mechanism to be possible: it must be real-valued and monotonically increasing in all objectives -- as to utility functions for any multi-objective decision problem -- and it must be bijective, i.e., given the bid attributes values and the result of an evaluation function, the cost corresponding to a bid can take only one possible value. This is necessary to be able to calculate the payments in a mechanism. Mechanisms for multi-objective decision problems are used in a variety of appli- cations, for example: Buettner and Landes [17] apply mechanism design in order to match employers looking for temporary workers, to workers looking for contracts. Because these contracts have several aspects that may lead to utility, as hourly salary, benefits, sick pay or overtime premiums, this is a multi-objective setting; Fard et al. [27] use mechanism design in for cloud work-flow management, where the agents have costs and completion time as objectives when trying to schedule tasks; and Kruse et al. [53] study multi-objective airline service procurement using mechanisms. 7 Conclusions and New Horizons In this paper, we analysed multi-objective multi-agent decision problems from a utility-based perspective. Starting from the execution phase and working back- wards, we derived when different solution concepts apply. We surveyed the liter- ature on the applicable solution concepts, methods that compute such solutions, and practical applications. The taxonomy of problem settings and solution meth- ods we propose structures this relatively new line of research from the perspective of user utility, and it is therefore our hope that this survey helps to place existing research papers in the larger multi-objective multi-agent decision problem context, 40 Roxana Radulescu et al. and informs and helps to inspire further research. To this end, in this last section we discuss what we consider to be the key new horizons and open problems in the field of multi-objective multi-agent decision making. 7.1 Optimisation Criteria and Solution Concepts In future work, it would be worthwhile to further explore the link between multi- objective optimisation criteria (ESR vs. SER) and solution concepts for MOMAS with non-linear utility functions. The body of literature on theory and experi- mental results is limited up until this point with respect to this topic, apart from an initial analysis conducted by Radulescu et al. [98] for multi-objective normal- form games under SER which proves by example that Nash equilibria need not exist, and that correlated equilibria can exist under certain conditions. This line of research should be extended to sequential settings (e.g., MOSGs), as well as to consider the other solution concepts discussed in Section 5. It is also possible that not all agents in a MOMAS would choose the same optimisation criterion; it is currently not known how mixing optimisation criteria would affect the collec- tive behaviour of MOMAS in practice. Developing stronger theoretical guarantees, as well as a better understanding of these issues using comprehensive empirical studies represents an important research direction one can pursue. 7.2 ESR Planning and Reinforcement Learning and SER Game Theory For multi-objective multi-agent decision problems, there is a large discrepancy between the game theory literature and the planning and reinforcement learning literature. The former focuses mostly on ESR settings, while the latter focuses almost exclusively on SER settings. Perhaps this is an artefact of the single-shot nature of most game-theoretic models and the sequential nature of planning and reinforcement learning models. However, as we recently noted [98], both optimality criteria are well-motivated, as they apply to different real-world decision problem settings, and lead to vastly different theoretical results as well as practical solutions in single-shot settings with non-linear utility functions. The same argument can be made for sequential decision making settings. Therefore, we believe that analysing sequential decision problems under ESR, and game-theoretic (single-shot) models under SER, is both highly important and almost entirely unstudied. 7.3 Opponent Modelling and Modelling Opponent Utility In single-objective reinforcement learning, an agent often aims to learn a model of the other agents' behaviours and uses this model when selecting or learning best responses. In multi-objective multi-agent settings, a good and possibly even sufficient predictor for this behaviour would be the utility function of the other agents. Therefore, explicitly estimating the utility functions of the other agents in a MOMAS is likely to be important in future research. In team-utility settings, i.e., when there is only one true utility function, Zintgraf et al. [133] show that this utility function can be estimated effectively by posing preference queries, and Multi-Objective Multi-Agent Decision Making 41 using monotonicity information about the utility function. However, this assumes that there is a single user to pose such queries to, who "owns" the utility function. In multi-agent settings, there may be multiple utility functions, and users, that have conflicting interests. Furthermore, if they can benefit from not revealing their true preferences, they might lie. This motivates two important open questions for future research: can we design mechanisms that force agents/users to be truthful about revealing their preferences over value/return vectors? And if not, can we es- timate their utility functions solely from the agent's behaviour in a multi-objective decision problem? Albrecht and Stone recently published a comprehensive survey on opponent modelling for single-objective MAS [3]; many of the methods they surveyed could plausibly be adapted or extended to model other agents' intentions and utilities in MOMAS. 7.4 Closing the Loop From our analysis of prior works on MOMAS in Section 6, it is apparent that the field to date has been quite fractured; some settings from our taxonomy (e.g. TRTU) have received much more attention than others, and a limited number of authors are currently active in the field. Consequently, there is not yet a standard- ised approach to identifying and completing all the steps which are necessary for a successful application of multi-objective multi-agent decision making. We propose the following sequence of steps: selecting an appropriate decision making model from among those listed in Section 3, identifying which setting from our taxonomy the decision making problem fits into (Section 4), defining the environment including the state and action spaces and reward and utility schemes, selecting an appropriate solution concept (Section 5), completing the planning/learning and/or negotiation phase, executing the policies found and fi- nally evaluating the outcome by measuring the utilities achieved. We hope that by following these main steps, it will become easier for other researchers to apply multi-objective multi-agent decision making theory in their work. 7.5 Interactive approaches In most of the survey we have assumed that there is a separate learning or plan- ning phase first, then a policy selection and/or negotiation phase, and finally an execution phase. However, it is also possible to elicit preferences from users while planning or learning, leading to a combined planning/learning and preference- elicitation/negotiation phase. In single-agent multi-objective systems this has been studied in [92, 91], and in cooperative game theory by [43]. Furthermore, the in- corporation of preference information during planning in [125] can also be seen in this line. This previous research however focuses either on eliciting preferences with respect to a team utility function [92, 91, 125] or individual utilities in the context of checking whether deviations from current coalitions are desired [43]. Parallel negotiation and learning or planning is, to our knowledge, still unexplored territory. 42 Roxana Radulescu et al. 7.6 Deep Multi-Objective Multi-Agent Decision Making Most of research discussed in this survey so far considers domains with discrete states and actions. For challenging real-world applications of MOMAS, it will be necessary to develop methods which consider continuous or high-dimensional state and action spaces. Considerable progress has been made on developing single- objective Deep RL methods for single-agent decision making. In the last couple of years, interest in Deep MORL has intensified, although primarily in single-agent settings (see e.g. [1, 33, 47, 70, 74, 82, 106, 111, 112]). Very recently, single-objective multi-agent RL has received considerable attention as well [30, 32, 39, 55, 97, 81, 109, 130]. An important next step is therefore to extend existing Deep RL methods for multi-objective multi-agent decision making settings. 7.7 Applications and Broader Applicability Now that we have identified the different settings and solution concepts which are relevant to MOMAS, significant opportunities exist to revisit problems that were initially modelled as single-objective multi-agent decision problems using a multi- objective perspective. This could provide a richer set of potential solutions for cooperative MAS using the concept of coverage sets (Section 6.1), or potentially improve performance by considering additional synthetic objectives of sub-tasks explicitly (a process known as multi-objectivisation [15, 14]. The possibility also exists to use MORL techniques to develop agents which may be tuned to adopt a range of different behaviours during deployment in MAS (e.g. cooperative vs. competitive), as recently demonstrated by Kallstrom and Heintz [47]. References 1. Abels, A., Roijers, D.M., Lenaerts, T., Now´e, A., Steckelmacher, D.: Dynamic weights in multi-objective deep reinforcement learning. In: ICML 2019: Proceedings of the 36th International Conference on Machine Learning (2019). To appear. 2. Ahmad, I., Ranka, S., Khan, S.U.: Using game theory for scheduling tasks on multi- core processors for simultaneous optimization of performance and energy. In: 2008 IEEE International Symposium on Parallel and Distributed Processing, pp. 1 -- 6. IEEE (2008) 3. Albrecht, S.V., Stone, P.: Autonomous agents modelling other agents: A comprehensive survey and open problems. Artificial Intelligence 258, 66 -- 95 (2018) 4. Alonso, E., D'inverno, M., Kudenko, D., Luck, M., Noble, J.: Learning in multi-agent systems. The Knowledge Engineering Review 16(3), 277 -- 284 (2001) 5. Altman, E.: Constrained Markov decision processes, vol. 7. CRC Press (1999) 6. Aoki, K., Kimura, H., Kobayashi, S.: Distributed reinforcement learning using bi- In: directional decision making for multi-criteria control of multi-stage flow systems. The 8th Conference on Intelligent Autonomous Systems, pp. 281 -- 290 (2004) 7. Aumann, R.J.: Subjectivity and correlation in randomized strategies. Journal of mathe- matical Economics 1(1), 67 -- 96 (1974) 8. Avigad, G., Eisenstadt, E., Cohen, M.W.: Optimal strategies for multi objective games and their search by evolutionary multi objective optimization. In: 2011 IEEE Conference on Computational Intelligence and Games (CIG'11), pp. 166 -- 173. IEEE (2011) 9. Baarslag, T., Kaisers, M.: The value of information in automated negotiation: A de- In: Proceedings of the 16th Conference on cision model for eliciting user preferences. Autonomous Agents and MultiAgent Systems, pp. 391 -- 400. International Foundation for Autonomous Agents and Multiagent Systems (2017) Multi-Objective Multi-Agent Decision Making 43 10. Babbar-Sebens, M., Mukhopadhyay, S.: Reinforcement learning for human-machine col- laborative optimization: Application in ground water monitoring. In: 2009 IEEE Inter- national Conference on Systems, Man and Cybernetics, pp. 3563 -- 3568. IEEE (2009) 11. Bellman, R.: Dynamic Programming. Princeton University Press, Princeton, NJ, USA (1957) 12. Bone, C., Dragi´cevi´c, S.: Gis and intelligent agents for multiobjective natural resource allocation: A reinforcement learning approach. Transactions in GIS 13(3), 253 -- 272 (2009) 13. Borm, P., Tijs, S., Van Den Aarssen, J.: Pareto equilibria in multiobjective games. Meth- ods of Operations Research 60, 302 -- 312 (1988) 14. Brys, T., Harutyunyan, A., Vrancx, P., Now, A., Taylor, M.E.: Multi-objectivization and ensembles of shapings in reinforcement learning. Neurocomputing 263, 48 -- 59 (2017). Multiobjective Reinforcement Learning: Theory and Applications 15. Brys, T., Harutyunyan, A., Vrancx, P., Taylor, M.E., Kudenko, D., Now´e, A.: Multi- objectivization of reinforcement learning problems by reward shaping. In: 2014 interna- tional joint conference on neural networks (IJCNN), pp. 2315 -- 2322. IEEE (2014) 16. Brys, T., Pham, T.T., Taylor, M.E.: Distributed learning and multi-objectivity in traffic light control. Connection Science 26(1), 65 -- 83 (2014). DOI 10.1080/09540091.2014. 885282 17. Buettner, R., Landes, J.: Web service-based applications for electronic labor markets: a multi-dimensional price vcg auction with individual utilities. In: ICIW 2012 Proceedings, pp. 168 -- 177 (2012) 18. Calvaresi, D., Marinoni, M., Sturm, A., Schumacher, M., Buttazzo, G.: The Challenge of Real-time Multi-agent Systems for Enabling IoT and CPS. In: Proceedings of the International Conference on Web Intelligence, WI '17, pp. 356 -- 364. ACM, New York, NY, USA (2017) 19. Chalkiadakis, G., Elkind, E., Wooldridge, M.: Computational aspects of cooperative game theory. Synthesis Lectures on Artificial Intelligence and Machine Learning 5(6), 1 -- 168 (2011) 20. Chung, J.J., Rebhuhn, C., Yates, C., Hollinger, G.A., Tumer, K.: A multiagent framework for learning dynamic traffic management strategies. Autonomous Robots pp. 1 -- 17 (2018) 21. Deb, K.: Multi-objective optimization. In: Search methodologies, pp. 403 -- 449. Springer (2014) 22. Delle Fave, F., Stranders, R., Rogers, A., Jennings, N.: Bounded decentralised coordina- tion over multiple objectives. In: Proceedings of the Tenth International Joint Conference on Autonomous Agents and Multiagent Systems, pp. 371 -- 378 (2011) 23. Dubus, J., Gonzales, C., Perny, P.: Choquet optimization using GAI networks for multia- gent/multicriteria decision-making. In: ADT 2009: Proceedings of the First International Conference on Algorithmic Decision Theory, pp. 377 -- 389 (2009) 24. Dubus, J., Gonzales, C., Perny, P.: Multiobjective optimization using GAI models. In: IJCAI 2009: Proceedings of the Twenty-third International Joint Conference on Artificial Intelligence, pp. 1902 -- 1907 (2009) 25. Dusparic, I., Cahill, V.: Distributed w-learning: Multi-policy optimization in self- In: 2009 Third IEEE International Conference on Self-Adaptive organizing systems. and Self-Organizing Systems, pp. 20 -- 29. IEEE (2009) 26. Eisenstadt, E., Moshaiov, A., Avigad, G.: Co-evolution of strategies for multi-objective games under postponed objective preferences. In: 2015 IEEE Conference on Computa- tional Intelligence and Games (CIG), pp. 461 -- 468. IEEE (2015) 27. Fard, H.M., Prodan, R., Moser, G., Fahringer, T.: A bi-criteria truthful mechanism for In: 2011 IEEE Third International Conference on scheduling of workflows in clouds. Cloud Computing Technology and Science, pp. 599 -- 605. IEEE (2011) 28. Fernandez, F.R., Hinojosa, M.A., Puerto, J.: Core solutions in vector-valued games. Jour- nal of Optimization Theory and Applications 112(2), 331 -- 360 (2002) 29. Flesch, J., Thuijsman, F., Vrieze, K.: Cyclic markov equilibria in stochastic games. In- ternational Journal of Game Theory 26(3), 303 -- 314 (1997) 30. Foerster, J., Assael, I.A., de Freitas, N., Whiteson, S.: Learning to communicate with deep multi-agent reinforcement learning. In: Advances in Neural Information Processing Systems, pp. 2137 -- 2145 (2016) 31. Foerster, J., Nardelli, N., Farquhar, G., Afouras, T., Torr, P.H., Kohli, P., Whiteson, S.: Stabilising experience replay for deep multi-agent reinforcement learning. In: Proceedings of the 34th International Conference on Machine Learning-Volume 70, pp. 1146 -- 1155. JMLR. org (2017) 44 Roxana Radulescu et al. 32. Foerster, J.N., Farquhar, G., Afouras, T., Nardelli, N., Whiteson, S.: Counterfactual multi-agent policy gradients. In: Thirty-Second AAAI Conference on Artificial Intelli- gence (2018) 33. Friedman, E., Fontaine, F.: Generalizing across multi-objective reward functions in deep reinforcement learning. arXiv preprint arXiv:1809.06364 (2018) 34. Golden, B., Perny, P.: Infinite order lorenz dominance for fair multiagent optimization. In: Proceedings of the 9th International Conference on Autonomous Agents and Multiagent Systems: volume 1-Volume 1, pp. 383 -- 390. International Foundation for Autonomous Agents and Multiagent Systems (2010) 35. Grandoni, F., Krysta, P., Leonardi, S., Ventre, C.: Utilitarian mechanism design for multi- objective optimization. In: Proceedings of the twenty-first annual ACM-SIAM symposium on Discrete Algorithms, pp. 573 -- 584. Society for Industrial and Applied Mathematics (2010) 36. Guestrin, C., Koller, D., Parr, R.: Multiagent planning with factored MDPs. In: Advances in Neural Information Processing Systems 15 (NIPS'02) (2002) 37. Hamidi, H., Kamankesh, A.: An approach to intelligent traffic management system using a multi-agent system. International Journal of Intelligent Transportation Systems Research 16(2), 112 -- 124 (2018) 38. Hansen, E.A., Bernstein, D.S., Zilberstein, S.: Dynamic programming for partially ob- servable stochastic games. In: Proceedings of the 19th National Conference on Artifical Intelligence, AAAI'04, pp. 709 -- 715. AAAI Press (2004) 39. He, H., Boyd-Graber, J., Kwok, K., Daum´e III, H.: Opponent modeling in deep rein- forcement learning. In: International Conference on Machine Learning, pp. 1804 -- 1813 (2016) 40. Hernandez-Leal, P., Kartal, B., Taylor, M.E.: Is multiagent deep reinforcement learning the answer or the question? a brief survey. arXiv preprint arXiv:1810.05587 (2018) 41. Houli, D., Zhiheng, L., Yi, Z.: Multiobjective reinforcement learning for traffic signal con- trol using vehicular ad hoc network. EURASIP journal on advances in signal processing 2010(1), 724035 (2010) 42. Hurtado, C., Ramirez, M.R., Alanis, A., Vazquez, S.O., Ramirez, B., Manrique, E.: To- wards a multi-agent system for an informative healthcare mobile application. In: KES International Symposium on Agent and Multi-Agent Systems: Technologies and Appli- cations, pp. 215 -- 219. Springer (2018) 43. Igarashi, A., Roijers, D.M.: Multi-criteria coalition formation games. In: International Conference on Algorithmic DecisionTheory, pp. 197 -- 213. Springer (2017) 44. Inja, M., Kooijman, C., de Waard, M., Roijers, D.M., Whiteson, S.: Queued pareto local search for multi-objective optimization. In: International Conference on Parallel Problem Solving from Nature, pp. 589 -- 599. Springer (2014) 45. Jennings, N.R., Faratin, P., Lomuscio, A.R., Parsons, S., Wooldridge, M.J., Sierra, C.: Automated negotiation: prospects, methods and challenges. Group Decision and Nego- tiation 10(2), 199 -- 215 (2001) 46. Jonker, C.M., Aydogan, R., Baarslag, T., Fujita, K., Ito, T.: Automated negotiating agents competition (anac). (2017) 47. Kallstrom, J., Heintz, F.: Tunable dynamics in agent-based simulation using multi- objective reinforcement learning. In: Proceedings of the Adaptive and Learning Agents Workshop (ALA-19) at AAMAS (2019) 48. Kawamura, T., Kanazawa, T., Ushio, T.: Evolutionarily and neutrally stable strategies in multicriteria games. IEICE Transactions on Fundamentals of Electronics, Communi- cations and Computer Sciences 96(4), 814 -- 820 (2013) 49. Khan, M.W., Wang, J.: The research on multi-agent system for microgrid control and optimization. Renewable and Sustainable Energy Reviews 80, 1399 -- 1411 (2017) 50. Kitano, H., Asada, M., Kuniyoshi, Y., Noda, I., Osawa, E.: Robocup: The robot world In: Proceedings of the First International Conference on Autonomous cup initiative. Agents, AGENTS '97, pp. 340 -- 347. ACM, New York, NY, USA (1997) 51. Kok, J.R., Vlassis, N.: Sparse cooperative Q-learning. In: Proceedings of the twenty-first international conference on Machine learning, ICML '04. ACM, New York, NY, USA (2004) 52. Kok, J.R., Vlassis, N.: Collaborative multiagent reinforcement learning by payoff propa- gation. Journal of Machine Learning Research 7(Sep), 1789 -- 1828 (2006) 53. Kruse, S., Brintrup, A., McFarlane, D., Lopez, T.S., Owens, K., Krechel, W.E.: Designing automated allocation mechanisms for service procurement of imperfectly substitutable Multi-Objective Multi-Agent Decision Making 45 services. IEEE Transactions on Computational Intelligence and AI in Games 5(1), 15 -- 32 (2013) 54. Lee, C.S.: Multi-objective game-theory models for conflict analysis in reservoir watershed management. Chemosphere 87(6), 608 -- 613 (2012) 55. Lowe, R., Wu, Y., Tamar, A., Harb, J., Abbeel, O.P., Mordatch, I.: Multi-agent actor- critic for mixed cooperative-competitive environments. In: Advances in Neural Informa- tion Processing Systems, pp. 6379 -- 6390 (2017) 56. Lozovanu, D., Solomon, D., Zelikovsky, A.: Multiobjective games and determining pareto- nash equilibria. Buletinul Academiei de S¸tiint¸e a Republicii Moldova. Matematica (3), 115 -- 122 (2005) 57. Madani, K., Lund, J.R.: A monte-carlo game theoretic approach for multi-criteria decision making under uncertainty. Advances in water resources 34(5), 607 -- 616 (2011) 58. Mannion, P.: Knowledge-based multi-objective multi-agent reinforcement learning. Ph.D. thesis, National University of Ireland Galway (2017) 59. Mannion, P., Devlin, S., Duggan, J., Howley, E.: Reward shaping for knowledge-based multi-objective multi-agent reinforcement learning. The Knowledge Engineering Review 33 (2018) 60. Mannion, P., Devlin, S., Mason, K., Duggan, J., Howley, E.: Policy invariance under reward transformations for multi-objective reinforcement learning. Neurocomputing 263 (2017) 61. Mannion, P., Duggan, J., Howley, E.: An experimental review of reinforcement learning algorithms for adaptive traffic signal control. In: L.T. McCluskey, A. Kotsialos, P.J. Muller, F. Klugl, O. Rana, R. Schumann (eds.) Autonomic Road Transport Support Systems, pp. 47 -- 66. Springer International Publishing (2016) 62. Mannion, P., Duggan, J., Howley, E.: A theoretical and empirical analysis of reward trans- formations in multi-objective stochastic games. In: Proceedings of the 16th International Conference on Autonomous Agents and Multiagent Systems (AAMAS) (2017) 63. Mannion, P., Mason, K., Devlin, S., Duggan, J., Howley, E.: Dynamic economic emissions dispatch optimisation using multi-agent reinforcement learning. In: Proceedings of the Adaptive and Learning Agents workshop (at AAMAS 2016) (2016) 64. Mannion, P., Mason, K., Devlin, S., Duggan, J., Howley, E.: Multi-objective dynamic dispatch optimisation using multi-agent reinforcement learning. In: Proceedings of the 15th International Conference on Autonomous Agents and Multiagent Systems (AAMAS) (2016) 65. Marinescu, R.: Exploiting problem decomposition in multi-objective constraint optimiza- In: CP 2009: Principles and Practice of Constraint Programming, pp. 592 -- 607 tion. (2009) 66. Marinescu, R.: Efficient approximation algorithms for multi-objective constraint opti- mization. In: ADT 2011: Proceedings of the Second International Conference on Algo- rithmic Decision Theory, pp. 150 -- 164 (2011) 67. Marinescu, R., Razak, A., Wilson, N.: Multi-objective constraint optimization with trade- offs. In: International Conference on Principles and Practice of Constraint Programming, pp. 497 -- 512. Springer (2013) 68. Mirrokni, V.S., Vetta, A.: Convergence issues in competitive games. In: Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques, pp. 183 -- 194. Springer (2004) 69. Moradi, M.H., Razini, S., Hosseinian, S.M.: State of art of multiagent systems in power engineering: A review. Renewable and Sustainable Energy Reviews 58, 814 -- 824 (2016) 70. Mossalam, H., Assael, Y.M., Roijers, D.M., Whiteson, S.: Multi-objective deep reinforce- ment learning. arXiv preprint arXiv:1610.02707 (2016) 71. Mouaddib, A.I., Boussard, M., Bouzid, M.: Towards a formal framework for multi- objective multiagent planning. In: Proceedings of the 6th international joint conference on Autonomous agents and multiagent systems, p. 123. ACM (2007) 72. Nash, J.: Non-cooperative games. Annals of Mathematics 54(2), 286 -- 295 (1951) 73. Ng, A.Y., Harada, D., Russell, S.J.: Policy invariance under reward transformations: Theory and application to reward shaping. In: Proceedings of the Sixteenth International Conference on Machine Learning, ICML '99, pp. 278 -- 287 (1999) 74. Nguyen, T.T.: A multi-objective deep reinforcement learning framework. arXiv preprint arXiv:1803.02965 (2018) 75. Nguyen, T.T., Nguyen, N.D., Nahavandi, S.: Deep reinforcement learning for multi- arXiv preprint agent systems: A review of challenges, solutions and applications. arXiv:1812.11794 (2018) 46 Roxana Radulescu et al. 76. Nisan, N., Roughgarden, T., Tardos, E., Vazirani, V.V.: Algorithmic game theory. Cam- bridge University Press (2007) 77. Perny, P., Weng, P., Goldsmith, J., Hanna, J.: Approximation of lorenz-optimal solutions in multiobjective markov decision processes. In: Proceedings of the 27th AAAI Conference on Artificial Intelligence, pp. 92 -- 94 (2013) 78. Pla, A., Lopez, B., Murillo, J.: Multi criteria operators for multi-attribute auctions. In: International Conference on Modeling Decisions for Artificial Intelligence, pp. 318 -- 328. Springer (2012) 79. Qu, S., Ji, Y., Goh, M.: The robust weighted multi-objective game. PloS one 10(9), e0138970 (2015) 80. Ramos, G.de.O., Radulescu, R., Now´e, A.: A budged-balanced tolling scheme for efficient equilibria under heterogeneous preferences. In: Proceedings of the Adaptive and Learning Agents Workshop (ALA-19) at AAMAS (2019) 81. Rashid, T., Samvelyan, M., de Witt, C.S., Farquhar, G., Foerster, J., Whiteson, S.: Qmix: Monotonic value function factorisation for deep multi-agent reinforcement learning. In: ICML 2018: Proceedings of the Thirty-Fifth International Conference on Machine Learn- ing (2018) 82. Reymond, M., Now´e, A.: Pareto-DQN: Approximating the Pareto front in complex multi- objective decision problems. In: Proceedings of the Adaptive and Learning Agents Work- shop (ALA-19) at AAMAS (2019) 83. Roijers, D.M.: Multi-objective decision-theoretic planning. Ph.D. thesis, University of Amsterdam (2016) 84. Roijers, D.M., Steckelmacher, D., Now´e, A.: Multi-objective reinforcement learning for In: Adaptive and Learning Agents Workshop (at the expected utility of the return. AAMAS/IJCAI/ICML 2018) (2018) 85. Roijers, D.M., Vamplew, P., Whiteson, S., Dazeley, R.: A survey of multi-objective se- quential decision-making. Journal of Artificial Intelligence Research 48, 67 -- 113 (2013) 86. Roijers, D.M., Whiteson, S.: Multi-objective decision making. Synthesis Lectures on Artificial Intelligence and Machine Learning 11(1), 1 -- 129 (2017) 87. Roijers, D.M., Whiteson, S., Ihler, A.T., Oliehoek, F.A.: Variational multi-objective co- In: MALIC 2015: NIPS Workshop on Learning, Inference and Control of ordination. Multi-Agent Systems (2015) 88. Roijers, D.M., Whiteson, S., Oliehoek, F.A.: Computing convex coverage sets for multi- In: International Conference on Algorithmic Decision objective coordination graphs. Theory, pp. 309 -- 323 (2013) 89. Roijers, D.M., Whiteson, S., Oliehoek, F.A.: Linear support for multi-objective coordina- tion graphs. In: Proceedings of the 2014 international conference on Autonomous agents and multi-agent systems, pp. 1297 -- 1304. International Foundation for Autonomous Agents and Multiagent Systems (2014) 90. Roijers, D.M., Whiteson, S., Oliehoek, F.A.: Computing convex coverage sets for faster multi-objective coordination. Journal of Artificial Intelligence Research 52, 399 -- 443 (2015) 91. Roijers, D.M., Zintgraf, L.M., Libin, P., Now´e, A.: Interactive multi-objective reinforce- In: ALA workshop at ment learning in multi-armed bandits for any utility function. FAIM, vol. 8 (2018) 92. Roijers, D.M., Zintgraf, L.M., Now´e, A.: Interactive Thompson sampling for multi- In: International Conference on Algorithmic Decision objective multi-armed bandits. Theory, pp. 18 -- 34. Springer (2017) 93. Roll´on, E.: Multi-objective optimization for graphical models. Ph.D. thesis, Universitat Polit`ecnica de Catalunya, Barcelona (2008) 94. Roll´on, E., Larrosa, J.: Bucket elimination for multiobjective optimization problems. Journal of Heuristics 12, 307 -- 328 (2006) 95. Rollon, E., Larrosa, J.: Multi-objective russian doll search. In: AAAI, pp. 249 -- 254 (2007) 96. Rollon, E., Larrosa, J.: Constraint optimization techniques for multiobjective branch and bound search. In: International Conference on Logic Programming, ICLP (2008) 97. Radulescu, R., Legrand, M., Efthymiadis, K., Roijers, D.M., Now´e, A.: Deep multi-agent reinforcement learning in a homogeneous open population. In: Proceedings of the 30th Benelux Conference on Artificial Intelligence (BNAIC 2018), pp. 177 -- 191 (2018) 98. Radulescu, R., Mannion, P., Roijers, D.M., Now´e, A.: Equilibria in multi-objective games: a utility-based perspective. In: Proceedings of the Adaptive and Learning Agents Work- shop (ALA-19) at AAMAS (2019) Multi-Objective Multi-Agent Decision Making 47 99. Scharpff, J., Roijers, D.M., Oliehoek, F.A., Spaan, M.T., de Weerdt, M.M.: Solving transition-independent multi-agent MDPs with sparse interactions. In: AAAI 2016: Pro- ceedings of the Thirtieth AAAI Conference on Artificial Intelligence (2016). To Appear. 100. Scharpff, J., Spaan, M.T., Volker, L., de Weerdt, M.M.: Coordinating stochastic multi- agent planning in a private values setting. Distributed and Multi-Agent Planning p. 17 (2013) 101. Sen, S., Weiss, G.: Multiagent systems. chap. Learning in Multiagent Systems, pp. 259 -- 298. MIT Press, Cambridge, MA, USA (1999). URL http://dl.acm.org/citation.cfm? id=305606.305612 102. Shapley, L.S.: Stochastic games. Proceedings of the national academy of sciences 39(10), 1095 -- 1100 (1953) 103. Shapley, L.S., Rigby, F.D.: Equilibrium points in games with vector payoffs. Naval Re- search Logistics Quarterly 6(1), 57 -- 61 (1959) 104. Shelton, C.R.: Importance sampling for reinforcement learning with multiple objectives. Ph.D. thesis, Massachusetts Institute of Technology, Cambridge, MA, USA (2001) 105. Shoham, Y., Powers, R., Grenager, T.: If multi-agent learning is the answer, what is the question? Artificial Intelligence 171(7), 365 -- 377 (2007) 106. Si, W., Li, J., Ding, P., Rao, R.: A multi-objective deep reinforcement learning approach In: 2017 10th International Symposium on for stock index futures intraday trading. Computational Intelligence and Design (ISCID), vol. 2, pp. 431 -- 436 (2017) 107. Silver, D., Huang, A., Maddison, C.J., Guez, A., Sifre, L., van den Driessche, G., Schrit- twieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., Dieleman, S., Grewe, D., Nham, J., Kalchbrenner, N., Sutskever, I., Lillicrap, T., Leach, M., Kavukcuoglu, K., Graepel, T., Hassabis, D.: Mastering the game of go with deep neural networks and tree search. Nature 529, 484 -- 503 (2016) 108. Song, J., Ren, H., Sadigh, D., Ermon, S.: Multi-agent generative adversarial imitation learning. In: Advances in Neural Information Processing Systems, pp. 7461 -- 7472 (2018) 109. Sunehag, P., Lever, G., Gruslys, A., Czarnecki, W.M., Zambaldi, V., Jaderberg, M., Lanctot, M., Sonnerat, N., Leibo, J.Z., Tuyls, K., et al.: Value-decomposition networks for cooperative multi-agent learning based on team reward. In: Proceedings of the 17th International Conference on Autonomous Agents and MultiAgent Systems, pp. 2085 -- 2087. International Foundation for Autonomous Agents and Multiagent Systems (2018) 110. Sutton, R., Barto, A.: Reinforcement Learning: An Introduction. MIT Press, Cambridge, MA (1998) 111. Tajmajer, T.: Multi-objective deep q-learning with subsumption architecture. arXiv preprint arXiv:1704.06676 (2017) 112. Tajmajer, T.: Modular multi-objective deep reinforcement learning with decision values. In: 2018 Federated Conference on Computer Science and Information Systems (FedCSIS), pp. 85 -- 93 (2018) 113. Tanino, T.: Multiobjective cooperative games with restrictions on coalitions. In: Multi- objective Programming and Goal Programming, pp. 167 -- 174. Springer (2009) 114. Tanino, T.: Vector optimization and cooperative games. In: Recent Developments in Vector Optimization, pp. 517 -- 545. Springer (2012) 115. Taylor, A., Dusparic, I., Galv´an-L´opez, E., Clarke, S., Cahill, V.: Accelerating learning in multi-objective systems through transfer learning. In: Neural Networks (IJCNN), 2014 International Joint Conference on, pp. 2298 -- 2305. IEEE (2014) 116. Tesauro, G.: Td-gammon, a self-teaching backgammon program, achieves master-level play. Neural Computing 6(2), 215 -- 219 (1994) 117. Tsimpoukis, D., Baarslag, T., Kaisers, M., Paterakis, N.G.: Automated negotiations un- der user preference uncertainty: A linear programming approach (2018) 118. Vamplew, P., Dazeley, R., Barker, E., Kelarev, A.: Constructing stochastic mixture poli- cies for episodic multiobjective reinforcement learning tasks. In: Australasian Joint Con- ference on Artificial Intelligence, pp. 340 -- 349. Springer (2009) 119. Van Moffaert, K., Brys, T., Chandra, A., Esterle, L., Lewis, P.R., Now´e, A.: A novel adap- tive weight selection algorithm for multi-objective multi-agent reinforcement learning. In: Neural Networks (IJCNN), 2014 International Joint Conference on, pp. 2306 -- 2314 (2014) 120. Vinyals, O., Ewalds, T., Bartunov, S., Georgiev, P., Vezhnevets, A.S., Yeo, M., Makhzani, A., Kuttler, H., Agapiou, J., Schrittwieser, J., et al.: Starcraft ii: A new challenge for reinforcement learning. arXiv preprint arXiv:1708.04782 (2017) 121. Vlassis, N.: A concise introduction to multiagent systems and distributed artificial intel- ligence. Synthesis Lectures on Artificial Intelligence and Machine Learning 1(1), 1 -- 71 (2007) 48 Roxana Radulescu et al. 122. Voorneveld, M., Vermeulen, D., Borm, P.: Axiomatizations of pareto equilibria in multi- criteria games. Games and economic behavior 28(1), 146 -- 154 (1999) 123. White, D.: Multi-objective infinite-horizon discounted markov decision processes. Journal of mathematical analysis and applications 89(2), 639 -- 647 (1982) 124. Wiggers, A.J., Oliehoek, F.A., Roijers, D.M.: Structure in the value function of two- player zero-sum games of incomplete information. In: Proceedings of the Twenty-second European Conference on Artificial Intelligence, pp. 1628 -- 1629. IOS Press (2016) 125. Wilson, N., Razak, A., Marinescu, R.: Computing possibly optimal solutions for multi- objective constraint optimisation with tradeoffs. In: IJCAI 2015: Proceedings of the Twenty-Fourth International Joint Conference on Artificial Intelligence, pp. 815 -- 821 (2015) 126. Wolpert, D.H., Wheeler, K.R., Tumer, K.: Collective intelligence for control of distributed dynamical systems. EPL (Europhysics Letters) 49(6), 708 (2000) 127. Wooldridge, M.: Introduction to Multiagent Systems. John Wiley & Sons, Inc., New York, NY, USA (2001) 128. Yliniemi, L., Tumer, K.: Multi-objective multiagent credit assignment in reinforcement learning and nsga-ii. Soft Computing 20(10), 3869 -- 3887 (2016) 129. Yliniemi, L.M.: Multi-objective optimization in multiagent systems. Ph.D. thesis, Oregon State University, Corvallis, OR (2015) 130. Zheng, Y., Meng, Z., Hao, J., Zhang, Z.: Weighted double deep multiagent reinforcement learning in stochastic cooperative environments. In: Pacific Rim International Conference on Artificial Intelligence, pp. 421 -- 429. Springer (2018) 131. Zinkevich, M., Greenwald, A., Littman, M.L.: Cyclic equilibria in markov games. In: Advances in Neural Information Processing Systems, pp. 1641 -- 1648 (2006) 132. Zintgraf, L.M., Kanters, T.V., Roijers, D.M., Oliehoek, F.A., Beau, P.: Quality assessment In: Benelearn 2015: Proceedings of the of morl algorithms: A utility-based approach. Twenty-Fourth Belgian-Dutch Conference on Machine Learning. (2015) 133. Zintgraf, L.M., Roijers, D.M., Linders, S., Jonker, C.M., Now´e, A.: Ordered preference elicitation strategies for supporting multi-objective decision making. In: Proceedings of the 17th International Conference on Autonomous Agents and Multi-Agent Systems, pp. 1477 -- 1485. International Foundation for Autonomous Agents and Multiagent Systems (2018)
1702.08529	1	1702	2017-02-27T21:03:26	Multi-agent systems and decentralized artificial superintelligence	[ "cs.MA" ]	Multi-agents systems communication is a technology, which provides a way for multiple interacting intelligent agents to communicate with each other and with environment. Multiple-agent systems are used to solve problems that are difficult for solving by individual agent. Multiple-agent communication technologies can be used for management and organization of computing fog and act as a global, distributed operating system. In present publication we suggest technology, which combines decentralized P2P BOINC general-purpose computing tasks distribution, multiple-agents communication protocol and smart-contract based rewards, powered by Ethereum blockchain. Such system can be used as distributed P2P computing power market, protected from any central authority. Such decentralized market can further be updated to system, which learns the most efficient way for software-hardware combinations usage and optimization. Once system learns to optimize software-hardware efficiency it can be updated to general-purpose distributed intelligence, which acts as combination of single-purpose AI.	cs.MA	cs	Multi-agent systems and decentralized artificial superintelligence Ponomarev S. (1), Voronkov A. E. (2) (1) Moscow Power Engineering Institute (2) Moscow Institute of Physics and Technology Abstract Multi-agents systems communication is a technology, which provides a way for multiple interacting intelligent agents to communicate with each other and with environment. Multiple- agent systems are used to solve problems that are difficult for solving by individual agent. Multiple-agent communication technologies can be used for management and organization of computing fog and act as a global, distributed operating system. In present publication we suggest technology, which combines decentralized P2P BOINC general-purpose computing tasks distribution, multiple-agents communication protocol and smart-contract based rewards, powered by Ethereum blockchain. Such system can be used as distributed P2P computing power market, protected from any central authority. Such decentralized market can further be updated to system, which learns the most efficient way for software-hardware combinations usage and optimization. Once system learns to optimize software-hardware efficiency it can be updated to general-purpose distributed intelligence, which acts as combination of single-purpose AI. Introduction This article outlines the theoretical prerequisites of creating intelligent multi-agent system of Decentralized Artificial Intelligence «SONM», as well as discusses the practical realization of these ideas. If you are interested in the theoretical questions, then see chapter one. If you are only interested in the practical realization, skip to chapter two. 1 Chapter 1. The theoretical rationale for the existence of decentralized artificial intelligence Agent in Informatics and artificial intelligence How is interpreted the term "agent" in modern Informatics and AI? To date, several different interpretations of this term have been formed (Fig. 1.1). ARTIFICIAL AGENT Artificial organism Personal assistant Active object Virtual agent Fig. 1.1. Possible interpretations of the concept of "artificial agent". Firstly, J. Holland [1] was first who introduced an idea about agent as an artificial organism, developing in the population of its kind, tending to learn and adapt to the environment in order to survive in it (and to defeat the competitors). This interpretation of agent is based on the theoretical approaches and models of artificial evolution (mutations and modifications of agents, and their struggle for survival, the selection of the strongest or rejection of the weakest ) and the principles of artificial life (self-reproduction, self-preservation, self-determination, self-control of agents, and so on). It is closely related to robotics (the construction of integral robots and functioning of robot groups), the problems of "armor" and "weapon" relations in computer networks, information security and information attacks, computer virology and creating Liveware tools (evolving software, which is constructed considering principles and mechanisms of the behavior of living organisms). Secondly, a metaphor appeared which interpreted agent as a personal assistant of user or, later, as intelligent intermediary between the user and the environment in which he works [2]. In particular, the strategy of artificial agent development, outlined in the IBM White Paper, is based on this idea of a "personal assistant", where the agent is any software or hardware system that can operate in order to achieve the objectives set by the user. 2 The idea of personal assistants in work of users with computers in its simplest form was embodied in a number of popular software products in late 90s. For example, Microsoft has built Wizards and System Agent in Windows95 and paperclip-assistant Clippit has appeared in Microsoft Office. Mac S includes a learning agent Open Sesame!, Lotus Notes V4 also has built- in agents. Among the modern personal assistants a voice helper Siry, existing in Apple's products can be distinguished. Thus, agents are often understood as autonomous (or semi-autonomous) software modules that can collaborate with the user and adapt to him. "Semi-autonomous" in this context means a software agent dependence on user, particularly, the user's ability to change the level of autonomy of his agent. This ensures not only friendly, but also personified character of the user interface. The origins of this concept of agent relate to the theory of dialogue "human- computer" and means of intelligent interface development. Real boom in the field of software agents began with the development of the Internet and appearance of chat bots. Information agents, such as PointCast, deliver news to users and report changes in selected sites. Shopping agents, such Bargain Finder, working for users, compare prices in electronic shops. Robotic spiders (crawlers) roam in the links and index an information for the search engines. Robots - chat bots are able conclude a transaction with the users to exchange cryptocurrency, etc. The greatest prospects for further applications of personalized "agents – user assistants" are associated with targeted information search in Internet, given its semantic and pragmatic characteristics, as well as the support of multi-criterial hard-to-formalize decision. We should expect a recently appeared concept "agentware" that characterizes the new architectural principles of information processing based on agent, will become widespread. Thirdly, the agents may be considered as active objects or meta-objects endowed with a certain degree of subjectness, i.e. able to manipulate other objects (e.g., smart contracts), create and destroy them, and interact with the environment and other agents. In this context, they can be created on the basis of constraint programming using the Active Object technologies [3]. Thus, the software technology of agents and agent-oriented programming are understood in general as a natural development of the ideas of object-oriented programming (OOP). The agent is a "self-contained programmed process, which includes some state and has the opportunity to interact with other agents via exchanging messages" [4]. Accordingly, the agent-oriented programming (AOP) [5] is a new programming paradigm based on "social points of view" on the calculations. In total − personalized assistants like Siry; programs-"demons" in the systems of UNIX and Windows utilities; softbots - chat-bots in Telegram, crawlers in search engines, programs- 3 advisors in Forex and Metatrader, bot players in computer games, smart contracts in Ethereum, etc. may be provided as the examples of agents. Multi-agent systems Solving of the task by one agent on the basis of knowledge engineering is a point of view of classical AI. According to it, the agent (for example, the intelligent system), having a global vision of the problem, has all the necessary skills, knowledge and resources for solving the task. In contrast, when creating multi-agent systems (MAS) we assume that a single agent can have only a partial understanding of the task and can solve only some of its subtasks. Therefore, to solve any complex problem, as a rule, the interaction of agents is required, which is inseparable from the formation of MAS. The tasks in MAS are distributed between the agents, each of which is considered as a member of the group or organization. Distribution of tasks involves assigning roles to each of the agents, the definition of measure of its responsibility and requirements to its experience. Depending on whether the distribution comes from the set task of or the ability of the particular agent, can be distinguished the systems of a distributed solution of the tasks and systems of decentralized AI. In the first case, the process of decomposition of the original task and the reverse process of composition of the obtained solutions is centralized. MAS is rigidly projected downward on the basis of partitioning of the general task into separate, relatively independent subtasks and preliminary determination of the agents' roles (or pre-formulated requirements to them). In the second case, the distribution of tasks happens largely spontaneously, directly in process of the interaction of the agents. Activity of artificial (computer) systems and organization of their joint work related to the collective and concerted solution of the tasks in virtual communities are fundamental characteristics of the conceptual novelty of advanced information technology and network organizations, built on the principles of the MAS. The main directions of development of multi-agent systems, distributed artificial intelligence Synergistic content of MAS conception is based on the processes of interaction of individual and collective agents, leading to the formation of artificial groups and communities, i.e., social computing systems with fundamentally new features. Depending on the number of 4 interacting agents and the inherent characteristics of their interactions, the various directions of development and types of MAS can be distinguished. Fig. 1.2. Classification of multi-agent systems. A distributed artificial intelligence and artificial life (in the narrow sense of the term) and the main directions in the development of MAS (Fig. 1.2). The studies of interaction and cooperation of a small number of intelligent agents, for example, the classical intelligent systems, including knowledge bases and solvers, compose a kernel of Distributed Artificial Intelligence (DAI) [6−11]. The main problem in DAI lies in the development of intellectual groups and organizations, capable to solve tasks by reasoning, which is related to the treatment of symbols. In other words, group intellectual behavior in DAI is based on individual intellectual behaviors. This means a congruence of the objectives, interests and strategies of different agents, coordination of their actions, the resolution of conflicts through negotiations; theoretical base in this process consists of the results obtained in the psychology of small groups and the sociology of organizations. DAI systems are defined by three main characteristics: 1) a method for the distribution of tasks between agents; 2) a method of distribution of powers; 3) method of communication of the agents. Typical scheme of distributed solution of the tasks by several agents includes the following steps: 5 MULTI-AGENT SYSTEMS DISTRIBUTED ARTIFICIAL NTELLIGENCE MAS is formed for solving the task. Centralized management and coordination of work of several intelligent agents are carried out DECENTRALISED ARTIFICIAL INTELLIGENCE The activity of autonomous agent in dynamic multi-agent environment is investigate ARTIFICIAL LIFE The processes of a decentralized management, evolution, adaptation and cooperation in MAS, consisting of a large number of reactive agents, are investigated and simulated 1) agent-subordinator (Head, the central body) decomposes the original problem into separate tasks; 2) these tasks are distributed between the agents-executants; 3) each agent-executant solves the task, sometimes also dividing it into sub-tasks; 4) in order to obtain the overall result a composition, integration of particular results corresponding to the selected task is produced. Agent-Integrator is responsible for the overall result (often, this is the same agent-subordinator). BOINC as a system of distributed artificial intelligence The BOINC (Berkeley Open Infrastructure for Network Computing.) is an open software platform (Berkeley University for GRID computing) − a non-profit middleware for organization of distributed computing. It is used for volunteer computing organization. Fig. 1.3. Scheme of BOINC work. 6 BOINC architecture (Fig. 1.3) is based on the idea of a distributed AI - its server consists of a set of individual subsystems (agents), each of which is responsible for its own well-defined task, such as performing of calculations, file transfer, etc. Each subsystem checks the status of sub-tasks produces some actions and changes the state of subtasks --- so they work in an infinite loop. In general, the system consists of BOINC server, a plurality of clients, performing the tasks of the server and, possibly, additional components in the form of GRID-affiliated networks. As we can see from the previous section, BOINC system, as well as any other system of distributed AI, existing at the moment, are the centralized systems, strictly managed by a central server, which, of course, is a significant drawback. SONM system uses the BOINC system, which is a system of distributed intelligence as the basis for the creation of a decentralized intelligence. Swarm Intelligence and artificial life The second direction − artificial life (AL) – is associated to a greater extent with the interpretation of intellectual behavior in the context of survival, adaptation and self-organization in dynamic, hostile environments, which goes back to the works by Piaget (see [12]). V.M. Bekhterev noticed that the more elementary goals and objectives of the collective, the more sizes the collective can reach. For example, man in a crowd of people loses inhibition ability, but wins in the imitative ones. In the tideway of AL a global intelligent behavior of the entire system is considered as a result of local interactions of a large number of simple and not necessarily intelligent agents. Terms such as "collective intelligence" (see, e.g., [13]; Fig. 1.4) or "swarm intelligence" [14, 15] are also used for AL. Adherents of this direction, in particular, R. Bruks, J. Deneubourg, L. Steels, etc. [16−20] rest on the following provisions: 1) the MAS is a population of simple and mutually dependent agents; 2) each agent independently determines its reaction to the events in the local environment and the interactions with other agents; 3) interrelations between agents are horizontal, i.e., there is no an agent-supervisor, managing the interaction of other agents; 4) there is no precise rules to define the global behavior of agents; 5) the behavior, properties and structure on the collective level are generated by only the local interactions of agents. Here, mechanisms of reaction to the impact of the environment and local interactions in general case do not include aspects such as forecasting, planning, processing knowledge, but sometimes allow to solve complex problems. Typical biological examples of such collective intelligence include ant colonies, beehives, bird flocks, etc. 7 Fig. 1.4. Collective Intelligence. In program form, blockchain can serve as the most typical example of the existence of swarm intelligence − a lot of miners work as reactive agents, who without any control of agent- supervisor carry out the work on maintenance of the network, moving only in accordance with their own motivation. The disadvantages of such systems consist of the inability of agents to the more complex organization, planning, and solution of the tasks, requiring sequential execution or data analysis, as well as the excess parallelism of tasks' execution. Decentralized artificial intelligence as a prologue of the future Many authors show principal differences between distributed and decentralized AI (Fig. 1.1) [21]. The ideology of distributed task solving [6, 22, 23] assumes mainly the separation of knowledge and resources between the agents and, to a lesser extent, distribution of management and power; as a rule, it postulates the existence of a common governing body that provides decision-making in critical (conflict) situations. At the same time, an overall complex problem for whose solution a group of agents is formed, a common conceptual model is constructed and the global criteria for achieving the goal are introduced, serves as an original object of the study. 8 In a fully decentralized systems management takes place only because of local interactions between agents. Here, the basic object of study is not a distributed solution of some general task, but an activity of autonomous agent in dynamic multi-agent world (as well as the coordination of different agents). At the same time, local tasks of individual agents, solved on the basis of local conceptual models and local criteria, are described along with the distributed knowledge and resources (Fig. 1.5). Fig. 1.5. Model of multi-agent system. We, Founders of SONM, believe that it is time to create such a decentralized AI based on existing models of multi-agent systems, such as the BOINC, which serves as an example of distributed AI and blockchain systems, which serve as examples of "swarm intelligence". Combining these two technologies will allow to compensate for the shortcomings of these systems and to enhance their benefits, thus creating a completely new AI model that will work for the benefit of all humankind. 9 Chapter 2. The practical implementation of the system Therefore, in the previous chapter, we found that an agent within the AI will be called a bot (robot, chat bot, softbot, etc.), which is motivated, able to perform actions with the external environment, and most importantly, able to communicate with other bots and shape their behavior. Multi-agent systems are systems consisting of these agents, which, however, are not always artificial. In addition, we found that the main directions of development of multi-agent systems are following:    Swarm intelligence; Distributed intelligence; Decentralized intelligence. Swarm intelligence is made up of simple reactive agents, which are able to perform only very simple functions (hash bruteforce, for example). Besides, swarm intelligence as a system, despite the fact that it is decentralized, is able to solve a very narrow range of tasks, because every its agent performs the same operation in one moment of time (the principle of redundant paralleling of the tasks). In contrast, the distributed intelligence (in the BOINC example) is able to solve a very wide range of tasks and currently it is actively used in different areas, but at the same time, it is strictly centralized and limited by design pattern. Decentralized intelligence should combine the two systems, compensating for their disadvantages and increasing their benefits. Decentralized intelligence SONM will use BOINC parts for distribution and integration of the tasks, as well as Ethereum smart contracts and p2p technology for decentralization. Scheme of implementation of decentralized solution "miner hub" Consider the process how miners and hubs communicate with each other when they need to establish cooperation, i.e., the phase when the miner has not yet decided whether to participate in the hub and receive tasks from him or not (Fig. 2.1). 10 Fig. 2.1. Scheme of construction of the trust. 11 At the start of the hub administrator disposes in the blockchain network a smart wallet contract, where SONM tokens, which he will pay to miners, will be located. Then, address of this smart contract, as well as the address of the organizer and IP of hub are recorded in a special SONM smart contract "Hub Pool List", which includes unconfirmed (unverified) hubs in the form of events, and the proven, i.e., introduced to the white list (in the first version it will be prepared by SONM, and further it will be formed only by the miners) hubs are included as an array. In any case, information about the hub in SONM's smart contract will contain information about the address of the hub owner, address of hub wallet and IP address. In case of change of IP address, hub owner can change the record, as well as in case of changing of the wallet address. That is, the hub registers the contract of the wallet, which contains the funds, which the hub pays to miners (so that the miners can check that the funds really exist) and registers basic information about itself, including the address of the owner and address of a wallet in a separate contract. Then, the agent on the side of the hub starts broadcasting in the network of p2p messenger protocol, sending a message about itself in the format «IP, organizer address, wallet address, and name» into a common channel. Agent on the side of the miner listens to the channel, and receives the message data from the servers, and then makes the request to the smart contract and compares data from the messages with data in the "white list". Miner may customize the agent so that he would accept messages from all servers or only from those, which are registered in smart contract as "proven". After that, miner's blockchain agent asks the information about the contract-wallet of the hub, the number of money in the contract and a list of recent transactions of the wallet. Intelligent analysis is conducted to compare the obtained data with the conditions, which were made by miner − whether funds in the hub wallet are sufficient and payments are regular, besides, the average amount of cash paid by hub is checked. Then messenger agent sends a direct message to the hub to request additional meta-data, and records full information about the hub in its table with the mark "not confirmed". At the same time, messenger agent regularly asks the questions to the common channel of miners with information about the hub, the average amount of remuneration paid to them, and so on, agents of other miners in the channel answer positively, if this information is correlated with their information, or negatively, if they believe this hub is not reliable. In the case if the miner agent receives a sufficient amount of confirmations, it makes a mark against the hub in its table as "checked", and if the transaction, which has been received by a 12 miner from this hub corresponds to the original agreements, then the status of a hub changes to "safe" After that, depending on the settings of miner's software, the miner can either manually select a hub to which he can connect to perform tasks, or miner's agent will automatically select and connect to a hub, which offers the maximum profit. A scheme of exchanging messages "miner-hub" is below (Fig. 2.2). Fig. 2.2. Scheme of interrelations between miner and hub. 13 Task distribution itself is carried out, considering maximizing the efficiency of the equipment used, so the miner will always get the task, which is most suitable for him. The distribution of tasks and validation of the results is performed by standard means of BOINC. Validation using a method of supernodes occurs in the channel, which is created by miners of one hub, working on the same task. Within the channel, the miners broadcast hashes of the results obtained, and simultaneously listen to messages from other miners. In the case if the result is the same, then they rebroadcast the message, including in it a confirmation of the result. In the case of reaching a consensus, the result is sent to the server with a mark of consensus, while in case of conflict detection, all versions of the results are sent to the server with a mark of conflict. In the case of conflict, the payment of remuneration for work is delayed, and a hub in the expert mode checks the result for compliance with the canonical approximation method. More information about the expert mode of the results testing can be found in BOINK documentation. By default (in the normal operation mode without supernodes), hubs always work in expert mode. Scheme of the implementation of decentralized solution "client-hub" Work of client agents (the buyers) with the hubs is substantially similar to the miner agents, excluding the intellectual parsing of the results, which in this case prefers the hubs with the maximally lowest price (and vice versa in the case of miners). It is more probable that the clients will use "Application Pool" (stated in the white paper), than a contract "Hub Pool" (Fig. 2.3). A method of content delivery is the only significant difference that we make in this process. As you might expect, there no difference between the rendering of 6-hour video in the local computer and uploading of this video to the server with waiting of the end of video rendering in the server, because most part of the time will be used for uploading. We have also the solution for this problem: when a client wants to upload a large file of raw data to the server, it simply creates a new torrent-file and sends a message indicating the operation to the selected hub. The hub receives the message and creates a task sequence for the uploading of this torrent, implementation of the work with the file and creation of a new torrent of the obtained file. The hub sends information about new file to the buyer, who only has to download the received file from the miners. We expect that this will be the most rapid solution of all existing ones at the moment. 14 Fig. 2.3. Scheme of work with clients in the case of decentralized solution "client-hub". Some intermediate messages in the scheme are omitted. 15 Money is sent to the contract of hub wallet at the opening of the tasks where it is deposited. When the buyer receives the result, he confirms the transfer of money by means of smart contract function (similar to a contract of Multisignature Wallet). Note that all transactions between the agents, blockchain, etc, go "under the hood", so the end users see only what they need. That is, the same buyer can simply select the desired application and run it, and the rest of the process − from the hub search to an implementation of the work by miners and the assembly of the result − may occur invisibly for him. You give the command "Make for me a part of this form!" and the system searches hubs that work with modeling applications and give them the task of calculating the parts with specific functions. Further, the hub distributes the task between miners, those obtains the results and the client receives a drawing. You say, "Give a forecast of dollar course for two months forward" and the system with enormous power calculates the probability of currency courses' movement and gives you a prediction. You say, "Give me a drawing of three-storey house in the Japanese style" and the system gives you a drawing of a Japanese three-storey house. You say, "Who will be killed by J. Martin next" and the system searches for an application for the calculation of the probabilities (or even neural network!), feeds it with all of the books "Game of Thrones" and you get a ready answer in a few minutes. You point − I punch! In other words, such a system would be able to solve absolutely any problems with virtually unlimited power (and with the appropriate application on the hubs, of course) and maximum efficiency. References 1. Holland, J.H. Adaptation in natural and artificial systems (2nd ed.). Cambridge, MA: The MIT Press, 1992 (1st ed., 1975). 2. Maes, P. (1994) Agent that Reduce Work and Information Overload, Communications of the ACM, 37, pp. 31-40. 3. Shvetsov, I., Nesterenko, T., Starovit, S. Technology of Active Objects. AAAI Workshop on Constraints and Agents, Providence, 1997. 4. Agha, G. (1986) ACTORS: A Model of Concurrent Computation in Distributed Systems. The MIT Press: Cambridge, MA. 5. Shoham, Y. (1993) Agent-oriented programming. Artificial Intelligence, 60 (1): 51−92. 6. Huhns, M.N. (editor). Distributed Artificial Intelligence. Pitman/Morgan Kaufmann, 1987. 16 7. Bond A.H. and Gasser L. Readings in Distributed Artificial Intelligence. Morgan Kaufmann Publishers, San Mateo, CA, 1988. 8. Pospelov, G.S. Artificial intelligence is basis of new information technology. Moscow, USSR: Nauka, 1988. 9. Parunak, H.V.D. (1997) 'Go to the Ant': Engineering Principles from Natural Agent Systems. Annals of Operations Research, 75: 69−101. 10. Chaib-draa, B., Moulin, B., Mandiau, R., and Millot, P. (1992) Trends in distributed artificial intelligence. Artificial Intelligence Review, 6: 35−66 11. O'Hare, G.M.P. and Jennings, N.R., editors, Foundations of Distributed Artificial Intelligence, John Wiley & Sons Inc., New York, 1996. 12. Piaget, J. Sociological Studies. London: Routledge, 1995. 13. Bonabeau, E. and Theraulaz, G. Intelligence collective. Hermès, Paris, 1994. 14. Deneubourg, J.L. et al. (1992) Swarm made architectures. 1st European Conference on Artificial Life, 123−133. 15. Adamatsky, A.I. and Holland, O. (1998) Swarm intelligence: conceptual representations and algorithms. Information technology and computer systems, 1: 45−53. [Russian]. 16. Brooks, R.A. (1991) Intelligence without Representation. Artificial Intelligence 47: 139– 160. 17. Heudin, J.C. La vie artificielle. Paris: Hermes, 1994. 18. Steels, L. Cooperation between distributed agents through self-organization. In Demazeau Y. and Müller J.-P. (editors). Decentralized AI − Proceedings of the First European Workshop on Modelling Autonomous Agents in a Multi-Agent World (MAAMAW-89), Elsevier Science Publishers B.V.: Amsterdam. The Netherlands, 1990, pp. 175−196. 19. Steels, L. (1994) The artificial life roots of artificial intelligence. Artificial Life Journal, Vol 1,1. MIT Press, Cambridge. 20. Ferber, J. Les systemes multi-agents. Vers une intelligence collective. Paris: InterEditions, 1995. 21. Demazeau, Y. and Müller, J.-P. (editors). Decentralized Artificial Intelligence. Proceedings of the First European Workshop on Modelling Autonomous Agents in a Multi-Agent World (MAAMAW'89). North-Holland, 1990 17 22. Durfee, E.H. Coordination of Distributed Problem Solvers, Kluwer Academic Press, Boston, 1988. 23. Rasmussen, S. et al. Dynamics of programmable matter. In Langton, C., Taylor, C., Farmer, J.D., and Rasmussen, S. (editors). Artificial Life II, SFI Studies in the Sciences of Complexity. Volume X. Redwood City, CA: Addison-Wesley 1991, pp. 211−254. 18
1601.04094	1	1601	2016-01-15T22:39:26	Efficient and Flexible Crowdsourcing of Specialized Tasks with Precedence Constraints	[ "cs.MA" ]	Many companies now use crowdsourcing to leverage external (as well as internal) crowds to perform specialized work, and so methods of improving efficiency are critical. Tasks in crowdsourcing systems with specialized work have multiple steps and each step requires multiple skills. Steps may have different flexibilities in terms of obtaining service from one or multiple agents, due to varying levels of dependency among parts of steps. Steps of a task may have precedence constraints among them. Moreover, there are variations in loads of different types of tasks requiring different skill-sets and availabilities of different types of agents with different skill-sets. Considering these constraints together necessitates the design of novel schemes to allocate steps to agents. In addition, large crowdsourcing systems require allocation schemes that are simple, fast, decentralized and offer customers (task requesters) the freedom to choose agents. In this work we study the performance limits of such crowdsourcing systems and propose efficient allocation schemes that provably meet the performance limits under these additional requirements. We demonstrate our algorithms on data from a crowdsourcing platform run by a non-profit company and show significant improvements over current practice.	cs.MA	cs	Efficient and Flexible Crowdsourcing of Specialized Tasks with Precedence Constraints Avhishek Chatterjee, Michael Borokhovich, Lav R. Varshney, and Sriram Vishwanath 1 6 1 0 2 n a J 5 1 ] A M . s c [ 1 v 4 9 0 4 0 . 1 0 6 1 : v i X r a Abstract Many companies now use crowdsourcing to leverage external (as well as internal) crowds to perform specialized work, and so methods of improving efficiency are critical. Tasks in crowdsourcing systems with specialized work have multiple steps and each step requires multiple skills. Steps may have different flexibilities in terms of obtaining service from one or multiple agents, due to varying levels of dependency among parts of steps. Steps of a task may have precedence constraints among them. Moreover, there are variations in loads of different types of tasks requiring different skill-sets and availabilities of different types of agents with different skill-sets. Considering these constraints together necessitates the design of novel schemes to allocate steps to agents. In addition, large crowdsourcing systems require allocation schemes that are simple, fast, decentralized and offer customers (task requesters) the freedom to choose agents. In this work we study the performance limits of such crowdsourcing systems and propose efficient allocation schemes that provably meet the performance limits under these additional requirements. We demonstrate our algorithms on data from a crowdsourcing platform run by a non-profit company and show significant improvements over current practice. I. INTRODUCTION The nature of knowledge work has changed to the point nearly all large companies use crowdsourcing approaches, at least to some extent [2]. The idea is to draw on the cognitive energy of people, either within a company or outside of it [3]. A particularly notable example is the non-profit impact sourcing service provider, Samasource, which relies on a marginalized population of workers to execute work, operating under the notion give work, not aid [4], [5]. There are multifarious crowdsourcing structures [6], [7] that each require different strategies for matching work to agents [8]. Contest-based platforms such as TopCoder and InnoCentive put out open calls for participation, and best submissions win prizes [9]. Microtask platforms such as Amazon Mechanical Turk allocate simple tasks on a first-come-first-serve basis to any available crowd agent. When considering platforms with skilled crowds and A. Chatterjee and L. R. Varshney are with the Coordinated Science Laboratory, University of Illinois at Urbana-Champaign, Urbana, Illinois, USA. (email: {avhishek,varshney}@illinois.edu). M. Borokhovich is with the AT&T Labs, New Jersey, USA. (email: [email protected]). S. Vishwanath is with the Wireless Networking and Communication Group, The University of Texas at Austin, Austin, Texas, USA. (email: [email protected]). Part of the material in this paper will be presented at IEEE INFOCOM 2016, San Francisco, USA [1]. 2 specialized work, such as oDesk (now upWork) [7], IBM's Application Assembly Optimization platform [10], and to a certain extent Samasource's SamaHub platform [4], efficient allocation algorithms are needed. In these skill-based crowdsourcing platforms, the specialized tasks have multiple steps, each requiring one or more skills. For example, software development tasks may first be planned (architecture), then developed (programming), and finally tested (testing and quality assurance), perhaps with several iterations. Even in skilled microtasking platforms like SamaHub, most jobs have more than one step. Task steps often have precedence constraints between them, implying that a particular step of a task can only be performed after another set of steps has been completed. To serve a step requiring multiple skills, we need either a single agent that has all of the skills or a group of agents that collectively do so. Whether multiple agents can be pooled to serve a step or not depends on the flexibility of the step: if there are strong interdependencies between different parts of a step, the step may require a single agent. Notions of flexibility and precedence constraints are central to this paper. Allocating tasks to servers is a central problem in computer science [11], communication networks [12], and operations research [13]. The skill-based crowdsourcing setting, however, poses new challenges for task allocation in terms of vector-valued service requirements, random and time-varying resource (agents) availability, large system size, a need for simple decentralized schemes requiring minimal actions from the platform provider, and the freedom of customers (task requesters) to choose agents without compromising system performance. Some of these issues have been addressed in recent work [14], [15], but previous work does not address precedence constraints or step flexibility. The notion of flexibility in [15] is based on agent-categories and is different from here. Task allocation with precedence constraints has been studied in theoretical computer science, as follows. Given several tasks, precedence constraints among them, and one or more machines (either same or different speed), allocate tasks to minimize the weighted sum of completion times or maximum completion time [16]. In crowdsourcing, we have a stream of tasks arriving over time and so we are interested in dynamics. Dynamic task allocation with precedence constraints has recently been studied in [17] for Bernoulli task arrivals. This is different from crowdsourcing scenarios, and the optimal scheme is required to search over the set of possible allocations, which is not suitable for crowdsourcing systems due to their inherent high-dimensionality (many types of tasks). Additional challenges in a crowdsourcing platform are: (i) random and time-varying agent availability; (ii) vector-valued service requirements; (iii) fast computation requirements for scaling; and (iv) freedom of choice for customers. Here we address the above issues for various flexibilities of steps and agents, to characterize limits of crowd systems and develop optimal, computationally-efficient, centralized allocation schemes. Based on insights garnered, we further present fast decentralized greedy schemes with good performance guarantees. To complement our theoretical results, we also present numerical studies on real and synthetic data, drawn from Samasource's SamaHub platform. The remainder of the paper is organized as follows. Sec. II describes the system model for crowdsourcing platforms with different precedence and flexibility constraints. Sec. III presents a generic characterization of the system limits and a generic centralized optimal allocation scheme. Secs. V -- VII address particular systems with 3 different flexibility constraints to yield fast decentralized schemes that meet crowdsourcing platform requirements mentioned above. Sec. VIII presents numerical studies on real and synthetic data. Detailed proofs of theoretical results are given in Appendix A. II. SYSTEM MODEL There are a total of S kinds of skills available in the crowdsourcing system, numbered [S] = {1, 2, . . . , S}. We define types of agents by skills, and denote the total number of types of agents by M. An agent of type m has skills Sm ⊂ [S]. Tasks posted on the platform are of N types. Each type of task j has one or multiple steps associated with it, denoted Kj. A step k ∈ {1, 2, . . . , Kj} of a job type j -- a (j, k)-step -- needs a skill-hour service vector rj,k ∈ RS + (non-negative orthant), i.e. rj,k,s hours of skill s. A part of a step of type (j, k) involving skill s is called a (j, k, s)-substep if rj,k,s > 0, the size of this substep. In the platform, allocations of work to available agents happen at regular time intervals, t = 1, 2, . . . . Tasks that arrive after an epoch t are considered for allocation at epoch t + 1, based on the available agents at that epoch. Tasks or parts of tasks that remain unallocated due to insufficient available skilled agents are considered again in the next epoch. We assume that for any substep (j, s), the time requirement is less than the duration between two allocation epochs. Tasks arrive according to a stochastic process in ZN + (non-negative orthant), A(t) = (A1(t), A2(t), . . . , AN (t)), where Ai(t) is the number of tasks of type i that arrive between epochs t − 1 and t. The stochastic process of available agents at epoch t is U(t) = (U1(t), U2(t), . . . , UM (t)). We assume A(t) and U(t) are independent of each other and that each of the processes are i.i.d. for each t, with bounded second moments. Let Γ(·) be the distribution function of U(t), and let λ = E[A(t)] and µ = E[U(t)] be the means of the processes. An agent is inflexible if it has pre-determined how much time to spend on each of its skills. Inflexible agents bring a vector hm = (hm,1, hm,2, . . . , hm,S) where hm,s > 0 if and only if s ∈ Sm and an inflexible agent spends no more than hm,s time for skill s. Contrariwise, flexible agents bring a total time hm which can be arbitrarily split across skills in Sm. A step is flexible if it can be served by any collection of agents pooling their service-times. All substeps of inflexible steps must be allocated to one agent. At any epoch t only an integral allocation of a step is possible. Hence, in any system for a step to be allocated, all of its substeps must be allocated. A set of flexible substeps sst1, sst2, . . . , sstn of size x1, x2, . . . , xn with skill s can be allocated to agents 1, 2, . . . , m if the available skill-hours1 for skill s of these agents, y1, y2, . . . , ym, satisfy the following for some {vpq ≥ 0 : 1 ≤ p ≤ n, 1 ≤ q ≤ m}, n(cid:88) m(cid:88) vip ≤ yi i ∈ [m], vqj ≥ xj j ∈ [n], (1) p=1 q=1 1Available skill-hour is determined by the availability of the agent, system state, and whether agents are flexible or inflexible. where the {vpq} capture how agents split their available skill hours across substeps. 4 For inflexible steps, a set of steps st1, st2, . . . , stn of size (vectors) x1, x2, . . . , xn can be allocated to an agent with available skill-hours (vector) y if n(cid:88) i=1 xi ≤ y. (2) There may also be precedence constraints on the order in which different steps of a task of type j can be served. For any task of type j, this constraint is given by a directed rooted tree Tj on Kj nodes where a directed edge (k → k(cid:48)), k, k(cid:48) ∈ [Kj] implies step k(cid:48) of a task of type j can only be served after step k of the same task has been completed. Scalings of several crowd system parameters are as follows. Task arrival rate λ(N ) =(cid:80)N skills of an agent is d(cid:48) = O(1) implying M =(cid:80) j=1 λj scales faster than number of task types N, i.e. limN→∞ N/λ(N ) = 0. Number of skills S scales slower than N, i.e. S = o(N ). In practice, a task requires a constant number of skills d, which implies Ω(Sd) possible types of steps. Number of 1 ) and M = Ω(N α2 ) for 0 < α2 < α1. Further, the length of tasks and availability of agents do not scale with the size of crowdsourcing systems. l M l = O(Sd(cid:48) ), implying M = O(N α λs(N )) =(cid:80) Beyond these practical system scalings, we make the following mild assumptions: λj = ω(1) for all j and j:rj,1,s>0 λj(N ) = Ω (N c) ∀s ∈ [S], for some c > 0. These assumptions mean the arrival rate of every type of job and the total number of jobs requiring a particular skill scale with the system. We call these scaling patterns crowd-scaling. III. NOTIONS OF OPTIMALITY To formally characterize the maximal supportable arrival rate of tasks, we introduce some more notation and invoke some well-accepted notions used in this regard. For each j ∈ [N ], let the number of unfinished tasks in the system just after allocation epoch t − 1 be Qj(t). Aj(t) is the number of tasks of type j arriving between epochs t−1 and t. The number of tasks of type j completely allocated (all steps) at epoch t is Dj(t). Thus Qj(t) evolves as: (3) Clearly Dj(t) ≤ Qj(t) + Aj(t) at any epoch t, since at most Qj(t) + Aj(t) type j tasks are available. Hence Qj(t) ≥ 0 for all t. Note that due to additional precedence constraints, typically Dj(t) < Qj(t) + Aj(t). Qj(t + 1) = Qj(t) + Aj(t) − Dj(t). Definition 1. A scheme of allocation of tasks is called a policy if it allocates tasks at a time epoch t based on knowledge of statistics of the processes A and U and their realizations up to time t, but does not depend on future values. Definition 2. A crowd system is stable under policy P if the process Q(t) = (Qj(t), j ∈ [N ]) has a finite expectation in steady-state under that policy, i.e., lim supt→∞ E[Qj(t)] < ∞, for all j for any initial condition. 5 Definition 3. An arrival rate λ is stabilizable if there exists a policy P under which Q(t) = (Qj(t), j ∈ [N ]) is stable. Definition 4. The capacity region of a crowd system for a given distribution Γ of the agent-availability process U(t) is the closure of the set CΓ = {λ : λ is stabilizable}. We aim to propose statistics-agnostic, computationally simple and decentralized schemes that offer customers freedom of choice while stabilizing any arrival rate in the system's capacity region. Stronger than stability, often we give high probability bounds on number of unallocated tasks. IV. CAPACITY AND CENTRALIZED ALLOCATION ROUTINE Here we present a generic characterization of the capacity region of a crowd system for all combinations of agent- and task-flexibility. We also give a generic centralized allocation routine that can be easily adapted to a particular system. For any given set of available agents u = (ui : 1 ≤ i ≤ M ), define the number of different types of steps ({aj,k}) . When we say {aj,k} is the number of steps that can potentially be allocated in a crowd system by C(u) ⊂ R of different types that can potentially be allocated, we consider the following scenario that satisfies the allocation constraints in Sec. II. A1 An infinite number of steps of each type (j, k), k ∈ [Kj] for a j ∈ [N ] are available for allocation, i.e., the j Kj (cid:80) + limitation only comes from the available resource u. A2 Precedence constraints among the steps are already satisfied, i.e., all corresponding (j, k)-steps of the available (j, k(cid:48))-steps have already been allocated previously. This is equivalent to an absence of precedence constraints. A3 Integral steps must be allocated, i.e., all substeps of a step need to be allocated for allocation of the step. A4 To allocate aj,k steps of different types to a collection of R agents of type {mr : r ∈ [R]} and available hours {ymr,s : r ∈ [R]} (which is a function of {hmr : r ∈ [R]} depending on the system), we need to satisfy either (1) or (2) depending on system type. Let Ccvx(u) be the convex hull of the set C(u), and define another set C ⊂ R j Kj as follows. (cid:40)(cid:88) C = Γ(u)a(u) : a(u) ∈ Ccvx(u) (cid:80) + (cid:41) Based on this we define another set C ⊂ RN . . . , (aN , aN , . . . , KN times)). Then C = {a : aE ∈ C}. This set characterizes the capacity region of the crowd system. + , aE := ((a1, a1, . . . , K1 times), (a2, a2, . . . , K2 times), + . Let for any a ∈ RN u Theorem 1. Any arrival rate λ is stabilizable if for some > 0, λ + 1 ∈ C and no arrival rate λ can be stabilized if λ is outside the closure of the set C. Note that we ignore the precedence constraint in defining C(u). This does not conflict with the fact the capacity region is a subset of C, but it may not be obvious C is in fact the capacity region. A fortiori, we show this with a scheme that respects precedence constraints and stabilizes any rate in the interior of C. 6 A. Centralized Allocation Let us develop a statistics-agnostic scheme that stabilizes any arrival rate λ. Let Qj,k(t) be the number of unallocated (j, k) steps just before allocation epoch t. This includes steps not allocated at epoch t − 1 and steps that became available for allocation between t − 1 and t. Thus, if for any (j, k), Dj,k(t) (j, k)-steps were allocated at epoch t and Aj,k(t + 1) new (j, k)-steps became available between t and t + 1, Qj,k(t + 1) = Qj,k(t) − Dj,k(t) + Aj,k(t + 1). Note that, for any j and Kj ≥ k > 1, new (j, k)-steps become available only when some (j, k− 1) steps have been completed. Service times {rj,k,s} are strictly less than the duration between two allocation epochs. So, any step allocated at epoch t is completed before epoch t + 1. Hence, for any j and Kj ≥ k > 1: Aj,k(t + 1) = Dj,k−1(t). On the other hand, for any j and k = 1, we have an external arrival Aj(t + 1) between epoch t and t + 1. At any time t, for a given resource availability, an allocation rule determines resources to be allocated for certain number of (j, k)-steps. We denote this by Sj,k(t). Note that Dj,k(t) = min(Qj,k(t), Sj,k(t)). Our goal is to design a scheme that finds a good {Sj,k(t)} for a given {Qj,k(t)} and U(t) = u. Centralized Allocation Input: {Qj,s(t) : j ∈ [N ], s ∈ [S]} and U(t) at t j,k(t)} and allocation of steps to agents Output: {S∗ 1: Define: lj,r : number of leaves in the subtree of Tj rooted at r 2: Obtain {S∗ sj,klj,r(Qj,k(t) − Qj,r(t)) (4) j,k(t)} = arg maxsj,k∈C(U(t))(cid:88) Kj(cid:88) (cid:88) j k=1 r:k→r∈Tj 3: For each (j, k) allocate S∗ j,k(t) (j, k)-steps This allocation scheme is statistics-agnostic and explicit in terms of system state. Also, note that by the design of the scheme the precedence constraint is automatically satisfied. One important thing to note is that the allocation scheme is generic, in the sense that this policy can be easily adapted for different agent- and step-flexibility. Note that C(U(t)) comes from the allocation constraints of the system. If in (4) we replace C(U(t)) by the corresponding allocation set, the centralized algorithm becomes a generic allocation routine. In fact, the generic statistics-agnostic routine for centralized allocation scheme described above is optimal, in the sense that any arrival rate that can possibly be stabilized by any policy can also be stabilized by this scheme. Theorem 2. The centralized allocation routine described above stabilizes any λ if λ + 1 ∈ C, the capacity region of the corresponding system for any > 0. 7 Though the scheme has similarity with back-pressure algorithms [12], [18], [19]; unlike the back-pressure scheme it also uses graph parameter (lj,r) in computing the weights. Proof is using a Lyapunov function involving {lj,r} and queue-lengths. Instead of directed rooted tree if the precedence constraint is a directed acyclic graph the same results extend. It would be apparent from proof of Theorem 1 that the converse (outer-bound on capacity) depend on the precedence graph. On the other hand, for any precedence constraint given by a directed acyclic graph, there exists a precedence constraint given by a directed rooted tree such that the tree constraint does not violate the directed acyclic graph constraint. Then, by applying the above centralized algorithm for this directed rooted tree capacity can be achieved. V. INFLEXIBLE AGENTS AND FLEXIBLE STEPS Here we characterize the limits of tasks allocation where all steps are flexible and agents are inflexible. Sec. III presented a generic capacity characterization and algorithm; this section investigates computational aspects of the generic algorithm for this particular system and also proposes a simple decentralized scheme that works well under a broad class of assumptions. Consider CI,F (u), the set of possible allocations with inflexible agents (I) and flexible tasks (F ) for availability of agents, u. Recall the allocation scenario in Sec. III to determine a generic C(u): A1 -- A3 are the same for any system flexibility, but A4 is specific. For an (I, F ) system we have the following. To allocate aj,k,s tasks of different types to a collection of R agents of type {mr : r ∈ [R]} and available hours {hmr,s : r ∈ [R]} we must satisfy (1):(cid:88) aj,k,srj,k,s ≤ R(cid:88) hmr,s for all s ∈ [S]. (5) j,k r=1 Note that whenever a step is allocated, all tasks in it must be allocated simultaneously. Hence, we can only allocate aj,k,s tasks with aj,k,s = aj,k,s(cid:48) ∀s, s(cid:48) ∈ [S] when satisfying (5). Given CI,F (u), the capacity region CI,F is obtained in the same way C was obtained from C(u) in Sec. III. The generic centralized allocation routine can be similarly specialized for (I, F ) systems: C(U(t)) in (4) of the routine is replaced by CI,F (U(t)). The centralized scheme is computable since CI,F (U(t)) can be written explicitly in terms of U(t), rj,k, and hm, but it cannot always be computed in polynomial time. Since any allocation in CI,F (u) must satisfy constraint (5), optimization problem (4) can be written as: (cid:88) Kj−1(cid:88) sj,krj,k,s ≤(cid:88) s.t. (cid:88) wj,ksj,k k=1 max sj,k∈Z+ j umhm,s for all s ∈ [S]. (6) j,k m Note that the solution to the problem does not change if we replace wj,k by max(vj,k, 0), as optimal schemes never allocate resources to negative wj,k. Thus, we assume wj,k ≥ 0. Note that (6) is a multi-dimensional knapsack problem, where the number of available items of a given weight and value are unbounded [20]. This problem is known to be NP-hard and without any fully polynomial-time 8 approximation scheme (FPTAS). A polynomial-time approximation scheme (PTAS) is known, but the complexity is exponential in dimension. Recently extended linear programming (LP) relaxations have been proposed, but have the same issues (see [21] and references therein). We aim to find a simple and fast distributed scheme, but first propose the following LP relaxation-based, polynomial-time (in N and M) scheme that gives nearly optimal centralized allocation for a large crowd system (under crowd scaling). { Sj,k(t)} = (6) with relaxing of {sj,k ∈ R+} Allocate {Sr j,k(t) = (cid:98) Sj,k(t)(cid:99)} (7) We cannot give performance guarantees for this scheme at each allocation epoch for arbitrary Qj,k, but for a sufficiently large crowd system, this scheme stabilizes almost any arrival rate that can be stabilized. Theorem 3. Under crowd scaling, for any α < 1 there is an N0 such that for any system with N ≥ N0, the LP-based scheme (7) stabilizes any arrival rate in αC = {a : a α ∈ C}. A. Decentralized Allocation In this section we develop a simple decentralized scheme with good performance guarantees. As discussed before, often one of the main reasons for customers to go to a crowd platform is the ability to choose workers themselves. As such, we propose a simple greedy scheme that allows customers the freedom of choice with minimal intervention from platform operators. This also reduces the platform's operational cost. In greedy allocation, each step competes against others to find an allocation for all of its tasks. Contention can be resolved arbitrarily, e.g., random, pre-ordered, or age-based. The Prioritized Greedy algorithm below performs greedy allocation among all steps across all types of tasks that are in the same order. It starts with steps that are in the beginning of the precedence tree and once these steps find an allocation (or cannot be allocated), only then are steps lower in the corresponding precedence trees allowed to allocate themselves. Algorithm 1 Prioritized Greedy Define D = maxj depth of Tj 1: Sj = ∅ for all j ∈ [N ] 2: for d=1:D do 3: Sj = {kj : depth of kj in Tj = d} Greedy allocation among ∪j{j, kj : kj ∈ Sj} steps 4: 5: end for This algorithm can be efficiently implemented on a crowdsourcing platform with minimal intervention from the platform operator. The operator need only tag unallocated steps in the system based on their depth in the rooted precedence tree and only show available workers to them after steps at lower depth have exercised their allocation choice. This may be implemented by personalizing the platform's search results. The algorithm is fast and has good performance guarantees under certain broadly-used assumptions on arrival and availability processes. 9 Definition 5. A random variable X is Gaussian-dominated if E[X 2] ≤ E[X]2 + E[X] and for all θ ∈ R, E[eθ(X−E[X])] ≤ e 2 ((E[X 2]−E[X]2)θ2, and Poisson-dominated if for all θ ∈ R, E[eθ(X−E[X])] ≤ eE[X](eθ−θ−1). 1 These domination definitions, commonly assumed in bandit problems [22], imply that variation around the mean is dominated in a moment generating function sense by that of a Gaussian (Poisson) random variable. Such a property is satisfied by many distributions used to model arrival processes, including in crowdsourcing systems [23]. Theorem 4. Consider inflexible agents and flexible steps crowdsourcing systems (size N) where for any s, s(cid:48) m µmhm,s(cid:48) is sub-poly(N ), i.e., o(N δ),∀δ > 0, arrival and availability processes are Poisson- dominated (and/or Gaussian-dominated), and system scales as per crowd-scaling. Then, for any α ∈ (0, 1), ∃Nα s.t. ∀N ≥ Nα, any arrival rate λ ∈ αCI,F can be stabilized by Prioritized Greedy, and at the steady state the total (cid:80) m µmhm,s −(cid:80) number of unallocated steps in the system across all types is O(log N ) w.p. 1 − O(cid:0) 1 (cid:1). N 2 This implies Prioritized Greedy can stabilize almost any stabilizable arrival rate while ensuring the number of unallocated tasks scales more slowly than the system size. VI. FLEXIBLE AGENTS AND FLEXIBLE STEPS Now consider systems with flexible agents and flexible steps (F, F ), and characterize capacity regions. For a given availability of agents u, the set of possible step allocations are CF,F (u). As for C(u) in Sec. III this satisfies A1 -- A3 in the allocation scenario; A4 for (F, F ) systems is as follows. A certain number of steps {aj,k} of each type can be allocated to a set of agents {1, 2, . . . , R} of types {mr : r ∈ [R]} if there exists a set of R vectors in RS, hmr = (hmr,s ≥ 0 : s ∈ [S]), such that: (cid:88) hmr,s = 0 if s (cid:54)∈ Smr for all s, mr, hmr,s ≤ hmr for all r ∈ [R], and (cid:88) aj,krj,k,s ≤ R(cid:88) hmr,s for all s ∈ [S]. s (8) j,k r=1 Based on the set of possible allocations CF,F (u), the capacity region CF,F can be characterized just as in Sec. III. Similar to Sec. V, if we replace C(U(t)) by CF,F (U(t)) in the centralized allocation routine we obtain an optimal policy for the (F, F ) system. It is not hard to see that for the instance where each agent has exactly one skill, the problem is again a multi-dimensional knapsack problem and therefore NP-hard. We develop a computationally- efficient scheme. optimize: This is a mixed ILP with(cid:80)N max (sj,k∈Z+),(hmr ,s∈R+) (cid:88) j,k (cid:88) max wj,ksj,k j,k (sj,k∈Z+),(αm,s∈R+) s.t. αm,s = 0 if s (cid:54)∈ Sm for all s, m, αm,s ≤ 1 for all m ∈ [M ], and sj,krj,k,s ≤ (cid:88) (cid:88) (cid:88) s If there are R agents of type {mr : r ∈ R} available, then the centralized allocation problem at time t is to 10 wj,ksj,k s.t. constraints in (8). (9) j=1 Kj integer variables and RS real variables. The complexity of this problem scales with the number of available agents in the system, R. We would like to avoid such a scaling as R may be much larger than M and N in a crowd system. Hence, we pose another optimization problem where the number of variables scales with M and N. Given U (t) = u, umhmαm,s for all s ∈ [S]. (10) j,k m=1:M Note that this optimization problem yields an allocation satisfying all constraints for flexible agent allocation. This is because αm,s is the fraction of time of an agent of type m that has been given to skill s, which can be positive only when agent of type m has skill s. The last inequality ensures that the skill-hour constraint per skill is satisfied. Hence, this is a feasible allocation procedure. This is again a mixed ILP, but with M + N variables. Note that this problem is also NP-hard, corresponding to a multi-dimensional knapsack problem if Sm = 1, for all m ∈ [M ]. We design a centralized scheme that allocates steps based on the following LP relaxation. Given U (t) = u, (sj,k : j, k) = arg max (sj,k∈R+),(αm,s∈R+) wj,ksj,k (cid:88) j,k s.t. constraints in (10) and allocate {(cid:98)sj,k(cid:99)} steps. (11) This scheme has the following performance guarantee. Theorem 5. Under crowd scaling, for any α < 1 there is an N0 s. t. for any system with N ≥ N0, the LP-based scheme (11) stabilizes any arrival rate in α CF,F = {a : a α ∈ CF,F}. Proof of this theorem is based on the equivalence of (9) and (10). A. Decentralized Allocation Now we develop a decentralized allocation scheme that requires minimal centralized operation, and gives cus- tomers the option to choose from a pool of multiple agents. Algorithm 2 Prioritized Greedy with Flexibility Initialize: {γ(t − t0) ∈ [0, 1], t ≥ t0}, at starting time t0 ¯A(t0) = 1, > 0 1: Update at each t: ¯A(t) = (1 − γ(t − t0)) ¯A(t − 1) + γ(t − t0)A(t) ¯U(t) = (1 − γ(t − t0)) ¯U(t − 1) + γ(t − t0)U(t) 2: Solve for {ψm,s(t) : s ∈ [S], m ∈ [M ]} (cid:88) max 1 s.t. ¯Aj(t)rj,k,s ≤ (1 − ) (cid:88) m:s∈Sm ¯Um(t)ψm,s(t) j,k ψm,s(t) ≥ 0, ψm,s(t) ≤ 1 (cid:88) s 11 (12) ψm,s(t) > 0 only if s ∈ Sm, if no solution pick ψm(t) randomly from a simplex in RS. 3: Initialize sets: Bs = ∅,∀s ∈ [S] 4: For each type m: put each available agent in one of {Bs} w. p. {ψm,s(t)} (independent rolls of loaded dices) 5: Create inflexible agents: an agent of type m in Bs has hm available time only for skill s 6: Run Prioritized Greedy for this (I,F) system This algorithm is amenable to crowdsourcing platform implementation. Note ¯A(t) is available from recent history. Creating the set Bs is simple: for any agent of type m we just randomly tag (as per ψ) with a particular skill and it is shown only tasks with this particular skill. Similarly customers are only shown that the agent has only the particular skill. The rest of the algorithm is exactly like Prioritized Greedy where we create classes of steps and priorities among them and then within each class the allocation is arbitrarily greedy. We can guarantee Alg. 2 performance when γ satisfies: γ(x) = O(cid:0)x−1(cid:1) and γ(x) = Ω are Poisson (and/or Gaussian) dominated with restricted asymmetry, i.e., maxs,s(cid:48) (cid:80) the total number of unallocated steps in the system across all types is O(log N ) w.p. 1 − O(cid:0) 1 Theorem 6. Consider a flexible agents and flexible steps crowdsourcing system with availability processes that j,k λjrj,k,s(cid:48), being O (subpoly(N )). For any α ∈ (0, 1), ∃Nα s.t. ∀N ≥ Nα in such systems of size N that follow crowd scaling any arrival rate λ ∈ αCF,F can be stabilized by Alg. 2 and at the steady state (i.e., for any finite t when t0 = −∞) , > 0. x− 1 2 +(cid:17) (cid:16) j,k λjrj,k,s −(cid:80) (cid:1). N 2 VII. FLEXIBLE AGENTS AND INFLEXIBLE STEPS Now consider the setting where agents may split their available service-time across their skills, but a step must be allocated to one agent. Multiple agents cannot pool their service time to serve a step. As before, for an agent availability vector u, there is a set of possible allocations of steps (of different (j, k)-types) to agents, denoted CF,I (u). Given CF,I (u) and the distribution of agent availability Γ(u), we can define a capacity region CF,I in the same way as CI,F is defined in Sec. V based on CI,F . Similarly, the generic centralized routine can be adapted by changing the optimization over C(U(t)) to an optimization over CF,I (U(t)) while ensuring optimality of the modified scheme for (F, I) system. 12 Allocation constraint (2) is for allocation of steps to a particular agent. For a given set of agents of different types u = (u1, u2, . . . , uM ) the allocation constraint can be written based on (2). Note that for inflexible steps agents cannot pool service-times to serve a step. Consider a set of available agents a1, a2, . . . , aR, of types 1, 2, . . . , mr respectively. An allocation of {sj,k ∈ Z+ : k ∈ [Kj], j ∈ [N ]} steps to these agents is possible if and only if there are integers {zj,k,r ∈ Z+} such that zj,k,r (j, k)-steps are allocated to agent ar and all {sj,k} steps are allocated to some agent, i.e., for each r there is an αr in an S-dimensional simplex so that: rj,kzj,k,r ≤ αrhmr , αr,s = 0 if s (cid:54)∈ Smr , r ∈ [R] Hence, the optimization problem in the centralized allocation routine for (F, I) system is an integer LP of the form: zj,k,r ≥ sj,k∀j, k. (cid:88) Kj(cid:88) wj,ksj,k (cid:88) (cid:88) j,k r max sj,k∈Z+ s.t. (cid:88) (cid:88) j,k r j k=1 rj,kzj,k,r ≤ αrhmr , αr,s = 0 if s (cid:54)∈ Smr , r ∈ [R], zj,k,r ∈ Z+ for all j, k, r, R = um. (13) zj,k,r ≥ sj,k for all j, k, and (cid:88) m (cid:80)Kj−1 Note that like (6), the objective can be written as(cid:80) (cid:88) (cid:88) Kj(cid:88) This problem has a special structure which leads to a computationally-efficient algorithm. Consider the following. k=1 vj,ksj,k where vj,k = max(wj,k, 0). j vj,k zj,k,r j k=1 r max zj,k,r∈Z+ s.t. (cid:88) j,k rj,kzj,k,r ≤ αrhmr , αr,s = 0 if s (cid:54)∈ Smr , r ∈ [R], and (cid:88) m zj,k,r ∈ Z+ for all j, k, r, R = um. (14) When operating at the optimum of (13),(cid:80) (cid:80) r zj,k,r = sj,k, and so we see that (13) and (14) have the same optimal value. Hence, we solve problem (14) instead of problem (13). Note that as there is no constraint between {zj,k,r : j, k} and {zj,k,r(cid:48) : j, k}, problem (14) decomposes into m um optimization problems, each for an available agent. Consider the optimization problem for an agent of type m. Kj(cid:88) k=1 vj,kzj,k (cid:88) s.t. (cid:88) max zk,j j j,k rj,kzj,k ≤ αrhmr , αr,s = 0 if s (cid:54)∈ Smr, 13 (15) which is again equivalent to the following problem, expressed in terms of the set Sm of skills of type m agent: (cid:88) (cid:88) max zk,j s.t. j,k:rj,k,s=0 if s(cid:54)∈Sm vj,kzj,k (cid:88) zj,k rj,k,s ≤ hm. (16) j,k:rj,k,s=0 if s(cid:54)∈Sm s This is a one-dimensional knapsack problem, and there are dynamic programming (DP) pseudo-polynomial time algorithms for solving. Since the sack size hm is finite (does not scale with the system), the DP has computational m um problems, each of which can j Kj). This implies the centralized scheme decomposes into(cid:80) complexity O((cid:80) be solved in polynomial time. Thus, the centralized scheme naturally leads to a decentralized scheme where each available agent solves (16) and uses the optimal solution as its potential allocated steps. Agents may use an arbitrary contention mechanism among themselves to decide which agent allocates first. Upon resolving contention, agents pick steps greedily by solving (16). Since the decentralized scheme follows directly from the centralized one (13), performance guarantees from Thm. 2 hold. Although this simple decentralized scheme is optimal, it does not give customers freedom of choosing agents. Thus, we propose another decentralized scheme where customers get to pick any agent from a subset of available agents. Alg. 3 allows the different types of steps to pick agents greedily, but in a restricted manner. It prioritizes steps with lower depth like Prioritized Greedy. Among steps with the same priority (in terms of depth), it gives preference to steps requiring more skills to ensure an agent with multiple skills is not used unwisely for a step with lesser requirements. Theorem 7. Consider a flexible agent and inflexible steps crowdsourcing system where each type of agent has O(1) skills, and arrival as well as availability processes are Poisson (and/or Gaussian) dominated, {Sm : m ∈ [M ]} is s rj,k,s are same for all (j, k). For this system for any α ∈ (0, 1), ∃Nα s.t. ∀N ≥ Nα in such systems of size N that follows crowd scaling any arrival rate λ ∈ αCF,I can be stabilized by Alg. 3 and at a partition of [S] and(cid:80) steady-state the total number of unallocated steps in the system across all types is O(log N ) w.p. 1 − O(cid:0) 1 (cid:1). N 2 For many systems the total sizes of steps are nearly identical and so the assumption on total size is not restrictive, though results can be extended to the case where the total sizes are random with the same mean. The assumption {Sm : m ∈ [M ]} is a partition is required for proving the performance guarantee, but the algorithm (actually a simpler version) works well on simulations. The above performance guarantee can be extended for the following Algorithm 3 Restricted Greedy Compute and Store: one-time 14 1: for d = 1 : D do 2: Ad = ∪j{set of required skills for steps at depth d of Tj} 3: pd = 0 while Ad (cid:54)= ∅ do pd = pd + 1 Pick maximal subsets from the collection Ad, say Ld pd. Ad = Ad\Ld pd 4: 5: 6: 7: 3: 4: end while 8: 9: end for Allocation at time t Define D = maxj depth of Tj 1: for d = 1 : D do while k ≤ pd do 2: Steps in Ld k = k + 1 k allocate themselves greedily (ties are broken arbitrarily) end while 5: 6: end for conditions. {Si : i ∈ I} is a partition of [S] for some I ⊂ [M ] and for any m Sm ⊂ Si for some i ∈ I, D = 1, and for any (j, k), (j(cid:48), k(cid:48)) pair, {s : rj,k,s > 0} and {s : rj(cid:48),k(cid:48),s > 0} either have no intersection or one is a subset of the other. VIII. EVALUATION Secs. III -- VII characterized limits of different types of crowdsourcing systems, proposed efficient policies for optimal centralized allocation and designed decentralized schemes with provable bounds on backlog while giving customers freedom of choice. This section complements theoretical results by studying real data from Samasource, a non-profit crowdsourcing company and realistic Monte Carlo simulations. We study performance of simplified (in implementation and computation) versions of proposed decentralized algorithms above. Let us first describe evaluating allocation using real data. The dataset contains 9.3M tasks and each belongs to a specific project. Some projects are regarded as real-time which means they have higher priority. The overall number of tasks that belong to the real-time projects is about 4.2M. Each task comprises 1 or 2 steps which in turn comprises a single substep. Some tasks have strict step ordering, i.e., the previous step must be completed before the next could be scheduled. Average substep working time requirement is 340 sec. From the data, we calculate 15 Fig. 1. CDF of tasks turn-around time (TAT) using real dataset. Current allocation on the platform "current" vs our algorithm "STEP FLEX". (a) (b) Fig. 2. Performance of our STEP FLEX algorithm on real data, as a function of number of workers. (a) Tasks turn-around time (TAT). (b) Average backlog (number of unallocated steps in the system). the turn-around time (TAT) for each task, i.e., the time since the task arrived to the system until the time its last step was completed. The cdfs of TAT for all projects and for real-time projects only are given in Fig. 1. SamaHub, the platform of Samasource considers both agents and steps to be flexible. We implement a simplified version of the relevant decentralized algorithm, called STEP FLEX, where we prioritize the steps with higher precedence to choose agents greedily with random tie-breaking. To compare current allocation on SamaHub with our approach, we use real data as input to STEP FLEX. Since we lack exact knowledge of worker availability, we make the following assumption in consultation with Samasource. 0.00.51.01.52.02.5TAT [sec]1e50.20.30.40.50.60.70.80.91.0Fraction of tasksall projects, current, avg=6.3e+04all projects, step_flex, avg=9.6e+03rt projects, current, avg=4.5e+04rt projects, step_flex, avg=5.4e+030.40.60.81.01.21.41.6Num of workers1e32000400060008000100001200014000TAT [sec]all projectsrt projects0.40.60.81.01.21.41.6Num of workers1e38001000120014001600180020002200Average backlog [number of unscheduled steps]16 The number of active workers in the system is 625, evenly distributed across four time zones: −4, 0, 3, 5.5, where each worker works every day from 9am to 5pm. Each worker possesses the skills required for any substep in the dataset. Fig. 1 compares the cdf of TAT of our approach STEP FLEX (simulated with the data as input) with currently deployed scheme. Our algorithm substantially outperforms current scheme: average TAT for all projects is ×6.5 better and more than ×8 better for real-time projects. This improvement is also influenced by our implementation, which is not restricted by the currently-practiced organizational structure. Fig. 2 shows how STEP FLEX performs as a function of number of workers. As the number of workers grows, TAT decreases (see Fig. 2(a)). The benefit of adding more workers can be seen even more clearly when analyzing backlog, i.e., the average number of steps that entered the system but not yet scheduled, see Fig. 2(b). We also evaluate our algorithms on synthetic data, considering flexible agents and flexible steps, and flexible agents and inflexible steps. Algorithm STEP FLEX is used for the first system and a simplified version of the Restricted Greedy scheme, STEP INFLEX, where we prioritize steps with higher skill requirements and allocate among them greedily is used for the second. We also consider a scenario in between flexible and inflexible steps, where each substep is allocated to a single agent, but different substeps of a step can be allocated to different agents. For this, we develop STEP SEMIFLEX where steps allocate themselves greedily while ensuring a substep gets all service from an agent. We expected STEP FLEX to outperform STEP INFLEX, but we found somewhat surprisingly that STEP FLEX and STEP SEMIFLEX perform very similarly. The first set of generated data has tasks with up to three steps in each and with strict ordering. Each step comprises one to three random substeps out of five possible types. Working time requirement for each substep is uniformly distributed between 60 and 600 sec. Each worker in the system has daily availability from 9am to 5pm, evenly distributed across four time zones: −4, 0, 3, 5.5. A worker possesses a random set of skills that enables her to work on up to three (out of five) substep types. For each of our three algorithms we compare three metrics: TAT, backlog queue, and worker utilization. The experiment simulated a single run over a timespan of 40 days. Fig. 3 shows algorithms STEP FLEX and STEP SEMIFLEX outperform STEP INFLEX for both cases: 500 workers in the system and 700 workers. When the load on the system is 150 tasks/hour and the number of workers is 500, algorithm STEP INFLEX is substantially worse since it becomes unstable for this load. Notice that STEP FLEX and STEP SEMIFLEX perform very similarly, which can be explained by relatively short substep work time requirement (in which case splitting becomes a rare event). Also note that worker utilization of STEP INFLEX is not much worse than of the other algorithms. This can be explained by the long backlog queue of STEP INFLEX. Though it is harder for STEP INFLEX to find a worker capable of working on the whole step, when the backlog becomes large, the probability that a given worker will be assigned to some whole step grows. The last set of results uses the same synthetic data as before, but the working time requirement for each substep is now uniformly distributed between 600 and 6000 sec. Fig. 4 shows a slight advantage of STEP FLEX over STEP SEMIFLEX. Due to the longer working time requirements per substep, cases in which a substep may be split to improve allocation are more probable. In this scenario, the disadvantage of STEP INFLEX is more obvious: for a load of 50 tasks/hour and 1200 workers, its TAT and backlog are very large and unstable. 17 (a) (b) (c) Fig. 3. Performance of our algorithms on synthetic data with short sub-steps (60 − 600 sec), as a function of load. (a) Tasks turn-around time (TAT). (b) Average backlog (number of unscheduled steps in the system). (c) Workers utilization. 50 [tasks/hour]100 [tasks/hour]150 [tasks/hour]0200040006000800010000120001400016000TAT [sec]42872step_flex, 500 wstep_semiflex, 500 wstep_inflex, 500 wstep_flex, 700 wstep_semiflex, 700 wstep_inflex, 700 w50 [tasks/hour]100 [tasks/hour]150 [tasks/hour]0100200300400500600700800Average backlog [number of unscheduled steps]3163step_flex, 500 wstep_semiflex, 500 wstep_inflex, 500 wstep_flex, 700 wstep_semiflex, 700 wstep_inflex, 700 w50 [tasks/hour]100 [tasks/hour]150 [tasks/hour]0510152025303540Workers utilization [%]step_flex, 500 wstep_semiflex, 500 wstep_inflex, 500 wstep_flex, 700 wstep_semiflex, 700 wstep_inflex, 700 w18 (a) (b) (c) Fig. 4. Performance of our algorithms on synthetic data with long sub-steps (600 − 6000 sec), as a function of load. (a) Tasks turn-around time (TAT). (b) Average backlog (number of unscheduled steps in the system). (c) Workers utilization. 20 [tasks/hour]50 [tasks/hour]80 [tasks/hour]0500010000150002000025000300003500040000TAT [sec]5405815165877517step_flex, 500 wstep_semiflex, 500 wstep_inflex, 500 wstep_flex, 700 wstep_semiflex, 700 wstep_inflex, 700 w20 [tasks/hour]50 [tasks/hour]80 [tasks/hour]0200400600800Average backlog [number of unscheduled steps]114568622977step_flex, 500 wstep_semiflex, 500 wstep_inflex, 500 wstep_flex, 700 wstep_semiflex, 700 wstep_inflex, 700 w20 [tasks/hour]50 [tasks/hour]80 [tasks/hour]0102030405060708090Workers utilization [%]step_flex, 500 wstep_semiflex, 500 wstep_inflex, 500 wstep_flex, 700 wstep_semiflex, 700 wstep_inflex, 700 w19 To summarize, our approach substantially outperforms Samasource's current allocation scheme. While STEP FLEX achieves best performance in terms of TAT and backlog, STEP SEMIFLEX may be a good alternative. Its performance is almost the same but does not require splitting substeps among different workers, and is computationally lighter. IX. CONCLUSION Inspired by skilled crowdsourcing systems, we have developed new algorithms for allocating tasks to agents while handling novel system properties such as vector-valued service requirements, precedence and flexibility constraints, random and time-varying resource availability, large system size, need for simple decentralized schemes requiring minimal actions from the platform provider, and the freedom of customers to choose agents without compromising system performance. We have provided capacity regions, asymptotic performance guarantees for decentralized algorithms, and demonstration of efficacy in practical regimes, via large-scale data from a non-profit crowdsourcing company. REFERENCES [1] A. Chatterjee, M. Borokhovich, L. R. Varshney, and S. Vishwananth, "Efficient and flexible crowdsourcing of specialized tasks with precedence constraints," in Proc. 2016 IEEE INFOCOM, Apr. 2016 (to appear). [2] A. Cuenin, "Each of the top 25 best global brands has used crowdsourcing," Jun. 2015. [Online]. Available: http: //www.crowdsourcing.org/editorial/each-of-the-top-25-best-global-brands-has-used-crowdsourcing/50145 [3] D. Tapscott and A. D. Williams, Wikinomics: How Mass Collaboration Changes Everything, expanded ed. New York: Portfolio Penguin, 2006. [4] F. Gino and B. R. Staats, "The microwork solution," Harvard Bus. Rev., vol. 90, no. 12, pp. 92 -- 96, Dec. 2012. [5] A. Marcus and A. Parameswaran, "Crowdsourced data management: Industry and academic perspectives," Foundations and Trends in Databases, vol. 6, no. 1-2, pp. 1 -- 161, Dec. 2015. [6] T. W. Malone, R. Laubacher, and C. Dellarocas, "The collective intelligence genome," MIT Sloan Manage. Rev., vol. 51, no. 3, pp. 21 -- 31, Spring 2010. [7] K. J. Boudreau and K. R. Lakhani, "Using the crowd as an innovation partner," Harvard Bus. Rev., vol. 91, no. 4, pp. 60 -- 69, Apr. 2013. [8] S. Dustdar and M. Gaedke, "The social routing principle," IEEE Internet Comput., vol. 15, no. 4, pp. 80 -- 83, July-Aug. 2011. [9] D. DiPalantino and M. Vojnovi´c, "Crowdsourcing and all-pay auctions," in Proc. 10th ACM Conf. Electron. Commer. (EC'09), Jul. 2009, pp. 119 -- 128. [10] L. R. Varshney, S. Agarwal, Y.-M. Chee, R. R. Sindhgatta, D. V. Oppenheim, J. Lee, and K. Ratakonda, "Cognitive coordination of global service delivery," arXiv:1406.0215v1 [cs.OH]., Jun. 2014. [11] J. Kleinberg and ´E. Tardos, Algorithm Design. Addison-Wesley, 2005. [12] R. Srikant and L. Ying, Communication Networks: An Optimization, Control and Stochastic Networks Perspective. Cambridge University Press, 2014. [13] M. L. Pinedo, Scheduling: Theory, Algorithms, and Systems. Springer, 2012. [14] G. Pang and A. L. Stolyar, "A service system with on-demand agent invitations," Queueing Systems, Nov. 2015. [15] A. Chatterjee, L. R. Varshney, and S. Vishwananth, "Work capacity of freelance markets: Fundamental limits and decentralized schemes," in Proc. 2015 IEEE INFOCOM, Apr. 2015, pp. 1769 -- 1777. [16] F. A. Chudak and D. B. Shmoys, "Approximation algorithms for precedence-constrained scheduling problems on parallel machines that run at different speeds," J. Algorithms, vol. 30, no. 2, pp. 323 -- 343, Feb. 1999. [17] R. Pedarsani, "Robust scheduling for queueing networks," Ph.D. dissertation, University of California, Berkeley, Berkeley, CA, 2015. [18] L. Tassiulas and A. Ephremides, "Stability properties of constrained queueing systems and scheduling policies for maximum throughput in multihop radio networks," IEEE Trans. Autom. Control, vol. 37, no. 12, pp. 1936 -- 1948, Dec. 1992. 20 [19] M. J. Neely, Stochastic Network Optimization with Application to Communication and Queueing Systems. Morgan & Claypool, 2010. [20] H. Kellerer, U. Pferschy, and D. Pisinger, Knapsack Problems. Springer, 2004. [21] D. A. G. Pritchard, "Linear programming tools and approximation algorithms for combinatorial optimization," Ph.D. dissertation, University of Waterloo, 2009. [22] S. Bubeck and N. Cesa-Bianchi, "Regret analysis of stochastic and nonstochastic multi-armed bandit problems," Found. Trends Mach. Learn., vol. 5, no. 1, pp. 1 -- 122, Dec. 2012. [23] M. Vukovic and O. Stewart, "Collective intelligence applications in IT services business," in Proc. IEEE 9th Int. Conf. Services Comput. (SCC), Jun. 2012, pp. 486 -- 493. [24] A. Chatterjee, L. R. Varshney, and S. Vishwananth, "Work capacity of freelance markets: Fundamental limits and decentralized schemes," arXiv:1508.00023 [cs.MA], Jul. 2015. [25] V. S. Borkar, Stochastic Approximation: A Dynamical Systems Viewpoint. Cambridge University Press, 2008. 21 In this section we present proofs of the main results. APPENDIX A PROOFS A. Proof of Theorem 1 Here we only prove that any λ outside the closure of C cannot be stabilized by any policy. To prove achievability, it is sufficient to show that there exists a policy that stabilizes any λ in the interior of C. Hence, it is sufficient to prove Thm. 2, which we do later. This proof consists of the following steps. We first compare two systems, the original system in question and another in which there is no precedence constraint among different steps of a job. We claim that on any sample path under any policy for the first system, there exists a policy in the second system so that the total number of incomplete jobs across all job types in the second is a lower bound (sample path-wise) for that in the first. Then we show that the second system cannot be stabilized for a λ outside the closure of C and so the result follows for the first system. Note that the claim regarding the number of incomplete jobs across all types in the second system being a lower bound on the first system follows by considering the same policy for the second system as for the first system. To proceed, consider the second system, for which we denote the number of unallocated steps of type (j, k) at epoch t by Qj,k(t). Now consider the set C. We claim that this set is coordinate convex, i.e., it is a convex set and + then a ∈ C. To prove this claim, we first show that the set C is coordinate convex. if a + ∈ C for some , a ∈ RN + . If a, a(cid:48) ∈ C, then there exist(cid:0)a(u) ∈ Ccvx(u) : u ∈ ZM First we prove that C is a convex subset of RN (cid:1) and + (cid:0)a(cid:48)(u) ∈ Ccvx(u) : u ∈ ZM + (cid:1) such that(cid:88) Γ(u)a(u) = λ, (cid:88) Γ(u)a(cid:48)(u) = λ(cid:48). Thus for any γ ∈ [0, 1], u γa + (1 − γ)a(cid:48) = (cid:88) u Γ(u)(γa(u) + (1 − γ)a(cid:48)(u). Note that Ccvx(u) is convex since it is the convex hull of C(u); hence γa(u) + (1 − γ)a(cid:48)(u) ∈ C(bu), which in turn implies γa + (1 − γ)a(cid:48) ∈ C. u For coordinate convexity note that any a is a Γ(u) combination of some {a(u) ∈ Ccvx(u)} and any a(u) is some convex combination of elements of C(u). Also, from the allocation constraints it is apparent that if a ∈ C(u) then also a(cid:48) ∈ C(u) if a(cid:48) ≤ a. These two imply that for any a ∈ ¯C, if there exists an a(cid:48) ≤ a (component-wise) and a(cid:48) ≥ 0, then a(cid:48) ∈ ¯C. Hence, C is coordinate convex. Note that C = {a : aE ∈ C}. Note that if a ≤ a(cid:48) (coordinate-wise) then the same is true for aE and a(cid:48)E. Also, if a(cid:48)(cid:48) = γa + (1 − γ)a(cid:48) for any γ ∈ [0, 1], then a(cid:48)(cid:48)E = γaE + (1 − γ)a(cid:48)E. This proves that C is coordinate convex subset of RN + . As C is coordinate convex, for any λ outside the closure of C there exists h ∈ RN + s.t. hT λ > supx∈C hT x. Consider the following. Let Q = ( Qj,k), A = (Aj,k(t)) and D(t) = (Dj,k(t)). Note that as jobs in this system 22 do not have precedence constraints, for all k, Aj,k = Aj(t). type, and · + is shorthand for max(·, 0). As x+ is a convex function of x, hT(cid:12)(cid:12)(cid:12) Q(t) + A(t) − ∆(t) (cid:12)(cid:12)(cid:12)+ where ∆(t) is the number of possible departures under the scheme if there were infinite number of steps of each is a convex function of Q(t), A(t), and ∆(t). Thus by Jensen's inequality: (cid:104) E E hT Q(t + 1) (cid:105) hT(cid:16) Q(t) + A(t) − D(t) (cid:17)(cid:105) (cid:104) (cid:12)(cid:12)(cid:12)+(cid:21) (cid:20) hT(cid:12)(cid:12)(cid:12) Q(t) + A(t) − ∆(t) = E = E (cid:12)(cid:12)(cid:12)+(cid:21) (cid:20) hT(cid:12)(cid:12)(cid:12) Q(t) + A(t) − ∆(t) ≥ hT(cid:12)(cid:12)(cid:12)E (cid:105) (cid:104) Q(t) (cid:105) (cid:104) Q(t) (cid:104) (cid:105) ≥ hT E hT Q(t + 1) ≥ hT E E (cid:104) Q(t) (cid:105) (cid:12)(cid:12)(cid:12)+ + E [A(t)] − E [∆(t)] + hT E [A(t)] − hT E [∆(t)] . Note that hT E [∆(t)] ≤ supx∈C hT x and hT E [A(t)] = hT λ, hence, + , > 0. (17) So, we have E[hT Q] to be unbounded (i.e., for any constant B, there exists a t s.t. E[hT Q(t)] > B) under any policy. As h ≥ 0, we have that QT 1 to be unbounded and hence, the system is not stable under any policy. B. Proof of Theorem 2 Before proceeding, we state small lemma that will be useful. Lemma 1. For any x, y, z ≥ 0, (x − y+ + z)2 ≤ x2 + y2 + z2 + 2x(z − y). Proof: (x − y+ + z)2 = (x − y+)2 + z2 + 2zx − y+ ≤ (x − y+)2 + z2 + 2xz ≤ (x − y)2 + z2 + 2xz, = x2 + y2 + z2 + 2x(z − y), where the last inequality follows because (max(0, a))2 ≤ a2. (cid:80) Now for the proof of Thm. 2 The process {Qj,k(t)} is a discrete-time Markov chain on Z j Kj under the centralized scheme. This is because arrival and availability processes are i.i.d. and the centralized allocation at epoch t does not depend on process values before t. We show that for this chain, all closed classes are positive recurrent and that the chain enters one of the closed classes almost surely. Note that this implies that starting with any initial distribution, the Markov + 23 (cid:80) chain reaches a stationary distribution (which may depend on the initial condition). This is in the sense that there exists a d ∈ {1, 2, . . .} (as there may be a closed class which is not aperiodic), and a distribution π on Z j Kj such that Q(td) → π in distribution. + To show stability we need lim supt→∞ E[Qj,k(t)] < ∞, for all (j, k). Towards this, note it is sufficient to show Eπ[(cid:80) j,k Qj,k] is finite, because this implies limt→∞ E[(cid:80) Qj,k(td + τ ) ≤(cid:88) (cid:88) j,k j,k d(cid:88) t(cid:48)=1 [Qj,k(t) + Aj(t(cid:48))]. j,k Qj,k(td)] is finite. Note that for any 1 < τ < d: Since arrivals have finite expectation, lim supt→∞ E[Qj,k(t)] < ∞, for all (j, k). Q(td) → π in distribution and Eπ[(cid:80) Now, it is sufficient to prove that starting with any initial distribution, there exists a d ∈ {1, 2,···} such that j,k Qj,k] is finite. To prove the convergence in distribution we use a variation of the Foster-Lyapunov theorem presented in [18]. When Tj is a directed rooted tree, we have to consider a Lyapunov function: (cid:88) (cid:88) L(Q) = lj,kQ2 j,k, where lj,k is the number of leaves in the subtree of Tj rooted at k. j k Before proceeding with the proof of this case, we prove a reordering lemma for the Lyapunov function. Lemma 2. For any allocation {Sj,k}, Qj,k(t)S∗ j,pj (k)(t) − Qj,k(t)Sj,k(t) (cid:17)(cid:33) (cid:16) lj,k (cid:88) k>1 lj,1Qj,1(Aj,1 − Sj,1) + Kj(cid:88) (cid:88) j k=1 r∈cj (k) lj,r(Qj,k(t) − Qj,r(t))(Aj,1 − Sj,k). (cid:88) j (cid:32) (cid:88) = (cid:16) (cid:88) (18) (19) Proof: First we claim that for any j, (cid:17) j,pj (k)(t) − Qj,k(t)Sj,k(t) Qj,k(t)S∗ Kj(cid:88) (cid:88) k>1 lj,k This can be seen by comparing coefficients of Qj,k, k > 1 on both sides of the expression. Note that(cid:80) lj,k. Also, note that in the sum in the right side, Qj,k appears twice, once in the sum −(cid:80) Qj,r(t))Sj,k where the coefficient is −Sj,k and again in the sum −(cid:80) r∈cj (k) lj,r = r∈cj (k) lj,r(Qj,k(t) − r∈cj (pj (k)) lj,r(Qj,pj (k)(t) − Qj,r(t))Sj,pj (k) r∈cj (k) k=2 = lj,r(Qj,k(t) − Qj,r(t))Sj,k where the coefficient is Sj,pj (k). This implies that for any j, (cid:16) lj,k (cid:88) k>1 (cid:17) j,pj (k)(t) − Qj,k(t)Sj,k(t) Qj,k(t)S∗ lj,r(Qj,k(t) − Qj,r(t))Sj,k. (20) − lj,1Qj,1Sj,1 + Kj(cid:88) (cid:88) k=1 r∈cj (k) = − Note that since lj,k =(cid:80) r∈cj (k) lj,r, lj,1Qj,1 = (cid:88) r∈cj (1) 24 (cid:88) r∈cj (1) lj,rQj,r. lj,r(Qj,1 − Qj,r) + Again applying the same restructuring of the terms for the subtrees rooted at r, we eventually obtain: lj,1Qj,1 = lj,r(Qj,k − Qj,r). The result follows by combining this (multiplied by Aj,1) with (20). Now for the main proof. E[L(Q(t + 1))Q(t)] = E (Qj,k(t) − Dj,k(t) + Aj,k(t))2Q(t)  . Let pj(k) denotes the parent of k in Tj and cj(k) denote the set of children of k in Tj. For all t and k = 1, Aj,1(t) = Aj(t). For k > 1, Aj,k(t) = Dj,pj (k)(t). Note that S∗ j,k ≥ Dj,k for all t. So, (cid:88) r∈cj (k) j,k (cid:88)  E j,k (cid:88) (cid:88) (cid:88) j,k = E ≤ E (cid:32) lj,k(Qj,k(t) − Dj,k(t) + Aj,k(t))2Q(t)  j,k(t)+ + Aj,k(t))2Q(t) (cid:88) j,1(t)+ + Aj,1(t))2 + lj,k(Qj,k(t) − S∗ lj,1(Qj,1(t) − S∗ j By Lem. 1, k>1 ≤ C2 + 2E lj,1Qj,1(Aj,1 − S∗ j,1) + lj,kQj,k(S∗ j,pj (k) − S∗ j,k) Q(t) Eq. (21) follows because arrival processes have bounded second moments and are i.i.d. and the fact (S∗ j,k(t))2 ≥ 0 j k>1 lj,k(Qj,k(t) − S∗ j,k(t)+ + S∗ j,pj (k)(t))2 = E j,1(t) + (S∗ j,1(t))2 + Q2 j,1 + 2Qj,1(Aj,1 − S∗ j,1) lj,k((S∗ j,k(t))2 + (S∗ j,pj (k)(t))2 + Q2 j,k + 2Qj,k(S∗ j,pj (k) − S∗ j,k)) ≤ C1 + 2E lj,1(S∗ j,k(t))2Q(t) + 2E lj,1Qj,1(Aj,1 − S∗ j,1) + lj,kQj,k(S∗ j,pj (k) − S∗ j,k) Q(t) E[L(Q(t + 1))Q(t)] (cid:20)(cid:88) (cid:88) j + k>1 lj,1 (cid:0)A2 (cid:88) (cid:88) (cid:32) (cid:88) j,k j (cid:32)  (cid:88) k>1 (cid:88) (cid:33)  . Q(t) (21) (22) (cid:33) (cid:21) Q(t)  (cid:33) (cid:33)  (so, over-counting them gives an upper bound). Eq. (22) is due to the following (where K = maxj Kj): (cid:88) E  ≤ KE 2 S∗ j,k(t) Q(t) lj,k(S∗ j,k(t))2Q(t) j,k ≤ K ≤ K ≤ K j,k  (cid:88) maxj,k((cid:80) maxj,k((cid:80) maxm((cid:80) maxj,k((cid:80) 1 1 s rj,k,s)2 E s rj,k,s)2 E s rm,s)2 s rj,k,s)2 E   (cid:88) (cid:32)(cid:88) (cid:34)(cid:88) m,s j,k,s U 2 m m (cid:35) 2 (cid:33)2   S∗ j,k(t)rj,k,s Q(t) Um(t)hm,s Q(t) 25 (23) < ∞. Eq. (23) comes from the task allocation constraint and the last step follows as availability processes have bounded second moment. Consider the last term of (22), as C2 plus this is the upper bound for Lyapunov drift E[L(Q(t + 1))Q(t)] − L(Q(t)). Then by Lem. 2 and the fact that {Aj(t)} are i.i.d., (cid:88) (cid:32) Kj(cid:88) (cid:88) j E = j k=1 r∈cj (k) (cid:88) Kj(cid:88) (cid:88) j k=1 r∈cj (k) (cid:88) Kj(cid:88) (cid:88) j k=1 r∈cj (k) (cid:88) (cid:16) (cid:17)(cid:33) j,k(t) Q(t)  lj,1Qj,1(Aj,1 − S∗ j,1) + lj,k Qj,k(t)S∗ (cid:88) k lj,r(Qj,k(t) − Qj,r(t))λj − E j,pj (k)(t) − Qj,k(t)S∗ Kj(cid:88) (cid:88) (cid:88) j k=1 r∈cj (k) lj,r(Qj,k(t) − Qj,r(t))S∗ j,kQ(t)  . Note that for any Q(t) and U(t): lj,r(Qj,k(t) − Qj,r(t))S∗ Note that in the optimal allocation {S∗ j,k}, S∗ just setting them to 0 gives a better allocation). So, lj,r(Qj,k(t) − Qj,r(t))S∗ a∈C(U(t)) (cid:88) j,k ≥ max lj,r(Qj,k(t) − Qj,r(t))aj,k. (cid:88) Kj(cid:88) j,k ≥ 0 only if (cid:80) r∈cj (k) lj,r(Qj,k(t) − Qj,r(t))) ≥ 0 (otherwise, Kj(cid:88) (cid:88) r∈cj (k) k=1 lj,r(Qj,k − Qj,r)+aj,k. (cid:88) j j,k ≥ max a∈C(U(t)) j k=1 r∈cj (k) 26 Hence, E = E ≥ sup a∈C j j k=1 k=1 r∈cj (k) r∈cj (k) (cid:88) (cid:88) Kj(cid:88) (cid:88) (cid:88) Kj(cid:88) (cid:88) (cid:88) Kj(cid:88) (cid:88) Kj(cid:88) (cid:88) (cid:88) (cid:88) (cid:88) Kj(cid:88) aj,k=aj 1≤k≤Kj sup a∈C: k=1 k=1 E E E j j = sup a∈C ≥ E ≥ r∈cj (k) r∈cj (k) because λ + ∈ C. j k=1 lj,r(Qj,k(t) − Qj,r(t))S∗ j,kQ(t) lj,rQj,k − Qj,r+S∗ j,kQ(t) lj,r(Qj,k − Qj,r)+aj,kQ(t) r∈cj (k) Kj(cid:88) (cid:88) j k=1 r∈cj (k) lj,rQj,k − Qj,r+aj,kQ(t)       + lj,rQj,k − Qj,r+ajQ(t) lj,rQj,k − Qj,r+λjQ(t) (cid:88) Kj(cid:88) j k=1 (cid:88) r∈cj (k) lj,r(Qj,k − Qj,r)+, As, we have (cid:88) Kj(cid:88) j k=1 (cid:88) r∈cj (k) lj,rQj,k − Qj,r+λj ≥(cid:88) j E[L(Q(t + 1)) − L(Q(t))Q(t)] ≤ C2 − Note that for {Qj,k} sufficiently large,(cid:80) (cid:80)Kj k=1 (cid:80) Kj(cid:88) (cid:88) k=1 r∈cj (k) (cid:88) (cid:88) Kj(cid:88) j k=1 r∈cj (k) lj,r(Qj,k(t) − Qj,r(t))λj lj,r(Qj,k − Qj,r)+. r∈cj (k) lj,r(Qj,k(t)−Qj,r(t))+ is also large. This is because if {Qj,k} is larger than B (in max-norm) then there exists a j such that maxk Qj,k > B. Now consider two cases, if Qj,Lj ≥ B 2 which can be made strictly negative by choosing B appropriately. 2 for some leaf node Lj then we have the drift ≤ C2 − B j If Qj,Lj < B 2 for all leaf nodes, then there exists a k0 such that Qj,k0 > B. Note the following for the set of nodes Tk0 in the subtree rooted at k0 and Lk0 being the leaves of Tk0: Hence, we have that(cid:80) k∈Tk0 (cid:88) (cid:88) (cid:80) r∈cj (k) r∈cj (k) lj,r(Qj,k − Qj,r) ≥ lj,k0 lj,r(Qj,k − Qj,r) = k∈Tk0 (cid:88) l∈Lk0 B 2 ≥ B 2 . (Qj,k0 − Qj,l). (24) Thus we show strictly negative drift for sufficiently large {Qj,k} and the drift is bounded by C2 < ∞. Hence, by the Foster-Lyapunov theorem in [18] we have that for any initial distribution, there exists a d ∈ {1, 2, . . .} such that Q(td) → π in distribution. To prove finite expectation we consider the following. E[L(Q(t + 1)) − L(Q(t))Q(t)] ≤ C2 − lj,r(Qj,k(t) − Qj,r(t))+ 27 .  . (cid:19) lj,r(Qj,k(t) − Qj,r(t))+ r∈cj (k) j k=1 (cid:88) (cid:88) Kj(cid:88) (cid:88) Kj(cid:88) (cid:88)  ≤ 1 (cid:18) r∈cj (k) C2 − 1 T k=1 j lj,r(Qj,k(t) − Qj,r(t))+  < C3,  < C3.  < C3. lj,r(Qj,k(td) − Qj,r(td))+ lj,r(Qj,k(td) − Qj,r(td))+ r∈c(k(cid:48)) (cid:88) (cid:88) (cid:104)(cid:80) j r∈c(k(cid:48)) lj,r(Qj,k(cid:48) − Qj,r) lj,r(Qj,k(cid:48) − Qj,r)+. (cid:80)Kj (cid:105) lj,rQj,k(t) − Qj,r(t)+ E[L(Q(T + 1))] + E[L(Q(0))] which implies that E[L(Q(t + 1)) − L(Q(t))] ≤ C2 − E Summing both sides from 0 to T , we get: (cid:88) Kj(cid:88) T(cid:88) E t=1 j k=1 1 T As E[L(Q(0))] finite, for any initial condition we have for all T . As all terms are positive, for any d ∈ {1, 2, . . .}, (cid:88) T(cid:88) r∈cj (k) 1 T t=1 E (cid:88) r∈cj (k) j k=1 Kj(cid:88) (cid:88) (cid:88) Kj(cid:88) (cid:88) Kj(cid:88) j lim T→∞ (cid:2)Qj,Lj Eπ d T E t=1 k=1 r∈cj (k) (cid:88) T(cid:88) (cid:88) (cid:3) < C3 for any leaf node lj. (cid:88) (cid:88) (cid:88) ≤(cid:88) Qj,lj + r∈cj (k) k(cid:48)∈Tk Qj,lj + k=1 lj j lj,kQj,k = This proves that Eπ By (24) we have that for any k ∈ Tj, By the ergodicity of a Markov chain in a positive recurrent class this implies that k(cid:48)∈Tk Hence, it follows that Eπ [Qj,k] < ∞. This implies that Eπ lj k=1 Qj,k < ∞ and so the proof is complete. C. Proof of Theorem 3 In deriving (22) we did not use any property of the allocation {S∗ j,k} other than the fact that it has to satisfy the step allocation constraint. Hence, this upper bound for Lyapunov drift is valid for any arbitrary feasible allocation {Sj,k}. Hence, under the LP-relaxation base allocation {SR j,k} by Lem. 2 we have: E[L(Q(t + 1)) − L(Q(t))Q(t)] ≤ C2 + 2E lj,r (Qj,k − Qj,r) SR j,kQ(t) (25)  . (cid:88) (cid:88) j r∈cj (k) 28 Note that for the optimum of the problem in (7), {SR j,k(t)}, the following is true. (cid:88) j r∈cj (k) k=1 Kj(cid:88) (cid:88) ≤(cid:88) = j (cid:88) Kj(cid:88) Kj(cid:88) k=1 j k=1 (cid:88) (cid:88) r∈cj (k) r∈cj (k) lj,r(Qj,k(t) − Qj,r(t))S∗ j,k lj,r(Qj,k(t) − Qj,r(t))+S∗ j,k lj,r(Qj,k(t) − Qj,r(t))+ Sj,k This is because (7) solves a relaxed problem and the optimal allocation has S∗ Qj,r(t)). As SR (cid:88) j,k = (cid:98) Sj,k(cid:99) we have that Kj(cid:88) (cid:88) lj,r(Qj,k(t) − Qj,r(t))+S∗ j,k ≤(cid:88) Kj(cid:88) (cid:88) j k=1 r∈cj (k) j k=1 r∈cj (k) j,k = Sj,k = 0 for(cid:80) r∈cj (k) lj,r(Qj,k(t)− lj,r(Qj,k(t) − Qj,r(t))+(SR j,k + 1) Hence, for any λ such that λ + 1(1 + ) ∈ C using the same proof as above we can show that the system is stable. D. Proof of Theorem 4 This proof has the following structure. As the total number of incomplete jobs is equal to the total number of unallocated steps, we first show that the total number of unallocated steps at depth 0 (i.e., at the root of each Tj) across all types have the desirable property. Then we show that this property propagates. Proof for Depth-0 Steps: This part is same as the proof [24] of performance guarantee of GreedyJob algorithm in [15]. We present it here for the sake of completeness. Consider the different types of unallocated steps at depth 0. These are given by {Qj,1(t) : j ∈ [N ]}. Consider the following processes: for each s ∈ [S], Qs j:rj,1,s>0 Qj,1rj,1,s which represent the number 1(t) =(cid:80) 1. So, if we can upper bound Q1, 1. Hence, in turn we get a bound for {Qj,1(t)} (since min{rj,k,s > 0} = Θ(1) s Qs of unserved hours of skills s for all steps at depth 0. by the assumption that {rj,k,s} do not scale with the system size). We now construct another process Q1 such that it dominates the process(cid:80) then the same bound applies for(cid:80) time t,(cid:80) At time t,(cid:80) j Aj,1(t)rj,1,s amount of s skill-hour is brought to add to Qs m Um(t)hm,s s-skill hour of service is brought by the agents. on the available agent hours. s Qs To construct a suitable Q1, we make the following observation about the dynamics of Qs 1 and {Qj,1}. At each 1. Also, this queue gets some service depending For a step to be allocated, all of its tasks must find an allocation. Hence, for a step in type j-job to find an allocation it must get rj,1,s hours of service from each skill s. Thus at any time t any skill s queue gets a service of at least (cid:88) m min s∈[S] Umhm,s − ¯r, where ¯r = max{rj,k,s}, due to the following. For each skill,(cid:80) s Umhm,s hours are available. Note that a step can be allocated if all its tasks find allocations, the converse of which is also true. That is, if all tasks of a step find (cid:80) allocation, then the step can be allocated. As mins∈[S] s Umhm,s hours of service are brought by the agents for s Umhm,s − ¯r of s-skill hours are served (because a maximum of ¯r can be wasted, each skill, at least mins∈[S] as no task is of size more than ¯r). Note that as depth d steps have priority in Priority Greedy algorithm over steps at depth ≥ d + 1, they do not have to share resource with higher-depth steps. So at depth d, mins∈[S] is available for service to steps at depth ≤ d. s Umhm,s (cid:80) (cid:80) 29 Also, note that the amount of required service brought to the queue Qs 1 at time t is upper-bounded by Consider a process Qs 1 with evolution: max s∈[S] Qs 1(t + 1) = max( Qs 1(t) + max s∈[S] Aj,1(t)rj,1,s. Aj,1(t)rj,1,s − min s∈[S] (cid:88) m Umhm,s + ¯r, 0). Note that given Qs 1(t0) ≥ Qs 1(t0) at some t0, the same holds true for all t ≥ t0. This is because for x, a, b ≥ 0 and x(cid:48), a(cid:48), b(cid:48) ≥ 0, with x ≥ x(cid:48), a ≥ a(cid:48) and b ≤ b(cid:48) max(x + a − b, 0) ≥ max(x(cid:48) + a(cid:48) − b(cid:48), 0), (cid:88) (cid:88) j j and so the monotonicity propagates over time. 1, it is sufficient to bound(cid:80) s s Qs To bound(cid:80) which bounds(cid:80) us consider 1. s Qs From the evolution, Q1(t + 1) = max( Q1(t) + S max s∈[S] Qs 1(t). Note that each Qs 1 has exactly the same evolution, so let Q1 := S Q1 1, (cid:88) j Aj,1(t)rj,1,s − S min s∈[S] (cid:88) m Umhm,s + ¯r, 0), and we can write the Loynes' construction for this process which has the same distribution as the following process (and for simplicity we use the same notation, as we are interested in the distribution): (cid:88) (cid:88) Let us define Xs(t) and Ys(t) as follows: Xs(t) :=(cid:80) τ≤t≤0 assuming that the process started at −∞. 1(0) = max τ≤0 (S max s∈[S] Q1 j Aj,1(t)rj,1,s − S min s∈[S] Umhm,s + ¯r), m j Aj,1(t)rj,1,s and Ys(t) :=(cid:80) m Umhm,s. Then, (cid:88) (cid:88) τ≤t≤0 Q1 1(0) = max τ≤0 S(max s Xs(t) − min s Ys(t) + ¯r). Now, for any θ > 0: (cid:88) j P( (cid:88) Qj,1 > ¯rq) ≤ P( Qs 1 > q) s ≤ P( Q1(0) > q) = P(θ Q1(0) > θq) Now, E[exp(θ Q1(0))] = E = P(exp(θ Q1(0)) > exp(θq)) ≤ E[exp(−θq)]E[exp(θ Q1(0))]. max exp(θS (cid:88) exp(θS (cid:88) ≤(cid:88) τ≤0 E τ≤t≤0 τ≤0 τ≤t≤0 (max s Xs(t) − min s Ys(t) + ¯r) (max s Xs(t) − min s Ys(t) + ¯r)) )   , 30 (26) (27) where inequality (27) follows because for any random variables {Zj}, exp(θZj) are positive random variables and the sum of positive values is more than their maximum. Next, we bound the term within the summation over τ ≤ 0 in (27). E[exp(θS (max s Xs(t) − min s E[exp(θ(max s Xs(t) − min s Ys(t) + ¯r)))], (28) Ys(t) + ¯r))] ≤ (cid:89) τ≤t≤0 (cid:88) τ≤t≤0 which follows because Xs(t), Ys(t) are i.i.d. over time. Next we bound the term within the product in (28), (cid:104) eθS(maxs Xs(t)−mins Ys(t)+¯r)(cid:105) ≤(cid:88) (cid:104) eθS(Xs(t)−Ys(cid:48) (t)+¯r)(cid:105) , E (29) E where this follows for the same reason as (27). s,s(cid:48) The following lemma regarding an outer bound to the capacity region will be useful later. Lemma 3. Let CO I,F = {λ :(cid:80) j,k λjrj,k,s ≤(cid:80) m umhm,s}. Then CI,F ⊂ CO I,F Proof. Consider any λ ∈ CI,F . Then by definition of CI,F , λE ∈ C and there exists c(u) ∈ convCI,F (u) such that Also, note that for each c(u) ∈ convCI,F (u), there exists an {αv(u) ≥ 0, 1 ≤ v ≤ Vu :(cid:80)Vu Γ(u)c(u). u v=1 αk = 1} and λE ≤(cid:88) av(u) ∈ CI,F (u), 1 ≤ v ≤ Vu such that Note that as av(u) ∈ CI,F (u) so by allocation constraint ak(u)αv(u) = c(u). Vu(cid:88) av,j,k(u)rj,k,s ≤(cid:88) v=1 (cid:88) j,k m umhm,s, 31 where av,j,k are the number of (j, k) steps that have been allocated under av(u) allocation. This in turn implies that: j,k (cid:88) av,j,k(u)αv(u)rj,k,s ≤(cid:88) =⇒ (cid:88) cj,krj,k,s ≤(cid:88) λjrj,k,s ≤(cid:88) =⇒ (cid:88) j,k m m umhm,s. umhm,s j,k m umhm,s Hence, λ ∈ CI,F implies that λ ∈ CO I,F . Let Aj(θ) = E(cid:2)eθAj (t)(cid:3) and Um(θ) = E(cid:2)eθUm(t)(cid:3) for j ∈ [N ] and m ∈ [M ]. For θ ∈ R, then, (cid:104) eθ(Xs(t)−Ys(cid:48) (t))(cid:105) E E E E (cid:104) = E = E i Ui(t)hi,s eθAj (t)rj,1,s e−θYs(cid:48) (t)(cid:105) eθXs(t)(cid:105) (cid:104) j Aj (t)rj,1,s+¯r(cid:105) (cid:104) (cid:104) (cid:105) eθ(cid:80) e−θ(cid:80) (cid:104) (cid:105)(cid:89) (cid:104) e−θUi(t)hi,s(cid:48)(cid:105) = eθ¯r(cid:89) = eθ¯r(cid:89) (cid:89) θ¯r + (cid:88) I,F ,(cid:80) j λjrj,1,s < α(cid:80) log Aj(θrj,1,s) + Ui(−θhi,s(cid:48)) Aj(θrj,1,s) = exp E j j j i i (cid:88)  . m µmhm,s and by assumption (cid:80) log Ui(−θhi,s(cid:48)) i (30) (31) (cid:80) Note that as λ ∈ αC, by the definition of CO m µmhm,s(cid:48) ≤ subpoly(N ) which is used in the following. First consider the Gaussian-dominated case. Since the process variance is no more than the mean and the moment m µmhm,s− generating function of the variance is upper-bounded by that of a zero-mean Gaussian: log Aj(θrj,s) ≤ λjθrj,1,s + λj log Ui(−θhi,s) ≤ −µjθhi,s + µj (θrj,1,s)2 2 (θhi,s)2 2 . Note that for any two functions k1x2 and k2x, limx→0 k2x/k1x2 = ∞, and hence for any ∈ (0, 1) there exists i,s > 0, for all i, j, s x∗ > 0 such that for all x < x∗, k1x2/k2x < . Hence for any ∈ (0, 1), there exist θ∗ such that for all θ < θ∗ = mini,j,s(θ∗ j,s, θ∗ j,s, θ∗ i,s), log Aj(θrj,1,s) ≤ λjθ∗rj,1,s(1 + ) log Ui(−θhi,s) ≤ −µiθhi,s(1 − ). Note that since N, S, and M are finite and θ∗ (32) i,s > 0, for all i, j, s, θ∗ > 0. Moreover, note that θ∗ does not depend on λ, µ since the ratio of the linear and quadratic terms in the log moment generating functions are independent of λ and µ. j,s, θ∗ 32 As eθ − 1 =(cid:80)∞ k=1 θk k! , for the Poisson-dominated case we have: log Aj(θrj,s) ≤ λj log Ui(−θhi,s) ≤ µj (cid:88) (cid:88) k k (θrj,1,s)k k! (−θhi,s)k . k! E (cid:104) i µihi,s − λjθ∗rj,1,s(1 + ) −(cid:88) (cid:33)(cid:33) Again, by the same argument, we can have a θ∗ for which (31) and (32) are satisfied. As (cid:80) (cid:80) i µihi,s(cid:48) = o(N δ),∀δ > 0, and(cid:80)  θ∗¯r + i µihi,s = Ω(N c), c > 0, for all θ < θ∗ we have: eθ(Xs(t)−Ys(cid:48) (t))(cid:105) ≤ exp (cid:88) (cid:32) (cid:32) ¯r −(cid:88) Since(cid:80) i µihi,s >(cid:80) j(1 − α)λjrj,s and(cid:80) all j, we have ¯r −(cid:80) i µihi,s(α − ) ≤ −γ(cid:80) (cid:104) eθS(Xs(t)−Ys(cid:48) (t))(cid:105) ≤ exp (−θSK(N )) , where K(N ) scales with N no slower than(cid:80) Note (33) follows from the fact λ ∈ (1− α)CO. As > 0 can be chosen arbitrarily small, we can have α− 2 > 0. j λjrj,1,s scales with λ(N ), for sufficiently large λα with λj ≥ λα for i µihi,s(α − ), for some γ > 0. Thus, we have for some θ > 0, µiθhi,s(1 − ) + θ∗o(µiθhi,s) µihi,s(α − 2) ≤ exp (34) (33) i . j i E θ Thus, Hence, from (27), (28), and (29) we have that E s:rj,1,s>0 λj(N ) = Ω(N c), c > 0. (cid:104) eθS(maxs Xs(t)−mins Ys(t))(cid:105) ≤ S2 exp (−θSK(N )) . E[exp(θ∗(cid:88) 1(0))] = E[exp(θ∗S Q1 Qs 1(0))] s = E[exp(θ∗S Q1(0))] ≤(cid:88) S2τ exp(−θ∗SK(N )τ) τ≤0 ≤ c(cid:48), because S2 < exp(θ∗SK(N )) for sufficiently large N. Note that though we proved E[exp(θQ1(t))] < c(cid:48) for t = 0, this holds for any finite t (exactly the same proof). This in turn implies that the number of unallocated steps in depth 0 have bounded exponential moment for some θ > 0. This will be used in the remainder of the proof where we show that the same is true for all depths. Induction over Depths, d to d + 1: Now we show that if the total number of unallocated steps at depth d satisfies E[exp(θQ(0))] < c(cid:48), then the same is true for d + 1. To show the same result for steps at all depths we consider the following process. Let dj(k) j,k:dj (k)≤d+1 Qj,krj,k,s represents the number of unserved hours of be the depth of k in Tj, then Qs skills s for all steps in the system. d+1(t) = (cid:80) (cid:80) Like in the case of the proof for depth 0, we construct process Qs s Qs d+1. Using the same argument as before, at any time t any skill s queue gets a service of at least d+1 such that(cid:80) 33 Qs d+1 dominates the process s Umhm,s − ¯r, min s∈[S] (cid:88) (cid:88) m max s∈[S] j,k:dj (k)≤d+1 Aj,k(t)rj,k,s. and the amount of required service brought to the queue Qs at time t is upper-bounded by Then using the same argument, the process (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)( Qd+1(t) + S max (cid:88) s :=(cid:80) s∈[S] j,k:dj (k)≤d+1 Qd+1(t + 1) = upper-bounds the process(cid:80) s :=(cid:80) Aj,k(t)rj,k,s − min s∈[S] (cid:88) m (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)+ Umhm,s + ¯r s Qs j,k:dj (k)≤d+1 Aj,k(t)rj,k,s and Y (cid:48) d+1. Then we can follow the steps that we followed using Xs and Ys previously. Let X(cid:48) m Umhm,s, respectively. But note that Aj,k for k > 1 is not an external i.i.d. process, rather it is the number of steps of type Aj,pj (k) that were completed. Hence, we cannot follows the exactly same steps. Note that (cid:88) τ≤t≤0 E[exp(θS (max s s(t) − min X(cid:48) s (cid:88) τ≤t≤0 exp(θ max s,s(cid:48)∈[S] (cid:88) (cid:88) τ≤t≤0 τ≤t≤0 (X(cid:48) s(t) − Y (cid:48) s(cid:48)(t) + ¯r))] (X(cid:48) s(t) − Y (cid:48) s(cid:48)(t) + ¯r))] (X(cid:48) s(t) − Y (cid:48) (cid:88) E[exp(θS s(cid:48)(t) + ¯r))]. s,s(cid:48)∈[S] ≤ E[ s (t) + ¯r))] ≤ E[exp(θ Y (cid:48) (cid:88) (cid:88) s(cid:48)(t) + ¯r))] ≤ (cid:88) s,s(cid:48)∈[S] = (cid:88) E[exp(θS s,s(cid:48)∈[S] τ≤t≤0 τ≤t≤0 (X(cid:48) s(t) − Y (cid:48) s(cid:48)(t) + ¯r))]. E[exp(θS (cid:88) Also note that,(cid:88) So, we investigate E[exp(θ(cid:80) (cid:88) max s,s(cid:48)∈[S] τ≤t≤0 τ≤0 E[exp(θ τ≤t≤0 (X(cid:48) s(t) − Y (cid:48) s(t) − Y (cid:48) τ≤t≤0(X(cid:48) s(t) − Y (cid:48) (X(cid:48) (cid:88) (cid:88) (cid:88) ( τ≤t≤0 j,k s(cid:48)(t) + ¯r))]. s(cid:48)(t) + ¯r))] Aj,k(t)rj,k,s −(cid:88) = exp(¯rθ)E[exp(θ (cid:88) Note that (cid:80) and 0 (with a similar interpretation for agents in case of (cid:80) = exp(¯rθ)E[exp(θ( j,k:dj (k)≤d+1 rj,k,s τ≤t≤0 Um(t)hm,s(cid:48)))] m Aj,k(t) −(cid:88) (cid:88) τ≤t≤0 m Um(t)hm,s(cid:48)))]. τ≤t≤0 Aj,k(t) represent the creation (or appearance/arrival) of steps of type (j, k) between time τ τ≤t≤0 Um(t)), which we denote by Aj,k(τ : 0) (and Um(τ : 0)), respectively. Now there is an important observation about Aj,k(τ : 0): Aj,k(τ : 0) ≤ Qj,pj (k)(τ − 1) + Aj,pj (k)(τ − 1 : −1), (35) where pj(k) is the parent of k in Tj, due to the following. As each job takes one slot to be served, no job whose step (j, pj(k)) completed after −1 can have its step (j, k) be available for service at or before 0. Thus by induction on the function pj we can write Qj,w(τ − 1 − d + w) + Aj,1(τ − dj(k) : −dj(k)), (36) Aj,k(τ : 0) ≤ d(cid:88) w=1 34 as dj(k) = d + 1 by the inductive assumption. Note that from 1 to k (at depth d + 1) there is a unique d-length path and hence, on that path w.l.o.g. we denote the respective steps by (j, w) where w is its depth on that path. (cid:88) Um(t)hm,s(cid:48)))] τ≤t≤0 m Hence, Note that(cid:80) m E[exp(θ( rj,k,s τ≤t≤0 j,k:dj (k)≤d+1 ≤ E[exp(θ( (cid:88) (cid:88) (cid:88) Aj,k(t) −(cid:88) d(cid:88) + Aj,1(τ − dj(k) : −dj(k))) −(cid:88) (cid:88) (cid:32) d(cid:88) j,k:dj (k)≤d+1 (cid:88) τ≤t≤0 rj,k,s( w=1 m τ≤t≤0 Um(t)hm,s(cid:48) is independent of (cid:80) Qj,w(τ − 1 − d + w) Um(t)hm,s(cid:48)))] rj,k,s Qj,w(τ − 1 − d + w) + Aj,1(τ − dj(k) : −dj(k)) j,k:dj (k)≤d+1 w=1 (cid:33) , because Aj,1 are i.i.d. (independent of Um) and Qj,w(τ − d + w) does not depend on Um(τ : 0) for d ≥ w ≥ 1. Hence, (cid:88) Aj,k(t) −(cid:88) (cid:32) d(cid:88) m τ≤t≤0 (cid:88) τ≤t≤0 E[exp(θ( rj,k,s j,k:dj (k)≤d+1 Um(t)hm,s(cid:48)))] ≤ E[exp(θ( Qj,w(τ − 1 − d + w) + Aj,1(τ − dj(k) : −dj(k)) (cid:88) (cid:88) (cid:88) j,k:dj (k)≤d+1 w=1 rj,k,s (cid:88) × E[exp(−θ We use the previously derived bound for E[exp(−θ(cid:80) Um(t)hm,s(cid:48))]. τ≤t≤0 m (cid:80) with E[exp(θ( (cid:88) j,k:dj (k)≤d+1 (cid:18) d(cid:88) w=1 (cid:33) ))] (cid:19) τ≤t≤0 Um(t)hm,s(cid:48))]. So, we only concern ourselves m rj,k,s Qj,w(τ − 1 − d + w) + Aj,1(τ − dj(k) : −dj(k)) ))]. Consider any Qj,w(τ − 1 − d + w) at depth w, then Aj,1(τ − d − 1) is independent of it. As Aj,1 are i.i.d. and future arrivals in a queue are independent of present and past queue-lengths, we have rj,k,s( Qj,w(τ − 1 − d + w) + Aj,1(τ − dj(k) : −dj(k)))] (cid:88) (cid:88) E[exp(θ( j,k:dj (k)≤d+1 = E[exp(θ( j,k:dj (k)≤d+1 d(cid:88) d(cid:88) w=1 w=1 rj,k,s Qj,w(τ − 1 − d + w)))] × E[exp( rj,k,sAj,1(τ − dj(k) : −dj(k)))] (cid:88) j,k:dj (k)≤d+1 35 For the second term we obtain a bound using previous techniques and note that since λ ∈ αC, j,k:dj (k)≤d+1 rj,k,sE[Aj,1] ≤(cid:88) (cid:88) rj,k,sAj,1(τ − dj(k) : −dj(k)) −(cid:88) m µmhm,s(cid:48), (cid:88) (cid:88) Um(t)hm,s(cid:48)))] τ≤t≤0 m which in the same way as above will imply that for some K(N ) and some θ > 0, E[exp(θ( j,k:dj (k)≤d+1 ≤ exp(−θK(N )τ ). (cid:88) τ≤t≤0 (cid:88) Note that E[exp(θ( (cid:88) d(cid:88) w=1 rj,k,s Qj,w(τ − 1 − d + w)))] < ∞ j,k:dj (k)≤d+1 by the inductive assumption that the number of unallocated steps at depth ≤ d have finite exponential moments. So we have that E[exp(θ (Xs(t) − Ys(t) + ¯r))] < c1 exp(−θK(N )τ ), and so, in turn (using the same steps as above) Qs has finite exponential moment for some θ. The rest of the steps are similar to above and we get the desired result that (cid:88) E[exp(θ j,k:dj (k)≤d+1 Qj,k)] < ∞. By induction on d, we have proven that the total number of unallocated steps over all types of jobs have finite exponential moment (say c(cid:48)). Therefore, P( Qj,k > q) ≤ exp(−θq)E[exp(θ Qj,k)] (cid:88) j,k j,k ≤ c(cid:48) exp(−θq). So for q = 3 log N θ , we have the result (as c(cid:48) is constant). E. Proof of Theorem 5 The following lemma is useful for the proof. Lemma 4. Any feasible solution of problem (9) is a feasible solution of problem (10). Proof: Among the total R available agents, let Rm be of type m, and let them be denoted i1, i2, . . . , iRm. For an allocation a in the formulation (9), let hik,s be the time that agent i1 is assigned for skill s. Then among all these type m agents, the total contribution to skill s is(cid:80)Rm Note that the total time for these Rm agents is Rmhm. Now by the allocation constraint we have(cid:80) (cid:80)Rm hik,s/Rmhm, this is a valid allocation in formulation (10), as (cid:80) Rmhm. If we choose αm,s ≥(cid:80)Rm hik,s ≤ s αm,s ≤ 1 and it also meets the allocation constraint. So, for this αm,s the allocation a is a valid allocation in problem (10). hik,s. i=1 i=1 s i=1 This proves that every valid allocation in (9) is also a valid allocation in (10). Since solving (10) yields a feasible allocation, the lemma implies the two problems are actually alternate formulations of one another. The rest of the proof follows the same steps as the proof of Thm. 3. F. Proof of Theorem 6 The result can be derived in the same way as Thm. 4, through the use of the following lemmata. 36 Lemma 5. Let {Ri} be i.i.d. Bernoulli random variables with P(Ri = 1) = p ∈ (0, 1) and N be a random variable independent of {Ri} with a moment generating function MN (θ). Then E[exp(θ Ri)] = MN (log (p exp(θ) + (1 − p))). N(cid:88) (cid:34) (cid:35) (p exp(θ) + (1 − p))N(cid:105) (cid:104) Ri)N ] E[exp(θ i=1 = E = E [exp(log (p exp(θ) + (1 − p)) N )] = MN (log (p exp(θ) + (1 − p))) N(cid:88) N(cid:88) i=1 i=1 Proof: E[exp(θ Ri)] = E F,F = {λ : ∃bm ∈ [0, 1]S for all m, s.t. (cid:80) Lemma 6. Let CO (cid:80) j,k λjrj,k,s for all s}. Then, CF,F ⊂ CO F,F . s bm,s ≤ 1, bm,s > 0 only if s ∈ Sm,(cid:80) m:s∈Sm bm,shmµm > Proof: This follows from the constraints in (10). Because constraints in (10) are per sample realization, and the above constraints are in expectation. So for constraints in (10) to be satisfied, the above constraints must be satisfied. Lemma 7. For any τ, P{∩∞ multiple solutions are broken deterministically.2 t=τ ∩s {ψm,s(t) = pm,s}} = 1 such that pm,s solves (12), assuming ties between Proof: For the choice of γ(t) (and t0 = −∞) it follows from the convergence of stochastic approximation update equations [25] and the facts that λ ∈ αC and t0 = −∞. By stochastic approximation updates ¯A and ¯U converges almost surely to λ and µ respectively at any finite τ. The rest follows from Lem. 6. Let Bs,m be the set of agents of type m that has been put into Bs. Bs,m is Bernoulli sampling from Um agents with probability ψm,s(t). 2Extends to random tie breaking also, but involves more details. 37 We follow the same steps as in the proof of Thm. 4. Consider the work and service time brought at time t (as j Aj,1(t)rj,1,s and service time brought by agents for skill s is hmBs,m(t). Hence, we can construct a queue Q1 (as before) m:s∈Sm before). Note that work brought for skill s is (cid:80) (cid:80) (cid:88) (cid:88) E[exp(θ Q1)] ≤(cid:88) 1(t + 1) = (Qs Following the same steps to obtain (27) we can have 1(t) + S max E[exp(θS Qs s j (max τ≤0 τ≤t≤0 s (cid:88) j (cid:88) m:s∈Sm (cid:88) m:s∈Sm Aj,1(t)rj,1,s − S min s hmBs,m(t))+ Aj,1(t)rj,1,s − min s hmBs,m(t))] But a result like (27) does not follow immediately, since Bs,m(t) are not independent over time and Bs,m(t) depends on Aj(t) via {ψs(t)}. j Aj,1(t)rj,1,s and Ys(t) =(cid:80) Consider the following. Let Xs(t) =(cid:80) (cid:88) Xs(t) − min (cid:88) τ≤t≤0 ≤ E[E[exp(θS Xs(t) − min E[exp(θS Ys(t))] (max (max s s τ≤t≤0 s s Now by Lem. 7, for any finite τ1, P{{ψs(t(cid:48)) = ps,∀t(cid:48) ≥ τ1,∀s, m} = 1. m:s∈Sm hmBs,m(t), then Ys(t)){ψm,s(t(cid:48)), s, mt(cid:48) ≥ τ}]] Hence, for any finite τ1, τ2, {ψm,s(t(cid:48)), s, mt(cid:48) ≥ τ1} are independent of {Aj,1(t) : tτ2}. Also, by the above argument, Xs(t) and Ys(t) are independent of each other, given {ψm,s(t(cid:48)), s, mt(cid:48) ≥ τ} and they are also independent over time. E[ E[exp(θS(max τ≤t≤0 Xs(t) − min s Ys(t)){ψm,s(t(cid:48)), s, mt(cid:48) ≥ τ})]] s = E[ = E[ E[exp(θS max s,s(cid:48) (Xs(t) − Ys(cid:48)(t)){ψm,s(t(cid:48)), s, mt(cid:48) ≥ τ})]] E[exp(θS(Xs(t) − Ys(cid:48)(t)){ψm,s(t(cid:48)), s, mt(cid:48) ≥ τ})]] s,s(cid:48) E[E[exp(θSXs(t){ψm,s(t(cid:48)), s, mt(cid:48) ≥ τ})] s,s(cid:48) τ≤t≤0 × E[exp(−θSYs(cid:48)(t){ψm,s(t(cid:48)), s, mt(cid:48) ≥ τ})]] E[exp(θSXs(t){ψm,s(t(cid:48)) = ψm,s(τ ), s, mt(cid:48) ≥ τ})] s,s(cid:48) τ≤t≤0 × E[exp(−θSYs(cid:48)(t){ψm,s(t(cid:48)) = ψm,s(τ ), s, mt(cid:48) ≥ τ})] τ≤t≤0 τ≤t≤0 (cid:89) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) τ≤t≤0 s,s(cid:48) = = = E[exp(θSXs(t)] × E[exp(−θSYs(cid:48)(t){ψm,s(t(cid:48)) = ψm,s(τ ), s, mt(cid:48) ≥ τ})] The first equality follows because maxx∈X f (x)−minx∈X g(x) = maxx,x(cid:48)∈X (f (x)−g(x)), for finite X . The second equality follows due to independence of Xs(t) − Ys(t) from t ≥ τ which is due to Lem. 7. The third equality follows due to independence of Xs(t) and Ys(t) given {ψs(t)}. The fourth equality is again due to Lem. 7, as {ψs(t(cid:48)) = ps, s, mt(cid:48) ≥ τ} is an almost sure event. The last equality follows because Aj(t), t ≥ τ1 are independent of {ψs(t(cid:48)) = ps, s, mt(cid:48) ≥ τ2} for any finite τ1 and τ2. Note that E[exp(θSXs(t)] can be evaluated exactly as in the proof of Thm. 4. Consider for τ ≤ t ≤ 0, 38 E[exp(−θSYs(cid:48)(t){ψm,s(t(cid:48)) = ψm,s(τ ), s, mt(cid:48) ≥ τ})] = E[exp(−θS hmBs,m{ψm,s(t(cid:48)) = ψm,s(τ ), s, mt(cid:48) ≥ τ})] (cid:88) (cid:89) = m:s∈Sm m:s∈Sm E[exp(−θShmBs,m{ψm,s(t(cid:48)) = ψm,s(τ ), s, mt(cid:48) ≥ τ})]. (37) Now Bs,m is Bernoulli sampling of Um agents with probability ψm,s(t(cid:48)) = ψm,s(τ ) := pm,s. Then, by Lem. 5, for a θm,s = log(exp(θ)pm,s + 1 − pm,s): E[exp(−θShmBs,m{ψs(t(cid:48)) = ps, s, mt(cid:48) ≥ τ})] = E[exp(−θm,shmUm)]. Now following the same steps as in the proof of Thm. 4, we can show that (for Poisson and Gaussian dominated cases) for a sufficiently large Nα, for all N ≥ Nα, and δ < α 2 , E[exp(−θm,shmUm)] ≤ exp(−(1 − δ)θm,shmµm). Now, by concavity of logarithms, θm,s = log (pm,s exp(θ) + (1 − pm,s)) ≥ pm,sθ. This implies As λ ∈ αCF,F , by Lem. 6 and 7, we have(cid:80) Also, as by assumption (cid:80) j,k λjrj,k,s −(cid:80) E[exp(−θm,shmUm)] ≤ exp(−(1 − δ)θ pm,shmµm). pm,shmµm > (1− α)(cid:80) (38) j,k λjrj,k,s for all s and < 1− α. (cid:88) m:s∈Sm m:s∈Sm j,k λjrj,k,s(cid:48) is sub-poly(N ), hence ps(cid:48),mhmµm > (1 − α) λjrj,k,s,∀s, s(cid:48) (cid:88) j,k This along with (38) and (37) gives the final result by following the same steps as the proof of Thm. 4. G. Proof of Theorem 7 The assumption {Sm : m ∈ [M ]} is a partition implies that there exists a partition of [S], say {Kl : 1 ≤ l ≤ L} such that for m ∈ [M ], Sm = Kl for some l ∈ [L]. Note that L ≤ S. As λ ∈ CF,I, for any step (j, k) with λj > 0, the set of required skills is a subset of some Kl. Otherwise, due to inflexibility of the steps, that step can never be allocated which contradicts that λ ∈ CF,I. for all l (cid:80) λj <(cid:80) m:Sm=Kl j:(j,k)−skills ⊂Kl 39 µm(cid:98) hm r1,1,1 (cid:99)}. Then, under the conditions Lemma 8. Let CO in Thm. 7, CF,I ⊂ CO F,I. F,I = {λ : Proof: Follows by noting the fact that under the conditions in Thm. 7 the steps with skill requirements in Kl has can only be served by agents of type m with Sm = Kl. Also, note that agents with skills in Kl cannot serve any other kinds of steps. Also, as each step is of same size r1,1,1 a type m agent can serve at most (cid:98) hm r1,1,1 (cid:99) steps with skill requirements in Sm. This implies that if the agent availability is u then aj,k steps can be served only if The rest follows because the sample path constraint is true only if the constraint is true in expectation. aj,k < j:(j,k)−skills ⊂Kl um(cid:98) hm r1,1,1 (cid:99). (cid:88) (cid:88) (cid:88) (cid:88) m:Sm=Kl (cid:88) (cid:88) m:Sm=Kl Notice that the condition can be written as This will be useful later. (cid:88) λj < j:(j,k)−skills ⊂Kl µm(cid:98) hm r1,1,1 (cid:99)} d j:(j,k)-skill ⊂Kl∩Ad λj < m:Sm=Kl µm(cid:98) hm r1,1,1 (cid:99). Coming back to the main proof, we consider a queue for each l ∈ [L] and depth d (≤ D), Ql d(t). This queue represents the number of unallocated steps with skill requirements in Kl that are at depth ≤ d on respective precedence trees. Note that such steps with skill requirements in Kl can only be served by agents of types {m : Sm = Kl}. Note that just as in Restricted Greedy, steps at depth ≤ d have priority (allocate themselves before) over steps at higher depth: Ql d(t + 1) = ( Ql d(t) + (cid:98) hm r1,1,1 (cid:99)Um)+. (39) Note that allocation of steps with skill requirements in Kl and Kl(cid:48) for l (cid:54)= l(cid:48) are independent. Also, the agents with skills in Kl and Kl(cid:48) for l (cid:54)= l(cid:48) are independent. So, if we define Qd(t): (cid:88) Aj,k(t) − (cid:88) (j,k)-skills⊂Kl m:Sm=Kl (cid:88) Aj,k(t) − (cid:88) m:Sm=Kl ( (j,k)-skills⊂Kl Ql d(t). l (cid:98) hm r1,1,1 (cid:99)Um))+, (40) Consider the depth d = 1 first. Then we can follow the same steps as in the proof of Thm. 4 for the depth d = 1 l Qd(t + 1) =( Qd(t) + L max then Qd(t) is a path-wise upper-bound on(cid:80) (cid:88) case. Note that λ ∈ αCF,I implies that for all l D(cid:88) d=1 j:(j,k)-skill ⊂Kl∩Ad λj < (1 − α) (cid:88) m:Sm=Kl µm(cid:98) hm r1,1,1 (cid:99), and hence, for any d ≤ D d(cid:88) d(cid:48)=1 (cid:88) j:(j,k)-skill ⊂Kl∩Ad(cid:48) 40 λj < (1 − α) (cid:88) m:Sm=Kl µm(cid:98) hm r1,1,1 (cid:99). This along with the same steps as in the proof of Thm. 4 gives that for some θ1 > 0 and θ1 = Ω(1), ∀θ < θ1 E[exp(θ1 Q1)] < ∞, for all l. Then, like the proof of Thm. 4 we can perform induction over d to prove that for some θD > 0 and θD = Ω(1), then for all θ < θD: where QD(t) is an upper bound on(cid:80) E[exp(θ QD)] < ∞, for all l, l Ql d(t), which is again the total number of unallocated steps in the system. We obtain the desired bound from this. Induction from d to d + 1 is similar to the proof of Thm. 4.
1708.07607	3	1708	2018-02-27T06:17:29	Reinforcement Mechanism Design for e-commerce	[ "cs.MA", "cs.AI" ]	We study the problem of allocating impressions to sellers in e-commerce websites, such as Amazon, eBay or Taobao, aiming to maximize the total revenue generated by the platform. We employ a general framework of reinforcement mechanism design, which uses deep reinforcement learning to design efficient algorithms, taking the strategic behaviour of the sellers into account. Specifically, we model the impression allocation problem as a Markov decision process, where the states encode the history of impressions, prices, transactions and generated revenue and the actions are the possible impression allocations in each round. To tackle the problem of continuity and high-dimensionality of states and actions, we adopt the ideas of the DDPG algorithm to design an actor-critic policy gradient algorithm which takes advantage of the problem domain in order to achieve convergence and stability. We evaluate our proposed algorithm, coined IA(GRU), by comparing it against DDPG, as well as several natural heuristics, under different rationality models for the sellers - we assume that sellers follow well-known no-regret type strategies which may vary in their degree of sophistication. We find that IA(GRU) outperforms all algorithms in terms of the total revenue.	cs.MA	cs	Reinforcement Mechanism Design for e-commerce Qingpeng Cai Tsinghua University Beijing, China [email protected] Pingzhong Tang Tsinghua University Beijing, China [email protected] Aris Filos-Ratsikas University of Oxford Oxford, United Kingdom [email protected] Yiwei Zhang University of California, Berkeley Berkeley, United States of America [email protected] ABSTRACT We study the problem of allocating impressions to sellers in e- commerce websites, such as Amazon, eBay or Taobao, aiming to maximize the total revenue generated by the platform. We employ a general framework of reinforcement mechanism design, which uses deep reinforcement learning to design efficient algorithms, taking the strategic behaviour of the sellers into account. Specifically, we model the impression allocation problem as a Markov decision process, where the states encode the history of impressions, prices, transactions and generated revenue and the actions are the possible impression allocations in each round. To tackle the problem of continuity and high-dimensionality of states and actions, we adopt the ideas of the DDPG algorithm to design an actor-critic policy gradient algorithm which takes advantage of the problem domain in order to achieve convergence and stability. We evaluate our proposed algorithm, coined IA(GRU), by com- paring it against DDPG, as well as several natural heuristics, under different rationality models for the sellers - we assume that sellers follow well-known no-regret type strategies which may vary in their degree of sophistication. We find that IA(GRU) outperforms all algorithms in terms of the total revenue. CCS CONCEPTS • Theory of computation → Algorithmic mechanism design; • Computing methodologies → Reinforcement learning; • Applied computing → Electronic commerce; KEYWORDS e-commerce; impression allocation; mechanism design; reinforce- ment learning ACM Reference Format: Qingpeng Cai, Aris Filos-Ratsikas, Pingzhong Tang, and Yiwei Zhang. 2018. Reinforcement Mechanism Design for e-commerce. In WWW 2018: The 2018 Web Conference, April 23–27, 2018, Lyon, France. ACM, New York, NY, USA, 10 pages. https://doi.org/10.1145/3178876.3186039 This paper is published under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. Authors reserve their rights to disseminate the work on their personal and corporate Web sites with the appropriate attribution. WWW 2018, April 23–27, 2018, Lyon, France © 2018 IW3C2 (International World Wide Web Conference Committee), published under Creative Commons CC BY 4.0 License. ACM ISBN 978-1-4503-5639-8/18/04.. https://doi.org/10.1145/3178876.3186039 1 INTRODUCTION A fundamental problem that all e-commerce websites are faced with is to decide how to allocate the buyer impressions to the potential sellers. When a buyer searches a keyword such as "iPhone 7 rose gold", the platform will return a ranking of different sellers providing an item that fits the keyword, with different prices and different historical sale records. The goal of the platform is to come up with algorithms that will allocate the impressions to the most appropriate sellers, eventually generating more revenue from the transactions. This setting can be modeled as a resource allocation problem over a sequence of rounds, where in each round, buyers arrive, the algorithm inputs the historical records of the sellers and their prices and outputs such an allocation of impressions. The sellers and the buyers carry out their transactions and the historical records are updated. In reality, most e-commerce websites employ a class of heuristic algorithms, such as collaborative filtering or content based filtering [34], many of which rank sellers in terms of "historical scores" calculated based on the transaction history of the sellers with buyers of similar characteristics. However, this approach does not typically take into account the fact that sellers strategize with the choice of prices, as certain sub-optimal prices in one round might affect the record histories of sellers in subsequent rounds, yielding more revenue for them in the long run. Even worse, since the sellers are usually not aware of the algorithm in use, they might "explore" with their pricing schemes, rendering the system uncontrollable at times. It seems natural that a more sophisticated approach that takes all these factors into account should be in place. In the presence of strategic or rational individuals, the field of mechanism design [29] has provided a concrete toolbox for manag- ing or preventing the ill effects of selfish behaviour and achieving desirable objectives. Its main principle is the design of systems in such a way that the strategic behaviour of the participants will lead to outcomes that are aligned with the goals of the society, or the objectives of the designer. Cai et al. [10] tackle the problem of faking transactions and fraudulent seller behaviour in e-commerce using the tools from the field of mechanism design. A common de- nominator in most of the classical work in economics is that the participants have access to either full information or some distri- butional estimate of the preferences of others. However, in large and constantly evolving systems like e-commerce websites, the participants interact with the environment in various ways, and adjust their own strategies accordingly and dynamically [32]. In addition to that, their rationality levels are often bounded by either computational or financial constraints, or even cognitive limitations [35]. For the reasons mentioned above, a large recent body of work has advocated that other types of agent behaviour, based on learning and exploration, are perhaps more appropriate for such large-scale online problems encountered in reality [13, 18–21, 28, 32, 33]. In turn, this generates a requirement for new algorithmic techniques for solving those problems. Our approach is to use techniques from deep reinforcement learning for solving the problem of the impres- sion allocation to sellers, given their selfish nature. In other words, given a rationality model for the sellers, we design reinforcement learning algorithms that take this model into account and solve the impression allocation problem efficiently. This general approach is called reinforcement mechanism design [11, 36, 40], and we can view our contribution in this paper as an instance of this framework. No-regret learning as agent rationality As mentioned earlier, the strong informational assumptions of clas- sical mechanism design are arguably unrealistic in complex and dynamic systems, like diverse online marketplaces. Such repeated game formulations typically require that the participants know the values of their competitors (or that they can estimate them pretty accurately based on known prior distributions) and that they can compute their payoff-maximizing strategies over a long sequence of rounds. Such tasks are usually computationally burdensome and require strong cognitive assumptions, as the participants would have to reason about the future, against all possible choices of their opponents, and in a constantly evolving environment. Given this motivation, an alternative approach in the forefront of much of the recent literature in algorithmic mechanism design is to assume that the agents follow some type of no-regret strategies; the agent picks a probability mixture over actions at each round and based on the generated payoffs, it updates the probabilities accordingly, minimizing the long-term regret. This is more eas- ily conceivable, since the agents only have to reason about their own strategies and their interaction with the environment, and there is a plethora of no-regret algorithms at their disposal. Pre- cisely the same argument has been made in several recent works [13, 21, 28, 32, 33] that study popular auction settings under the same rationality assumptions of no-regret, or similar types. In fact, there exist data from Microsoft's Ad Actions which suggest that advertisers do use no-regret algorithms for their actions [41]. For a more detailed discussion on related rationality models, the reader is referred to [18]. The seller rationality model: To model the different sophistica- tion levels of sellers, we consider four different models of rationality, based on well-established no-regret learning approaches. The first two, ϵ-Greedy [43] and ϵ-First are known as semi-uniform meth- ods, because they maintain a distinction between exploration and exploitation. The later is often referred to as "A/B testing" and is widely used in practice [9, 12]. The other two approaches, UCB1 [2, 5] and Exp3 [5, 6] are more sophisticated algorithms that differ in their assumptions about the nature of the rewards, i.e. whether they follow unknown distributions or whether they are completely adversarial. Note that all of our rationality models employ algo- rithms for the multi-arm bandit setting, as in platforms like Taobao or eBay, the impression allocation algorithms are unknown to the sellers and therefore they can not calculate the payoffs of unused auctions. The update of the weights to the strategies is based solely on the observed payoffs, which is often referred to as the bandit feedback setting [16]. We note here that while other related rationality models can be used, the goal is to choose a model that real sellers would conceivably use in practice. The semi-uniform algorithms are quite simpler and model a lower degree of seller sophistication, whereas the other two choices correspond to sellers that perhaps put more effort and resources into optimizing their strategies - some examples of sophisticated optimization services that are being used by online agents are provided in [28]. Note that both UCB1 and Exp3 are very well-known [9] and the latter is perhaps the most popular bandit feedback implementation of the famous Hedge (or Multi- plicative Weights Update) algorithm for no-regret learning in the fully informed feedback setting. The impression allocation problem We model the impression allocation problem as a Markov decision process (MDP) in which the information about the prices, past transactions, past allocations of impressions and generated revenue is stored in the states, and the actions correspond to all the different ways of allocating the impressions, with the rewards being the im- mediate revenue generated by each allocation. Given that the costs of the sellers (which depend on their production costs) are private information, it seems natural to employ reinforcement learning techniques for solving the MDP and obtain more sophisticated im- pression allocation algorithms than the heuristics that platforms currently employ. In our setting however, since we are allocating a very large num- ber of impressions, both the state space and the action space are extremely large and high-dimensional, which renders traditional re- inforcement learning techniques such as temporal difference learn- ing [38] or more specifically Q-learning [14] not suitable for solving the MDP. In a highly influential paper, Mnih et al. [31] employed the use of deep neural networks as function approximators to es- timate the action-value function. The resulting algorithm, coined "Deep Q Network" (DQN), can handle large (or even continuous) state spaces but crucially, it can not be used for large or continuous action domains, as it relies on finding the action that maximizes the Q-function at each step. To handle the large action space, policy gradient methods have been proposed in the literature of reinforcement learning with actor-critic algorithms rising as prominent examples [7, 15, 39], where the critic estimates the Q-function by exploring, while the actor adjusts the parameters of the policy by stochastic gradient ascent. To handle the high-dimensionality of the action space, Silver et al. [37] designed a deterministic actor-critic algorithm, coined "Deterministic Policy Gradient" (DPG) which performs well in stan- dard reinforcement-learning benchmarks such as mountain car, pendulum and 2D puddle world. As Lillicrap et al. [27] point out however, the algorithm falls short in large-scale problems and for that reason, they developed the "Deep-DPG" (DDPG) algorithm which uses the main idea from [31] and combines the deterministic policy gradient approach of DPG with deep neural networks as function approximators. To improve convergence and stability, they employ previously known techniques such as batch normalization [23], target Q-networks [30], and experience replay [1, 22, 31]. The IA(GRU) algorithm: We draw inspiration from the DDPG algorithm to design a new actor-critic policy gradient algorithm for the impression allocation problem, which we refer to as the IA(GRU) algorithm. IA(GRU) takes advantage of the domain properties of the impression allocation problem to counteract the shortcomings of DDPG, which basically lie in its convergence when the number of sellers increases. The modifications of IA(GRU) to the actor and critic networks reduce the policy space to improve convergence and render the algorithm robust to settings with variable sellers, which may arrive and depart in each round, for which DDPG per- forms poorly. We evaluate IA(GRU) against DDPG as well as several natural heuristics similar to those usually employed by the online platforms and perform comparisons in terms of the total revenue generated. We show that IA(GRU) outperforms all the other algo- rithms for all four rationality models, as well as a combined pool of sellers of different degrees of sophistication. 2 THE SETTING In the impression allocation problem of e-commerce websites, there are m sellers who compete for a unit of buyer impression.1 In each round, a buyer2 searches for a keyword and the platform returns a ranking of sellers who provide an item that matches the keyword; for simplicity, we will assume that all sellers provide identical items that match the keyword exactly. Each seller i has a private cost ci for the item, which can be interpreted as a production or a purchasing cost drawn from an i.i.d. distribution Fs. all items offered are identical. We let qi =n m Typically, there are n slots (e.g. positions on a webpage) to be allocated and we let xij denote the probability (or the fraction of time) that seller i is allocated the impression at slot j. With each slot, there is an associated click-through-rate αj which captures the "clicking potential" of each slot, and is independent of the seller, as j=1 xij αj denote the probability that the buyer will click the item of seller i. Given this definition (and assuming that sellers can appear in multiple slots in each page), the usual feasibility constraints for allocations, i.e. for all i, for all j, t holds that 0 ≤ xij ≤ 1 and or all j, it holds that i =1 xij ≤ 1 can be alternatively written as for all i, qi ≥ 0, it holds that qi ≤ n m i =1 n j=1 αj and αj = 1. j=1 That is, for any such allocation q, there is a feasible ranking x that realizes q (for ease of notation, we assume that the sum of click- through rates of all slots is 1) and therefore we can allocate the 1Since the buyer impressions to be allocated is a huge number, we model it as a continuous unit to be fractionally allocated. Even if we used a large integer number instead, the traditional approaches like DDPG fall short for the same reasons and furthermore all of the performance guarantees of IA(GRU) extend to that case. 2As the purchasing behavior is determined by the valuation of buyers over the item, without loss of generality we could consider only one buyer at each round. buyer impression to sellers directly instead of outputting a ranking over these items when a buyer searches a keyword.3 Let hit = (vit , pit , nit , ℓit) denote the records of seller i at round t, which is a tuple consisting of the following quantities: (1) vit is the expected fraction of impressions that seller i gets, (2) pit is the price that seller i sets, (3) nit is the expected amount of transactions that seller i makes, (4) ℓit is the expected revenue that seller i makes at round t. Let Ht = (h1t , h2t , ..., hit) denote the records of all sellers at round t, and let Hit = (hi1, hi2, ..., hit) denote the vectors of records of seller i from round 1 to round t, which we will refer to as the records of the seller. At each round t + 1, seller i chooses a price pi(t +1) for its item and the algorithm allocates the buyer impression to sellers. MDP formulation: The setting can be defined as a Markov decision process (MDP) defined by the following components: a continuous state space S, a continuous action space A, with an initial state distribution with density p0(s0), and a transition distribution of states with conditional density p(st +1st , at) satisfying the Markov property, i.e. p(st +1s0, a0, ..., st , at) = p(st +1st , at). Furthermore, there is an associated reward function r : S × A → R assigning payoffs to pairs of states and actions. Generally, a policy is a function π that selects stochastic actions given a state, i.e, π : S → P(A), ∞ where P(A) is the set of probability distributions on A. Let Rt denote the discounted sum of rewards from the state st , i.e, Rt(st) = k =t γ k−t r(sk , ak), where 0 < γ < 1. Given a policy and a state, the value function is defined to be the expected total discounted reward, i.e. V π(s) = E[Rt(st)st = s; π] and the action-value function is defined as Qπ(s, a) = E[Rt(st)st = s, at = a; π]. For our problem, a state st of the MDP consists of the records of all sellers in the last T rounds, i.e. st = (Ht−T , ..., Ht−1), that is, the state is a (T , m, 4) tensor, the allocation outcome of the round is the action, and the immediate reward is the expected total revenue generated in this round. The performance of an algorithm is defined as the average expected total revenue over a sequence of T0 rounds. Buyer Behaviour: We model the behaviour of the buyer as being dependent on a valuation that comes from a distribution with cu- mulative distribution function Fb. Intuitively, this captures the fact that buyers may have different spending capabilities (captured by the distribution). Specifically, the probability that the buyer pur- chases item i is nit = (1 − Fb(pit)) · vit , that is, the probability of purchasing is decided by the impression allocation and the price seller i sets. For simplicity and without loss of generality with re- spect to our framework, we assume that the buyer's valuation is drawn from U(0, 1), i.e. the uniform distribution over [0, 1]. Seller Rationality As we mentioned in the introduction, following a large body of recent literature, we will assume that the sellers employ no-regret type strategies for choosing their prices in the next round. Generally, a seller starts with a probability distribution over all the possible prices, and after each round, it observes the payoffs that these 3The framework extends to cases where we need return similar but different items to a buyer, i.e, the algorithm outputs a ranking over these items. Furthermore, our approach extends trivially to the case when sellers have multiple items. strategies generate and adjusts the probabilities accordingly. As we already explained earlier, it is most natural to assume strategies in the bandit feedback setting, where the seller does not observe the payoffs of strategies in the support of its strategy which were not actually used. The reason is that even if we assume that a seller can see the prices chosen in a round by its competitors, it typically does not have sufficient information about the allocation algorithm used by the platform to calculate the payoffs that other prices would have yielded. Therefore it is much more natural to assume that the seller updates its strategy based on the observed rewards, using a multi-arm bandit algorithm. More concretely, the payoff of a seller i that receives vit im- pressions in round t when using price pij(t), is given by uij(t) = nit(pij(t) − ci) = vit(1 − Fb(pit))(pij(t) − ci). For consistency, we normalize the costs and the prices to lie in the unit interval [0, 1] and we discretize the price space to a "dense enough" grid (of size 1/K, for some large enough K). This discretization can either be enforced by the platform (e.g. the sellers are required to submit bids which are multiples of 0.05) or can be carried out by the sell- ers themselves in order to be able to employ the multi-arm bandit algorithms which require the set of actions to be finite, and since small differences in prices are unlikely to make much difference in their payoffs. We consider the following possible strategies for the sellers, based on well-known bandit algorithms. ε-Greedy [43]: With probability ε, each seller selects a strategy uniformly at random and with probability 1 − ε, the strategy with the best observed (empirical) mean payoff so far. The parameter ε denotes the degree of exploration of the seller, whereas 1 − ε is the degree of exploitation; here ε is drawn i.i.d. from the normal distribution N(0.1, 0.1/3). ε-First: For a horizon of T rounds, this strategy consists of an ex- ploration phase first, over ε · T rounds, followed by an exploitation phase, for the remaining period. In the exploration phase, the seller picks a strategy uniformly at random. In the remaining rounds, the sellers picks the strategy that maximizes the empirical mean of the observed rewards. For each seller, we set T = 200 and ε = 0.1. Exponential-weight Algorithm for Exploration and Exploita- tion (Exp3) [5, 6]: We use the definition of the algorithm from [6]. Let γ ∈ (0, 1] be a real number and initialize wi(1) = 1 for i = 1, . . . , K + 1 to be the initial weights of the possible prices4. In each round t, K +1 • For i = 1, . . . , K + 1, let πi(t) = (1 − γ) wi(t) j =1 wj(t) + γ K +1, where wi(t) is the weight of price pi in round t. • Select a price pj(t) according to the probability distribution defined by π1(t), . . . , πK +1(t). • Receive payoff uj(t) ∈ [0, 1]. 4For ease of notation, we drop the subscript referring to a specific seller, as there is no ambiguity. • For ℓ = 1, . . . , K + 1, let (cid:40) uℓ(t) = uℓ(t)/πℓ(t), 0, if ℓ = j otherwise and wℓ(t + 1) = wℓ(t)eγ · uℓ(t)/(K +1). We remark here that since the payoff of each seller in some round t actually takes values in [−1, 1], we scale the payoff to [0, 1] by applying the transformation f (u) = (u + 1)/2 to any payoff u. (cid:40) Upper Confidence Bound Algorithm (UCB1) [2, 4]: For each price pj ∈ [0, 1/K, 2/K, . . . , 1], initialize xj(1) = 0. At the end of each round t, update xj(t) as: xj(t − 1)/t + uj(t)/t, xj(t) = xj(t − 1), For any round t ∈ {0, . . . , K}, the seller chooses a price pj that has not been used before in any of the previous rounds (breaking ties arbitrarily). For any round t ≥ K + 1, the seller chooses the price pj with the maximum weighted value xj, i.e, j∈{0,1/K, ...,1} xj(t) + if j was chosen in this round t otherwise pj(t) ∈ arg t log2 t τ =1 Ijτ max , where Ijτ is the indicator function, i.e. (cid:40)1, if pj was chosen in round τ Ijτ = 0, otherwise. ε-Greedy and ε-First are simple strategies that maintain a clear distinction between exploration and exploitation and belong to the class of semi-uniform strategies. Exp3 is the most widely used bandit version of perhaps the most popular no-regret algorithm for the full information setting, the Hedge (or Multiplicative Weight updates) algorithm [17] and works in the adversarial bandit feedback model [6], where no distributional assumptions are being made about the nature of the rewards. UCB1, as the name suggests, maintains a certain level of optimism towards less frequently played actions (given by the second part of the sum) and together with this, it uses the empirical mean of observed actions so far to choose the action in the next round. The algorithm is best suited in scenarios where the rewards do follow some distribution which is however unknown to the seller. For a more detailed exposition of all these different algorithms, [9] provides a concise survey. The point made here is that these choices are quite sensible as they (i) constitute choices that a rel- atively sophisticated seller, perhaps with a research team at its disposal could make, (ii) can model sellers with different degrees of sophistication or pricing philosophies and (iii) are consistent with the recent literature on algorithmic mechanism design, in terms of modeling agent rationality in complex dynamic environments. 3 ALLOCATION ALGORITHMS In this section, we will briefly describe the algorithms that we will be comparing IA(GRU) against - two natural heuristics similar to those employed by platforms for the impression allocation problem, as well as the DDPG algorithm of Lillicrap et al. [27]. Heuristic Allocation Algorithms As the strategies of the sellers are unknown to the platform, and the only information available is the sellers' historical records, the platform can only use that information for the allocation. Note that these heuristics do not take the rationality of the sellers into account, when deciding on the allocation of impressions. The first algorithm is a simple greedy algorithm, which allocates the impressions proportionally to the revenue contribution. ℓiτ /m Greedy Myopic Algorithm: At round 0, the algorithm allocates a 1/m-fraction of the buyer impression to each seller. At any other round τ + 1 (for τ ≥ 0), the algorithm allocates a fraction of j=1 ℓjτ of the buyer impression to each seller, i.e. propor- tionally to the contribution of each seller to the total revenue of the last round. The second algorithm is an algorithm for the contextual multi-arm bandit problem, proposed by [26], based on the principles of the family of upper confidence bound algorithms (UCB1 is an algo- rithm in this family). The algorithm is among the state of the art solutions for recommender systems [9] and is an example of contex- tual bandit approaches, which are widely applied to such settings [3, 8, 25, 26]. To prevent any confusion, we clarify here that while we also used bandit algorithms for the seller rationality models, the approach here is fundamentally different as the Linear UCB Algorithm is used for the allocation of impressions - not the choice of prices - and the arms in this case are the different sellers. Linear UCB Algorithm [26]: We implement the algorithm as described in [26] - in the interest of space, we do not provide the definition of the algorithm, but refer the reader to Algorithm 1 in [26]. We model each seller as an arm and set hit as the feature of each arm i in each round t. The parameter α is set to 1. Deep Deterministic Policy Gradient Here, we briefly describe the DDPG algorithm of [27], which we we draw inspiration from in order to design our impression allocation algorithm. Before describing the algorithm, we briefly mention the main ingredients of its predecessor, the DPG algorithm of Silver et al. [37]. Deterministic Policy Gradient: The shortcoming of DQN [31] is that while it can handle continuous states, it can not handle continu- ous actions or high-dimensional action spaces. Although stochastic actor-critic algorithms could handle continuous actions, they are hard to converge in high dimensional action spaces. The DPG algo- rithm [37] aims to train a deterministic policy µθ : S → A with parameter vector θ ∈ Rn. This algorithm consists of two compo- nents: an actor, which adjusts the parameters θ of the deterministic policy µθ(s) by stochastic gradient ascent of the gradient of the discounted sum of rewards, and the critic, which approximates the action-value function. Deep Deterministic Policy Gradient: Directly training neural net- works for the actor and the critic of the DPG algorithm fails to achieve convergence; the main reason is the high degree of tempo- ral correlation which introduces high variance in the approximation of the Q-function by the critic. For this reason, the DDPG algorithm uses a technique known as experience replay, according to which the experiences of the agent at each time step are stored in a replay buffer and then a mini-batch is sampled uniformly at random from this set for learning, to eliminate the temporal correlation. The other modification is the employment of target networks for the regularization of the learning algorithm. The target network is used to update the values of µ and Q at a slower rate instead of updating by the gradient network; the prediction yt will be relatively fixed and violent jitter at the beginning of training is absorbed by the target network. A similar idea appears in [42] with the form of double Q-value learning. 4 THE IMPRESSION ALLOCATION (GRU) ALGORITHM In this section, we present our main deep reinforcement learning algorithm, termed IA(GRU) ("IA" stands for "impression allocation" and "GRU" stands for "gated recurrent unit") which is in the center of our framework for impression allocations in e-commerce plat- forms and is based on the ideas of the DDPG algorithm. Before we present the algorithm, we highlight why simply applying DDPG to our problem can not work. Shortcomings of DDPG: First of all, while DDPG is designed for settings with continuous and often high-dimensional action spaces, the blow-up in the number of actions in our problem is very sharp as the number of sellers increases; this is because the action space is the set of all feasible allocations, which increases very rapidly with the number of sellers. As we will show in Section 5, the direct application of the algorithm fails to converge even for a moderately small number of sellers. The second problem comes from the in- ability of DDPG to handle variability on the set of sellers. Since the algorithm uses a two-layer fully connected network, the position of each seller plays a fundamental role; each seller is treated as a different entity according to that position. As we show in Section 5, if the costs of sellers at each round are randomly selected, the per- formance of the DDPG algorithm deteriorates rapidly. The settings in real-life e-commerce platforms however are quite dynamic, with sellers arriving and leaving or their costs varying over time, and for an allocation algorithm to be applicable, it should be able to handle such variability. We expect that each seller's features are only af- fected by its historical records, not some "identity" designated by the allocation algorithm; we refer to this highly desirable property as "permutation invariance". Based on time-serial techniques, our algorithm uses Recurrent Neural Networks at the dimension of the sellers and achieves the property. The IA(GRU) algorithm: Next, we explain the design of our algo- rithm, but we postpone some implementation details for Section 5. At a high level, the algorithm uses the framework of DDPG with different network structures and different inputs of networks. It maintains a sub-actor network and a sub-critic network for each seller, i.e, the expected discounted sum of revenues from the sub- state following the policy. Then, it sums up the estimated Q-value of all sub-critic networks to output the final estimated Q-value, with the assumption that the strategy of each seller is independent of the records of other sellers, which is the case in all of our rationality models. The framework of the critic network is similar to Figure 1. 5 EXPERIMENTAL EVALUATION In this section, we present the evaluation of our algorithms in terms of convergence time and revenue performance against several benchmarks, namely the direct application of the DDPG algorithm (with a fully connected network) and the heuristic allocation algo- rithms that we defined in Section 3. We use Tensorflow and Keras as the engine for the deep learning, combining the idea of DDPG and the techniques mentioned in Section 4, to train the neural network. Designed experiments: First, we will compare IA(GRU) and DDPG in terms of their convergence properties in the training phase and show that the former converges while the latter does not. Next, we will compare the four different algorithms (Greedy Myopic, Linear UCB, DDPG and IA(GRU)) in terms of the generated revenue for two different settings, a setting with fixed sellers and a setting with variable sellers. The difference is that in the former case, we sample the costs ci once in the beginning whereas in the latter case, the cost ci of each seller is sampled again in each round. This can ei- ther model the fact that the production costs of sellers may vary based on unforeseeable factors or simply that sellers of different capabilities may enter the market in each round. For each one of these two settings, we will compare the four algorithms for each one of the four different rationality models (ϵ-Greedy, ϵ-First, UCB1 and Exp3) separately as well as in a com- bined manner, by assuming a mixed pool of sellers, each of which may adopt a different rationality model from the ones above. The latter comparison is meant to capture cases where the population of sellers is heterogeneous and may consist of more capable sellers that employ their R&D resources to come up with more sophisti- cated approaches (such as UCB1 or Exp3) but also on more basic sellers that employ simpler strategies (such as ϵ-Greedy). Another interpretation is that the distinction is not necessarily in terms of sophistication, but could also be due to different market research, goals, or general business strategies, which may lead to different decisions in terms of which strategy to adopt. Our experiments are run for 200 sellers, a case which already captures a lot of scenarios of interest in real e-commerce platforms. A straightforward application of the reinforcement learning algo- rithms for much larger numbers of sellers is problematic however, as the action space of the MDP increases significantly, which has dras- tic effects on their running time. To ensure scalability, we employ a very natural heuristic, where we divide the impression allocation problem into sub-problems and then solve each one of those in par- allel. We show at the end of the section that this "scale-and-solve" version of IA(GRU) clearly outperforms the other algorithms for large instances consisting of as many as 10.000 sellers. Experimental Setup: In the implementation of DDPG, the actor network uses two full connected layers, a rectified linear unit (ReLu) Figure 1: The framework of the actor network of the IA(GRU) algorithm. seller and employs input preprocessing at each training step, to en- sure permutation invariance. Input Preprocessing: In each step of training, with a state ten- sor of shape (T , m, 4), we firstly utilize a background network to calculate a public vector containing information of all sellers: it transforms the state tensor to a (T , m×4) tensor and performs RNN operations on the axis of rounds. At this step, it applies a permuta- tion transformation, i.e. a technique for maintaining permutation invariance. Specifically, it first orders the sellers according to a certain metric, such as the weighted average of their past generated revenue and then inputs the (state, action) pair following this order to obtain the public vector (pv). On the other hand, for each seller i, it applies a similar RNN operation on its history, resulting in an individual temporal feature called (fi). Combining those two features, we obtain a feature vector (pv, fi) that we will use as input for the sellers' sub-actor and sub-critic networks. Actor network: For each seller, the input to the sub-actor network is (pv, fi) and the output is a score. This algorithm uses a softmax function over the outputs of all sub-actor networks in order to choose an action. The structure of the policy which is shown in Figure 1 ensures that the policy space is much smaller than that of DDPG as the space of inputs of all sub-actor networks is restricted, and allows for easier convergence, as we will show in Section 5. Critic network: For the critic, we make use of a domain-specific property, namely that the immediate reward of each round is the sum of revenues of all sellers and the record of each seller has the same space. Each sub-critic network inputs the expected fraction of buyer impression the seller gets (the sub-action) and (pv, fi) (the sub-state) as input and outputs the Q-value of the corresponding (a) Rewards of DDPG in training (b) Rewards of IA(GRU) in training (a) Loss of DDPG in training (b) Loss of IA(GRU) in training Figure 2: Rewards of DDPG and IA(GRU) in training for ra- tional sellers. Figure 3: Loss of DDPG and IA(GRU) in training for rational sellers. as the activation function, and outputs the action by a softmax func- tion. The critic network inputs a (state,action) pair and outputs the estimation of the Q-value using similar structure. The algorithm IA(GRU) uses the same structure, i.e. the fully connected network in the sub-actor and sub-critic networks, and uses a Recurrent Neural Network with gate recurrent units (GRU) in cyclic layers to obtain the inputs of these networks. For the experiments we set T = 1, i.e, the record of all items of the last round is viewed as the state.5 We employ heuristic algorithms such as the Greedy Myopic Algorithm for exploration, i.e. we add these samples to the replay buffer before training. Experimental Parameters: We use 1000 episodes for both train- ing and testing, and there are 1000 steps in each episode. The valua- tion of the buyer in each round is drawn from the standard uniform distribution U(0, 1) and the costs of sellers follow a Gaussian distri- bution with mean 1/2 and variance 1/2. The size of the replay buffer is 105, the discount factor γ is 0.99, and the rate of update of the target network is 10−3. The actor network and the critic network are trained via the Adam algorithm, a gradient descent algorithm presented in [24], and the learning rates of these two networks are 10−4. Following the same idea as in [27], we add Gaussian noise to the action outputted by the actor network, with the mean of the noise decaying with the number of episodes in the exploration. Convergence of DDPG and IA(GRU) First, to show the difference in the convergence properties of DDPG and IA(GRU), we train the algorithms for 200 sellers using the ϵ- greedy strategy as the rationality model with variable costs for the sellers. Figure 2 shows the comparison between the rewards of the algorithms and Figure 3 shows the comparison in terms of the training loss with the number of steps. The gray band shows the variance of the vector of rewards near each step. From the figures, we see that DDPG does not converge, while IA(GRU) converges, as the training loss of the algorithm decreases with the number of steps. The convergence properties for the other rationality models are very similar. 5We found out that training our algorithms for larger values of T does not help to improve the performance. Performance Comparison In this subsection, we present the revenue guarantees of IA(GRU) in the setting with 200 sellers and how it fairs against the heuristics and DDPG, for either each rationality model separately, or for a heterogeneous pool of sellers, with a 1/4-fraction of the sellers following each strategy. As explained in the previous page, we con- sider both the case of fixed sellers and variable sellers. Performance Comparison for Fixed Sellers: We show the per- formance of DDPG, IA(GRU), Greedy Myopic and Linear UCB on sellers using • the ϵ-Greedy strategy (Figure 4), • the ϵ-First strategy (Figure 5), • the UCB1 strategy (Figure 6), • the Exp3 strategy (Figure 7). We also show the performance of the four different algorithms in the case of a heterogeneous population of sellers in Figure 8. Every point of the figures shows the reward at the corresponding step. We can conclude that the IA(GRU) algorithm is clearly better than the other algorithms in terms of the average reward on all rationality models. We also note that DPPG does not converge with 200 sellers and this is the reason for its poor performance. Performance Comparison for Variable Sellers: We show the performance of DDPG, IA(GRU), Greedy Myopic and Linear UCB on sellers using • the ϵ-Greedy strategy (Figure 9), • the ϵ-First strategy (Figure 10), • the UCB1 strategy (Figure 11), • the Exp3 strategy (Figure 12). We also show the performance of the four different algorithms in the case of a heterogeneous population of sellers in Figure 13. Again here, we can conclude that the IA(GRU) algorithm clearly outperforms all the other algorithms in terms of the average reward on all rationality models. Also, IA(GRU) fairs better in terms of stability, as the other algorithms perform worse in the setting with variable sellers, compared to the setting with fixed sellers. Scalability In this subsection, we present the revenue guarantees of IA(GRU) in the setting with 10000 fixed sellers and how it fairs against the Figure 4: Rewards for fixed sellers and ϵ-Greedy strategies. Figure 7: Rewards for fixed sellers and Exp3 strategies. Figure 5: Rewards for fixed sellers and ϵ-First strategies. Figure 8: Rewards for fixed sellers and heterogeneous strate- gies. Figure 6: Rewards for fixed sellers and UCB1 strategies. heuristics and DDPG to show the scalability properties of IA(GRU) with the number of sellers. For IA(GRU) and DDPG, we will employ a simple "scale-and-solve" variant, since applying either of them directly to the pool of 10.000 sellers is prohibitive in terms of their running time. We design 50 allocation sub-problems, consisting of 200 sellers each, and divide the total number of impressions in 50 sets of equal size, reserved for each sub-problem. We run IA(GRU) and DDPG algorithms in parallel for each sub-problem, which is Figure 9: Rewards for variable sellers and ϵ-Greedy strate- gies. feasible in reasonable time. For the heuristics, we run the algorithms directly on the large population of 10.000 sellers. The results for the case of ϵ-Greedy seller strategies are show in Figure 14 (the results for other strategies are similar). We can see that even though we are applying a heuristic version, the performance of IA(GRU) is Figure 10: Rewards for variable sellers and ϵ-First strategies. Figure 13: Rewards for variable sellers and heterogeneous strategies. Figure 11: Rewards for variable sellers and UCB1 strategies. Figure 12: Rewards for variable sellers and Exp3 strategies. still clearly superior to all the other algorithms, which attests to the algorithm being employable in larger-case problems as well. 6 CONCLUSION In this paper, we employed a reinforcement mechanism design framework for solving the impression allocation problem of large e-commerce websites, while taking the rationality of sellers into Figure 14: Rewards for 10.000 fixed sellers and ϵ-Greedy strategies. account. Inspired by recent advances in reinforcement learning, we designed a deep reinforcement learning algorithm which outper- forms several natural heuristics under different realistic rationality assumptions for the sellers in terms of the generated revenue, as well as state-of-the-art reinforcement learning algorithms in terms of performance and convergence guarantees. Our algorithm can be applied to other dynamical settings for which the objectives are similar, i.e. there are multiple agents with evolving strategies, with the objective of maximizing a sum of pay- ments or the generated revenue of each agent. It is an interesting future direction to identify several such concrete settings and ap- ply our algorithm (or more generally our framework), to see if it provides improvements over the standard approaches, as it does here. ACKNOWLEDGMENTS Qingpeng Cai and Pingzhong Tang were supported in part by the National Natural Science Foundation of China Grant 61561146398, a Tsinghua University Initiative Scientific Research Grant, a China Youth 1000-talent program and Alibaba Innovative Research pro- gram. Aris Filos-Ratsikas was supported by the ERC Advanced Grant 321171 (ALGAME). REFERENCES [1] Sander Adam, Lucian Busoniu, and Robert Babuska. 2012. Experience replay for real-time reinforcement learning control. IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews) 42, 2 (2012), 201–212. [2] Rajeev Agrawal. 1995. Sample mean based index policies by O (log n) regret for the multi-armed bandit problem. Advances in Applied Probability 27, 4 (1995), 1054–1078. [3] Shipra Agrawal and Navin Goyal. 2013. Thompson sampling for contextual bandits with linear payoffs. In International Conference on Machine Learning. 127–135. [4] Peter Auer, Nicolo Cesa-Bianchi, and Paul Fischer. 2002. Finite-time analysis of the multiarmed bandit problem. Machine learning 47, 2-3 (2002), 235–256. [5] Peter Auer, Nicolo Cesa-Bianchi, Yoav Freund, and Robert E Schapire. 1995. Gambling in a rigged casino: The adversarial multi-armed bandit problem. In Foundations of Computer Science, 1995. Proceedings., 36th Annual Symposium on. IEEE, 322–331. [6] 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. [7] Shalabh Bhatnagar, Richard S Sutton, Mohammad Ghavamzadeh, and Mark Lee. 2007. Incremental Natural Actor-Critic Algorithms.. In NIPS. 105–112. [8] Djallel Bouneffouf, Amel Bouzeghoub, and Alda Lopes Gançarski. 2012. A contextual-bandit algorithm for mobile context-aware recommender system. In International Conference on Neural Information Processing. Springer, 324–331. [9] Giuseppe Burtini, Jason Loeppky, and Ramon Lawrence. 2015. A survey of online experiment design with the stochastic multi-armed bandit. arXiv preprint arXiv:1510.00757 (2015). [10] Qingpeng Cai, Aris Filos-Ratsikas, Chang Liu, and Pingzhong Tang. 2016. Mech- anism Design for Personalized Recommender Systems. In Proceedings of the 10th ACM Conference on Recommender Systems. ACM, 159–166. [11] Qingpeng Cai, Aris Filos-Ratsikas, Pingzhong Tang, and Yiwei Zhang. 2018. Reinforcement Mechanism Design for Fraudulent Behaviour in e-Commerce. (2018). [12] Shuchi Chawla, Jason Hartline, and Denis Nekipelov. 2016. A/B testing of auctions. In Proceedings of the 2016 ACM Conference on Economics and Computation. ACM, 19–20. [13] Constantinos Daskalakis and Vasilis Syrgkanis. 2016. Learning in auctions: Regret is hard, envy is easy. In Foundations of Computer Science (FOCS), 2016 IEEE 57th Annual Symposium on. IEEE, 219–228. [14] Peter Dayan and CJCH Watkins. 1992. Q-learning. Machine learning 8, 3 (1992), [15] Thomas Degris, Patrick M Pilarski, and Richard S Sutton. 2012. Model-free reinforcement learning with continuous action in practice. In American Control Conference (ACC), 2012. IEEE, 2177–2182. [16] Dylan J Foster, Zhiyuan Li, Thodoris Lykouris, Karthik Sridharan, and Eva Tardos. 2016. Learning in games: Robustness of fast convergence. In Advances in Neural Information Processing Systems. 4734–4742. [17] Yoav Freund and Robert E Schapire. 1995. A desicion-theoretic generalization of on-line learning and an application to boosting. In European conference on computational learning theory. Springer, 23–37. [18] Sergiu Hart. 2005. Adaptive heuristics. Econometrica 73, 5 (2005), 1401–1430. [19] Sergiu Hart and Andreu Mas-Colell. 2000. A simple adaptive procedure leading to correlated equilibrium. Econometrica 68, 5 (2000), 1127–1150. [20] Sergiu Hart and Andreu Mas-Colell. 2001. A general class of adaptive strategies. Journal of Economic Theory 98, 1 (2001), 26–54. [21] Jason Hartline, Vasilis Syrgkanis, and Eva Tardos. 2015. No-regret learning in Bayesian games. In Advances in Neural Information Processing Systems. 3061– 3069. 279–292. [22] Nicolas Heess, Jonathan J Hunt, Timothy P Lillicrap, and David Silver. 2015. arXiv preprint Memory-based control with recurrent neural networks. arXiv:1512.04455 (2015). [23] Sergey Ioffe and Christian Szegedy. 2015. Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167 (2015). [24] Diederik Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimiza- tion. arXiv preprint arXiv:1412.6980 (2014). [25] Andreas Krause and Cheng S Ong. 2011. Contextual gaussian process bandit optimization. In Advances in Neural Information Processing Systems. 2447–2455. [26] Lihong Li, Wei Chu, John Langford, and Robert E Schapire. 2010. A contextual- bandit approach to personalized news article recommendation. In Proceedings of the 19th international conference on World wide web. ACM, 661–670. [27] Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. 2015. Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971 (2015). [28] Thodoris Lykouris, Vasilis Syrgkanis, and Éva Tardos. 2016. Learning and effi- ciency in games with dynamic population. In Proceedings of the Twenty-Seventh Annual ACM-SIAM Symposium on Discrete Algorithms. Society for Industrial and Applied Mathematics, 120–129. [29] Eric S Maskin. 2008. Mechanism design: How to implement social goals. The American Economic Review 98, 3 (2008), 567–576. [30] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller. 2013. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602 (2013). [31] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. 2015. Human-level control through deep reinforcement learning. Nature 518, 7540 (2015), 529–533. [32] Denis Nekipelov, Vasilis Syrgkanis, and Eva Tardos. 2015. Econometrics for learning agents. In Proceedings of the Sixteenth ACM Conference on Economics and Computation. ACM, 1–18. [33] A Blum PI, M Blum, M Kearns, T Sandholm, and MT Hajiaghayi. [n. d.]. Machine Learning, Game Theory, and Mechanism Design for a Networked World. ([n. d.]). [34] Francesco Ricci, Lior Rokach, and Bracha Shapira. 2011. Introduction to recom- mender systems handbook. Springer. [35] Ariel Rubinstein. 1998. Modeling bounded rationality. MIT press. [36] Weiran Shen, Binghui Peng, Hanpeng Liu, Michael Zhang, Ruohan Qian, Yan Hong, Zhi Guo, Zongyao Ding, Pengjun Lu, and Pingzhong Tang. 2017. Rein- forcement mechanism design, with applications to dynamic pricing in sponsored search auctions. arXiv preprint arXiv:1711.10279 (2017). [37] David Silver, Guy Lever, Nicolas Heess, Thomas Degris, Daan Wierstra, and Martin Riedmiller. 2014. Deterministic policy gradient algorithms. In International Conference on Machine Learning (ICML. [38] Richard S Sutton. 1988. Learning to predict by the methods of temporal differences. Machine learning 3, 1 (1988), 9–44. [39] Richard S Sutton, David A McAllester, Satinder P Singh, Yishay Mansour, et al. 1999. Policy gradient methods for reinforcement learning with function approxi- mation.. In NIPS, Vol. 99. 1057–1063. [40] Pingzhong Tang. 2017. Reinforcement Mechanism Design. In Early Carrer High- lights at Proceedings of the 26th International Joint Conference on Artificial Intelli- gence (IJCAI. 5146–5150. [41] Eva Tardos. 2017. Learning and Efficiency of Outcomes in Games. (2017). Seminar [42] Hado Van Hasselt, Arthur Guez, and David Silver. 2016. Deep Reinforcement Learning with Double Q-Learning.. In AAAI. 2094–2100. [43] Christopher John Cornish Hellaby Watkins. 1989. Learning from delayed rewards. Ph.D. Dissertation. King's College, Cambridge. Slides.
1807.02987	1	1807	2018-07-09T08:45:47	Fair Task Allocation in Crowdsourced Delivery	[ "cs.MA", "cs.HC" ]	Faster and more cost-efficient, crowdsourced delivery is needed to meet the growing customer demands of many industries, including online shopping, on-demand local delivery, and on-demand transportation. The power of crowdsourced delivery stems from the large number of workers potentially available to provide services and reduce costs. It has been shown in social psychology literature that fairness is key to ensuring high worker participation. However, existing assignment solutions fall short on modeling the dynamic fairness metric. In this work, we introduce a new assignment strategy for crowdsourced delivery tasks. This strategy takes fairness towards workers into consideration, while maximizing the task allocation ratio. Since redundant assignments are not possible in delivery tasks, we first introduce a 2-phase allocation model that increases the reliability of a worker to complete a given task. To realize the effectiveness of our model in practice, we present both offline and online versions of our proposed algorithm called F-Aware. Given a task-to-worker bipartite graph, F-Aware assigns each task to a worker that minimizes unfairness, while allocating tasks to use worker capacities as much as possible. We present an evaluation of our algorithms with respect to running time, task allocation ratio (TAR), as well as unfairness and assignment ratio. Experiments show that F-Aware runs around 10^7 x faster than the TAR-optimal solution and allocates 96.9% of the tasks that can be allocated by it. Moreover, it is shown that, F-Aware is able to provide a much fair distribution of tasks to workers than the best competitor algorithm.	cs.MA	cs	IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL. XX, NO. XX, XXX 2018 1 Fair Task Allocation in Crowdsourced Delivery Fuat Basık†∗, Bugra Gedik†, Hakan Ferhatosmanoglu§, Kun-Lung Wu‡ † Department of Computer Engineering, Bilkent University, Turkey § Department of Computer Science, University of Warwick, UK ‡ IBM Research, New York, USA [email protected], [email protected], [email protected], [email protected] 8 1 0 2 l u J 9 ] A M . s c [ 1 v 7 8 9 2 0 . 7 0 8 1 : v i X r a Abstract-Faster and more cost-efficient, crowdsourced delivery is needed to meet the growing customer demands of many industries, including online shopping, on-demand local delivery, and on-demand transportation. The power of crowdsourced delivery stems from the large number of workers potentially available to provide services and reduce costs. It has been shown in social psychology literature that fairness is key to ensuring high worker participation. However, existing assignment solutions fall short on modeling the dynamic fairness metric. In this work, we introduce a new assignment strategy for crowdsourced delivery tasks. This strategy takes fairness towards workers into consideration, while maximizing the task allocation ratio. Since redundant assignments are not possible in delivery tasks, we first introduce a 2-phase allocation model that increases the reliability of a worker to complete a given task. To realize the effectiveness of our model in practice, we present both offline and online versions of our proposed algorithm called F-Aware. Given a task-to-worker bipartite graph, F-Aware assigns each task to a worker that minimizes unfairness, while allocating tasks to use worker capacities as much as possible. We present an evaluation of our algorithms with respect to running time, task allocation ratio (TAR), as well as unfairness and assignment ratio. Experiments show that F-Aware runs around 107× faster than the TAR-optimal solution and allocates 96.9% of the tasks that can be allocated by it. Moreover, it is shown that, F-Aware is able to provide a much fair distribution of tasks to workers than the best competitor algorithm. Index Terms-spatial crowdsourcing, crowdsourced delivery, fairness. ! 1 INTRODUCTION Spatial crowdsourcing has emerged as a viable solution for delivery logistics, such as on-demand local delivery, online shopping, and on-demand transportation [10]. As such, it has attracted significant attention from both the academia and the industry in recent years. For instance, Amazon utilizes the crowd to provide same day shipment of packages from warehouses to customers 1. Postmates2, a company offering on demand food and delivery, is available all around the US. Enormous growth of the crowdsourced taxi services Uber3 and Lyft4 has attracted significant inter- est, resulting in numerous research studies being conducted about them. Crowdsourced delivery applications have three stake- holders: customers, workers and the matching platform. Customers submit tasks of spatial deliveries to the plat- form. The platform matches the tasks with the workers' availabilities, and allocates workers to tasks considering the spatio-temporal requirements. To support faster and cheaper delivery, spatial crowdsourcing platforms require a critical mass of workers. The workers should be attracted by a high income potential which is possible with a large number of customers. This situation drives these platforms into a chicken and egg problem [29], in which a powerful network is necessary to attract customers and customers are necessary to engage a powerful network. * Part of this work was done while the author was an intern at IBM. 1. http://flex.amazon.com 2. http://www.postmates.com 3. http://www.uber.com 4. http://www.lyft.com A negative correlation between job satisfaction and worker turnover is naturally expected in crowdsourcing environments. According to a study with MTurk workers, a common indicator of positive behavior of the employer, hence the job satisfaction, is fairness [7]. Fairness can be defined in the context of anti discrimination laws, equity of opportunity and equality of outcome [37], [3]. In the context of crowdsourcing, the distributive fairness is particularly relevant [16], [17]. This definition seeks fairness based on the proximity between a worker's own input/output ratio and the input/output ratio of a referent [1]. For example, the workers would expect to be assigned a fair number of tasks that is proportional to their spatio-temporal matching qualities/availabilities for the tasks. Effect of such fairness expectations on the likelihood of participation is more than that of considerations of self-interest [17]. Hence, fairness needs to be considered as an essential concept for sustaining a powerful crowd with significant participation of work- ers [5]. In this paper, we study the problem of fair allocation of delivery tasks to workers within the context of spatial crowdsourcing. The tasks are associated with receive and delivery locations and time constraints. The workers inform the platform about their working status using availabilities, i.e., the location and time period they are willing to serve. While the primary objective is to maximize the task allocation ratio (TAR), which is the ratio of number of allocated tasks to number of all tasks, we aim to achieve this via a fair distribution of tasks to workers. Current approaches focus only on the first objective of maximizing the number of tasks under certain constraints from workers [21]. While this reduces the use of third party services or employing IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL. XX, NO. XX, XXX 2018 2 full time couriers [29], it ignores fairness and worker satis- faction. This can result in lower engagement and migration to other services. [16]. There is a number of challenges to achieve an effective and fair crowdsourced delivery. First, unlike other crowd- sourcing applications, a redundant allocation of tasks is not possible in the crowdsourced delivery. Redundant allocation improves the reliability of task completion by increasing the number of workers the task is assigned to [10][21]. In delivery tasks, however, only one worker can be allocated to complete the task. Hence, to increase the reliability of the worker selection in such tasks, we employ a 2-phase allocation model. In the first phase, the platform selects a set of nominees among available workers and the task is offered to a subset of them. In the second phase, the platform selects one worker among those who accepted the offer. This avoids the broadcasting of the offer to all nominees in the first phase to forestall spamming of the workers. The second challenge is to provide distributional fairness among workers to ensure participation. Unless workers and their availabilities are identical, assigning each worker an equal number of tasks is not a fair distribution, as it does not take the user input into consideration. To address this issue, we call the input/output ratio local assignment ratio and set the output of a worker as the amount of revenue she gets from the system, while the input is the total reward of the offers she has accepted (not necessarily allocated). This view allows us to define a technical measure of the global fairness as the coefficient of variation, a statistical measure of relative variability, of all local assignment ratios. A low coefficient of variation is associated with the fairness of allocation. Without considering fairness, the task allocation problem can be reduced to the minimum cost flow (MCF) prob- lem [21]. However, the MCF-based solutions fall short to capture fairness, since every assignment needs to update its cost matrix. We introduce F-Aware as a solution to assign tasks to workers in a bipartite graph, by minimizing the un- fairness locally and allocating tasks to fill worker capacities. The third challenge is to handle online allocation. In applications such as online shopping (e.g., with a 3-hour delivery guarantee), the platform can allocate multiple tasks in mini-batches with no global knowledge of all tasks and availabilities in advance. In contrast, in on-demand trans- portation services, like Uber and Lyft, customers want to know whether the vehicle is on the way, almost instantly. Therefore the allocation should be done at the very moment of the task arrival. F-Aware with the 2-phase model is shown to be applicable for offline, online, and mini-batch allocation strategies. In summary, this paper makes the following contribu- tions: • Model. We introduce a generic task allocation model to cover a variety of crowdsourced delivery sce- narios. The 2-phase allocation model increases the reliability of task completion by double-checking a worker's willingness to complete the tasks. This model handles the case where a potential worker may refuse the task even though she is available. • Algorithm. We introduce a fairness-aware solution called F-Aware, which locally minimizes unfairness, while targeting maximum task allocation. The MCF- based algorithms fall short on modeling the dynamic fairness metric, and are not feasible for the online scenarios. • Online Allocation. We enhance our 2-phase model to perform online task allocation. We show that F- Aware is effective also for online and mini-batch allocation scenarios, as it is for offline allocation. We provide a comprehensive experimental study using real-world datasets to showcase the effectiveness and effi- ciency of our 2-phase model and of the F-Aware algorithm in terms of running time, task allocation ratio, and fairness it achieves. The rest of this paper is organized as follows. Section 2 gives the preliminaries of the problem. Section 3 explains the details of our 2-phase task allocation model. Section 4 extends our approach to online task allocation. Section 5 presents the experimental evaluation. Section 6 discusses the related work. Finally, Section 7 concludes the paper. 2 PROBLEM DEFINITION We aim to develop a new strategy on allocation of tasks to workers in a crowdsourced environment. The overall goal is to maximize the task allocation ratio (TAR), the ratio of number of allocated task over the number of all tasks, while distributing tasks to workers fairly. We now give the preliminaries of the domain and formalize this multi-criteria optimization problem. Definition 1. Time period. A time period, h, is a pair of date- time values b and e, representing the beginning and end times, respectively. Definition 2. Delivery task. Tasks are in the form of spatio- temporal deliveries, such as workers need to move to the source of the delivery to receive the item and deliver it to the recipient. In this manner, one can consider a task as a composite of receive and deliver steps. We represent the set of all tasks with T , and ith task with ti. A task is a quintuple {hs, ls, hr, lr, m}. Here, h and l represents a time period and a geo-spatial point such as a latitude/longitude pair, respectively. Subscripts s and r stands for the source and the destination of the task. In other words, ti.ls stands for location of the source while ti.lr stands for location to deliver the item for ith task ti. Note that these steps are associated with different time periods, as receive and deliver steps have their own validity periods hs and hr. Lastly, m represents the reward of the task. Definition 3. Worker. Workers are people who participate in the platform to make money. We represent the set of all workers with W, and ith worker with wi. Each worker, is a triple {A, c, f}. wi.A is the set of availabilities of worker wi, wi.c is her capacity and wi.f is the local assignment ratio. Local assignment ratio, which will be detailed later, is a dynamic metric used to determine how fair the system treated a worker so far, defined as the ratio of the worker's revenue over the total reward of the offers she has accepted. The revenue of the worker equals to sum of rewards of the tasks allocated to her. IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL. XX, NO. XX, XXX 2018 3 Fig. 1: Tasks and availabilities are the inputs of the platform. For each task, the nominees are identified. During this calculation, the system also finds out the acceptance probability, based on Equation 8. The task is multicasted to k of those nominees. Workers who accept the offer are referred to as candidates and one worker among the candidates is selected for allocation. Definition 4. Availability. Workers inform the system about their working status using availabilities. aip is the pth availability of ith worker wi, such that aip ∈ wi.A. Each availability is a triple, {h, l, r}. aip.l is a geo-spatial point. It is the center of the region in which wi is willing to accept tasks. aip.r is the radius of the same region. Let us represent this region with the Cr(l, r) function. The worker is ready to serve during the time period aip.h. For a task, ti, to be completed, a worker has to move to the receive location, ti.ls, during its validity period ti.hs; and after that, she has to move to the deliver location, ti.lr, during its validity period ti.hr. An example task ti in crowdsourced delivery could be to pick-up a gift item from a local shop (ti.ls) between 12:00 and 14:00 on 23rd of December 2016 (ti.hs), and deliver it to a home address (ti.lr) between 18:00 and 22:00 on the same day (ti.hr). In return for completing this task, the worker will be paid ti.m amount of money. In the real world, multiple parameters effect the ti.m value, including the distance between ti.ls and ti.lr, the size of the package, and the sensitivity of the content, etc. 2.1 Fairness An effective network is key to building a powerful crowd- sourcing platform, therefore, providing continuous partici- pation of workers and avoiding worker turnover are crucial. A negative correlation is naturally expected between job satisfaction and worker turnover in crowdsourcing envi- ronments. According to a study with MTurk 5 workers, 11 to 26 percent of turnover in crowdsourcing environment is explained with job satisfaction. In the same study, fairness is listed as one of the most common indicator of the job satisfaction [7], [30]. Fairness needs to be considered as a first class citizen in designing crowdsourcing applications to ensure long term commitment and participation [5]. There are three major forms of fairness defined in the social psychology literature, namely: procedural, interactional, and distributive. Procedural fairness is the perception of jus- tice on the procedures, policies, and the criteria used by the decision maker [16]. Interactional fairness is the interper- sonal aspect of the procedural fairness. Distributive fairness 5. http://www.mturk.com/ is defined as the proximity between a worker's own in- put/output ratio and the input/output ratio of a referent [1]. Prior research on the relationship between fairness and job satisfaction shows that when fairness is regressed along all three dimensions, the job satisfaction gets impacted the most due to the loss of distributed fairness [30]. Note that, unless the workers are identical, assigning each worker to an equal number of tasks is not a fair distribution by this definition, as it does not take user input into consideration. Therefore, we define a new fairness model that captures distributive fairness, which will be detailed shortly. 2.2 Formalization With the given definitions, let us first define the problem before discussing each component separately. Fair allocation of delivery tasks in a crowdsourcing environment: Given the set of delivery tasks T and the set of workers W, represented with their availabilities, the problem is allocating tasks to workers with the goals of maximizing the task allocation ratio (T AR) and minimizing the unfairness (F) objectives, under the candidacy, capacity and assignment ratio (AR) constraints which will be explained next. Task Allocation Ratio (TAR): To reduce the dependency of the businesses to using a third party service we set maximizing task allocation ratio (TAR) as the first component of our objective function. T AR is defined as the ratio of the number of allocated tasks over the number of all tasks. Formally, let Tall be the set of allocated tasks and T be the set of all tasks. The T AR, defined as: Tall T T AR = (1) Unlike other crowdsourcing applications, redundant al- location of tasks is not possible in crowdsourced delivery. Therefore, in order to increase the reliability of a worker completing a given task, the allocation is done via a 2-phase procedure, illustrated in Figure 1. In the first phase, the system nominates a set of workers whose availabilities are suitable to complete the task. The task is offered to these IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL. XX, NO. XX, XXX 2018 4 nominees, and they have an opportunity to accept or reject it. A worker may refuse the offer, even though she is available. For brevity, we leave the details of nomination and offering strategy to the next section. Workers who have accepted the offer are called candidates. In the second phase, one worker among the candidates is selected and is assigned to the task. Fairness: Recall that distributive fairness is defined as the proximity between a worker's own input/output ratio and the input/output ratio of a referent. In the 2-phase allocation model, the input of a worker is the total reward of the offers she accepted. Note that availabilities cannot be used as input since a worker might reject offers even though she is available. The output of a worker, on the other hand, is the amount of money she earned. Intuitively, not all tasks have the same complexity, yet the reward of each task is proportional to its hardness. To capture the hardness of the tasks while determining a worker's input/output ratio, instead of counting the number of tasks a worker accepted or got assigned, we use the reward of each task. Each worker wi is associated with a local assignment ratio (LAR), w.fi, defined as the ratio of the total reward of tasks allocated to a worker (output) over the total reward of tasks she has accepted (input). Formally, let wi.Tall be the set of tasks allocated to worker wi, and wi.Tacc is the set of offers she has accepted, the LAR of wi is defined as: (cid:80)x (cid:80)y j=1 tj.m k=1 tk.m wi.f = where , ∀j,k tj ∈ wi.Tall, tk ∈ wi.Tacc (2) x = wi.Tall and y = min(wi.Tacc, wi.c) (3) The number of tasks a worker accepts can exceed her ca- pacity, since she will be assigned to a subset of the tasks she has accepted. However, since the capacity limits the number of allocations, we consider minimum of {wi.Tacc, wi.c} acceptances when calculating the denominator. The system is considered more fair as the proximity of the LAR values will be higher. Although the standard deviation of the set of LAR values represents this proximity, using it as the evaluation metric would be misleading since the different allocation schemes will have different number of tasks allo- cated. Hence, we evaluate the overall fairness of the system, F, using coefficient of variation of the set of LAR values, i.e., standard deviation of the LAR values divided by their mean. Let F be the set of all local assignment ratio values of users system fairness, F, is formalized as: F = σ(F ) µ(F ) (4) Candidacy Constraint: Recall wi.Tacc is the set of tasks worker wi has accepted. A given task tj can be assigned to worker wi only if the system nominated her for the task, and she accepted the offer: ti ∈ wi.Tacc. Capacity Constraint: Definition of the capacity constraint is intuitive. The number of tasks assigned to a worker cannot exceed her capacity: wi.Tall ≤ wi.c. Assignment Ratio Constraint: Note that each task is offered TABLE 1: Commonly Used Notation Notation h [b, e] w{A, c, f} t{hs, ls, hr, lr, m} aip{h, l, r} T , W, A, F Tall, wi.Tall, wi.Tacc Explanation time period; beginning from b, ending at e worker; consists of set of availabilities (w.A), capacity (w.c) and local assignment ratio (w.f) task; consists of source location (t.ls), source time period (t.hs), destination location (t.lr), destination time period (t.hr) and a reward (t.m). pth availability of ith worker wi consists of a time period h, a location l and radius r set of all tasks, workers, availabilities and local assignment ratios respectively. set of allocated tasks, allocated tasks to worker wi and tasks accepted by wi. to a set of nominees, yet among the candidates who accept the offer, only one worker will be selected. To forestall spamming of workers and avoid unnecessary communi- cation costs, we avoid broadcasting offers to nominees. At the one extreme, the task can be repeatedly unicasted until one nominee accepts it. However, this approach would cause potentially long wait times. Therefore, we present a hybrid solution: multicasting the offer to k workers, which avoids spamming of the workers while increasing the probability of at least one nominee accepting the offer. The value of k is calculated for each task independently. In the next section, we show that higher values of k will result in higher number of candidates, but it leads to a large set of spammed workers. To be able to define an upper limit, we introduce a system- wide metric called assignment ratio (AR), which is the ratio of the number of allocated tasks over the number of accepted offers. Higher AR indicates more accurate nominee selection, or less number of spammed workers. Therefore, we constraint assignment ratio to be higher than a predefined threshold θ. We formalize the AR as follows: AR = Tall (cid:80)W i=1 wi.Tacc (5) And define assignment ratio constraint as AR ≥ θ. Later, we will discuss relaxing this constraint to decrease the wait time of the customers. 3 ALLOCATION MODEL In this section, we describe the details of the 2-phase allo- cation model. Inputs to the system are the tasks from the customers, and the availabilities from the workers, both with time and location components. In many practical cases, both the tasks and the availabilities are registered in advance. Hence, one needs to check if a worker is still willing to do the job. Our model is able to do this check to increase the reliability of the worker with respect to the completion of the given task. In the next section, we also cover the online allocation scenario, in which tasks and availabilities can appear anywhere, anytime. In the remaining parts of this section, we give the details of the 2-phase model in three steps: (i) nomination of workers, (ii) batched progressive offer strategy, and (iii) task allocation. IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL. XX, NO. XX, XXX 2018 5 Fig. 2: Movement cost calculation of a task is different de- pending on the number of availabilities satisfying it. Roads are numbered with respect to travel order. 3.1 Nomination of Workers Initial step of the 2-phase allocation procedure is to find out the appropriate set of workers to offer the task. These workers are called nominees. For a worker wi to be nomi- nated to a task tj, her availability set, wi.A, should contain the necessary availabilities to satisfy both steps of tj. Recall since tasks are in the form of spatio-temporal deliveries, one can divide them into two: receive and deliver steps. Satisfying one of these parts, lets assume receive, means that worker wi should have at least one availability aip, such that time period of task tj.hs intersects with time period of availability aip.h and source location tj.ls lies inside the region Cr(aip.l, aip.r). One can formalize the satisfaction relation between a step, receive or deliver, of a task tj and an availability aip of a worker as a function S: S(tj.hs, tj.ls, aip) ≡ (tj.hs ∩ aip.h (cid:54)= ∅)∧ (tj.ls ∈ Cr(aip.l, aip.r)) (6) The intuition behind dividing the tasks into two steps is that, a worker can have separate availabilities such that one satisfies the requirements of the receive step and the other satisfies the deliver step. Given the set of workers, the system searches for those workers that have satisfying availabilities for both steps of a task. To formulate this relation, we employ the S function to locate the pair of availabilities of a worker wi that satisfy the receive and deliver steps, respectively. We denote the resulting function as N (tj, wi). If no such pair can be located, then the function produces an empty set. {aip, aiq} ∅ N (tj, wi) = ∃ aip ∈ wi.AS(tj.hs, tj.ls, aip)∧ ∃ aiq ∈ wi.AS(tj.hr, tj.lr, aiq) otherwise (7) We should note that, in Equation 7, p and q values can be equal, which means a single availability might satisfy both requirements. It would be unrealistic to assume that all nominated workers will have the same probability to accept the of- fered task. Besides availability, there are many factors that influence such a decision. Existing research showed that workers are willing to accept the tasks that are less costly for them and closer to their home locations [2]. Therefore, the acceptance probability of each nominated worker for a given task is negatively correlated with the task's cost, which is the movement cost in our case. Many existing local delivery systems use the distance between the source and the destination as an indicator of the payment amount. It means that, tasks which require longer traveling pay more to their workers. On the other hand, the movement cost of a task might differ between workers, as they should move towards the source from their current location or move back to their previous location from the destination. Since we do not track the workers' locations, we assume that a worker is at the location provided as part of her valid availability and she tends to go back to that location, after completing the deliver step. We also assume that the acceptance does not depend on the previous acceptances or rejections. Figure 2 shows two different scenarios regarding the calculation of the movement cost of a task. Let us define the movement cost of a task as α + β, and let d be the func- tion that calculates the distance between two geo-spatial points. The distance between source and the destination, given by α = d(tj.ls, tj.lr), remains the same no matter which worker is assigned to the task. On the other hand, the distance traveled towards the source and from the destination, β, depends on the worker availabilities. In the first scenario, single availability of the worker, aip, satisfies both of the steps of the task. Therefore, the worker will only move towards the source, and from the destination (β = d(aip.l, tj.ls) + d(tj.lr, aip.l)). In the second scenario however, after moving towards the source, as well as after delivering the item, worker will go back her initial posi- tion. Moreover, distances between the locations of those availabilities should be considered as well. Consequently, workers movement in the second scenario would be equal to β = 2 × d(aip.l, tj.ls) + 2 × d(tj.lr, aiq.l) + d(aip.l, aiq.l). The reward of the task is positively correlated with α and indifferent to who completes the task. Minimum value for the total travel distance is 2 × α, which happens when the worker is already at the source location, or at the destination location, or lies on the linear line connecting these two points, i.e. α = β. Intuitively, workers are will- ing to accept tasks with high income and less movement. Therefore, acceptance probability of a task tj by worker wi, denoted as R(tj, wi), is positively correlated with α, but negatively correlated with β. For the ideal case, α = β acceptance probability should be 1. We formalize this model in Equation 8. To take into account the probability of a worker refusing an offer even when she is a perfect match, we use a constant c, where 0 ≤ c ≤ 1. R(tj, wi) = eα−β × c (8) Since α ≤ β and 0 ≤ c ≤ 1 the value of R is guaranteed to lay between 0 and 1. Selection of nominees for a task tj outputs the set L(tj,W), which contains the worker wi and acceptance probability R(tj, wi) pair. This set is used in our batched progressive offer strategy, a technique we implement to minimize the waiting time, while maximizing the assignment ratio. L(tj,W) = {wi, R(tj, wi)wi ∈ W ∧ N (tj, wi) (cid:54)= ∅} (9) Let A be the set of all availabilities. Brute force approach to construct L(tj,W) for a given task tj ∈ T iterates IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL. XX, NO. XX, XXX 2018 through the availabilities set A and calculates R(tj, wi) if N (tj, wi) (cid:54)= ∅, for each worker wi ∈ W. For all tasks, complexity of this calculation is O(T A). To decrease the computation time, we index availabilities on the temporal dimension using interval trees. For each task, we query this index two times, one for the receive step, and one for the deliver step. We check the spatial intersection only for the resulting availabilities. In the Online Task Allocation section, we discuss how long to wait between task arrival and nominee selection processes. 3.2 Batched Progressive Offer Strategy There are three different approaches regarding offering a task to a predefined set of nominees. The first approach is broadcasting the offer. To forestall spamming of workers and avoid unnecessary communication costs, we avoid broadcasting offers to nominees. Moreover, we want to minimize the number of cases where a worker accepts the incoming offer but is not allocated to the task, as this might cause churn over time. Therefore, the ideal case is when there is exactly one candidate. Reaching this ideal case is only possible with unicasting the offer. However, there is the probability of the nominee rejecting the offer, even though she is available. Therefore, the system should follow a progressive approach, by offering the task one by one, until somebody accepts, and waiting for a preset time between each round. Obviously, this will result in a long waiting time for the customer to see if his task is going to be served or not. To avoid both situations, we multicast the offer to k nominees in batches, progressively, until there is at least one candidate. With this approach, we decrease the waiting time of the customer, while limiting the number of candidates. For each task, once the value of k is calculated, it is used for further batching, if necessary. Let us call the probability of the task being accepted by at least one nominee as probability of response. For each task tj, the probability of response, when offered to k nominees, can be calculated using the probability of all k nominees refusing it. Recall that L(tj,W) is the set of nominee and acceptance probability pairs, and assume that it is sorted by the acceptance probability values. Probability of response for a batch of k workers is calculated as follows. P (k, L(tj,W)) = 1 − k(cid:89) (1 − (R(tj, wi))) (10) i=1 By keeping the value of P (k, L(tj,W)) above a tuning parameter , the lower bound of the k can be defined. We call , the threshold of probability of response. The upper bound, however, is calculated using the as- signment ratio as a constraint. To limit the value of k, first we should be able to predict how many nominees are likely to accept the offer. In the worst case for assignment ratio all k nominees accept the offer. As we try to maximize the assignment ratio, the number of candidates should not exceed a certain number. Although it is not realistic to expect that all k workers will accept the offer, it is still useful to limit the k value. Given a task tj, let us define the probability of ith worker accepting the offer as a random variable xi. Then 6 1 x1+x2+...+xk the expected value of it is E[xi] = R(tj, wi). Transforming into multiple workers, assignment ratio would be one over the number of candidates, therefore, it would be equal to ]. From probability theory, it is known that E[ ] ≥ E[x1+x2+...+xk] , thus the latter could E[ be used as a lower bound for the expected value, where, E[x1 + x2... + xk] = R(tj, w1) + R(tj, w2) + ...R(tj, wk). With this at hand, let us define a function E(k, L(tj)) as a lower bound for the expected value of the assignment ratio: x1+x2+...+xk 1 1 E(k, L(tj,W)) = 1 E[x1 + x2 + ... + xk] (11) Recall one of the constraints is keeping the assignment ratio of the system above a predefined threshold θ. We use the same threshold, as a lower bound to function E(k, L(tj,W)) to satisfy assignment ratio constraint locally, for each allocation. Although local satisfaction is stronger than the global constraint, as we will discuss next, the assignment ratio constraint is relaxed when it contradicts with the probability of response. While the probability of response P (k, L(tj,W)) is pos- itively correlated with k, as it is shown above, the expected value of the assignment ratio (value of E(k, L(tj,W))) de- creases with it. Therefore, bounding E function from below, sets an upper bound on the value of k. With the above defined thresholds, the value of k should guarantee that probability of response P (k, L(tj,W)) is above , while the assignment ratio constraint is satisfied for each task, i.e. E(k, L(tj,W)) is above θ. In summary, k is selected using following inequality: k ≥ min{k(P (k, L(tj,W)) ≥ } k ≤ max{kE(k, L(tj,W)) ≥ θ} (12) k is set to the maximum value that satisfies both of those inequalities. However, the given inequality might be invalid with respect to selection of the and θ values. Consider that, even the smallest k value satisfying the upper inequality might not satisfy the lower one. In that case, we relax the assignment ratio constraint, and use the k that satisfies the upper inequality. After this calculation, the task is offered to the first k workers and the system waits for a predefined time. Recall that, workers are sorted in the decreasing order of the acceptance probability. In case of all nominees refuse the offer, the task is offered to the next k workers, until there is at least one candidate or all nominees are asked. 3.3 Task Allocation The last step of the 2-phase model is to select one worker among the candidates to allocate the task. Figure 3 shows an example scenario in which there are three tasks and three candidates. The edges are from tasks to their candidates. For example t1 is accepted by w1 and w2, while w1 is the only candidate for t2. The simplest version of this problem is finding a one-to-one assignment scheme of tasks to workers, assuming all the edges have equal weight. By definition, this is the bipartite graph assignment prob- lem. In our specific case, there is a one-to-many relationship between the tasks and the workers. Moreover, to capture the spatial aspect of the problem, one might want to use IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL. XX, NO. XX, XXX 2018 7 Alg. 1: F-Aware Task Allocation Data: G(E, V ): Task-Worker Bipartite Graph s.t. V = T ∪ W Result: G(cid:48)(E, V ):Updated Bipartite Graph G(cid:48)(E, V ) ← G(E, V ) (cid:46) Copy the original graph T ← sort(T ) (cid:46) Tasks sorted in increasing order of in-degree for t ∈ T do (cid:46) For each task in task list (cid:46) Get the subset of workers sharing an edge with t W(cid:48) ← G(E, V ).get(t) W(cid:48) ← sort(W(cid:48)) (cid:46) Workers sorted in increasing LAR order isAssigned ← F alse index ← 0 while ¬isAssigned && index < W(cid:48) do (cid:46) Assignment is not done but there are still candidates w ← W (cid:48).get(index) if w.c > 0 then (cid:46) If worker has capacity G(cid:48)(E, V ) ← G(cid:48)(E, V ) ∪ assign(t, w) (cid:46) Mark edge prof it(w) ← prof it(w) + t.m (cid:46) Update profit w.c ← w.c − 1 (cid:46) Decrease capacity isAssigned ← T rue index ← index + 1 (cid:46) Return the modified bipartite graph else return G(cid:48)(E, V ) one. Given the task-to-worker bipartite graph G(E, V ), s.t. V = T ∪ W, it considers tasks in the increasing order of the node in-degree. Tasks with less candidates are placed first, as the ones with higher degree have more flexibility. For each task, each worker with remaining capacity higher than 0 is considered as a candidate and the one with the lowest local assignment ratio is selected. If there are more than one worker with the same LAR, the one with the higher denominator is selected. Recall that if the capacity of a worker, wi.c, is lower than the number of tasks she has accepted, we only consider the first wi.c offers. The process continues until all tasks are visited. When a task tj is assigned to a worker wi, the input graph, the total earnings, and the residual capacity of wi are updated accordingly. Our experimental results confirm the effectiveness of F- Aware in terms of running time, task allocation ratio, and fairness it achieves. 4 ONLINE TASK ALLOCATION While in offline allocation based applications, all the tasks and availabilities are known in advance, in real-time en- vironments, they can appear at anytime, anywhere [36]. Consider an example scenario, in which a customer would like to travel to airport and asks the crowd for a ride. Overnight reply to this request would be too late, as the user would take a taxi after waiting for a relatively small amount of time. In fact, these dynamically arriving requests require online processing. Therefore, the problem of allocating tasks to workers in a dynamically changing environment raises. While aforementioned example requires instant response to a customer, there are also some crowdsourced delivery applications that allow decision maker system to wait for a period, before allocating the task. Online shopping is an example of such applications. In this case, a seller can wait to group deliveries by their destinations and allocate only one worker for multiple packages. During this waiting period, system collects mini-batches of tasks and worker availabilities for a period and they are processed against Fig. 3: Allocation of tasks to candidates edge weights to represent movement cost and add capacity to workers as well. Including these additional constraints, the task-to-worker bipartite graph can be represented as a flow graph by adding a source and a sink. With this repre- sentation, task to worker allocation can be reduced to the minimum cost flow (MCF) problem and solved optimally with well known algorithms, i.e. successive shortest path or cycle canceling [21]. Even though they reach optimal result on task allocation, minimum cost flow based solutions have their own draw- backs. First of all, running time of the optimal solution is far from being feasible because of high complexity. Given the task to candidate bipartite graph G(E, V ), s.t. V = T ∪ W and E represents the acceptances, the complexity of a care- ful implementation of successive shortest path algorithm is O(V ElogV ). Second, with the aforementioned defini- tion of fairness, it is hard to integrate it into the MCF-based algorithms. This is because at each iteration LAR values are updated, which updates the cost-matrix as well. Assuming that LAR values are static, introducing fair- ness as a new constraint transforms the problem to mini- mum cost maximum flow problem. The goal of this prob- lem is to select the minimum cost flow among multiple maximum flows. In our setting, this corresponds to max- imizing task allocation ratio while minimizing unfairness objectives. This problem could be solved with Hungarian algorithm, however, similar to the previous discussion, the high complexity of the algorithm, O(n3), makes it inefficient for online allocation scenarios [24]. The naıve approach for allocation is random selection of one candidate, which is used as one of the baselines in our experimental evaluation. On the other hand, to capture the spatial aspect of crowdsourced delivery, it is beneficial to allocate the nearest worker [21]. This approach can be extended by proactive allocation of workers, if distribution of tasks is known in advance [25]. Last but not the least, inspired from file allocation techniques from the operating systems literature, selecting the Least Allocated Worker first could increase the Task Allocation Ratio by reserving room for further allocations. While all approaches are feasible in terms of their running time, our experimental evaluation shows that they either fell short in terms of modeling the fairness or have subpar performance with respect to task allocation when the capacity is constrained. To cope with the aforementioned challenges, we intro- duce an algorithm called F-Aware, given in Algorithm 1. It is a greedy algorithm that allocates tasks to workers one by IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL. XX, NO. XX, XXX 2018 8 each other when the period expires. In this section, we present our modifications on the 2-phase allocation model to cover both instant and mini-batch allocation scenarios. Figure 1 shows the illustration of the workflow for the 2-phase allocation model. Our first modification for adapt- ing this workflow to online task allocation is adding two windows on the selection of nominees. These windows are for the tasks and for the availabilities, respectively. They are neither sliding nor tumbling windows. When a task or availability arrives, it is appended into the corresponding window. An availability is removed from the window when it expires. A task on the other hand can be removed under two conditions: i) it is assigned to a worker, ii) it expires. However, waiting until expiration of a task before allocation might cause misses, as all satisfying availabilities might expire meanwhile. Therefore, we also define a window size, in terms of minutes. For every window expiration, the tasks are processed against availabilities and nominees are identified. After this step, batched progressive offer and task allocation steps are used as they are. There are two corner cases about tasks. First, if a task has no nominees at the time of the window expiration, it stays in the window and participates in the following window expirations, until its own expiration. Second, if the task has nominees but all of them rejected the offer, it again remains in the window for new nominees to arrive. The same task is never offered to the same nominee more than once. There are multiple constraints on the window size deci- sion. First of all, it should be shorter than the smallest time period of the set of all tasks and availabilities to guarantee processing. Second, the waiting time should not exceed reasonable response time of the application. An online shopping application that needs 2-hours delivery guarantee cannot define the window size as 3 hours. For applications that require instant reply, the window size can be set to 0. In instant task allocation, when a task arrives, it is processed against the availabilities window to identify nominees. If there are not any, it is added to the tasks window and stays there until its expiration time. When an availability arrives, all tasks and availabilities present in the window are processed, since this availability might be completing a partial match. A partial match is possible when a worker has a satisfying availability for only one of the steps of the task. We leave the detailed discussion of partial match processing as future work. In our experimental evaluation, we study the feasibility and effectiveness of our allocation model with various window sizes, including instant allocation, i.e. window size equals to 0. 5 EXPERIMENTAL EVALUATION In this section, we present the detailed evaluation of our proposed 2-phase allocation model and the fairness-aware task allocation algorithm, F-Aware. To easily determine the superiority of a solution over other solutions, we combine optimization goals into a single parametric objective function O and define it as: O = T AR × e−(ρ∗F ) (13) Since the goals are maximizing the TAR while minimiz- ing the unfairness, the objective function is proportional to TAR. But the exponential part of it is inversely proportional with the global unfairness metric F. To enable the system to prioritize one component of the objective over the other, we introduce the parameter ρ, where 0 ≤ ρ ≤ 1. By setting ρ to 0, one can simplify this objective to task allocation ratio only. Higher values of it will increase the importance of the unfairness in the overall objective. The evaluation includes four sets of experiments. In the first set, we compare F-Aware with 4 other competitor algorithms in terms of task allocation ratio (TAR), unfairness and value of the objective function. Recall that the objective function, as defined in Eq. 13, is the combined metric of TAR and unfairness. The naıve approach of assigning tasks to workers is random selection among the candidates, re- ferred to as Random in the performance graphs. The second approach is to select Least Allocated Worker First(LAF). The intuition behind this approach is trying to reserve room for further task allocations. Existing work of task allocation in spatial crowdsourcing mostly use Nearest Neighbor Priority strategy [21], [25] to capture the spatial-aspect of the prob- lem. [21] introduces allocation techniques based on location entropy, and [25] extends nearest worker priority technique with pro-active deployment of workers to geo-grids. We only prioritize the nearest worker since we do not make assumptions on distribution of tasks. This algorithm is re- ferred to as Nearest in this section. Lastly, we use successive shortest paths algorithm which solves the minimum cost flow(MCF) problem [23]. This algorithm is optimal on Task Allocation Ratio, hence it is used to evaluate TAR of all algo- rithms. The second set of experiments studies the efficiency of our batch incremental offer strategy, that is how different values of and θ affect the k value, thus assignment ratio, and unfairness. The third set studies online task allocation, presenting task allocation ratio and unfairness as a function of window size. Finally, the last set is the sensitivity study that presents task allocation ratio and unfairness of different time period lengths as a function of coefficient of mean. As we will detail soon, length of the time period and coefficient of the mean are two variables we use to adapt real-world data to our setup. We implemented all algorithms using Java 1.8. All exper- iments were executed on a Linux server with 2 Intel Xeon E5520 2.27GHz CPUs and 64GB of RAM. Dataset. Experiments are performed using two real- world datasets. The first dataset contains the Foursquare check-ins from New York City for the month of May 2012 [34]6. This dataset contains around 50,000 check-ins from 987 different users. The second dataset is a taxi-trip dataset for Manhattan, for the same time period7. We used up to 512,000 randomly sampled rows from the taxi dataset. Half million tasks for a city, for one month, is a fair work- load, considering assignment could be done independently for each city. Each row of the check-in data contains a user id, time of the check-in, and the location of it. Each row in the taxi dataset contains the time and the location of the pick-up and the drop-off. It also contains the cost of the trip. To simulate crowdsourced delivery behavior, we use taxi trips as tasks and check-ins as spatio-temporal availabilities 6. sites.google.com/site/yangdingqi/home/foursquare-dataset 7. www.nyc.gov/html/tlc/html/about/trip record data.shtml IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL. XX, NO. XX, XXX 2018 9 (a) Runtime (b) TAR (c) Unfairness (d) Objective Fig. 4: Scalability (a) Unfairness (b) Unfairness-small task set (c) # Of Completed Tasks (d) Objective Fig. 5: Effect of the Worker Capacity of the workers. For both datasets, we converted all time points to time ranges by adding time periods. Given a point in time, p, a period of length ∆T is created with beginning and end points [p, p + (∆T )]. We treat pick-up location and time of a taxi trip as the task's source location and source validity time. The same applies to drop-off location and time and tasks deliver location and validity. For each check-in, we use the location of it as the location of the availability and converted the time of it to a time period as described above. Radius of each availability is sampled from a Normal distribution with mean and standard deviation calculated using taxi trips. We study the effect of this conversion as part of our last experiment presented in this section, by varying the mean size of the time period. For all other experiments, the mean value for the time period is taken as 2 hours. Last, we used the cost of the trip as the reward of the task. 5.1 Scalability In this set, we present two subsets of experiments. In the first subset, we observe the running time performance, task allocation ratio, and unfairness as a function of the number of tasks. The number of tasks is doubled for every data point ranging from 4000 tasks up to 512,000. Capacities of the workers are assigned using a Normal distribution and the mean is set to the number of tasks over the number of workers. Standard deviation of the distribution is set to mean over 4 to make sure all capacities are at least 0. Using this tight capacity assignment for this particular experiment set, we ensure that the capacity is barely enough for allocating all the tasks. This gives a clear advantage to algorithms that can allocate tasks close to optimal. In the second subset, we observe the task allocation ratio and unfairness as a function of worker capacity. 128,000 tasks are used, and capacities are doubled for every data point ranging from 32 to 4096. To observe the behavior of the MCF algorithm, we also present unfairness as a function of capacity using 8000 tasks (Figure 5b). In this experiment, capacities are ranging from 4 to 256. Figures 4, and 5 present our scalability related results. In all figures, the x-axis represents either the number of tasks to allocate, or the capacity of workers, and y-axis represents the performance metric. Different series represent different assignment algorithms. Figure 4a plots the running time as a function of the number of the tasks. We make two observations from the figure. First, and most importantly, F-Aware runs 107 times faster than the Minimum Cost Flow MCF algorithm. For 32,000 tasks, the running time of the MCF is 1.89 × 108 milliseconds while F-Aware completes in 27 milliseconds. Because of the long evaluation time, we do not present MCF results for more than 32, 000 tasks. Secondly, the running time of the F-Aware algorithm is linear with the task count. Increasing from 4000 tasks to 512,000 tasks, the running time increases from 4.6 milliseconds to 13,130 milliseconds. The difference between the running times of Random, LAF and Nearest assignment algorithms are negligible. Figure 4b plots the task allocation ratio as a function it also of the number of tasks. To increase readability, includes the zoomed small figure of data points between [4000, 32,000]. Since any allocation algorithm will left tasks with no candidates unassigned, in this experiment we con- sider only tasks with at least one candidate. We observe that all allocation algorithms are able to hold their alloca- tion ratio with the increasing number of tasks. The most important observation is that F-Aware is able to assign 96.9% of the tasks that are allocated by MCF. For 32, 000 tasks, MCF reaches to 97% task allocation ration, while F-Aware allocates 94% of all tasks. When we double the number of tasks, F-Aware still allocates 96.5% of tasks, while Random and Nearest worker allocation algorithms stay at 80% and 84%, respectively. In terms of TAR, LAF is the best com- petitor of F-Aware. This is expected as its goal is to increase IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL. XX, NO. XX, XXX 2018 10 the number of allocated tasks. However, TAR of F-Aware is still higher. For 512,000 tasks, F-Aware reaches to 98.1% task allocation ratio, while LAF reaches to 95.3%. fairness. In Figure 4d, we observe that the difference between the objective values of F-Aware and MCF are negligible. F- Aware performs as good as MCF, in terms of our objective, with 107 times faster processing speed. Another observation from the same figure is that, F-Aware performs 4% better than LAF, even with limited capacity. For 128, 000 tasks, F- Aware has 0.86 objective value while LAF has 0.82. When there are more than enough room for assignments, we observe that the gap between F-Aware and LAF becomes even more significant, as shown by Figure 5d. For a capacity value of 4096 the objective value of F-Aware is 0.5, while LAF could reach only 0.39. Objective values of Nearest and Random are better than LAF, because of the lower unfairness values, but still they are far from performing as good as F-Aware. To summarize this experiment set, one can say that using MCF is impractical due to its long running time. In contrast, F-Aware runs 107 times faster. When only TAR is considered, LAF performs similar to F-Aware, but the other two approaches, Random and Nearest, leave 19% and 15% of all tasks unassigned, respectively. While LAF is the best competitor of F-Aware, in terms of TAR and runtime, its unfairness metric value is 2.5× that of F-Aware and its objective value is 80% that of our F-Aware algorithm. 5.2 Effect of the Batch Size We observe average k, unfairness, assignment ratio, and av- erage wait time as a function of the probability of response threshold . Figure 6 presents our batch size related results. In all figures x-axis represents the value of and y-axis represents a performance metric. Different series represent different assignment ratio thresholds, θ. For all series we use the F-Aware assignment algorithm. For this set of ex- periments capacity of workers is high enough to prevent a bottleneck. Figure 6a plots the change in the value of k as a function of . The series with circle marker is the unicast line, i.e. k is set to 1. Recall that the value of k is bounded from below by a function of and bounded from above by a function of θ. We select the largest k inside this range. Higher values of and θ imply tighter bounds. Since θ = 0.0 means unlimited upper bound, practically it is the broadcasting line. We observe that as long as ≤ 0.95 the value of k is limited by the assignment ratio threshold θ. Increasing θ from 0.2 to 0.4 decreases the average k value from 14 to 6. Since task completion is our primary goal, when lower bound is higher than the upper bound (possible in some cases based on the definition of Equation 12), we use the lower bound for deciding the k value. One can observe this behavior when ≥ 0.95, as all the average k values are closer to broadcasting. Figure 6b plots the unfairness as a function of . We observe that smaller k values provides a more fair systems. The unfairness of unicasting is 0, as whenever a worker accepts a task, she will be assigned to it. Whereas the unfair- ness of broadcasting is 0.36. Most importantly, unfairness of multicasting the offer to an average of 5.8 nominees is 0.20. There are two reasons behind this observation. First, local assignment ratio of a worker is negatively correlated with the number of her acceptances. Therefore, the mean of We discuss the Figures 4c, 5a, 5b and 5c together as they are complementary. First three figures plot the unfairness as a function of task count, capacity and capacity respectively. The last one represents the number of assigned tasks, as a function of capacity. In Figure 5b, we use only 8000 tasks to include MCF algorithm. In the first Figure, (4c), we use up to 512,000 tasks but exclude MCF after 32,000 tasks, because of the impractically long running time. For the remaining two, we use 128,000 tasks, and capacity of the workers varies from 32 to 4096. At this point, it would be useful to recall, LAR of a worker is the ratio of revenue she made from completed task over total reward of the offers she has accepted. How- ever, since capacity limits the assignable task count, if a worker wi accepts more offer than her capacity wc i , we only i offers when calculating LAR. Therefore, consider first wc when evaluating performance of an algorithm in terms of fairness, one must consider two cases when capacity limits assignments, and when there are more than enough room for assignments. When worker capacities are enough to serve all tasks, we observe a significant difference between unfairness values. In Figure 5b, for small dataset, Random, Nearest, and MCF perform similar. In contrast, unfairness metric of LAF 0.8× of the same metric of those three. We observe F-Aware, performs best among all. Unfairness metric MCF algorithm is 1.5× that of our F-Aware algorithm when capacity is 256. For the larger task set, the difference between LAF and F-Aware becomes even more significant (figures 5a, 5c). When the capacity is set to 1024 unfairness value of LAF is 0.41 while F-Aware has only 0.16 unfairness. For same data point Nearest, and Random have 0.30, 0.27 unfairness values respectively. We make two additional observations from this figure. First, up to 256 capacity, unfairness values increase. This is because capacity of workers less than number of ac- ceptances. In this set of experiments each worker accepts 215 offers in average. After this point number of accepted tasks used when calculating LAR. Using Figure 5c we observe all algorithms reach maximum number of allocated tasks at capacity 1024, which reflected as stabilized unfairness values in Figure 5a. Second, LAF performs poorer than Nearest and Random assignment approaches. Since it does not take user input into account, (i.e. accepting offers) when tasks are distributed evenly, workers who have accepted small number of offers have LAR = 1, while workers with large number of acceptances have too low LAR values. When capacity is set to a too low value, we cannot observe significant difference between unfairness values of MCF, LAF, and F-Aware. This is because all workers are fully allocated. On the other hand, in Figure 4c values of unfairness metric for Random, and Nearest is around 3.6× that of F-Aware. The reason behind is system could serve less number of tasks (Figure 4b), when one of these two assignment algorithm is used. Figures 4d and 5d present the performance in terms of our objective, Eq 13, as a function of task count and capacity respectively. In these experiments, the ρ parameter is set to 1, to observe the balanced outcome of task allocation ratio and IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL. XX, NO. XX, XXX 2018 11 (a) Average K (b) Unfairness (c) Assignment Ratio (d) Average Wait Time Fig. 6: Effect of the Batch Size (a) TAR Rel. to Offline (b) Unfairness Rel. to Offline (c) TAR (d) Unfairness Fig. 7: Online Task Allocation the set of LAR values increases, which leads to a decrease in the coefficient of variation metric. Second, and more importantly, when a worker accepts an offer, probability of her getting the job is higher with the smaller values of k. In the extreme case, that is unicasting, acceptance implies assignment, hence unfairness = 0. We can also observe a similar behavior in Figure 6c. Assignment ratio of the unicasting equals to 1. For the other series, we can see that assignment ratio is negatively correlated with the average k value. For broadcasting, it decreases up to 0.13. When the average k is 5.8, the assignment ratio is 0.38. Another important observation is theoretically θ = 1.0 should be the unicast line but it has 0.6 assignment ratio. The reason behind is is a stronger constraint than θ. We do not observe same behavior lines other than θ = 1.0, and θ = 0.8 Figure 6d plots the average wait time as a function of . y-axis shows the average number of rounds passed until there is at least one candidate. Recall that the task is offered to nominees in batches until there is at least one response. Between each round, the system waits for a predefined period to let nominees decide. For some scenarios, e.g. on-demand transportation, customers expect almost instant reply. Higher number of rounds before acceptance leads to late notification to a customer. In case of broadcasting, since all nominees are notified at once, number of rounds for response is always 1. However, as we have just seen, broadcasting leads to low assignment ratio and high unfair- ness. At the other extreme, that is unicasting, the average wait time is 1.8× of broadcasting. Multicasting is better than both approaches. Multicasting the task to an average of 5.08 ( = 0.8, θ = 0.95) nominees leads to only 1.01 average rounds, while providing only 0.13 unfairness. In summary, one can say that multicasting is better than broadcasting in terms of assignment ratio and fairness. More- over, it beats unicasting when average wait time is considered, while being almost as fair. 5.3 Online Allocation We observe the task allocation ratio and unfairness as a function of the window size and as a function of capacity. Figure 7 shows our online allocation related results. For Figures 7a and 7b, the x-axis represents the window size in terms of minutes. In addition to window size equals to 0, the x-axis also includes data points starting from 2 minutes, going up to 64, doubling at each step. Different series represent different task allocation algorithms. For Figures 7c and 7d, the x-axis represents worker capacities. The values are doubled for each data point, starting from 16, going up to 512. For all figures, the y-axis represents a performance metric. In this set, we used 96, 000 tasks. Figure 7a shows the task allocation ratio as a function of the window size. The red line with cross markers represents offline F-Aware algorithm. All the other series are drawn relative to this line, therefore it is the y = 1 line. We observe that for the instant task allocation, the difference between different series is negligible. This is because the number of nominees for each task is very small, and thus the decisions of the algorithms do not create any difference. On the other hand, when the window size is increased, we can see that F-Aware, and LAF performs better than the other two approaches and gets closer to the offline assignment. For the 64 minutes window, F-Aware, and LAF allocates 88%, and 85% of the tasks assignable by offline allocation, whereas Nearest, and Random allocation approaches stay at 79% and 76%, respectively. Figure 7b plots unfairness as a function of the window size using the same setup as Figure 7a. When smaller win- dows are used, since tasks are offered to only current nom- inees, the unfairness behavior is similar to using smaller k values. For example, unfairness is 0, when instant allocation strategy applied, as in unicasting. For longer window sizes, the results expected to become more similar to the offline setup. For example The unfairness of offline F-Aware is IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL. XX, NO. XX, XXX 2018 12 (a) Task Allocation Ratio (b) Unfairness (a) Earning Distribution Fig. 8: Sensitivity Experiment 0.13, while online F-Aware increases from 0.04 to 0.11 when window size increased to 64 minutes from 2 minutes. The most important observation is F-Aware performs best among all online algorithms. For the window size of 64 minutes online F-Aware has 0.11 unfairness while LAF, Nearest, and Random have 0.15, 0.32, 0.33 respectively. We also observe the other three algorithms are even less fair than offline F- Aware, for windows larger than 16 minutes. Figures 7c and 7d plot the task allocation ratio and unfairness as a function of worker capacity. Different series represent different window sizes, including offline and in- stant assignments. For all series, F-Aware algorithm is used. We make two main observations. First, the task allocation ratio for smaller window sizes is higher compared to larger window sizes. The reason behind this is that, some of the availabilities satisfying tasks expire before making a decision. At first sight, one might think larger window sizes should produce closer results to the offline scenario, but this is not the case. In the offline scenario all the information is known in advance, and decisions are made before expira- tion. In contrast, with large windows the availabilities might expire before processing. Second, larger windows result in less fair allocation. As we mentioned earlier, smaller windows behave like batches with smaller k values. As batch-size experiments showed, when a task offered to less number of workers, fairness increases since sum of total accepted offers for each worker decreases. 5.4 Sensitivity Experiments We observe the task allocation ratio and unfairness as a function of the coefficient of mean. Different series represent different ∆T values. Recall that these are the two parame- ters involved when we used the taxi trips as the tasks, and the check-ins as spatio-temporal availabilities of workers. The radius added to a check-in is sampled from a Normal distribution with mean and standard deviation set to that of the taxi trips. For each data point on the x-axis, the mean of this distribution is multiplied with the respective number. For all experiments, y-axis represents a performance metric. Figure 8a plots the task allocation ratio as a function of the coefficient of radius. We observe that ∆T has a great effect on the task allocation ratio. When coefficient of radius is 1 and ∆T = 1, 57% of the tasks are allocated, while this number is 82% when ∆T = 8 for the same coefficient. On the other hand, we cannot observe the same effect for larger radius values. Consider the ∆T = 4 line. The task allocation ratio increases by only 0.1% when the coefficient is increased from 1 to 4. (b) Earnings of Workers Pre- sented as Dollar per Kilometer Fig. 9: Discussions Figure 8b plots unfairness using the same setup from Figure 8a. We observe that the difference between unfairness values is negligible. The peaks are a results of randomness present in task acceptance. Therefore, we conclude that unfairness is not effected from the adaption of real-world data to our problem setup. 5.5 Discussions We present two additional experiments that provide in- sights of our 2-phase assignment model and the F-Aware algorithm. Figure 9a shows the distribution of workers as a function of the average earning per acceptance. We observe that 800 workers have average values between 2$ and 4$ when F-Aware algorithm is used. Although the most dense areas similar for other approaches as well, we can see that the standard deviation of them is much higher, which is an indicator of unfairness. Unlike traditional crowdsourcing, in spatial crowdsourc- ing workers have to physically travel to the source and destination of the task. Therefore, dollar earned per traveled kilometer is a good indicator of the what is the reward of a worker in return of her labor. In this point we would like to remind that, batched progressive offer strategy multicasts the offer to the workers who are most likely to accept it, implicitly workers who are closer to the task. Figure 9b presents dollar earned per kilometer as a function of the coefficient of mean values as in the sensitivity experiments. Solid lines represent different assignment algo- rithms when the batched progressive offer strategy is used, while dashed lines show the same but when the offers are broadcasted. Recall that the nearest worker assignment is beneficial for capturing the spatial aspect of the assignment problem. In case of broadcasting, represented with dashed lines, we observe the Nearest better than other approaches. Using this approach, workers could make more than 20 cents per kilometer more compared to other approaches. On the other hand, we observe the other three approaches benefit significantly from multicasting. The difference be- tween Nearest and F-Aware decreases from 22 cents to 5 cents when the system prefers to multicast the offers instead of broadcast. Therefore, we could state that F-Aware approach captures the spatial aspect of the problem, with the help of multicasting, as well as allocating each task to Nearest worker. In summary, our experimental evaluation shows that: IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL. XX, NO. XX, XXX 2018 13 • • F-Aware is efficient: It runs around 107 times faster than the MCF algorithm and allocates 98.1% of the assignable tasks. F-Aware is fair: Unfairness metric of the LAF is 2.5× that of F-Aware and, it maintains its fairness with increasing capacity, and increasing task count. • Batched progressive offering is useful to limit the assignment ratio, while not sacrificing the wait time. Moreover, it helps to capture the spatial aspect of the problem by helping to provide competitive dollar per kilometer ratio with Nearest worker approach. 6 RELATED WORK We discuss the related work in the areas of crowdsourcing, including spatial crowdsourcing and fairness. Crowdsourcing. A wide spectrum of crowdsourcing appli- cations are surveyed in [12]. Crowdsourcing is proposed also for technical tasks such as relational query process- ing [18], [28], [26]. In [20], three crucial aspects of crowd- sourcing, namely task design, marketplace dynamics, and worker behavior are analyzed. However, none of these methods from the literature have considered fairness among workers. To the best of our knowledge, this is the first work that applies the concept and findings of fairness from social psychology research to crowdsourcing applications. Spatial Crowdsourcing. Spatial crowdsourcing requires workers to physically travel to task locations. Earlier work [2], [21], [11] extended crowdsourcing to the physical world, with a variety of applications such as answering queries [18], [28] and serving micro-tasks (such as taking a photo of a monument) [10], [36], [9]. Numerous work has addressed the maximum task assignment problem [21], [11], [31] and its extensions that integrate the reliability of workers [22], [33]. In [10], maximizing the reliability and spatial diversity are considered together. Although these works study task allocation in spatial crowdsourcing, they mostly focus on offline scenarios. In [19], online allocation is performed when only the workers are dynamic. In [36], micro tasks are allocated when both tasks and workers can appear anywhere, anytime. Different from our work, [32] learns the workers' acceptance probability in dynamic tasks static workers setup. Recall, we assume acceptance depends on the hardness level of the tasks and independent from previous acceptances. Unlike these works, in crowdsourced delivery, redundant task allocation is not possible. Similar to our problem, in the context of crowdsensing, there is a trade off among quality of information (QoI), budget and time constraints, which requires multi-objective aware task allocation algorithms as well. A recent work solves this problem with a particle swarm optimization technique that maximizes the aggregated QoI/budget ratio. A delegation mechanism is used, in case the workers cannot finish their allocated task, they may recommend a set of workers from their social network to complete it [15]. This technique serves the same purpose with our 2-phase allocation model. Fairness is not in the scope of any of the aforementioned work. Fairness in Crowdsourcing. In various application do- mains, such as networking [27], staff scheduling [14], and resource allocation [4], fairness is considered to be one of the most important constraints. In [8], fairness among customers, but not workers, is considered. Its goal is making sure that the system is fair to the customer when it is not possible to serve all tasks. In [6] unfairness is defined as discrimination against individuals, while in [13], the au- thors differentiate between various perspectives on fairness and ethics in crowdsourcing. They consider distributional fairness as a subjective measure and discuss ethical implica- tions. In [35], in addition to the cost minimization objective, fair allocation of tasks to heterogeneous workers (workers with different capacities and costs to execute the task) is studied. The central idea of fairness in [35] is to maximize the minimum utility of all bidders (workers). Utility is defined as the number of allocations. They also assume that the set of all tasks is known in advance and workers are predefined. Unlike previous works in spatial crowdsourc- ing, [25] stresses both the cost incurred by the movement of the worker and the fairness of the assignment among the workers. A sequence of sets of spatial tasks are assigned to crowdsourced workers as they arrive. The one-to-one assignment of tasks to workers is done in mini-batches. In their setup, workers are not assigned to multiple tasks. On the other hand, to provide redundant task allocation they copy the task, and allocate one worker for each copy. Besides utilizing a redundant task allocation strategy, the fairness definition of this work is different from for our scenario. Yet, in the experimental evaluation we included the least allocated worker first, LAF, which is inspired from this work. We showed that, according to the state-of-the- art fairness definition, LAF is not fair unless all workers are homogeneous. 7 CONCLUSION In this paper, we created a strategy on allocation of delivery tasks. In this strategy, we use a combined objective of maxi- mum task allocation and fair distribution of tasks to work- ers. In our 2-phase allocation model, for each task a set of nominees are identified using availabilities of workers. The task is offered to nominees using our batched-progressive offer strategy. Once the candidates for each task are identified, we showed that the problem can be reduced to Minimum Cost Flow(MCF) problem if fairness is not considered. To cope with drawbacks of MCF-based solutions in terms of running time and lack of fairness handling, we introduce our F-Aware algorithm. We then adapt our model to online task allocation and mini-batch task allocation scenarios. Our evaluation showed that F-Aware runs around 107× faster than the TAR- optimal solution and assigns 96.9% of the tasks that can be assigned by it. Moreover, F-Aware assigns 18% more tasks than random assignment approach and is 2.5× more fair than the least allocated worker first assignment strategy. Our experimental evaluation showed that the dis- tributed fairness criteria can be satisfied with no significant changes in task allocation ratios. The proposed approach of fair allocation of tasks can lead to more sustainable crowdsourced delivery platforms. Human perspectives of fairness with quantitative and qualitative surveys, and long- term effects of fair allocation strategies in real crowdsourced IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL. XX, NO. XX, XXX 2018 14 delivery platforms are among potential future work in this area. REFERENCES [1] J. Adams. Inequity in social exchange. New York Academic Press, 2:267–299, 1965. [2] F. Alt, A. S. Shirazi, A. Schmidt, U. Kramer, and Z. Nawaz. Location-based crowdsourcing: Extending crowdsourcing to the In Proceedings of Nordic Conf. on HCI: Extending real world. Boundaries, pages 13–22, Reykjavik, Iceland, 2010. S. Barocas and A. D. Selbst. Big data's disparate impact. 2016. [3] [4] D. Bertsimas, V. F. Farias, and N. Trichakis. On the efficiency- fairness trade-off. Manage. Sci., 58:2234–2250, Dec. 2012. [5] L. G. Boiney. When efficient is insufficient: Fairness in decisions affecting a group. Management Science, 41:1523–1537, 1995. [6] R. M. Borromeo, T. Laurent, M. Toyama, and S. Amer-Yahia. Fairness and transparency in crowdsourcing. In Proc. of the EDBT, pages 466–469, 2017. [7] A. M. Brawley and C. L. Pury. Work experiences on mturk. Comput. Hum. Behav., 54(C):531–546, Jan. 2016. [8] A. M. Campbell, D. Vandenbussche, and W. Hermann. Routing for relief efforts. Transportation Science, 42:127–145, May 2008. [9] Z. Chen et al. gmission: A general spatial crowdsourcing platform. VLDB Conf., 7:1629–1632, Aug. 2014. [10] P. Cheng, X. Lian, Z. Chen, R. Fu, L. Chen, J. Han, and J. Zhao. Re- liable diversity-based spatial crowdsourcing by moving workers. VLDB Conf., 8:1022–1033, June 2015. [11] D. Deng, C. Shahabi, and U. Demiryurek. Maximizing the number of worker's self-selected tasks in spatial crowdsourcing. In Proc. of ACM SIGSPATIAL, pages 324–333, Orlando, Fl., 2013. [12] A. Doan, R. Ramakrishnan, and A. Y. Halevy. Crowdsourcing systems on the www. Commun. ACM, 54:86–96, Apr. 2011. [13] D. Durward, I. Blohm, and J. M. Leimeister. Is there papa in crowd work?: A literature review on ethical dimensions in crowdsourc- ing. In 2016 Intl IEEE Conf. on Ubiquitous Intelligence Computing, pages 823–832, July 2016. [14] A. Ernst, H. Jiang, M. Krishnamoorthy, and D. Sier. Staff schedul- ing and rostering: A review of applications, methods and models. European Journal of Operational Research, 153:3 – 27, 2004. [15] R. Estrada, R. Mizouni, H. Otrok, A. Ouali, and J. Bentahar. A crowd-sensing framework for allocation of time-constrained and location-based tasks. IEEE Transactions on Services Computing. [16] R. Faullant, J. Fueller, and K. Hutter. Fair play: perceived fairness in crowdsourcing communities and its behavioral consequences. Academy of Management Proceedings, 2013. [17] N. Franke, P. Keinz, and K. Klausberger. does this sound like a fair deal?: Antecedents and consequences of fairness expectations in the individuals decision to participate in firm innovation. Orga- nization Science, 24(5):1495–1516, 2013. [18] M. J. Franklin, D. Kossmann, T. Kraska, S. Ramesh, and R. Xin. Crowddb: Answering queries with crowdsourcing. In ACM Int. Conf. on Management of Data, pages 61–72, Athens, Greece, 2011. [19] C.-J. Ho and J. W. Vaughan. Online task assignment in crowd- sourcing markets. In Proc. of the AAAI Conf. on Artificial Intelligence, pages 45–51, Toronto, Ontario, Canada, 2012. [20] A. Jain, A. D. Sarma, A. Parameswaran, and J. Widom. Under- standing workers, developing effective tasks, and enhancing mar- ketplace dynamics: A study of a large crowdsourcing marketplace. Proc. of the VLDB Endow., 10(7):829–840, Mar. 2017. [21] L. Kazemi and C. Shahabi. Geocrowd: enabling query answering In Proc. of Int. Conf. on Advances in with spatial crowdsourcing. Geographic Information Systems, pages 189–198, 2012. [22] L. Kazemi, C. Shahabi, and L. Chen. Geotrucrowd: Trustworthy In Proceedings of query answering with spatial crowdsourcing. ACM SIGSPATIAL, pages 314–323, Orlando, Florida, 2013. [23] P. Kovcs. Minimum-cost flow algorithms: an experimental evalu- ation. Optimization Methods and Software, 30:94–127, 2015. [24] E. Lawler. Combinatorial optimization - networks and matroids. Holt, Rinehart and Winston, New York, 1976. [25] Q. Liu, T. Abdessalem, H. Wu, Z. Yuan, and S. Bressan. Cost Minimization and Social Fairness for Spatial Crowdsourcing Tasks, pages 3–17. Dallas, TX, USA, 2016. [26] A. Marcus, E. Wu, D. R. Karger, S. Madden, and R. C. Miller. Demonstration of qurk: A query processor for humanoperators. In ACM Int. Conf. on Management of Data, pages 1315–1318, Athens, Greece, 2011. [27] W. Ogryczak, M. Pi´oro, and A. Tomaszewski. Telecommunications Journal of network design and max-min optimization problem. Telecommunications and Information Technology, 3:43–56, 2005. [28] H. Park, H. Garcia-Molina, R. Polyzotis, A. Parameswaran, and J. Widom. Deco: A system for declarative crowdsourcing. VLDB Conf., 5:1990–1993, Aug. 2012. Pang, N. [29] J.-F. Rouges and B. Montreuil. Crowdsourcing delivery: New in- terconnected business models to reinvent delivery. In Int. Physical Internet Conf., Quebec, Canada, May 2014. [30] S. P. Schappe. Understanding employee job satisfaction: The im- portance of procedural and distributive justice. Journal of Business and Psychology, 12(4):493–503, 1998. [31] H. To, C. Shahabi, and L. Kazemi. A server-assigned spatial crowdsourcing framework. ACM Trans. Spatial Algorithms Syst., 1:2:1–2:28, July 2015. [32] U. Ul Hassan and E. Curry. A multi-armed bandit approach to online spatial task assignment. 2014 IEEE 11th Int. Conf on Ubiquitous Intelligence And Computing, pages 212–219, 2014. [33] U. Ul Hassan and E. Curry. Efficient task assignment for spatial crowdsourcing: A comb. fractional optimization approach with semi-bandit learning. Exp Systems With Apps., 58:36–56, 2016. [34] D. Yang, D. Zhang, V. W. Zheng, and Z. Yu. Modeling user activity preference by leveraging user spatial temporal characteristics in IEEE Trans. on Systems, Man, and Cybernetics: Systems, lbsns. 45:129–142, 2015. [35] Q. Ye, Y. Zhang, and R. Dekker. Fair task allocation in transporta- tion. Omega, pages –, Apr. 2016. [36] T. Yongxin, S. Jieying, D. Bolin, W. Libin, and C. Lei. Online mobile In Proc. of ICDE micro-task allocation in spatial crowdsourcing. 2016, Helsinki, Finland, May 2016. [37] M. B. Zafar, I. Valera, M. Gomez Rodriguez, and K. P. Gummadi. Fairness beyond disparate treatment disparate impact: Learning In Proc. of Int. classification without disparate mistreatment. Conf. on WWW, WWW, pages 1171–1180, Republic and Canton of Geneva, Switzerland, 2017. Fuat Basık is a graduate student in the Depart- ment of Computer Engineering, Bilkent Univer- sity, Turkey. He holds a M.Sc. degree in Com- puter Science from Bilkent University. His re- search interests are in scalable data integration. Bu gra Gedik is an Associate Professor in the Department of Computer Engineering, Bilkent University, Turkey. He holds a Ph.D. degree in Computer Science from Georgia Institute of Technology. His research interests are in data- intensive distributed systems. Hakan Ferhatosmano glu is a Professor in the Department of Science at the University of War- wick. His research is on scalable data manage- ment and analytics for multi-dimensional data. He holds a Ph.D. degree in Computer Science from University of California, Santa Barbara. He received research career awards from the US Department of Energy, US National Science Foundation, The Science Academy of Turkey, and Alexander von Humboldt Foundation. Kun-Lung Wu is a Manager at the IBM T. J. Watson Research Center. His research interests are in big data systems and applications.
1803.04932	1	1803	2018-03-13T16:45:33	Impacts of transport development on residence choice of renter households: An agent-based evaluation	[ "cs.MA", "cs.AI", "cs.CY" ]	Because of improving accessibility, transport developments play an important role in residence choice of renter households. In this paper, an agent-based model is developed to investigate impacts of different transport developments on residence choice of renter households in Tehran, the capital of Iran. In the proposed model, renter households are considered as agents who make a multi-objective decision and compete with each other to rent a preferred residential zone. Then, three transport development scenarios including construction a new highway, subway and bus rapid transit (BRT) line are simulated and resulting changes in residence choice of agents are evaluated. Results show that transport development scenarios significantly affect residence choice behavior of different socio-economic categories of renter households and lead to considerable changes in the residential demand, composition of residents, mean income level and mean car ownership in their vicinities.	cs.MA	cs	Impacts of transport development on residence choice of renter households: An agent-based evaluation Ali Shirzadi Babakan and Mohammad Taleai This is an Author's Original Manuscript (Preprint) of an Article Published by ELSEVIER in Habitat International, 2015, Vol. 49: 275-285. To link to this article: https://doi.org/10.1016/j.habitatint.2015.05.033 Abstract Because of improving accessibility, transport developments play an important role on residence choice of renter households. In this paper, an agent-based model is developed to investigate impacts of different transport developments on residence choice of renter households in Tehran, the capital of Iran. In the proposed model, renter households are considered as agents who make a multi-objective decision and compete with each other to rent a preferred residential zone. Then, three transport development scenarios including construction a new highway, subway and bus rapid transit (BRT) line are simulated and resulting changes in residence choice of agents are evaluated. Results show that transport development scenarios significantly affect residence choice behavior of different socio-economic categories of renter households and lead to considerable changes in the residential demand, composition of residents, mean income level and mean car ownership in their vicinities. Keywords Transport development, Residence choice, Agent-based modeling, NSGA-II algorithm, Renter household. 1. Introduction A mutual relationship between land uses and transportation has been shown by many researchers (Iacono et al. , 2008, Waddell, 2011, Wegener, 2014). Development and changes of land uses influence the transportation system and on the other hand, development and changes of the transportation system affect land use patterns. For example, development of a new highway leads to changes in surrounding land uses and reciprocally by altering travel demands, affects the transportation system. In order to study this relationship, various land use-transportation models have been developed using various mathematical, statistical, heuristic, and microsimulation methods. Detailed reviews of these models are available in (Iacono et al. , 2008, Waddell, 2011, Wegener, 2014). One of the most essential parts of land use-transport models is modeling of residence choice of households. This modeling includes a complex decision-making process in which households select their residence by considering various criteria. Households often choose their residence according to their socio-economic characteristics (e.g. income, the number of members, and the number of owned cars), neighborhood characteristics (e.g. housing price and accessibility to various services and opportunities), and accessibility to pre-specified destinations such as workplaces (Jordan et al. , 2012, Jun et al. , 2013, Wang and Li, 2006, Wu et al. , 2013, Yang et al. , 2013). Therefore, transportation system has a major role in this decision making, because it directly affects the accessibility to various services and pre- specified destinations. The impacts of transport developments on the housing market and residential decisions of households have been studied by many researchers. Pagliara et al. (2010) applied a bid-choice model and an hedonic model to study the impacts of different transport policies such as road user charging, changes to fuel duties and the provision of light rapid transit systems on residential location decisions, housing occupancy 1 rates and housing prices in the Greater Oxford area. Their research shows that transport policies have significant impacts on the housing market. For example, the road user charging might reduce the average housing price about 2%. Also, development of a new public transport system might increase the housing price by around 3% on average. Eliasson (2010) attempted to examine the influence of accessibility on residential location choice of households using TILT, a land use-transportation model for the Stockholm region. He found that the attractiveness of a location increases with the improvement of accessibilities to workplaces and different types of service. Cervero and Kang (2011) investigated the impacts of converting regular bus operations to bus rapid transit (BRT) on land use changes and land values in Seoul, Korea. Their findings showed that development of BRT leads to convert single-family residences to higher density apartments. Also an increase of more than 10% for values of residential land uses within 300 m of BRT stops is revealed. Calvo et al. (2013) evaluated the impacts of development of the subway system on population and land uses in Madrid, Spain. They found that considerable residential developments and population growth are observed in vicinities of the new subway stations. Also, urbanization and population settlement are significantly more dynamic in proximities of the subway developments. They suggested that the population density reduces by increasing distance from the new subway stations. Mathur and Ferrell (2013) estimated the impact of a light rail line on single-family home prices using a hedonic regression model in San Jose, CA. They found that the average home sale price within 1/8 mile of the light rail stations increases about 3.2% for every 50% reduction in the distance between the home and stations. In addition, the average housing price within 1/8 mile of the light rail line is 18.5% higher than the average price within distances of more than 1/8 mile from the light rail line. Zhang et al. (2014) examined the impacts of different transit systems including bus rapid transit (BRT), light rail transit (LRT) and metro rail transit (MRT) on the price of residential properties using hedonic price models. Their research showed that transit impacts on the housing market extend to 1 mile for MRT, 0.5 mile for LRT and are discernable for BRT stations. Also, the value of residential properties increases 39.41 and 17.57 USD/m2 for every 100 m closer to the stations of MRT and LRT, respectively. Hamersma et al. (2015) studied the trade-off between nuisances and accessibility in residential moving intentions of people living near highways using structural equation modeling in the Netherlands. Their study investigated that highway usage and other residential characteristics including satisfaction with buildings, traffic safety, and amount of green spaces may compensate perceived highway nuisances. Also, some groups of residents such as home owners are less tended to move without considering their residential satisfaction. However, a majority of these studies are aggregate and insensitive to the behavior of individual households (Benenson, 2004). This issue has prompted researchers to use disaggregate models which can represent decision of individual households. One of the most applicable methods which has attracted the attention of researchers for disaggregated modeling of residence choice of households is agent-based modeling. Agent-based model is a 'bottom-up' approach in which behaviors and interactions of agents are characterized and used to produce different aggregated results. Agents are autonomous entities who can perceive their environment, move through space and time, and take actions based on their objectives (Crooks and Heppenstall, 2012). Agent-based models have many capabilities for studying residence choice of households including clear representation and interpretation of behavior of households (Barros, 2004), better understanding of quantitative and qualitative changes in the urban system (Benenson, 2004), reduction of the computational complexity, explicit representation of heterogeneity and interdependencies among households and their environment, considering various conditions and constraints on households and flexible aggregation of results (Barros, 2004, Hunt, 2002). Therefore, a number of researchers have used agent-based models to study residence choice behavior of households (e.g. Benenson, 2004, Devisch et al. , 2009, Ettema, 2011, Gaube and Remesch, 2013, Haase et al. , 2010, Jordan et al. , 2012). A detailed review of these models can be found in (Huang et al. , 2014). In addition to these models, a new generation of land use-transportation models including RAMBLAS (Veldhuisen et al. , 2000), UrbanSim (Waddell et al. , 2003) and ILUTE (Salvini and Miller, 2005) have been developed using agent-based modeling. These comprehensive urban models have been applied by several researchers (e.g. (Farooq and Miller, 2012, Kakaraparthi and Kockelman, 2011, Kryvobokov et al. , 2015, Veldhuisen et al. , 2005, Waddell et al. , 2007)) to model the interactions 2 between land uses and transportation system and also to evaluate impacts of various land use and transport "what if" scenarios on housing market and residential decisions of households in many metropolitan regions. In this paper, an agent-based model is used to investigate role of different transport development scenarios in the residence choice process of renter households. Majority of previous studies have used discrete choice models based on the random utility maximization theory for determination of households' residence. In this theory, households calculate utility of a finite number of well-identified options and select the one with the maximum utility. But, in this paper, a multi-objective decision making method and a heuristic competition method are used to determine renter households' residence. In the proposed agent- based model, renter households are considered as agents who have different criteria and preferences depending on their socio-economic and demographic characteristics. They search among residential zones and select their appropriate residential options according to their criteria and preferences. Subsequently, they compete with each other to rent a residence among their preferred options in different time periods. Finally, three transport development scenarios are simulated and changes in residence choice of agents are evaluated. The proposed model has been applied in Tehran, the capital of Iran. Tehran with an area of about 750 square kilometers and a population of about 8.3 million is one of the largest and the most populated capitals in the world (Tehran Municipality, 2013b). This metropolis has been divided into 560 traffic analysis zones (TAZs) which have been used as the spatial units in this research. Tehran consists of a wide transit system including highway, subway, bus and BRT networks. This metropolis includes a large number of highways with the length of about 550 kilometers, a subway network composed of 4 active intraurban lines with the length of about 125 kilometers and a bus network comprised of 250 lines with the length of about 3000 kilometers. In addition to these networks, 6 bus rapid transit (BRT) lines with the length of about 102 kilometers are operational in Tehran. Although the length of BRT lines is about 3% of the total length of the bus lines, about 40% of bus passengers are transported with these lines (Tehran Municipality, 2013a). This shows the popularity and effectiveness of BRT network in transit system of Tehran. In recent decades, insufficient development of the public transit on one side and low prices of fuel on the other side have resulted in different problems including traffic congestion and air pollution in Tehran. Therefore, urban planners and policymakers attempt to resolve these problems by developing various urban plans such as improvement of public transit and highway networks in Tehran. However, these plans have side effects on the other urban activities including the residence choice process of households. In this paper, it is attempted to study these effects. The paper is structured in 6 sections. The proposed agent-based model is described in detail in section 2. Sections 3 and 4 highlight implementation and validation procedures of the proposed model in Tehran metropolis. In section 5, different transport development scenarios are simulated and their effects on the residence choice of households are evaluated. Finally, conclusions of the research are presented in section 6. 2. Residence choice model In the proposed model, renter households are considered as agents who look for an appropriate residence among residential zones. In reality, renters are unable to search all zones, but they usually select some zones to search a residence among them. The same procedure is used in this model. At the first step, a synthetic population of renter households is generated using the Monte Carlo simulation. Then, agents select their residential options according to their socio-economic characteristics and desired objectives. In fact, they face with a multi-objective decision making problem, because some of their objectives (e.g. minimizing housing rent and maximizing accessibility to different services) conflict with each other. A constrained NSGA-II algorithm is developed to find the optimal residential options with respect to their objectives. Finally, they compete with each other to rent a residence among their desired residential options. The main parts of the proposed model are described below. 3 2.1. Population synthesis In this section, the available aggregated demographic data in each zone (Tehran Municipality, 2013b) are used as references to generate renter households (agents) with different attributes. For this purpose, a sequential approach using the Monte Carlo simulation is developed in which attributes of agents are simulated in each step based on the determined attributes in the previous step. In this approach, at first, monthly income, number and age of the members of agents are randomly simulated in each zone. Then, number of owned cars and employees, and required residential area of agents are randomly generated according to the aforementioned attributes. In the next step, preferences and criteria of each agent for selecting a residence are randomly generated based on the previously determined attributes. Finally, the existing information of employment distribution in different zones and the general pattern of distances between workplace and residence of employees (TCTTS, 2012) are used to define the workplace of employed members of agents. Data provided by a field survey of the stated preferences of residence choice of sample renter households in Tehran are used to simulate the residential criteria and preferences of agents. These data were collected from 480 renter households with different socioeconomic and cultural characteristics by using questionnaires in April 2013. They were requested to state their residential criteria and preferences in the scale range of 0 (unimportant) to 9 (very important). In Table 1, these sample households have been classified by the household size, average monthly income and number of owned cars. Then, in each class, the percentage of households who stated a preference number of greater than 4 to each criterion has been specified. Table 1: A summary of preferred residential criteria of sample renter households in Tehran Attribute Class Percentage Housing Rent Distance from the Workplace Distance from the Former Residence Air Pollution Noise Pollution Accessibility to retail stores Size Average Monthly income (USD) Number of Cars Total Single Couple 3-4 > 4 < 500 500-1000 > 1000 0 1 > 1 6.9 35.6 44.4 13.1 29.4 51.9 18.8 11.7 66.5 21.9 100 100 100 100 100 100 100 100 100 100 100 100 92.9 90.7 84.2 79.4 92.8 85.8 80.7 97.1 86.1 81.5 86.4 46.4 52.6 62.6 61.8 68.1 56.4 52.6 55.9 58.4 58.5 58.2 50.0 63.9 61.4 64.7 40.6 64.2 77.2 38.2 62.8 69.2 61.5 64.3 74.2 69.6 70.6 47.8 73.5 87.7 52.9 70.6 80.0 70.6 46.4 61.9 62.6 55.9 47.8 64.7 59.6 64.7 63.2 47.7 60.3 to Educational centers 17.9 10.3 71.3 88.2 42.0 58.3 33.3 55.9 55.8 29.2 50.6 Preferred Residential Criteria (%) Accessibility Accessibility Accessibility to Green and Recreational spaces to health Centers Accessibility to Cultural Centers Traffic Restrictions Accessibility to Highway network Accessibility to Subway Stations Accessibility to Bus Stops 7.1 50.5 57.3 64.7 34.8 53.9 64.9 44.1 55.8 41.5 51.8 3.6 7.2 12.3 14.7 7.2 12.3 7.0 14.7 10.8 6.2 10.3 10.7 7.2 8.8 5.9 5.8 8.8 8.8 8.8 8.2 7.7 8.2 67.9 64.9 63.7 67.6 46.4 64.7 87.7 5.9 64.5 96.9 64.8 78.6 73.2 71.3 73.5 58.0 70.6 98.2 14.7 73.6 100 72.7 35.7 48.5 53.2 52.9 58.0 49.5 43.9 88.2 55.4 12.3 50.3 17.9 20.6 22.8 20.6 29.0 21.1 14.0 73.5 18.2 6.2 21.5 2.2. Determination of residential options of agents In this section, agents independently choose their desirable residential zones without considering choices of other agents and residential capacity of zones. They select their residential options according to various criteria including housing rent, distance from their former residence, accessibility to public and transport services, environmental pollutions, traffic restrictions, and distance from their workplaces. These are the most important criteria which are derived from the survey of stated preferences of residence choice of renter households in Tehran. Agents select maximum number of ten residential options using the non- dominated sorting genetic algorithm II (NSGA-II). NSGA-II proposed by Deb et al. (2002) is one of the fast and most efficient multi-objective evolutionary algorithms which has been successfully applied to solve a wide range of multi-objective optimization problems. The main steps of this algorithm are outlined in Table 2 (For more details refer to (Deb et al. , 2002)). 4 Table 2: The main steps of NSGA-II NSGA-II Framework 1. Random generation of an initial parent population (P0) of size N. 2. Generation of an offspring population (Q0) of size N using common genetic operators 3. Pt=P0 and Qt=Q0. 4. Generation of a combined population (Rt) of size 2N by Rt=Pt(cid:1)Qt. 5. Generation of non-dominated fronts (Fi) from solutions of Rt based on the definition of non- domination. The first front contains solutions which do not dominate each other and dominate all the other solutions. Similarly, this process is continued until all remaining solutions of Rt are assigned to a front. 6. Generation of Pt+1 of size N from the solutions assigning to the first (best) fronts of Rt. 7. Generation of Qt+1 of size N by applying binary tournament selection, crossover and mutation operators on Pt+1. In the binary tournament selection, the solution with lower domination rank or higher crowding distance is selected as the winner. The crowding distance of an individual solution is the perimeter of cuboid formed by its nearest neighboring solutions in the objective space that shows the density of solutions surrounding that solution. 8. Repetition of steps 4 through 7 until the convergence criterion is met. In this paper, the following objective functions are used in NSGA-II algorithm to find optimal residential options of agents. It should be noted that depending on their residential criteria and preferences determined by the Monte Carlo simulation, agents consider one, some or all of the following objectives in their residence choice process. Housing rent: All agents attempt to find a residence with minimum housing rent and in compatible with their income level using Eq. (1). (cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:7)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:9)(cid:10)(cid:11)(cid:12)(cid:8)(cid:13)(cid:14)(cid:15)(cid:16) (cid:12)(cid:8)(cid:17)(cid:14)(cid:15)(cid:18)(cid:8)(cid:5)(cid:19)(cid:13)(cid:20)(cid:21)(cid:17)(cid:7)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:5)(cid:22)(cid:23)(cid:24)(cid:25)(cid:24)(cid:26)(cid:8)(cid:8)(cid:27)(cid:13)(cid:28)(cid:17)(cid:29)(cid:20)(cid:30)(cid:13)(cid:31)(cid:8)(cid:19)(cid:13)(cid:20)(cid:21)(cid:17)(cid:31)(cid:27)(cid:13)(cid:28)(cid:13) (cid:20)(cid:30)(cid:13)(cid:7) where: Sa is the set of agents who the criterion is important for them; Sz is the set of zones; Aa is the required residential area of agent (a); Ri is the average housing rent per square meter in zone (i); Ia is the monthly income of agent (a); determined in the Monte Carlo simulation and restrict the search space of agents to zones in which they (cid:27)(cid:13)(cid:28)(cid:17)(cid:29), (cid:27)(cid:13)(cid:28)(cid:13) (cid:8)are the limits on the monthly income of agent (a) for renting a residence. These limits are afford to rent a residence, where 0 (cid:1) (cid:27)(cid:13)(cid:28)(cid:17)(cid:29) , (cid:27)(cid:13)(cid:28)(cid:13) (cid:8)(cid:1) 1; Accessibility to public services: Almost all agents attempt to find a residence with maximum accessibility to their preferred public services including educational, retail, green and recreational, cultural or health services using Eq. (2) developed by Tsou et al. (2005). (cid:2)(cid:3)(cid:4)(cid:5)!(cid:7)(cid:8)(cid:8)(cid:8)(cid:8)(cid:9)"#(cid:12)(cid:8)(cid:13)(cid:14)(cid:15)(cid:16) where: s is the type of public services; js is the case (j) of the public service (s); (cid:12)(cid:8)(cid:17)(cid:14)(cid:15)(cid:18)$%%&(cid:13)'(cid:20)(cid:22)()(cid:20)(cid:17)()+, &(cid:13)' is the preference of agent (a) to the public service (s), where .&(cid:13)'/(cid:6); (cid:22)() is the relative effect of (js) which is calculated by (cid:22)()/(cid:19)()0(cid:9)"#(cid:8)(cid:5)(cid:19)()(cid:7), where ((cid:19)()) is the area of (js); (cid:17)() is the distance between zone (i) and (js); - ' () 5 Accessibility to transport services: Almost all agents attempt to find a residence with maximum accessibility to their preferred transport services including bus stops, subway stations or highways using Eq. (3) developed by Currie (2010). (cid:2)(cid:3)(cid:4)(cid:5)1(cid:7)(cid:8)(cid:8)(cid:8)(cid:8)(cid:9)"#(cid:12)(cid:8)(cid:13)(cid:14)(cid:15)(cid:16) where: t is the type of transportation services; jt is the case (j) of the transportation service (t); (cid:12)(cid:8)(cid:17)(cid:14)(cid:15)(cid:18)$%%(cid:19)(2(cid:19)(cid:17) (cid:20)&(cid:13)3 (2 (cid:19)(2 is the area of service range of ( jt) which is inside zone (i); &(cid:13)3 is the preference of agent (a) to the transportation service (t), where .&(cid:13)3 /(cid:6); Ai is the area of zone (i); - 3 Distance from the workplace: A great number of agents attempt to find a residence with minimum distance from their workplaces using Eq. (4). (cid:2)(cid:3)(cid:4)(cid:5)4(cid:7)(cid:26)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8) (cid:9)(cid:10)(cid:11)(cid:12)(cid:8)(cid:13)(cid:14)(cid:15)(cid:16) (cid:12)(cid:8)(cid:17)5(cid:8)67(cid:16)(cid:14)(cid:15)(cid:18)8%(cid:8)(cid:17)67(cid:16) < (cid:29) 9:; where: (cid:17)67(cid:16) is the distance between zone (i) and the workplace of agent's employee ((cid:22)9(cid:16)); n is the number of employees of agent (a); Distance from the former residence: Because of different reasons such as familiarity and dependency to the former residential area, many agents attempt to find a residence in proximity of their former residence using Eq. (5). (cid:2)(cid:3)(cid:4)(cid:5)=(cid:7)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8) (cid:9)(cid:10)(cid:11)(cid:12)(cid:8)(cid:13)(cid:14)(cid:15)(cid:16) (cid:12)(cid:8)(cid:17)5(cid:8)>(cid:16)(cid:14)(cid:15)(cid:18)(cid:8)(cid:5)(cid:17)>(cid:16)(cid:7) Air and noise pollutions: Many agents attempt to find zones with minimum air and noise pollutions using Eqs. (6) and (7). where: (cid:17)>(cid:16) is the distance between zone (i) and the former residence of agent (fa); (cid:12)(cid:8)(cid:17)(cid:14)(cid:15)(cid:18)(cid:5)@(cid:17)(cid:7)(cid:8)(cid:8)(cid:8)(cid:5)(cid:22)(cid:23)(cid:24)(cid:25)(cid:24)(cid:26)A@(cid:17)(cid:31)!B5(cid:12)C(cid:14)DE(cid:7)(cid:8)(cid:8)(cid:8) (cid:12)(cid:8)(cid:17)(cid:14)(cid:15)(cid:18)(cid:5)G(cid:17)(cid:7)(cid:8)(cid:8)(cid:8)(cid:5)(cid:22)(cid:23)(cid:24)(cid:25)(cid:24)(cid:26)AG(cid:17)(cid:31)!B5(cid:12)C(cid:14)D(cid:29)(cid:7)(cid:8)(cid:8)(cid:8) (cid:2)(cid:3)(cid:4)(cid:5)?(cid:7)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:9)(cid:10)(cid:11)(cid:12)(cid:8)(cid:13)(cid:14)(cid:15)(cid:16) (cid:2)(cid:3)(cid:4)(cid:5)F(cid:7)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:9)(cid:10)(cid:11)(cid:12)(cid:8)(cid:13)(cid:14)(cid:15)(cid:16) where: Pi, Ni are the annual average of air and noise pollutions in zone (i); where they are classified in five categories varying from clean (0) to highly polluted (4); Sp, Sn are the sets of agents who air and noise pollutions are very important for them and their residential options are selected among clean (0) to medium (2) pollution categories; Traffic restrictions: Many agents attempt to find zones with minimum traffic restrictions on private cars including restrictions on all or odd-even private cars using Eq. (8). (cid:2)(cid:3)(cid:4)(cid:5)H(cid:7)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:9)(cid:10)(cid:11)(cid:12)(cid:8)(cid:13)(cid:14)(cid:15)(cid:16) (cid:12)(cid:8)(cid:17)(cid:14)(cid:15)(cid:18)(cid:5)I(cid:17)(cid:7)(cid:8)(cid:8)(cid:8)(cid:8)(cid:5)(cid:22)(cid:23)(cid:24)(cid:25)(cid:24)(cid:26)AI(cid:17)/JB5(cid:12)C(cid:14)D3(cid:7) where: Ti is the traffic restriction in zone (i) classified in 3 categories including (0) for no traffic restrictions; (1) for odd-even car restriction; and (2) for restriction on all private cars; 6 St is the set of agents who traffic restrictions are very important for them and their residential options are selected among zones with no traffic restrictions. 2.3. Determination of final residence of agents In this step, agents compete with each other in different months of the year to select their final residence among their preferred residential options. For this purpose, the relocation time of each agent is randomly assigned to a month of the year in accordance with the available statistical information of volume of relocation in different months of the year. Also, residential capacities of zones are limited in each month and computed by Eq. (9). (cid:2)(cid:3)(cid:4)(cid:5)K(cid:7)(cid:26)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)L(cid:17)3/G(cid:17)3MN3 (cid:19)(cid:17)O. (cid:19)(cid:17)O (cid:17) where: residence at month (t); L(cid:17)3 is the residential capacity of zone (i) at month (t); (cid:19)(cid:17)O is the residential area in zone (i); G(cid:17)3 is the number of agents whose former residences are located in zone (i) and want to change their N3 is the number of new dwellings entered into the rental housing market at month (t) which is estimated using the available statistical information. In each month, some agents search among their residential options according to the closeness of their options to their former residence or workplace and choose the first option having the capacity. If some agents simultaneously select an option which has not enough capacity for all of them, they compete with each other to reside in that zone. In this competition, agents with fewer members (except singles), higher income level and with no child will have a higher chance to succeed respectively, because most owners in Tehran prefer to rent their properties to these types of renters. In case of the similarity of qualifications of competitors, winners are selected randomly and defeated agents have to compete on their next options. This process is continued until either all agents reside in one of their options or the evaluation of all options of defeated agents is finished. Agents who are not able to reside in any of their residential options will have the opportunity to compete again with other agents in the next month. If they cannot reside in a zone in the next month, no residence would be considered for them. 3. Implementation For simulating residence choice of renter households in Tehran, 50,000 agents are generated using the Monte Carlo simulation and their residence are determined by the developed NSGA-II and competition modules in MATLAB 7.9 software. Spatial distribution of residential options and final residence of agents have been shown in Figures 1 and 2, respectively. As seen in these Figures, The density of residential options and residents in central areas is higher than the other areas. This suggests that the central areas of Tehran generally are more attractive for renter households due to some reasons such as low housing rents and high accessibilities to public transport services, various public facilities and employment opportunities. 7 (a) (b) Figure 1: Spatial distribution of (a) Agents' residential options and (b) density of Agents' residential options per square kilometer (a) (b) Figure 2: Spatial distribution of (a) Agents' final residence and (b) density of Agents' final residence per square kilometer 4. Validation The proposed model is validated using a sample data composed of 1350 renter households derived from the survey of residential and travel preferences of households in Tehran (TCTTS, 2012). These data are not used for calibration of the proposed model in the Monte Carlo simulation. Residence of these sample households is simulated by the model and results are compared with their actual residence. Results indicate that the proposed model is able to correctly simulate the actual residential zone of 59.3% of households. Figure 3 shows the simulation accuracy of the households' residence by distance between their actual and simulated residences. As shown in this Figure, residence of 81.4% of the sample households is simulated in distances of less than 5 km from their actual residence which is an indicator of a good performance of the proposed model. 100 80 60 40 20 0 ) % ( s d l o h e s u o h e l p m a S 2 1 Distance of simulated residence from the 3 4 9 10 8 5 6 7 actual residence (km) Figure 3: Simulation accuracy of residence of sample households by distance between their actual and simulated residences 8 5. Transport development scenarios In this section, three major scenarios of the future development plan of Tehran are considered. Undoubtedly, these scenarios would have various long-term and short-term impacts on traffic, environmental pollutions, land use changes, housing prices and accessibility. But, in this study, only impacts of these scenarios on residence choice of renter households are evaluated. The considered scenarios are as below: Scenario 1: A new highway construction In this study, effects of the southern part of Imam Ali highway on residence choice of renters are studied. As illustrated in Figure 4(a), this highway with the length of about 35 kilometers is one of the longest highways of Tehran that connects the northeast of Tehran to its southeast. Northern part of the highway has been completed and its southern part with the length of about 26 kilometers is under construction (Tehran Municipality, 2013a). Scenario 2: A new subway line construction In this study, effects of the under construction southern part of line 3 (Figure 4(b)) with the length of about 19 kilometers and 15 stations on residence choice of renters are evaluated. Scenario 3: A new BRT line construction In this study, effects of the under construction eastern part of line 5 (Figure 4(c)) with the length of about 9 kilometers on residence choice of renters are examined. (a) (b) (c) Figure 4: The existing and under construction (a) highways, (b) subway lines; and (c) BRT lines in Tehran 9 5.1. Changes in accessibility and housing rent The transport development scenarios significantly improve the accessibility in their neighborhoods. Therefore, new accessibility values of TAZs are updated using Eq. (3). In addition, due to improvement of the accessibility, the housing rent is changed in the neighborhoods of these scenarios. Changes of housing rent are estimated by consulting with real estate agencies and experiences gained from construction of other similar transportation projects in Tehran. Investigations show that in contrast to BRT which shows no considerable effects on the housing rent, highway and subway developments have significant effects on the housing rent. Table 3 represents the rate of changes in the housing rent within different neighborhoods of the highway and subway projects. These heuristic values are used in Eq. (10) as Pr to calculate the new housing rent of TAZs which are within neighborhoods of the new highway and subway. (cid:2)(cid:3)(cid:4)(cid:5)(cid:6)J(cid:7)(cid:26)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:21)(cid:17)(cid:29)/(cid:5)(cid:19)(cid:17)P.Q(cid:17)O(cid:7)(cid:21)(cid:17)M(cid:5).Q(cid:17)O(cid:7)@O(cid:21)(cid:17) (cid:8) (cid:19)(cid:17) where: r is the neighborhood radius; (cid:21)(cid:17)(cid:29) and Ri are the new and previous housing rent per square meter in zone (i), respectively; Ai and Q(cid:17)O respectively are the area of zone (i) and the buffer area of radius (r) inside zone (i); Pr is the rate of change of housing rent in the neighborhood of radius r. Table 3: the rate of changes in the housing rent within different neighborhoods of the highway and subway projects Highway Subway stations Neighborhood radius (km) Rate of change (%) 0-0.1 0.1-1 1-1.5 1.5-2 0-0.5 0.5-1.2 1.2-1.9 -15 15 10 5 15 10 5 5.2. Changes in residence choice of renter households For studying how residential demand and residents of zones are changed by each scenario, the proposed model is rerun after implementation of each scenario and resulting changes in residential options and residence of agents are evaluated (Figures 5, 6 and 7). As indicated in Figures 5(a), 6(a) and 7(a), residential demand generally increases in neighboring zones of the new transport developments and decreases in farther zones. This illustrates that some agents have preferred to move to neighboring zones of the new transport facilities due to improvement of accessibility to transport services in these zones. In other words, neighboring zones of the new transport facilities attract some residential demands from farther zones. Also, because of high residential demand and subsequently more intense competition for residence choice in neighboring zones of the new transport facilities, considerable changes are occurred in the composition of residents in these zones (Figures 5(b), 6(b) and 7(b)). 10 (a) (a) (b) Figure 5: Changes in (a) residential options and (b) residence of agents after implementation of the new highway (b) Figure 6: Changes in (a) residential options and (b) residence of agents after implementation of the new subway line (a) (b) Figure 7: Changes in (a) residential options and (b) residence of agents after implementation of the new BRT line Changes of residential demand and residents vary in different geographical directions around the new transport developments. In scenario 1, these changes are greater in southern and central neighboring zones of the new highway (Figure 5). The main reason for this may be that the accessibility of these zones to the highway network was very low before implementation of the new highway. Therefore, after implementation of this highway, residential demand significantly increases in these zones due to considerable improvements in accessibility of these zones to the highway network. In addition, because of dominant pattern of daily trips in these areas which is from south to north, the new highway may be more 11 attractive for residents of the southern neighborhoods. As illustrated in Figure 6, the new subway line significantly improves the accessibility of many deprived areas of the southwest of Tehran to the subway network. Therefore, this line shows substantial effects on residential demand and composition of residents in these areas. In other words, because of the lack of subway stations in southwestern zones, residents of these zones pay higher attention to the new constructed subway line in their residence choice. Figure 7 shows that the new BRT line has more influences on residential demand and composition of residents in its southern and eastern neighboring zones. This suggests that residents of these zones, who generally have lower income levels, show higher interest to use the new BRT line, perhaps due to the prevailing pattern of daily trips in these areas which is from east to west and also the lower costs of traveling by BRT. Figure 8 represents changes in the mean income level of residents in residential zones after implementation of the scenarios. As illustrated in Figures 8(a) and (b), the mean income of residents is generally increased in neighboring zones of the new highway and subway line. There are some reasons for this; first, the housing rent generally increases in these zones due to improvement of the accessibility to transport services. Second, the competition for residing in these zones is more intensive due to the increase of residential demand which consequently results in success of agents with higher incomes to reside in these zones. However, the situation is completely different in neighborhoods of the new BRT line. As shown in Figure 8(c), the mean income level decreases in neighboring zones of the new BRT line, because high-income agents do not show interests to move to these zones. (a) (b) (c) Figure 8: Changes in the mean income in TAZs after implementation of the new (a) highway, (b) subway line, and (c) BRT line 12 As seen in Figure 9, the mean car ownership is meaningfully changed in neighborhoods of the new transport developments. This attribute is generally increased in neighboring zones of the new highway and subway line. This indicates that agents with one or more cars are more interested to reside in these zones. However, a part of this increase may result from the increase of the mean income level in these zones, because there is a strong correlation between the income level and the number of cars owned by households. Also, it should be noted that the increase of mean car ownership in neighborhoods of the new subway line generally is lower than those of the new highway neighborhoods. The increase of mean car ownership in neighboring zones of the new subway line shows that this public transport mode attracts agents with one or more cars. Therefore, this suggests that development of the subway network may reduce the use of private car among agents having one or more cars. On the other side, the new BRT line, as planned in scenario 3, does not show meaningful effects on the mean car ownership in its neighboring residential zones. Although the more intensive competition among agents results in residing of agents with higher incomes and thereby causes a slight increase in the mean car ownership in some zones, other zones show a decreasing trend, because agents with fewer cars generally are more interested to reside in neighborhoods of the BRT network. As a result, it can be said that agents with more cars do not show great interests to move to neighborhoods of the new BRT line. (a) (b) (c) Figure 9: Changes in the mean car ownership in TAZs after implementation of the new (a) highway, (b) subway line, and (c) BRT line 13 The transport development scenarios result in slight increase of the number of agents who are unable to reside in any zone. The number of these agents increases from 1.36% to 1.51% and 1.58% after implementation of scenarios 1 and 2, respectively. However, the number of these agents is not changed in scenario 3. The main reason of this is that after implementation of the new highway and subway line, the housing rent increases in the southern poor areas and therefore some of the low-income agents who previously resided in these areas cannot reside again there. A summary of changes in residence choice of agents after implementation of the transport development scenarios are presented in Table 4. Table 4: Changes in residential demand and residence of agents after implementation of the transport development scenarios Scenario 1 Scenario 2 Scenario 3 Percentage of zones in which residential demand is changed The maximum increase in residential demand of zones The maximum decrease in residential demand of zones Percentage of zones which their residents are changed (affected zones) The maximum change in residents of zones Percentage of agents who change their residence 1 Percentage of agents whose new and former residences are less than 2.5 km apart 1 Percentage of agents whose new and former residence are more than 5 km apart 2,3 Percentage of high-income agents who move to neighborhoods of the new transport services 2,3 Percentage of low-income agents who move to neighborhoods of the new transport services 2 Percentage of agents who own more than one car and move to neighborhoods of the new transport services 2 Percentage of agents who have no car and move to neighborhoods of the new transport services Changes in percentage of agents with no residence 1 This parameter is calculated relative to all agents who change their residence. 2 This parameter is calculated relative to all agents who move to neighborhoods of the new transport services. 3 The income level of agents is determined with respect to the mean income of residents in neighborhoods of the transport developments. (%) 46.8 19.9 10.8 39.5 33.3 5.6 48.8 5.5 43.7 18.6 16.0 45.1 +15.7 (%) 48.1 21.4 10.9 37.8 28.0 6.3 53.7 4.1 54.3 9.4 28.6 7.5 +9.1 (%) 27.1 17.8 8.0 20.1 18.5 1.4 61.1 2.7 11.0 59.5 6.1 62.9 0 6. Conclusion There are complex relationships between transportation and residence choice of households. By changing accessibility to various opportunities and services, transport developments may affect residence choice of different socio-economic categories of renter households. In this paper, an agent-based model was developed to study these effects. In this model, renter households were represented as agents who can individually decide, choose, compete and reside according to their residential criteria and preferences. Agent-based modeling is currently known as an efficient disaggregated approach to model large and complex socio-economic systems and processes. The main advantage of agent-based models is their ability to represent residence choice behavior of individual households. The developed agent-based model was used to evaluate effects of three transport development scenarios including construction of a new highway, subway, and BRT line on residence choice of renter households in Tehran. These scenarios lead to considerable effects on residential demand and composition of residents in many zones such that residents of some zones are changed up to 33 percent. Various socio- economic categories of households show different residence choice behaviors in resulting situations from implementation of each scenario. High-income households owning one or more cars show the highest sensitivity to development of the highway network. Whilst in the case of BRT line development, households with low incomes and without car are more affected. It is interesting that in development of the subway network, both categories of households show some interests to live near the new subway line. As a result, it can be said that subway is attractive for various socio-economic categories of residents in Tehran. Particularly, subway is a more attractive option than BRT for developing public transportation where residents have high incomes and more than one car. In addition, after development of the new highway and subway, the number of households who do not afford to reside in any zone slightly increases. This shows significant impacts of transport development policies on residence choice of tenants which even may result in loss of home for some renter households. However, because of using traffic analysis zones (TAZs) as the spatial units, findings of this paper are exposed to effects of the modifiable areal unit problem. These effects require close considerations by 14 testing the model with varying sizes and configurations of spatial units. Also, in this paper, only residence choice behavior of renter households was investigated, but the model can be extended to include all types of households. In addition, the time distance can be used to more accurately measure accessibilities to various opportunities in future studies, especially when large-scale units such as census blocks or parcels are used. The proposed model can be also extended to determine the housing rent in a price bidding framework. Finally, the model can be developed to evaluate effects of different transport policies (e.g. traffic restrictions on private cars and changes in the existing transportation networks), housing policies (e.g. housing assistance programs to low-income households, changes in the housing supply in some areas and changes in housing taxes), and land use policies (e.g. changes in distribution of workplaces and public facilities) on residence choice behavior of various households. References Barros J.X., Urban Growth in Latin American Cities: Exploring urban dynamics through agent-based simulation [Ph.D. Thesis] University College London (UCL), 2004. Benenson I., Agent-Based Modeling: From Individual Residential Choice to Urban Residential Dynamics, in: M.F. Goodchild, D.G. Janelle, (Eds), Spatially Integrated Social Science: Examples in Best Practice, New York, 2004, 67-95. Calvo F., J. de Oña, F. Arán, Impact of the Madrid subway on population settlement and land use, Land Use Policy 31 (2013) 627-639. Cervero R., C.D. Kang, Bus rapid transit impacts on land uses and land values in Seoul, Korea, Transport Policy 18(1) (2011) 102-116. Crooks A.T., A.J. Heppenstall, Introduction to Agent-Based Modelling, in: A.J. Heppenstall, A.T. Crooks, L.M. See, M. Batty, (Eds), Agent-Based Models of Geographical Systems, Dordrecht, 2012, 85-105. Currie G., Quantifying spatial gaps in public transport supply based on social needs, Journal of Transport Geography 18 (2010) 31-41. Deb K., A. Pratap, S. Agarwal, T. Meyarivan, A Fast and Elitist Multiobjective Genetic Algorithm: NSGA-II, IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION 6 (2) (2002) 182-197. Devisch O.T.J., H.J.P. Timmermans, T.A. Arentze, A.W.J. Borgers, An agent-based model of residential choice dynamics in nonstationary housing markets, Environment and Planning A 41(8) (2009) 1997-2013. Eliasson J., The Influence of Accessibility on Residential Location, in: F. Pagliara, J. Preston, D. Simmonds, (Eds), Residential Location Choice: Models and Applications, Verlag Berlin Heidelberg, 2010, 137-164. Ettema D., A multi-agent model of urban processes: Modelling relocation processes and price setting in housing markets, Computers, Environment and Urban Systems 35 (2011) 1-11. Farooq B., E.J. Miller, Towards integrated land use and transportation: A dynamic disequilibrium based microsimulation, Transportation Research Part A 46 (7) (2012) 1030-1053. Gaube V., A. Remesch, Impact of urban planning on household's residential decisions: An agent-based simulation model for Vienna, Environmental Modelling & Software 45 (2013) 92-103. Haase D., S. Lautenbach, R. Seppelt, Modeling and simulating residential mobility in a shrinking city using an agent-based approach, Environmental Modelling & Software 25 (2010) 1225-1240. Hamersma M., E. Heinen, T. Tillema, J. Arts, Residential moving intentions at highway locations: The trade-off between nuisances and accessibility in the Netherlands, Transportation Research Part D: Transport and Environment 35 (2015) 130-141. Huang Q., D.C. Parker, T. Filatova, S. Sun, A review of urban residential choice models using agent-based modeling, Environment and Planning B: Planning and Design 41(4) (2014) 661-689. Hunt J.D., Agent Behaviour Issues Arising with Urban System Micro-Simulation, European Journal of Transport and Infrastructure Research 2 (3/4) (2002) 233-254. 15 Iacono M., D. Levinson, A. El-Geneidy, Models of transportation and land use change: A guide to the territory, Journal of Planning Literature 22 (4) (2008) 323-340. Jordan R., M. Birkin, A. Evans, Agent-Based Modelling of Residential Mobility, Housing Choice and Regeneration, in: A.J. Heppenstall, A.T. Crooks, L.M. See, M. Batty, (Eds), Agent-based models of geographical systems, Dordrecht, 2012, 511-524. Jun M.-J., S.-K. Ha, J.-E. Jeong, Spatial concentrations of Korean Chinese and determinants of their residential location choices in Seoul, Habitat International 40 (2013) 42-50. Kakaraparthi S., K. Kockelman, Application of UrbanSim to the Austin, Texas, Region: Integrated-Model Forecasts for the Year 2030, Journal of Urban Planning and Development 137(3) (2011) 238-247. Kryvobokov M., A. Mercier, A. Bonnafous, D. Bouf, Urban simulation with alternative road pricing scenarios, Case Studies on Transport Policy In Press (2015). Mathur S., C. Ferrell, Measuring the impact of sub-urban transit-oriented developments on single-family home values, Transportation Research Part A 47 (2013) 42-55. Pagliara F., J. Preston, J.H. Kim, The Impact of Transport Policy on Residential Location, in: F. Pagliara, J. Preston, D. Simmonds, (Eds), Residential Location Choice: Models and Applications, Verlag Berlin Heidelberg, 2010, 115-136. Salvini P.A., E.J. Miller, ILUTE: An Operational Prototype of Comprehensive Microsimulation Model of Urban Systems, Network and Spatial Economics 5 (2005) 217-234. TCTTS, Statistical data of comprehensive transportation and traffic studies in Tehran, Tehran, Iran, 2012. Tehran Municipality, Tehran municipality statistical yearbook 2012-2013, Tehran, Iran, 2013a. Tehran Municipality, Tehran statistical yearbook 2012-2013, Tehran, Iran, 2013b. Tsou K.W., Y.T. Hung, Y.L. Chang, An accessibility-based integrated measure of relative spatial equity in urban public facilities, Cities 22 (6) (2005) 424-435. Veldhuisen K.J., H.J.P. Timmermans, L.L. Kapoen, RAMBLAS: a regional planning model based on the microsimulation of daily activity travel patterns, Environment and Planning A 32 (2000) 427-443. Veldhuisen K.J., H.J.P. Timmermans, L.L. Kapoen, Simulating the effects of urban development on activity-travel patterns: an application of Ramblas to the Randstad North Wing, Environment and Planning B 32(4) (2005) 567-580. Waddell P., Integrated land use and transportation planning and modelling: addressing challenges in research and practice, Transport Reviews 31 (2) (2011) 209-229. Waddell P., A. Borning, M. Noth, N. Freier, M. Becke, G.F. Ulfarsson, Microsimulation of urban development and location choices: Design and implementation of UrbanSim, Networks and Spatial Economics 3(1) (2003) 43- 67. Waddell P., G.F. Ulfarsson, J.P. Franklin, J. Lobb, Incorporating land use in metropolitan transportation planning, Transportation Research Part A 41 (2007) 382–410. Wang D., S.-m. Li, Socio-economic differentials and stated housing preferences in Guangzhou, China, Habitat International 30 (2006) 305-326. Wegener M., Land-Use Transport Interaction Models, in: M.M. Fischer, P. Nijkamp, (Eds), Handbook of Regional Science, Berlin Heidelberg, 2014, 741-758. Wu W., W. Zhang, G. Dong, Determinant of residential location choice in a transitional housing market: Evidence based on micro survey from Beijing, Habitat International 39 (2013) 16-24. Yang L., G. Zheng, X. Zhu, Cross-nested logit model for the joint choice of residential location, travel mode, and departure time, Habitat International 38 (2013) 157-166. Zhang M., X. Meng, L. Wang, T. Xu, Transit development shaping urbanization: Evidence from the housing market in Beijing, Habitat International 44 (2014) 545-554. 16
1402.6208	2	1402	2018-06-04T16:25:07	The Anatomy of a Modular System for Media Content Analysis	[ "cs.MA", "cs.AI", "cs.DC" ]	Intelligent systems for the annotation of media content are increasingly being used for the automation of parts of social science research. In this domain the problem of integrating various Artificial Intelligence (AI) algorithms into a single intelligent system arises spontaneously. As part of our ongoing effort in automating media content analysis for the social sciences, we have built a modular system by combining multiple AI modules into a flexible framework in which they can cooperate in complex tasks. Our system combines data gathering, machine translation, topic classification, extraction and annotation of entities and social networks, as well as many other tasks that have been perfected over the past years of AI research. Over the last few years, it has allowed us to realise a series of scientific studies over a vast range of applications including comparative studies between news outlets and media content in different countries, modelling of user preferences, and monitoring public mood. The framework is flexible and allows the design and implementation of modular agents, where simple modules cooperate in the annotation of a large dataset without central coordination.	cs.MA	cs	The Anatomy of a Modular System for Media Content Analysis Ilias Flaounas, Thomas Lansdall-Welfare, Panagiota Antonakaki, Nello Cristianini Intelligent Systems Laboratory, University of Bristol, Bristol BS8 1UB, United Kingdom Intelligent systems for the annotation of media content are increasingly being used for the automation of parts of social science research. In this domain the problem of integrating various Artificial Intelligence (AI) algorithms into a single intelligent system arises spontaneously. As part of our ongoing effort in automating media content analysis for the social sciences, we have built a modular system by combining multiple AI modules into a flexible framework in which they can cooperate in complex tasks. Our system combines data gathering, machine translation, topic classification, extraction and annotation of entities and social networks, as well as many other tasks that have been perfected over the past years of AI research. Over the last few years, it has allowed us to realise a series of scientific studies over a vast range of applications including comparative studies between news outlets and media content in different countries, modelling of user preferences, and monitoring public mood. The framework is flexible and allows the design and implementation of modular agents, where simple modules cooperate in the annotation of a large dataset without central coordination. 1 Introduction The ready availability of vast amounts of digital data and the creation of new powerful methods of analysis have started transforming many branches of science, opening the possibility for data-driven approaches and science automation to fields as diverse as biology, physics, chemistry and even the humanities and social sciences [1, 2]. In the social sciences, one area of particularly intense progress is the study of traditional news media and also the analysis of the new social media. Both of these applications generate large quantities of readily available data for media analysts and social scientists in general to process and investigate. Traditional research in this area has relied on the labour intensive step of human coding: the activity of human experts reading and annotating news or other media items. Recent works have shown that Artificial Intelligence (AI) algorithms can be deployed to automate many steps of this expensive process, therefore paving the way to the analysis of much larger sets of data. This automation has become possible because of the convergent effect of two trends: the availability of data and the technology to manage it; and the emergence of a new generation of powerful (mostly statistically-driven) algorithms for machine learning, data mining and text analysis. advances in AI algorithms with the recent trend to automate scientific research in social sciences: the basic tasks that are solved by the classical AI algorithms do not directly coincide with the tasks that are of value to social scientists. In other words, basic social science tasks such as measurement of gender bias in a text are the result of multiple elementary AI algorithms, such as the detection of named entities, co-reference resolution, topic detection and so on. This situation directly highlights the main technological challenge presented by automation of science in general, and of social sciences in the present case: the design of integrated intelligent systems that allow many AI algorithms to collaborate to extract value from data. Namely, a framework for the combination of multiple AI algorithms that is principled and "independent" of the specific algorithms used for the annotation. in The design and implementation of modular systems is a general challenge for AI, where significant progress has been achieved the optimization of single-task algorithms such as classification, but where most interesting applications call for complex tasks that require the coordinated usage of many such methods. The understanding of how complex tasks can be decomposed into modules is an important part of modern AI, as well as the inverse problem of integrating simple modules to generate complex behaviour. One important observation is key when combining recent In this paper we describe a solution for the design of large scale intelligent systems that combine multiple AI modules, and that allows the automatic annotation of large amounts of data for scientific purposes. These modules are combined in a coherent and scalable framework which we describe. Besides describing the system we have developed over the past few years for our own research, the method we present is general and likely to be useful in many other domains, as discussed in Section 5. All communication between modules in our approach is obtained by reading and writing on a shared blackboard, and decisions are made by the system without any centralised control. Intelligent goal driven behaviour emerges in our system as the result of the interaction of all these modules, calling each other, without central control. Of course the system is designed, but the behaviour is emergent. 2 System Overview Typically simple computational analysis systems found in AI literature are single purposed and usually implement a specific pipeline-based architecture: Data flow in a linear way from one analysis component to the next, and some specific output is produced at the final step. Our system is designed to be highly modular and each module performs a specific analysis task. It is designed to achieve a series of key properties including flexibility, robustness and no centralised decision making. Flexibility means that the processing lines are not hardwired but that they can easily change and modules can be added or removed without interrupting or needing to adapt current modules. Robustness is the property where we want the system to produce the most accurate results possible, even if a specific module fails. In a simple pipeline the failure of a module will break the processing line. Robustness is also related with the notion of no single point of failure. That means that the system should be able to handle hardware or network failures that are expected to happen in the long run. The no centralised decision making means that we want to avoid having a meta-module that will organise the behaviour of the existing modules. Some background on the ideas that affected our system design are provided in Section 5. An overview of the system is shown in Fig. 1. Modules are organised in three groups: input, output and analysis. Data live in a database and are processed by the modules through a common Application Programming Interface (API). The input modules create new data by crawling the web or, in the case of machine translation, by translating news articles into English; the analysis modules analyse the data by adding annotations to each item; and the output modules create reports or populate websites that provide insight to the final user. At the logical level the data are organised in a series of Fig 1. Simplified overview of framework, demonstrating the interaction of modules (circles) to annotate data via central blackboards. the Macsy Fig. 2. Data is organised at the logical level in blackboards (grey), and are annotated by modules (blue). blackboards as illustrated in Fig. 2. Each blackboard contains similar items like, for example, news articles or tweets. The items are accessible by modules which can alter them by adding or removing annotations. The modules do not communicate directly with each other, but only the blackboards. The background literature we used as a basis for the developed architecture is discussed in the last section of the paper. indirectly through 2.1 Blackboards In our system, a blackboard is a shared repository of data, and all data is organised in a set of blackboards. There are two main blackboards, one for storing news articles, and one for storing tweets. For the news analysis, we have two more blackboards, one for storing news feeds that are the sources of news stories, and one for storing outlets. Each outlet corresponds to an individual publisher, such as newspapers, broadcast stations or blogs. We use website domain names to define what constitutes a publisher. One outlet can have one or more news feeds, and each article can come from one or more news feeds. In a similar way a different blackboard holds the queries used to query the location of 54 cities of UK in Twitter. Other auxiliary blackboards include the Locations blackboard that holds the geographic position of cities, geographic regions or other places of interest and the URLs blackboard that is used by the FeedFinder module to explore the web in order to find novel news sources. Each blackboard holds its own set of tags, i.e. textual strings, for annotating the data. Tags are important for a number of reasons: they define overlapping sets of items, e.g. the topics covered in an article, they convey information like the language an article is written to, and they define the modules that will operate on the item as we discuss in the following section. 2.2 Modules in three categories: Modules are organised Input modules that create the basic content that populates blackboards; Analysis modules the blackboards' content and; Output modules that provide the output of the system. that annotate The input modules include the news crawler and the Twitter crawler. The first crawls the web for the items in the feeds blackboard and returns their content as individual articles in the articles blackboard. For each article the crawler provides the title, a short description, the publication date, and a link to the HTML page that contains the full content of the article. The articles are tagged with the language that they are written (if known), the location of the outlet that published them, and the ID of the feed that carried them. Also, the same module generates a hash based on title, description and outlet ID that is used to identify identical articles. This is very useful since it allows the module to quickly identify duplicate articles that should not be re-added to the blackboard. This happens when the crawler is relaunched and finds the same content again from the same sources; or when the same article is published in more than one feed. In the later case we add any extra information that the second feed provides to the article. Similar functionality is implemented in the Twitter crawler, which collects tweets from a predefined list of locations. Currently we monitor and collect data from 54 major UK cities and we collect 500K tweets per day. Modules are launched with some predefined frequency, for example every hour, or once per day. A module starts its operation by querying a blackboard in order to find a set of items that contain some specific set of tags and/or annotations. For example a topic tagger will search for items that have a special tag "FOR>SportsTagger". That tag is placed on items by some different module like the feature extractor. An upper limit of items processed per module-run in order to bound the module running time. The module processes each item, one after the other in an independent way. Of course, a module can implement a multithreaded approach so that multiple items can be processed in parallel. The result of the analysis is a set of new tags and annotations. The topic tagger for sports will annotate the relevant items with the tag "Sports". Also, it has the option to add extra tags that may be used to trigger the launch of another module, like a module that creates reports based on the sports articles of the day. Modules can store their own private information outside a blackboard but this information is not shared with other modules. For example, a module may store a data model that it uses to make predictions. If two modules need to communicate, i.e. exchange information, this can be achieved only by reading and writing the information on the items of a blackboard. A common issue when building a modular system is deciding on the appropriate size of a module. At the one end a small module can only perform some very trivial processing, while at the other end a large module can perform an overly complex task which could be divided in smaller parts. In our architecture we follow the idea that a module should be large enough to be able to create specific annotations to items that are useful to another module. Next, we summarise some of the basic implemented modules that constitute our system: • HTML Scraper. The scraper parses the HTML page that contains the full content of an article and returns the raw textual information of the article after removing irrelevant text from the page, along with images, menu items and any HTML code. • Feed Finder. This module implements a focused crawler that searches the web in order to find novel news sources in RSS or Atom format. Currently, the addition of news feeds to the watch is semi-automatic since human approval is needed. list • Machine Translation. Currently this module is able to machine-translate 21 European languages into English. The module is based on the popular Phase Based Model statistical machine translation approach. Every non-English article into machine translation and the result is written in the database as a new machine-translated article. We based the module implementation on the open source software Moses. is tagged to be fed • Feature Extractor. It creates a vector representation of the article based on TF/IDF features. The module implements a typical text processing pipeline stop-words removal and stemming. The output is used by a including series of analysis modules like the mood or topics classifiers. 3 Implementation and Data Management • Language Detector. It annotates an article based on the language used. • Sentiment Extractor. The module measures and annotates articles based on whether they contain adjectives that carry sentiment. • Mood Detector. The module computes the inner product of each article with a list of predefined mood-related words and it annotates items with a score. We track four moods: joy, anger, fear and sadness. • Topic Detector. The module annotates articles based on an SVM classifier trained on standard datasets like the Reuters and The New York Times corpora. Topics we track include politics, business, sports, crime, war and religion. • Geolocator. The module identifies the mention of locations in articles. It also deploys some location disambiguation algorithms to identify the correct location between locations that have the same name. • Readability Annotator. Provides scores based on how readable a document is. The score is based on the FLES readability score, and it is based on simple metrics like average length of sentences and average number of syllables per word. • Popularity Annotator. This module measures the popularity of an article. The popularity is based on a linear model built using an online ranking algorithm and data from specific outlets that publish a list of their most popular articles. This algorithm is trained by comparing the articles that managed to become popular versus the articles that did not become popular although they were published by the same outlet and on the same day. Some of the modules were developed in house, while others act like wrappers of existing NLP or machine learning libraries. For example, the topic classifiers are based on the LibSVM library while the machine translation is based on the Moses library. Finally, the output modules typically create reports or some XML files that can be used to populate the content of demo websites. In Sect. 4 we present some exemplary case studies that show the capabilities of the output modules. Modules are implemented in Java and are typically comprised of an executable and a settings file. All modules use the same API that was developed in order to guarantee the homogeneous use of blackboards across the system. The API also serves as an intermediate layer between the database and the modules allowing the change of the underlying database management system without the need of changing the modules' code. The settings file typically includes the module name; a short description of the module functionality; the name of the blackboard that is used as input; the set of tags and fields that items should have in order to be processed by the module; the name of the blackboard that is used as output (usually it is the same as input); and the name of the tags and annotations that the module will add to the items. The settings file also define the maximum number of items that the module will process in each launch and the number of threads that it is allowed to initiate (modules are typically multithreading for performance reasons). This allows the system to have multiple modules with similar functionality without the need of having more than one binary executable. For example, the topic detectors are all based on the same executable but each one has a different settings file. Modules are replicated and distributed across multiple physical machines for additional robustness. Currently our system is organised on eight physical machines. Modules are allocated specific time slots in machines, i.e. they are triggered in predefined times and they have an upper timeout limit enforced by operating system. This guarantees that no module can abuse computational resources. Also, a web interface we call SystemWatch has been developed, that allows some basic administration tasks and the monitoring of the system performance. For example, the interface allows observation of the status of the physical machines, and the monitoring of modules status' and their input/output, such as how many articles were collected and how many were analysed per module. Storing and management of data is an important factor in our system. The first version of the system was built around a MySQL database [3]. Recently the database system was replaced by MongoDB, a modern NoSQL solution. This database management system has a number of benefits that makes it quite attractive. First of all it is a schema-less document based database which allows enormous flexibility: all annotations of an item are stored with the item in the same table. Thus, a single query may return all the annotations about an item, without the need of having a central place to store which tables are available for which items. Also, the database improves is inherently distributed allowing reading and writing from multiple machines. This not only the performance of the database but it also highly increases the availability of the system since there is no single point failure. Currently our database is organised on four physical machines. The machines are logically organised in two pairs (shards). The data is split between the two shards and each shard is comprised of two machines that store the same data for extra availability and performance. Note that machines are physically located in two different buildings for avoiding external problems like power or network availability. 4 Case Studies – Application Scenarios Our system has been used successfully in various social science projects including the analysis of Twitter content [4], the analysis of traditional news media [5], the analysis of the EU mediasphere [6], etc. Here, we present two case studies as the current functionality of the system, one on the comparison of news outlets and one on the sentiment analysis of Twitter content. representative results of 4.1 Comparison of News Outlets on Topics, Writing Style and Gender Bias A popular topic of research in media studies is the detection of differences or biases among news outlets. We showed how similar studies can be performed using our system in our previous work [5]. We analysed a large corpus comprised of 2.5 million news articles collected from 498 news outlets over a period of 10 months. For each article we identified the general topics that it covered, as well as two basic writing style properties, namely the readability and linguistic subjectivity. The computation of the aforementioned quantities allowed the answering of a series of research questions. For example, for the articles of each topic we calculated the average readability, finding that articles about sports are the easiest to read while articles on politics are the hardest to read; and for linguistic subjectivity, finding that articles about fashion and arts are the most linguistically subjective, while business articles were the most objective. Furthermore, we directly compared 15 major US and UK newspapers on which topics they tend to cover more often and their writing style. In Fig. 3 we visualise the comparison of outlets based on their writing style: outlets with similar writing style are closer together. 4.2 Sentiment Analysis of Twitter Content Measuring the current public mood is a challenging task. The traditional approach would require questioning a large number of people about their feelings. However, social media, such as Twitter, can easily become a Fig. 3. Comparison of news outlets in the US and UK based upon their writing style. valuable source of information about the public due to the fact that people use them to express their feelings in public. As demonstrated in our study [7], it is feasible to capture the public mood by monitoring the stream of Twitter data. The dataset that was analysed was comprised of 484 million tweets that were generated by more than 9.8 million users in UK, between July 2009 and January 2012. We focused on tracking four moods and for each mood we generate one timeline of the volume of related tweets. The further analysis of these timelines reveals that each of the four emotions changes over time in a rather predictable manner. For example, we found a periodic peak of joy around Christmas and a periodic peak of fear around Halloween. More surprisingly, we found that negative moods started to dominate the Twitter content after the announcement of massive cuts in public spending on October 2010. In Fig. 4, we plot the mood levels for the period of study and we visualise them as a facial expression using the Grimace tool. 5 Discussion The architecture we have developed for the creation of our media-content analysis platform presents a general solution to the problem of integrating very diverse algorithms into a single modular system. We anticipate that this approach can be applied to many other scenarios where data needs to be processed in a collaborative fashion by multiple software modules. One of its most appealing features is the way in which the various modules cooperate to annotate the data, in a decentralised manner: communication between modules takes place only via annotation left by them on the items contained in the common blackboard. This is an instance of stigmergic communication. Stigmergy is a mechanism for indirect coordination between agents, obtained by described by Brooks [12]. In the literature there is a plethora of works that focus on extracting information from public news datasets using some specific methodology or tool. On the other hand, there are very few computational systems and approaches that are oriented to the automation of more than a single perspective of news analysis. Some interesting examples include the Europe Media Monitor where they provide a coherent summary of current news [13]; and the work by Castillo et al. for the analysis of television news programs [14]. Currently in our group work is under way for the creation of a computer vision system based on the same principles, where multiple modules cooperate in the extraction of information from large quantities of images collected from news sites. Finally, we want to note that core parts of the system are distributed as an open source library that we call Macsy ("Modular Architecture Github: for https://github.com/mediapatterns/Macsy. The library contains the API that is used to coordinate the modules and keep track of the blackboards, as well as a series of implemented modules. It is built on top of MongoDB and it is developed in Java. Systems"), Cognitive at 6 References 1. D. Lazer, A. Pentland, L. Adamic, S. Aral, A.-L. Barabasi, et al., "Computational Social Science", Science, vol. 323, 2009, pp. 721-723. 2. D. Watts, "A twenty-first century science", Nature, vol. 445, 2007, pp. 489. 3. I. Flaounas, O. Ali, M. Turchi, T. Snowsill, F. Nicart, T. De Bie, N. Cristianini, "NOAM: news outlets analysis and monitoring system", Proc. of SIGMOD Conference, ACM, 2011, pp. 1275-1278. 4. V. Lampos and N. Cristianini, "Nowcasting Events from the Social Web with Statistical Learning", ACM Transactions on Intelligent Systems and Technology, vol. 3, no. 4, 2011. 5. I. Flaounas, O. Ali, T. Lansdall-Welfare, T. De Bie, N. Mosdell, J. Lewis, N. Cristianini, "Research Methods in the Age of Digital Journalism: Massive-scale automated analysis of news-content – topics, style and gender", Digital Journalism, Routledge, vol. 1, no. 1, 2013, pp. 102-116. 6. I. Flaounas, M. Turchi, O. Ali, N. Fyson, T. De Bie, N. Mosdell, J. Lewis, N. Cristianini, "The Structure of EU Mediasphere", PLoS ONE, vol. 5, no. 12, 2010, pp. e14243. 7. T. Lansdall-Welfare, V. Lampos, N. Cristianini, "Effects Fig. 4. Visualising mood levels in UK Twitter over time with facial expressions. each agent leaving traces (stigmata) in the environment while performing an action, that can be read by another agent, and affect its behaviour, for example triggering another action. In nature this mechanism is used to generate coherent collective behaviour, and is one of the ways in which complex systems self-organise without need for central planning or control, or for direct communication between individual agents. This mechanism allows efficient collaboration to take place between very simple agents who lack memory, planning and other cognitive capabilities. It is the way in which ants coordinate their behaviour, for example, the same method is found across many natural and artificial systems [8]. the common shared We implemented the general approach to distributed, decentralised coordination in a modular system by using a blackboard architecture, whereby all communications between modules are forced to take place via the reading and writing on information (blackboard). This provides a simple solution to the notoriously difficult problem of segmenting complex behaviour into several simpler modules: by insisting that only limited information is passed among modules, via the blackboard, we find a natural way to decompose complex processing into modules. The idea of using a common blackboard where multiple agents can read and write is very old in Artificial Intelligence, going back at least to the influential Pandemonium system created by Oliver Selfridge in 1959 to coordinate the action of several daemons [9]. It is also worth noting that our approach to modularity is consistent with the basic axioms listed by Fodor [10, 11] to be expected in modular systems: Domain specificity, i.e., modules are specialised to operate on specific kinds of input; information encapsulation, i.e., modules do not need information from within other modules to operate, they only process the input they are provided using their own private i.e., modules process all their inputs without choice; and shallow outputs, i.e., they produce simple outputs, in our case typically 'tags'. The last two of his axioms are relative to biological systems and therefore are omitted here. Our approach to modularity is also related to the class of agent architectures known as "reactive robotics" information; mandatory operation, of the Recession on Public Mood in the UK", in Mining Social Network Dynamics (MSND) session on Social Media Applications in News and Entertainment at WWW '12., ACM, 2012, pp. 1221-1226. 8. E. Bonabeau, "Editor's Artificial Life, MIT Press, vol. 5, no. 2, 1999, pp. 95-96. introduction: stigmergy", 9. O.G. Selfridge, "Pandemonium: a paradigm for learning", the Symposium on Mechanisation of Thought Processes, D.V. Blake and A.M. Uttley (eds), 1959, pp. 513-526. In Proceedings of 10. J.A. Fodor, "The Modularity of Mind", MIT Press, 1983. 11. P. Robbins, "Modularity of Mind", The Stanford Encyclopedia of Philosophy, E.N. Zalta (eds.), 2010. 12. R.A. Brooks, "Intelligence Without Representation", . Artificial Intelligence, vol. 47, 1991, pp. 139-159. 13. R. Steinberger, B. Pouliquen, and E. Van Der Goot, "An introduction to the Europe Media Monitor family of applications", in Information Access in a Multilingual World- Proceedings of the SIGIR, 2009, pp. 1-8. 14. C. Castillo, GDF Morales, M. Mendoza, N. Khan, "Automatic Analysis of Television News: Media, People, Framing and Bias", arxiv.org preprint, arXiv:1307.4879, 2013. 7 Acknowledgments This research is funded by European Community's Seventh Framework Programme under grant agreement N° 270327 (CompLACS Project)
1112.1338	2	1112	2012-03-18T18:16:44	The Role of Persistent Graphs in the Agreement Seeking of Social Networks	[ "cs.MA" ]	This paper investigates the role persistent arcs play for a social network to reach a global belief agreement under discrete-time or continuous-time evolution. Each (directed) arc in the underlying communication graph is assumed to be associated with a time-dependent weight function which describes the strength of the information flow from one node to another. An arc is said to be persistent if its weight function has infinite $\mathscr{L}_1$ or $\ell_1$ norm for continuous-time or discrete-time belief evolutions, respectively. The graph that consists of all persistent arcs is called the persistent graph of the underlying network. Three necessary and sufficient conditions on agreement or $\epsilon$-agreement are established, by which we prove that the persistent graph fully determines the convergence to a common opinion in social networks. It is shown how the convergence rates explicitly depend on the diameter of the persistent graph. The results adds to the understanding of the fundamentals behind global agreements, as it is only persistent arcs that contribute to the convergence.	cs.MA	cs	The Role of Persistent Graphs in the Agreement Seeking of Social Networks ∗ Guodong Shi and Karl Henrik Johansson† Abstract This paper investigates the role persistent arcs play for a social network to reach a global belief agreement under discrete-time or continuous-time evolution. Each (directed) arc in the underlying communication graph is assumed to be associated with a time-dependent weight function which describes the strength of the information flow from one node to another. An arc is said to be persistent if its weight function has infinite L1 or ℓ1 norm for continuous- time or discrete-time belief evolutions, respectively. The graph that consists of all persistent arcs is called the persistent graph of the underlying network. Three necessary and sufficient conditions on agreement or ǫ-agreement are established, by which we prove that the persistent graph fully determines the convergence to a common opinion in social networks. It is shown how the convergence rates explicitly depend on the diameter of the persistent graph. The results adds to the understanding of the fundamentals behind global agreements, as it is only persistent arcs that contribute to the convergence. Keywords: Consensus, Persistent Graphs, Social Networks, Dynamical Systems 1 Introduction Recent years have witnessed wide research interest in opinion dynamics and information ag- gregation of social networks. Individuals are equipped with beliefs or opinions which updated as information is exchanged from time to time; how beliefs are propagated depends on the interactions between individuals. ∗This work has been supported in part by the Knut and Alice Wallenberg Foundation, the Swedish Research Council and KTH SRA TNG. †G. Shi and K. H. Johansson are with ACCESS Linnaeus Centre, School of Electrical Engineering, Royal Institute of Technology, Stockholm 10044, Sweden. Email: [email protected], [email protected] 1 DeGroot's model is a classical model on belief evolution [10]. It is simply formulated as a discrete-time linear system, where the state transition matrix is time-invariant and row stochas- tic. The ij-entry of the transition matrix of DeGroot's model represents the weight of the arc which marks the influence of j to i. The convergence to an agreement is equivalent to the conver- gence to a stationary distribution of the finite-state Markov chain given by the same transition matrix. Results from Markov chain analysis can therefore be used in the agreement analysis [10, 7, 8]. Variations of DeGroot's model are considered in [11, 12, 13, 14] for the study of opinion dynamics in social networks. Here if an asymptotic belief agreement can be reached or not has always been a central question. Consensus problems which are very related to DeGroot's model appear in many different contexts in the study of computer science and engineering, e.g., decentralized and parallel com- putations [15, 36, 37], coordinations of autonomous agents [16, 22, 20, 21] and sensor networks [38, 39, 32]. Agreement seeking has been extensively studied in the literature for both discrete- time and continuous-time models [16, 15, 40, 20, 18, 27, 28, 19, 17, 25, 30, 26]. The communication graph plays an important role in proper conditions to ensure a consensus. In most existing work, the arc weights, which reflect the strength of the influence from one node to another, are assumed to either be constant whenever two nodes meet with each other [10, 18, 17], or in a compact set with positive lower and upper bounds [40, 16, 27, 28, 13]. However, in reality, the arc weights may vary in a wide range, and may even fade away. Moreover, different arcs may have quite different persistency properties. For instance, the opinion of people may be heavily influenced over short time periods by political campaigns, but over long time periods persistent links to family and friends might be more important. This is to say, the weights of the opinions from different sources are in practice generally time-varying and highly irregular over the underlying communication graph, and therefore, links can be impulsive, vanishing, persistent, etc. Then an interesting question arises: are there certain arcs which are the ones that actually generate the convergence to a consensus and how do their graph properties influence the convergence rate? The central aim of the paper is to build a model to classify different arcs in the underlying communication graph, and then give a precise description on how the persistent arcs indeed determine the agreement seeking. We define the persistent graph as the graph having links whose weight functions have infinite L1 or ℓ1 norm for continuous-time or discrete-time belief dynamics, respectively. Global agreement and ǫ-agreement are defined as whether the max- 2 imum state difference converges to zero, and whether the convergence is exponentially fast, respectively. For the discrete-time case, a necessary and sufficient condition is obtained on ǫ- agreement under general stochasticity, self-confidence and arc balance assumptions. Then for the continuous-time case, two necessary and sufficient conditions are established on global agree- ment and ǫ-agreement, respectively. In this way, we precisely state how the persistent graph plays a fundamental role in consensus seeking. Additionally, comparisons of our new conditions are given with existing results and the relations between the discrete-time and continuous-time evolutions are highlighted. The rest of the paper is organized as follows. In Section 2, we introduce the network model and define the problem of interest. Then in Sections 3 and 4, the main results and convergence analysis are presented for discrete-time and continuous-time dynamics, respectively. Finally some discussions and concluding remarks are given in Sections 5 and 6. 2 Problem Definition In this section, we present the social network model and define the considered problem. To this end, we first introduce some basic graph theory [4]. A (simple) digraph G = (V, E) consists of a finite set V = {1, . . . , n} of nodes and an arc set E, where each arc (i, j) ∈ E is an ordered pair from node i ∈ V to another node j ∈ V. If the arcs are pairwise distinct in an alternating sequence v0e1v1e2v2 . . . ekvk of nodes vi and arcs ei = (vi−1, vi) ∈ E for i = 1, 2, . . . , k, the sequence is called a (directed) path with length k. A path from i to j is denoted i → j, and the length of i → j is denoted i → j. A path with no repeated nodes is called a simple path. If there exists a path from node i to node j, then node j is said to be reachable from node i. Each node is thought to be reachable by itself. A node v from which any other node is reachable is called a center (or a root) of G. G is said to be strongly connected if it contains path i → j and j → i for every pair of nodes i and j; G is said to be quasi-strongly connected if G has a center [5, 25]. The distance from i to j, d(i, j), is defined as the length of a shortest (simple) path i → j when j is reachable from i, and the diameter of G as d0 = max{d(i, j)i, j ∈ V, j is reachable from i}. In this paper, we consider a social network model with node set V = {1, . . . , n}. Let the digraph G∗ = (V, E∗) denote the underlying graph of the considered social network. The under- lying graph indicates all potential interactions between nodes. Node j is said to be a neighbor 3 of i at time t when there is an arc (j, i) ∈ E∗; each node is supposed to be a neighbor of itself. Let Ni = {i} ∪ {j : (j, i) ∈ E∗} denote the neighbor set of node i. Let xi(t) ∈ R be the belief of node i at time t. Time is either discrete or continuous. The initial time is t0 ≥ 0 in both cases and each node is equipped with an initial belief xi(t0). The belief updating rule is in discrete time: and in continuous time: xi(t + 1) = Xj∈Ni Wij(t)xj(t), i = 1, . . . , n xi(t) = Xj∈Ni Wij(t)(cid:2)xj(t) − xi(t)(cid:3), i = 1, . . . , n. (1) (2) Here Wij(t) : [0, ∞) → [0, ∞) is a nonnegative scalar function which represents the weight of arc (j, i). Clearly Wij(t) describes the strength of the influence of node j on i. Since Wij(t) = 0 may happen from time to time, the graph is indeed time-varying. We define ψ(t) . = min i∈V {xi(t)}, Ψ(t) . = max i∈V {xi(t)} as the minimum and maximum state value at time t, respectively. Then H(t) . = Ψ(t) − ψ(t) is a natural agreement measure marking the maximum distances between the individual beliefs. The considered global agreement and ǫ-agreement for both the discrete-time and continuous-time updating rules are defined as follows. Definition 1 (a) Global agreement is achieved if for any x(t0) . = (x1(t0) . . . xn(t0))T ∈ Rn, we have lim t→∞ H(t) = 0. (3) (b) Global ǫ-agreement is achieved if there exist two constants 0 < ǫ < 1 and T0 > 0 such that for any x(t0) ∈ Rn and t ≥ t0, we have H(t + T0) ≤ ǫH(t). (4) Remark 1 A global agreement only requires that H(t) will converge to zero as t tends to infinity. If it is further required that the convergence speed is at least exponentially fast, we use global ǫ-agreement. This definition of ǫ-agreement and other similar concepts have been widely used to characterize the convergence rate of consensus evolutions in the literature, e.g., [18, 40, 41, 42]. 4 The goal of this paper is to distinguish the arcs from the underlying graph that are persistent over a long time range and how they influence global agreement. To be precise, we impose the following definition for persistent arcs and persistent graphs based on the L1 or ℓ1 norms of the weight functions. Figure 1: The underlying graph consists of persistent arcs (solid) and vanishing arcs (dashed). The persistent graph is shown to play a fundamental role for the convergence to an agreement. Definition 2 (a) An arc (j, i) ∈ G∗ is a persistent arc of the discrete-time updating rule (1) if ∞ Xt=0 Wij(t) = ∞, and a persistent arc of the continuous-time updating rule (2) if Z ∞ s Wij(t)dt = ∞ for all s ≥ 0. (b) The graph Gp = (V, E p) that consists of all persistent arcs is called the persistent graph. Next, in Sections 3 and 4, we will investigate the discrete-time and continuous-time updating rules, respectively. We will establish sufficient and necessary conditions on global agreement and ǫ-agreement, which illustrate that the notion of persistent graphs is critical to the convergence. 3 Discrete-time Belief Evolution In this section, we focus on the discrete-time belief evolution (1). In order to obtain the main result, we need the following assumptions. A1 (Stochasticity) Pj∈Ni Wij(t) = 1 for all i ∈ V and t ≥ 0. 5 A2 (Self-confidence) There exists 0 < η < 1 such that Wii(t) ≥ η for i ∈ V and t ≥ 0. A3 (Arc Balance) There exists a constant A > 1 such that for any two arcs (j, i), (m, k) ∈ E p and t ≥ 0, we have A−1Wij(t) ≤ Wkm(t) ≤ AWij(t). The main result for the discrete-time updating rule (1) on global ǫ-agreement is as follows. Theorem 1 Suppose A1, A2 and A3 hold. Global ǫ-agreement is achieved for (1) if and only if (a) Gp is quasi-strongly connected; (b) there exist a constant a∗ > 0 and an integer T∗ > 0 such that Pt+T∗−1 s=t Wij(s) ≥ a∗ for all t ≥ 0 and (j, i) ∈ E p. In fact, if (a) and (b) hold, then we have H(t + d0T∗) ≤ (cid:16)1 − ηd0T∗ 2 a∗ T∗(cid:1)d0(cid:17)H(t) ·(cid:0) (5) for all t ≥ t0, where d0 represents the diameter of Gp. Remark 2 Consensus convergence for many variations of (1) has been extensively studied in the literature, e.g., [1, 13, 12, 7, 8, 16, 21, 20, 27]. As for convergence rate, a relatively con- servative bound is given in [1, 16], and then generalized in [28, 41]. Recently a sharper bound for convergence rate was obtained in [42]. The self-confidence condition A2 is generally not necessary to ensure a consensus, but the convergence properties may be quite different without A2, especially for the case with time-varying graphs. Remark 3 Most of existing results are based on the assumption that all weight functions Wij(t) in the underlying graph have a positive lower bound whenever they are not zero. Here we just need the self-loop weights, Wii(t), i = 1, . . . , n, to have a positive lower bound. As indicated by the proof below, the sufficiency statement of Theorem 1 relies on the self-confidence assumption A2, while the arc balance assumption A3 is used in the necessity part. Before we state the proof, we introduce some more notations, which will be used throughout the rest of the paper. For two sets S1 and S2, S1 \ S2 is defined as S1 \ S2 = {z : z ∈ S1, z /∈ S2}. For the underlying graph G∗ = (V, E∗) and the persistent graph Gp = (V, E p), we denote θ(t) = X(j,i)∈E∗\E p Wij(t), 6 (6) and ξ+(t; m) = Xj∈Nm\{m} Wmj(t), ξ+ 0 (t; m) = Xj∈Nm\{m},(j,m)∈E p Wmj(t). (7) In the following two subsections, we prove the necessity and sufficiency parts of Theorem 1, respectively. 3.1 Necessity We need to show that a global ǫ-agreement cannot be achieved without either condition (a) or (b). The upcoming analysis relies on the following well-known lemmas. Lemma 1 Suppose 0 ≤ pk < 1 for all k. Then P∞ k=0 pk = ∞ if and only if Q∞ k=0(1 − pk) = 0. Lemma 2 log(1 − t) ≥ −2t for all 0 ≤ t ≤ 1/2. We have the following proposition indicating that Gp being quasi-strongly connected is not only a necessary condition for (1) to reach global ǫ-agreement, but also necessary for (simple) global agreement, even in the absence of assumptions A2 and A3. Proposition 1 Suppose A1 holds. If global agreement is achieved for (1), then Gp is quasi- strongly connected. Proof. Suppose Gp is not quasi-strongly connected. Then there exist two distinct nodes u and w such that Vu ∩ Vw = ∅, where Vu = {nodes from which u is reachable in Gp} and Vw = {nodes from which w is reachable in Gp}. Moreover, there is no arc entering either Vu or Vw in the persistent graph Gp. Let xi(t0) = 0 for all i ∈ Vu, and xi(t0) = 1 for all i ∈ V \ Vu. Denote ℓ(t) = maxi∈Vu xi(t) and (t) = mini∈Vw xi(t). We define g+(t; m) = Pj∈Nm,j /∈Vu m ∈ Vu and f +(t; k) = Pj∈Nk,j /∈Vw Wkj(t) for k ∈ Vw. We further denote f +(t; k). Wmj(t) for g+(t; m); u (t) = Xm∈Vu ζ + ζ + w (t) = Xk∈Vw It is straightforward to see that ψ(t) is non-decreasing and Ψ(t) is non-increasing for (1). It follows that xi(t) ∈ [0, 1] for all i and t ≥ t0. There are two cases. 7 (i). First, for any m ∈ Vu, we have xm(t0 + 1) = Xj∈Nm Wmj(t0)xj(t0) ≤ 0 · (cid:0)1 − g+(t0; m)(cid:1) + 1 · g+(t0; m) ≤ ζ + u (t0), which yields ℓ(t0 + 1) ≤ ζ + u (t0) immediately. Then, for the next slot we have that for any m ∈ Vu, xm(t0 + 2) = Xj∈Nm Wmj(t0 + 1)xj(t0 + 1) ≤ Xj∈Nm,j∈Vu = ζ + Wmj(t0 + 1)ℓ(t0 + 1) + Xj∈Nm,j /∈Vu u (t0) · (cid:0)1 − g+(t0 + 1; m)(cid:1) + g+(t0 + 1; m) u (t0) + ζ + u (t0 + 1), ≤ ζ + Wmj(t0 + 1) · 1 (8) which leads to ℓ(t0 + 1) ≤ ζ + u (t0) + ζ + u (t0 + 1). Continuing we get that for any s = 1, 2, . . . , we have ℓ(t0 + s) ≤ t0+s−1 Xt=t0 ζ + u (t) ≤ ∞ Xj=t0 θ(t) < ∞ (9) because there is no arc entering Vu in the persistent graph Gp. (ii). Consider now Vw. According to the definition of θ(t), there exists T1 > 0 such that when θ(t) < 1, t ≥ T1. Let t0 ≥ T1. Then we have ζ + no arc entering Vw in the persistent graph Gp. w (t) ≤ θ(t) < 1 for all t ≥ t0 since there is Similarly we obtain (t0 + 1) ≥ 1 − ζ + w (t0) since for any k ∈ Vw, we have xk(t0 + 1) = Xj∈Nk Wkj(t0)xj(t0) ≥ 0 · f +(t0; k) + 1 · (cid:0)1 − f +(t0; k)(cid:1) ≥ 1 − ζ + w (t0). Furthermore, for any k ∈ Vw, one has xk(t0 + 2) ≥ 0 · f +(t0 + 1; k) +(cid:0)1 − f +(t0; k)(cid:1) ·(cid:0)1 − ζ + w (t0)(cid:1) (10) w (t0 + 1)(cid:1) ·(cid:0)1 − ζ + ≥ (cid:0)1 − ζ + w (t0 + 1)(cid:1) ·(cid:0)1 − ζ + w (t0)(cid:1), and thus (t0 + 2) ≥ (cid:0)1 − ζ + for any s = 1, 2, . . . , w (t0)(cid:1). Proceeding the analysis we know that (t0 + s) ≥ t0+s−1 ∞ Yt=t0 (cid:0)1 − ζ + w (t)(cid:1) ≥ Yt=t0(cid:0)1 − θ(t)(cid:1) ≥ ∞ Yt=T1(cid:0)1 − θ(t)(cid:1) . = σ∗ > 0, (11) where σ∗ exists from Lemma 1 and the definition of θ(t). 8 Because P∞ j=0 θ(t) < ∞, we can always choose t0 sufficiently large so that P∞ Therefore, (9) and (11) lead to H(t0 + s) ≥ (t0 + s) − ℓ(t0 + s) ≥ σ∗/2 > 0. A global agreement θ(t) ≤ σ∗/2. j=t0 is thus impossible. This completes the proof. (cid:3) We establish a lemma on the upper and lower bounds for some particular nodes. Lemma 3 Suppose A1 holds. Let xm(t) = µψ(t) + (1 − µ)Ψ(t) with 0 ≤ µ ≤ 1. Then for any integer T > 0, we have: xm(t + T ) ≤ µ t+T −1 Ys=t (cid:0)1 − ξ+(s; m)(cid:1) · ψ(t) +(cid:16)1 − µ t+T −1 Ys=t (cid:0)1 − ξ+(s; m)(cid:1)(cid:17) · Ψ(t). and xm(t + T ) ≥ µ t+T −1 Ys=t (cid:0)1 − ξ+(s; m)(cid:1) · Ψ(t) +(cid:16)1 − µ t+T −1 Ys=t (cid:0)1 − ξ+(s; m)(cid:1)(cid:17) · ψ(t). Proof. When xm(t) = µψ(t) + (1 − µ)Ψ(t), for time t + 1, we have xm(t + 1) = Xj∈Nm Wmj(t)xj(t) ≤ (cid:0)1 − ξ+(t; m)(cid:1) ·(cid:0)µψ(t) + (1 − µ)Ψ(t)(cid:1) + ξ+(t; m)Ψ(t) = µ(cid:0)1 − ξ+(t; m)(cid:1) · ψ(t) +(cid:16)1 − µ(cid:0)1 − ξ+(t; m)(cid:1)(cid:17)Ψ(t). For time t + 2, we obtain xm(t + 2) ≤ (cid:0)1 − ξ+(t + 1; m)(cid:1) · hµ(cid:0)1 − ξ+(t; m)(cid:1) · ψ(t) +(cid:16)1 − µ(cid:0)1 − ξ+(t; m)(cid:1)(cid:17)Ψ(t)i + ξ+(t + 1; m)Ψ(t) = µ t+1 Ys=t(cid:0)1 − ξ+(s; m)(cid:1) · ψ(t) +(cid:16)1 − µ t+1 Ys=t(cid:0)1 − ξ+(s; m)(cid:1)(cid:17) · Ψ(t). (12) (13) (14) (15) Continuing, we obtain (12). In equality (13) can be easily obtained using a symmetric analysis as for (12). (cid:3) We are now in a place to present the following conclusion, which shows the necessity of condition (b) in Theorem 1. Proposition 2 Suppose A1 and A3 hold. If global ǫ-agreement is achieved for (1), then there exist a constant a∗ > 0 and an integer T∗ > 0 such that Pt+T∗ (j, i) ∈ Gp. s=t Wij(s) ≥ a∗ for all t ≥ 0 and 9 Proof. We prove the conclusion by contradiction. Suppose the condition does not hold. Then ∀0 < ǫ < 1, T > 0, ∃t∗(T, ǫ) ≥ 0 and (j∗, i∗) ∈ E p such that t∗+T −1 Xs=t∗ Wi∗j∗(s) < 1 2 A−1(n − 1)−1 · log(cid:16) 1 + ǫ 2 (cid:17)−1 . (16) Since (j∗, i∗) ∈ Gp, it is straightforward to see that t∗(T, ǫ) → ∞ as T → ∞ for any fixed ǫ. Thus, we can assume that (16) also holds for the arcs in E∗ \ E p. Moreover, without loss of generality, we can also assume that ξ+(s; i) ≤ 1/2 for all i and t∗ ≤ s ≤ t∗ + T − 1. With arc balance assumption A3 and Lemma 2, (16) implies t∗+T −1 Ys=t∗ (cid:0)1 − ξ+(s; i)(cid:1) = ePt∗+T −1 s=t∗ log(cid:0)1−ξ+(s;i)(cid:1) ≥ e−2 Pt∗+T −1 s=t∗ ξ+(s;i) > e − log(cid:16) 1+ǫ 2 (cid:17)−1 = 1 + ǫ 2 (17) for all i ∈ V. Moreover, taking xm(t∗) = ψ(t∗) and xk(t∗) = Ψ(t∗), we know from Lemma 3 that xm(t∗ + T ) ≤ t∗+T −1 Ys=t∗ (cid:0)1 − ξ+(s; m)(cid:1) · ψ(t∗) +(cid:16)1 − t∗+T −1 Ys=t∗ (cid:0)1 − ξ+(s; m)(cid:1)(cid:17) · Ψ(t∗) and xk(t∗ + T ) ≥ t∗+T −1 Ys=t∗ (cid:0)1 − ξ+(s; k)(cid:1) · Ψ(t∗) +(cid:16)1 − t∗+T −1 Ys=t∗ (cid:0)1 − ξ+(s; k)(cid:1)(cid:17) · ψ(t∗). Therefore, with (17), (18) and (19), we eventually obtain H(t∗ + T ) ≥ xk(t∗ + T ) − xm(t∗ + T ) t∗+T −1 Ys=t∗ (cid:0)1 − ξ+(s; k)(cid:1) + ≥ h > (cid:0)2 · 1 + ǫ 2 − 1(cid:1)H(t∗) t∗+T −1 Ys=t∗ (cid:0)1 − ξ+(s; m)(cid:1) − 1i · H(t∗) = ǫH(t∗). The desired conclusion thus follows. The necessity claim in Theorem 1 follows from Propositions 1 and 2. 3.2 Sufficiency (18) (19) (20) (cid:3) We now present the sufficiency proof of Theorem 1. In fact, we are going to prove a stronger statement which does not rely on the arc balance assumption A3. 10 Proposition 3 Suppose A1 and A2 hold. Global ǫ-agreement is achieved for (1) if Gp is quasi-strongly connected and there exist a constant a∗ > 0 and an integer T∗ > 0 such that Pt+T∗−1 s=t Wij(s) ≥ a∗ for all t ≥ 0 and (j, i) ∈ Gp. Proof. Let i0 ∈ V be a center of Gp. Take t0 ≥ 0. Assume first that xi0(t0) ≤ 1 2 ψ(t0) + 1 2 Ψ(t0). Then from Lemma 3, one has t0+T −1 xi0(t0 + T ) ≤ 1 2 ≤ ηT 2 for all T = 0, 1, . . . . Ys=t0 (cid:0)1 − ξ+(s; i0)(cid:1) · ψ(t0) +(cid:16)1 − ψ(t0) +(cid:16)1 − ηT 2 (cid:17)Ψ(t0) 1 2 t0+T −1 Ys=t0 (cid:0)1 − ξ+(s; i0)(cid:1)(cid:17) · Ψ(t0) (21) (22) Denote V1 as the node set consisting of all the nodes of which i0 is a neighbor in Gp, i.e., V1 = {j : (i0, j) ∈ E p}. Note that V1 is nonempty because i0 is a center. For any i1 ∈ V1, there exists an instance ¯t1 ∈ [t0, t0 + T∗ − 1] such that Wi1i0(¯t1) ≥ a∗/T∗ because Pt0+T∗−1 t=t0 Wi1i0(t) ≥ a∗. Suppose ¯t1 = t0 + 1 with 1 ∈ [0, T∗ − 1]. Then with (22), we have xi1(¯t1 + 1) = xi1(t0 + 1 + 1) ≤ Wi1i0(t0 + 1)xi0(t0 + 1) +(cid:0)1 − Wi1i0(t0 + 1)(cid:1)Ψ(t0) T∗(cid:17)Ψ(t0) ·h η1 a∗ T∗ η1 a∗ ≤ ψ(t0) +(cid:0)1 − · ψ(t0) +(cid:16)1 − η1 · 2 (cid:1)Ψ(t0)i +(cid:16)1 − 2T∗(cid:17)Ψ(t0). a∗ 2 a∗ 2T∗ = η1 · (23) (24) (25) Based on Lemma 3, we can further conclude xi1(t0 + 1 + T ) ≤ η1+T −1 · a∗ 2T∗ · ψ(t0) +(cid:16)1 − η1+T −1 · a∗ 2T∗(cid:17)Ψ(t0) for all T = 1, 2, . . . , which implies xi1(t0 + T∗ + K) ≤ ηT∗+K · a∗ 2T∗ · ψ(t0) +(cid:16)1 − ηT∗+K · a∗ 2T∗(cid:17)Ψ(t0), K = 0, 1, . . . . Next, since Gp is quasi-strongly connected, we can denote V2 as the node set consisting of all the nodes each of which has a neighbor in {i0}∪V1 within Gp. For any i2 ∈ V2, there exist a node i∗ ∈ {i0} ∪ V1 and an instance ¯t2 = t0 + T∗ + 2 with 2 ∈ [0, T∗ − 1] such that Wi2i∗ (¯t1) ≥ a∗/T∗. Similarly we have xi2(¯t2 + 1) ≤ Wi2i∗ (t0 + T∗ + 2)xi∗ (t0 + T∗ + 2) +(cid:0)1 − Wi2i∗(t0 + T∗ + 2)(cid:1)Ψ(t0) T∗(cid:17)Ψ(t0) ·hηT∗+2 · a∗ 2T∗ a∗ · ψ(t0) +(cid:16)1 − ηT∗+2 · · (cid:0) ηT∗+2 T∗(cid:1)2 · ψ(t0) +(cid:16)1 − 2T∗(cid:17)Ψ(t0)i +(cid:16)1 − T∗(cid:1)2(cid:17)Ψ(t0), a∗ a∗ T∗ ηT∗+2 · (cid:0) a∗ ≤ = a∗ 2 2 11 (26) and therefore xi2(t0 + 2T∗ + K) ≤ η2T∗+K 2 a∗ T∗(cid:1)2ψ(t0) +(cid:16)1 − ·(cid:0) η2T∗+K 2 a∗ T∗(cid:1)2(cid:17)Ψ(t0), K = 0, 1, . . . · (cid:0) Proceeding the estimate, V3, . . . , Vk can be similarly defined until (cid:0) ∪k i=1 Vi(cid:1) ∪ {i0} = V. Moreover, it is not hard to see that i0 can be selected so that k = d0, where d0 is the diameter of Gp, and thus xi(t0 + d0T∗) ≤ ηd0T∗ 2 a∗ T∗(cid:1)d0 · ψ(t0) +(cid:16)1 − ·(cid:0) ηd0T∗ 2 a∗ T∗(cid:1)d0(cid:17)Ψ(t0), ·(cid:0) i = 1, . . . , n (27) which yields Ψ(t0 + d0T∗) ≤ ηd0T∗ 2 a∗ T∗(cid:1)d0 · ψ(t0) +(cid:16)1 − ·(cid:0) ηd0T∗ 2 a∗ T∗(cid:1)d0(cid:17)Ψ(t0). ·(cid:0) With (28), we eventually have H(t0 + d0T∗) ≤ ηd0T∗ 2 = (cid:16)1 − a∗ 2 ·(cid:0) ηd0T∗ T∗(cid:1)d0 · ψ(t0) +(cid:16)1 − T∗(cid:1)d0(cid:17)H(t0). ·(cid:0) a∗ ηd0T∗ 2 a∗ T∗(cid:1)d0(cid:17)Ψ(t0) − ψ(t0) ·(cid:0) For the opposite case of (21) with xi0(t0) > 1 2 ψ(t0) + 1 2 Ψ(t0), (28) (29) (30) (29) is obtained using a symmetric argument by bounding ψ(t0 + d0T∗) from below. Therefore, the desired conclusion follows with ǫ = 1 − ηd0 T∗ 2 holds independent with the choice of t0. T∗(cid:1)2 and T0 = d0T∗ since (29) ·(cid:0) a∗ (cid:3) 4 Continuous-time Belief Evolution In this section, we turn to the continuous-time updating rule. We need an assumption on the continuity of each weight function Wij(t) for the existence of trajectories of (2). A4 (Continuity) Each Wij(t), (j, i) ∈ E∗ is continuous except for a set with measure zero. With assumption A4, the set of discontinuity points for the right-hand side of equation (2) has measure zero. Therefore, the Caratheodory solutions of (2) exist for arbitrary initial conditions, and they are absolutely continuous functions that satisfy (2) for almost all t on the maximum interval of existence [3, 9]. In the following, each solution of (2) is considered in the sense of Caratheodory without explicit mention. 12 Let us first study the feasibility of the solutions of (2). Consider (2) with initial condition x(t0) = (x1(t0), . . . , xn(t0))T = x0 ∈ Rn, t0 ≥ 0. The upper Dini derivative of a function h : (a, b) → R at t is defined as D+h(t) = lim sup s→0+ h(t + s) − h(t) s The next result is useful for the calculation of Dini derivatives [6, 25]. Lemma 4 Let Vi(t, x) : R × Rm → R, i = 1, . . . , n, be C 1 and V (t, x) = maxi=1,...,n Vi(t, x). If I(t) = (cid:8)i ∈ {1, . . . , n} : V (t, x(t)) = Vi(t, x(t))(cid:9) is the set of indices where the maximum is reached at t, then D+V (t, x(t)) = maxi∈I(t) Vi(t, x(t)). The following lemma establishes the monotonicity of Ψ(t) and ψ(t). Lemma 5 For all t ≥ t0 ≥ 0, we have D+Ψ(t) ≤ 0 and D+ψ(t) ≥ 0. Proof. We prove D+Ψ(t) ≤ 0. The other part can be proved similarly. Let I0(t) represent the set containing all the agents that reach the maximum in the definition of Ψ(t) at time t, i.e., I(t) = {i ∈ V xi(t) = Ψ(t)}. Then according to Lemma 4, we obtain D+Ψ(t) = max i∈I0(t) xi(t) = max i∈I0(t)h Xj∈Ni Wij(t)(xj (t) − xi(t))i ≤ 0, which completes the proof. (31) (cid:3) Lemma 5 implies, H(t) is non-increasing for all t ≥ t0, and therefore each (Caratheodory) trajectory of (2) is bounded within the initial states of the nodes. As a result, the trajectories exist in [t0, ∞) for any initial condition. The main result on global consensus and ǫ-consensus is stated in the following two theorems. Theorem 2 Suppose A3 and A4 hold. Global agreement is achieved for (2) if and only if Gp is quasi-strongly connected. Theorem 3 Suppose A3 and A4 hold. Global ǫ-agreement is achieved for (2) if and only if (a) Gp is quasi-strongly connected; (b) there exists two constants a∗, τ0 > 0 such that R t+τ0 t Wij(s)ds ≥ a∗ for all t ≥ 0 and (j, i) ∈ Gp. Moreover, if (a) and (b) hold, then we have H(cid:16)t + τ0 ·l d0 log 2 a∗ m(cid:17) ≤ (cid:16)1 − md0 2 (cid:17)H(t), 0 (32) 13 2 (cid:1)2 where m0 = (cid:0) ω0 smallest integer which is no smaller than z. 1 (n−1)A with ω0 = e− R ∞ 0 θ(t)dt, d0 is the diameter of Gp, and ⌈z⌉ represents the Theorem 2 implies that the connectivity of the persistent graph Gp totally determines whether an agreement can be achieved globally. Furthermore, Theorem 3 implies that R T O(T ) is a critical condition to ensure a global ǫ-consensus. 0 Wij(t)dt = Remark 4 Consensus for (2) was first studied in [17], where the convergence rate was shown to be determined by the second largest eigenvalue of the Laplacian of the communication graph. Further discussions can be found in [21, 25, 30]. Remark 5 Theorems 2 and 3 still hold if assumption A3 is replaced by the following integral version. A5. (Integral Arc Balance) There exists a constant A > 1 such that for any two arcs (j, i), (m, k) ∈ E p, we have for all 0 ≤ a < b. A−1Z b a Wij(t)dt ≤ Z b a Wkm(t)dt ≤ AZ b a Wij(t)dt Remark 6 If we have R T t=t0 Wij(t)dt = ∞, (j, i) ∈ Gp for some finite T , it follows from the proof of Theorem 2 below that (2) will reach a global agreement in finite time when t tends to T . 4.1 Preliminaries In this subsection, we establish two lemmas which describe the boundaries of how much each individual arc affects the nodes' dynamics. Then the proof of Theorems 2 and 3 will be proposed in the next two subsections. Lemma 6 Suppose xm(s) ≤ µψ(s) + (1 − µ)Ψ(s) for some s ≥ t0 and m ∈ V with 0 ≤ µ ≤ 1 a giving constant. Then we have xm(t) ≤ µe− R t s ξ+(τ ;m)dτ ψ(s) +(cid:2)1 − µe− R t s ξ+(τ ;m)dτ(cid:3)Ψ(s) (33) for all t ≥ s. Proof. Based on Lemma 5, we see that xm(t) = Xj∈Nm Wmj(t)(cid:2)xj(t) − xm(t)(cid:3) ≤ Xj∈Nm Wmj(t)(cid:2)Ψ(s) − xm(t)(cid:3) = −ξ+(t; m)(cid:2)xm(t) − Ψ(s)(cid:3), t ≥ s. (34) 14 This implies xm(t) ≤ e− R t ≤ µe− R t s ξ+(τ ;m)dτ xm(s) +(cid:2)1 − e− R t s ξ+(τ ;m)dτ ψ(s) +(cid:2)1 − µe− R t s ξ+(τ ;m)dτ(cid:3)Ψ(s) s ξ+(τ ;m)dτ(cid:3)Ψ(s) by Gronwall's inequality. The proof is completed. We give a lemma investigating the dynamic evolution between two connected nodes. Lemma 7 Suppose (l, m) ∈ E∗ and there exists a constant 0 < µ < 1 such that xl(t) ≤ µψ(s0) + (1 − µ)Ψ(s0), t ∈ [s0, s] for t0 ≤ s0 < s. Then we have (35) (cid:3) u ξ+(τ ;m)dτ Wml(u)du · ψ(s0) xm(t) ≤ µZ t s0 e− R t +h1 − µZ t s0 e− R t u ξ+(τ ;m)dτ Wml(u)duiΨ(s0), t ∈ [s0, s]. (36) Proof. Similar to (34), for any t ∈ [s0, s], we have xm(t) ≤ (cid:2)ξ+(t; m) − Wml(t)(cid:3) · (cid:2)Ψ(s0) − xm(t)(cid:3) + Wml(t)(cid:2)µψ(s0) + (1 − µ)Ψ(s0) − xm(t)(cid:3). Therefore, noting the fact that Z t s0 we obtain e− R t u ξ+(τ ;m)dτ ξ+(u; m)du = Z t s0 d duhe− R t u ξ+(τ ;m)dτi = 1 − e− R t s0 ξ+(τ ;m)dτ , xm(t) ≤ e− R t s0 e− R t u ξ+(τ ;m)dτ(cid:2)ξ+(u; m) − Wml(u)(cid:3)du · Ψ(s0) u ξ+(τ ;m)dτ Wml(u)du ·hµψ(s0) + (1 − µ)Ψ(s0)i e− R t u ξ+(τ ;m)dτ(cid:2)ξ+(u; m) − Wml(u)(cid:3)du · Ψ(s0) s0 s0 e− R t ξ+(τ ;m)dτ xm(s0) +Z t +Z t ξ+(τ ;m)dτ Ψ(s0) +Z t +Z t s0 e− R t +h1 − µZ t e− R t e− R t s0 s0 s0 ≤ e− R t s0 = µZ t u ξ+(τ ;m)dτ Wml(u)du ·hµψ(s0) + (1 − µ)Ψ(s0)i u ξ+(τ ;m)dτ Wml(u)du · ψ(s0) u ξ+(τ ;m)dτ Wml(u)duiΨ(s0), t ∈ [s0, s] by Gronwall's inequality and some simple manipulations. This completes the proof. 15 (37) (cid:3) 4.2 Proof of Theorem 2 Sufficiency Let i0 ∈ V be a center of Gp. Assume first that xi0(t0) ≤ 1 2 ψ(t0) + 1 2 Ψ(t0). (38) Denote ω0 = e− R ∞ 0 θ(t)dt. Then we have 0 < ω0 ≤ 1. Thus, based on Lemma 6 and noting the fact that ψ(t0) ≤ Ψ(t0), we have xi0(t) ≤ ≤ Define t0 1 e− R t 2 ω0 2 e− R t t0 ξ+(τ ;i0)dτ ψ(t0) +(cid:2)1 − 0 (τ ;i0)dτ ψ(t0) +h1 − ξ+ 1 2 t0 e− R t ω0 2 ξ+(τ ;i0)dτ(cid:3)Ψ(t0) e− R t 0 (τ ;i0)dτiΨ(t0). ξ+ t0 t1 = inf nt ≥ t0 : e− R t t0 ξ+ 0 (τ ;i0)dτ = 1 2o. We see that t1 is finite from the definition of E p. As a result, we obtain xi0(t) ≤ ω0 4 ψ(t0) +(cid:2)1 − ω0 4 (cid:3)Ψ(t0), t ∈ [t0, t1]. (39) (40) Next, we denote the node set consisting of all the nodes of which i0 is a neighbor in Gp as V1, i.e., V1 = {j : (i0, j) ∈ E p}. Note that V1 is nonempty because i0 is a center. Then for any i1 ∈ V1, we see from Lemma 7 that t1 u ξ+(τ ;i1)dτ Wi1i0(u)du · ψ(t0) 4 Z t1 e− R u ξ+(τ ;i1)dτ Wi1i0(u)duiΨ(s0) t1 t0 ω0 t1 u ξ+ 0 (τ ;i1)dτ Wi1i0(u)du · ψ(t0) ω0 xi1(t1) ≤ ≤ t0 e− R 4 Z t1 +h1 − 4 Z t1 e− R +h1 − t0 ω2 0 ω2 0 4 Z t1 t0 t1 u ξ+ e− R 0 (τ ;i1)dτ Wi1i0(u)duiΨ(s0). (41) The arc balance assumption A3 implies that Z t1 u 0 (t; i1)dt ≤ Z t1 ξ+ u (n − 1)AWi1i0(t)dt, 16 which yields Z t1 t0 t1 u ξ+ e− R 0 (τ ;i1)dτ Wi1i0(u)du ≥ Z t1 t0 e−(n−1)A R t1 u Wi1i0 (τ )dτ Wi1i0(u)du = = On the other hand, we also have 1 (n − 1)A Z t1 t0 e−(n−1)A R t1 u Wi1i0 (τ )dτ d du 1 (n − 1)A · h1 − e−(n−1)A R t1 t0 Wi1 i0 (τ )dτi. (42) Z t1 t0 0 (t; i0)dt ≤ Z t1 ξ+ t0 (n − 1)AWi1i0(t)dt. Thus, we know from (42) and the definition of t1 that Z t1 t0 e− R t1 u ξ+ 0 (τ ;i1)dτ Wi1i0(u)du ≥ ≥ = Equations (41) and (43) result in t1 t0 Wi1i0 (τ )dτi · h1 − e−(n−1)A R · h1 − e− R t1 t0 ξ+ 0 (τ ;i0)dτi 1 (n − 1)A 1 (n − 1)A 1 2(n − 1)A . xi1(t1) ≤ 2 (cid:1)2 for all i1 ∈ V1, where m0 = (cid:0) ω0 1 (n−1)A . m0 2 ψ(t0) + (1 − m0 2 )Ψ(t0) We continue to estimate the upper bound of nodes in {i0} ∪ V1 when t ≥ t1. Define Y(t) = max i∈{i0}∪V1 xi(t). Then Y(t1) ≤ m0 2 ψ(t0) + (1 − m0 2 )Ψ(t0). Similar to Lemma 6, we find that D+Y(t) ≤ −β(t)[Y(t) − Ψ(t1)], t ≥ t1, where β(t) = Pi∈{i0}∪V1,j /∈{i0}∪V1 Wij(t). This implies Y(t) ≤ e− R t t1 t1 β(τ )dτ Y(t1) +h1 − e− R t β(τ )dτh m0 · ω0e− R t t1 ψ(t0) + (1 − β(τ )dτ ψ(t0) +h1 − 2 ≤ e− R t t1 m0 2 ≤ β(τ )dτiΨ(t1) )Ψ(t0)i +h1 − e− R t m0 2 t1 m0 2 · ω0e− R t t1 β(τ )dτiΨ(t0), β(τ )dτiΨ(t0) where β(t) = Pi∈{i0}∪V1,j /∈{i0}∪V1,(j,i)∈E p Wij(t). We can then define V2 = nj /∈ {i0} ∪ V1 : ∃i ∈ {i0} ∪ V1 s.t. (i, j) ∈ E po 17 (43) (44) (45) and t2 = inf nt ≥ t1 : e− R t t1 β(τ )dτ = 1 2o and similar analysis with (44) gives a bound to any node i2 ∈ V2 as xi2(t2) ≤ m2 0 2 ψ(t0) +(cid:0)1 − m2 0 2 (cid:1)Ψ(t0). (46) Moreover, (46) also holds for nodes in {i0} ∪ V1. Since Gp has a center, we can proceed the estimation to nodes in V2, . . . , Vk until (cid:0) ∪k j=1 Vj(cid:1) ∪ {i0} = V with t2, . . . , tk such that for all i ∈ V, which leads to xi(tk) ≤ mk 0 2 ψ(t0) + (1 − mk 0 2 )Ψ(t0) Ψ(tk) ≤ mk 0 2 ψ(t0) + (1 − mk 0 2 )Ψ(t0). (47) (48) We see that i0 can be chosen so that k ≤ d0 always holds, where d0 is the diameter of Gp. Denoting t1 = tk, we eventually arrive at H(t1) = Ψ(t1) − ψ(t1) ≤ md0 0 2 ψ(t0) +(cid:0)1 − md0 0 2 (cid:1)Ψ(t0) − ψ(t0) = (cid:16)1 − md0 2 (cid:17)H(t0). 0 (49) Although the analysis up to now is based on assumption (38), we see that (49) also holds for the other case with xi0(t0) > 1 2 ψ(t0) + 1 2 Ψ(t0) using a symmetric argument by investigating the lower bound of ψ(t1). Similar estimate can be carried out for tk, k = 2, 3, . . . , which leads to H(tk+1) ≤ (cid:16)1 − md0 2 (cid:17)H(tk) 0 for all tk, k = 1, 2, . . . , which yields H(tk) ≤ (cid:16)1 − md0 0 2 (cid:17)k H(t0). (50) (51) Therefore, we can now conclude that limt→∞ H(t) = 0 because H(t) is non-increasing and 0 < m0 < 1. The sufficiency statement of Theorem 2 is thus proved. Necessity We follow the same line as the proof of Proposition 1. Suppose Gp is not quasi-strongly connected. Let Vu, Vw, ℓ(t) and (t) follow the definitions in the proof of Proposition 1. Also 18 let xi(t0) = 0 for all i ∈ Vu, and xi(t0) = 1 for all i ∈ V \ Vu. According to Lemma 5, we have xi(t) ∈ [0, 1]. Based on Lemma 4, we have D+ℓ(t) = max Wij(t)(cid:0)xj(t) − xi(t)(cid:1)i Wij(t)(cid:0)xj(t) − xi(t)(cid:1)i ≤ max i∈I1(t)h Xj∈Ni i∈I1(t)h Xj∈Ni\Vu i∈I1(t)h Xj∈Ni, (j,i)∈E∗\E p ≤ θ(t) ·(cid:0)1 − ℓ(t)(cid:1) = max Wij(t)(cid:0)xj(t) − xi(t)(cid:1)i (52) where I1(t) is the index set that contains the nodes where the maximum is reached and θ(t) is defined in (6). Similarly we have D+(t) = min Wij(t)(cid:0)xj(t) − xi(t)(cid:1)i Wij(t)(cid:0)xj(t) − xi(t)(cid:1)i ≥ min i∈I2(t)h Xj∈Ni i∈I2(t)h Xj∈Ni\Vw i∈I2(t)h Xj∈Ni, (j,i)∈E∗\E p = min Wij(t)(cid:0)xj(t) − xi(t)(cid:1)i ≥ −θ(t) · (t) where I2(t) is the index set that contains the nodes where the minimum is reached. With (52) and (53), denoting L(t) = (t) − ℓ(t), we obtain D+L(t) ≥ −θ(t) · ((t) − ℓ(t) + 1) = −θ(t) · (L(t) + 1), which is equivalent to Therefore, we have D+heR t t0 θ(τ )dτ (L(t) + 1)i ≥ 0. L(t) ≥ 2e− R t t0 θ(τ )dτ − 1. (53) (54) (55) Since R ∞ 0 θ(t)dt < ∞, we can choose t0 sufficiently large to ensure e− R t 3 for all t ≥ t0. This leads to H(t) ≥ L(t) ≥ 1/3, t ≥ t0. The necessity part of Theorem 2 thus follows. θ(τ )dτ ≥ 2 t0 19 4.3 Proof of Theorem 3 We first prove the necessity statement. Based on Theorem 2, we only need to prove that condition (b) in Theorem 3 is necessary. Suppose (b) in Theorem 3 does not hold. Then ∀0 < ǫ < 1, T > 0, ∃t∗(T, ǫ) ≥ 0 and (j0, i0) ∈ E p such that Z t∗+T t∗ Wi0j0(τ )dτ < A−1(n − 1)−1 · log ǫ−1 2 . (56) Since (j0, i0) ∈ Gp, it is not hard to see that t∗(T, ǫ) → ∞ as T → ∞ for any fixed ǫ. Thus, without loss of generality, we can assume that (56) also holds for the arcs in E∗ \ E p. Moreover, assumption A3 implies Z t∗+T t∗ Wij(τ )dτ < (n − 1)−1 · log ǫ−1 2 for all (j, i) ∈ E p. From similar argument we used to obtain (54), D+H(t) ≥ −2h X(j,i)∈E∗ Wij(t)iH(t), t ≥ t0. (57) (58) Therefore, letting the system initial time be t0 = t∗ with H(t∗) > 0, where t∗ is defined in (56), we see from (56) and (57) that 2 X(j,i)∈E∗ Z t∗+T t∗ Wij(τ )dτ < log ǫ−1. Consequently, (58) and (59) lead to H(t∗ + T ) ≥ e−2 P(j,i)∈E∗ R t∗+T t∗ Wij (τ )dτ H(t∗) > ǫH(t∗). (59) (60) Then the necessity part of Theorem 3 holds because ǫ and T are arbitrarily chosen in (60). Next, we prove the sufficiency part of Theorem 3 based on the convergence analysis in the proof of Theorem 2. When there exist two constants a∗, τ0 > 0 such that R t+τ0 (j, i) ∈ Gp, we have t Wij(τ )dτ ≥ a∗ for all t ≥ 0 and Z t+τ0 t b0(τ )dτ ≥ a∗, t ≥ 0, (61) where b0(t) = min(j,i)∈Gp Wij(t). Let us revisit the proof of Theorem 2. The definition of t1 in (39) satisfies t1 = inf nt ≥ t0 : e− R t t0 ξ+ 0 (τ ;i0)dτ = 1 2o ≤ inf nt ≥ t0 : e− R t t0 b0(τ )dτ = 1 2o. (62) 20 Similarly, for tj, j = 2, . . . , k with k ≤ d0, we have tj ≤ inf nt ≥ tj : − R t e tj−1 b0(τ )dτ = 1 2o. (63) Thus, for t1 = tk in (49), it holds that b0(τ )dτ = (cid:16) 1 t1 ≤ inf nt ≥ t0 : e− R t t0 2(cid:17)d0o = inf nt ≥ t0 : Z t t0 b0(τ )dτ = d0 log 2o. (64) Based on (61), we have j t − t0 τ0 ka∗ ≤ Z t t0 b0(τ )dτ, where ⌊z⌋ represents the largest integer which is no larger than z. This immediately implies t1 ≤ t0 + τ0 ·l d0 log 2 a∗ m, where ⌈z⌉ represents the smallest integer which is no smaller than z. Therefore, it can be concluded from Lemma 5 and (49) that H(cid:16)t0 + τ0 ·l d0 log 2 a∗ m(cid:17) ≤ (cid:16)1 − md0 2 (cid:17)H(t0). 0 (65) (66) The desired conclusion follows since (66) holds independent with the choice of t0. Thus, we have now completed the proof of Theorem 3. 5 Discussions In this section, we present some comparisons between our results with existing work, and com- parisons between the discrete-time and continuous-time belief evolutions. 5.1 Relation to Cut-balanced Graphs In [43], a cut-balance condition is introduced in the sense that there exists a constant K ≥ 1 such that for all t and any nonempty subset S ⊆ V, it holds that K−1 Xi∈S,j /∈S Wji(t) ≤ Xi∈S,j /∈S Wij(t) ≤ K Xi∈S,j /∈S Wji(t). (67) If the persistent graph Gp is strongly connected, the arc balance assumption A3 implies condition (67) over Gp. Therefore, in this particular case, assumption A3 is a special case of the cut-balance condition in [43], though assumption (67) in [43] is over the underlying graph G∗ . 21 Except for this slight difference, the convergence statements in Theorem 1 and Theorem 2 are consistent with the results given in [43] for strongly connected graphs. On the other hand, when Gp is quasi-strongly connected, the cut-balance condition never holds even under assumption A3, because there may be no arc pointing to the center node. Hence, in general, the results given in this paper provides conditions for node agreement inde- pendent of the conditions in [43]. 5.2 Discrete-time vs. Continuous-time Theorems 1 and 3 share quite similar structure and statement. However, there are some essential differences between them. (a) The discrete-time result in Theorem 1 highly relies on the self-confidence condition A2. Without A2, oscillations among the nodes may become inevitable and periodic solutions of (1) may arise for almost all initial condition even under A1 and A3. Note that the arc balance condition A3 is only useful for the necessity part of Theorem 1. (b) For the continuous-time result in Theorem 3, each self weight Wii(t) does not even show up in the model (2). The arc balance condition A3 is essential for the dynamics. Without A3, oscillations may occur if the arc weights of the persistent graph alternatively become large. Therefore, we can conclude that the self-confidence condition is critical for discrete-time belief agreement, as is the arc balance condition for continuous-time case. An interesting question is whether a similar conclusion can be made for the discrete-time model (1) as the statement in Theorem 2. This question is open and needs additional explo- rations. More general discussion on this problem can be found in [44] on the ergodicity of stochastic chains. 6 Conclusions Individuals are equipped with beliefs in social activities. The evolution of the beliefs can be modeled as dynamical systems over graphs using for instance the widely studied consensus algo- rithms. This paper studied persistent graphs under discrete-time and continuous-time consensus algorithms. Sufficient and necessary conditions were established on the persistent graph for the 22 network to reach global agreement or ǫ-agreement. It was shown that the persistent graph essentially determines both the convergence and convergence rate to an agreement. Acknowledgment The authors would like to thank Prof. Julien Hendrickx, Universit´e Catholique de Louvain, for many helpful discussions. References [1] D. P. Bertsekas and J. N. Tsitsiklis. Parallel and Distributed Computation: Numerical Methods, Prentice Hall, 1989. [2] F. Clarke, Y. Ledyaev, R. Stern, and P. Wolenski. Nonsmooth Analysis and Control Theory. Speringer-Verlag, 1998. [3] A. F. Filippov. Differential Equations with Discontinuous Righthand Sides. Norwell, MA: Kluwer, 1988. [4] C. Godsil and G. Royle. Algebraic Graph Theory. New York: Springer-Verlag, 2001. [5] C. Berge and A. Ghouila-Houri. Programming, Games, and Transportation Networks. John Wiley and Sons, New York, 1965. [6] J. Danskin. The theory of max-min, with applications. SIAM J. Appl. Math., vol. 14, 641-664, 1966. [7] J. Wolfowitz, "Products of indecomposable, aperiodic, stochastic matrices," Proc. Amer. Math. Soc., vol. 15, pp. 733-736, 1963. [8] J. Hajnal, "Weak Ergodicity in Non-homogeneous Markov Chains," Proc. Cambridge Phi- los. Soc., no. 54, pp. 233-246, 1958. [9] J. Cort´es, "Discontinuous dynamical systems -- a tutorial on solutions, nonsmooth analysis, and stability," IEEE Control Systems Magazine, vol. 28, no. 3, 36-73, 2008. 23 [10] M. H. DeGroot, "Reaching a consensus," Journal of the American Statistical Association, vol. 69, no. 345, pp. 118-121, 1974. [11] P. M. DeMarzo, D. Vayanos, J. Zwiebel, "Persuasion bias, social influence, and unidimen- sional opinions," Quarterly Journal of Economics, vol. 118, no. 3, pp. 909-968, 2003. [12] B. Golub and M. O. Jackson, "Naıve learning in social networks and the wisdom of crowds," American Economic Journal: Microeconomics, vol. 2, no. 1, pp. 112-149, 2007. [13] 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, 2010. [14] D. Acemoglu, G. Como, F. Fagnani, and A. Ozdaglar, "Opinion fluctuations and persistent disagreement in social networks," 50th IEEE Conference on Decision and Control, Orlando, Florida, December 2011. [15] J. N. Tsitsiklis, D. Bertsekas, and M. Athans, "Distributed asynchronous deterministic and stochastic gradient optimization algorithms," IEEE Trans. Automatic Control, vol. 31, no. 9, pp. 803-812, 1986. [16] 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, no. 6, pp. 988-1001, 2003. [17] R. Olfati-Saber and R. Murray, "Consensus problems in the networks of agents with switch- ing topology and time dealys," IEEE Trans. Automatic Control, vol. 49, no. 9, pp. 1520- 1533, 2004. [18] S. Boyd, A. Ghosh, B. Prabhakar and D. Shah, "Randomized gossip algorithms," IEEE Trans. Information Theory, vol. 52, no. 6, pp. 2508-2530, 2006. [19] J. Fax and R. Murray, "Information flow and cooperative control of vehicle formations," IEEE Trans. Automatic Control, vol. 49, no. 9, pp. 1465-1476, 2004. [20] L. Moreau, "Stability of multiagent systems with time-dependent communication links," IEEE Trans. Automatic Control, vol. 50, no. 2, pp. 169-182, 2005. 24 [21] W. Ren and R. Beard, "Consensus seeking in multiagent systems under dynamically chang- ing interaction topologies," IEEE Trans. on Automatic Control, vol. 50, no. 5, pp. 655-661, 2005. [22] S. Martinez, J. Cort´es, and F. Bullo, "Motion coordination with distributed information," IEEE Control Systems Magazine, vol. 27, no. 4, pp. 75-88, 2007. [23] H. G. Tanner, A. Jadbabaie, G. J. Pappas, "Flocking in fixed and switching networks," IEEE Trans. Automatic Control, vol. 52, no. 5, pp. 863-868, 2007. [24] Y. Hong, L. Gao, D. Cheng, and J. Hu, "Lyapuov-based approach to multi-agent systems with switching jointly connected interconnection," IEEE Trans. Automatic Control, vol. 52, pp. 943-948, 2007. [25] Z. Lin, B. Francis, and M. Maggiore, "State agreement for continuous-time coupled non- linear systems," SIAM J. Control Optim., vol. 46, no. 1, pp. 288-307, 2007. [26] F. Xiao and L. Wang, "Asynchronous consensus in continuous-time multi-agent systems with switching topology and time-varying delays," IEEE Trans. Automatic Control, vol. 53, no. 8, pp. 1804-1816, 2008. [27] M. Cao, A. S. Morse and B. D. O. Anderson, "Reaching a consensus in a dynamically changing environment: a graphical approach," SIAM J. Control Optim., vol. 47, no. 2, pp. 575-600, 2008. [28] M. Cao, A. S. Morse and B. D. O. Anderson, "Reaching a consensus in a dynamically changing environment: convergence rates, measurement delays, and asynchronous events," SIAM J. Control Optim., vol. 47, no. 2, pp. 601-623, 2008. [29] D. Cheng, J. Wang, and X. Hu, "An extension of LaSalle's invariance principle and its application to multi-agents consensus," IEEE Trans. Automatic Control, vol. 53, pp. 1765- 1770, 2008. [30] G. Shi and Y. Hong, "Global target aggregation and state agreement of nonlinear multi- agent systems with switching topologies," Automatica, vol. 45, pp. 1165-1175, 2009. [31] G. Shi, Y. Hong and K. H. Johansson, "Connectivity and set tracking of multi-agent systems guided by multiple moving leaders," IEEE Trans. Automatic Control, to appear. 25 [32] S. Kar and J. M. F. Moura, "Distributed consensus algorithms in sensor networks: link failures and channel noise," IEEE Trans. on Signal Processing, vol. 57, no. 1, pp. 355-369, 2009. [33] S. Kar and J. M. F. Moura, "Distributed consensus algorithms in sensor networks: quantized data and random link failures, IEEE Trans. Signal Processing, vol. 58, no. 3, pp. 1383-1400, 2010. [34] S. Patterson, B. Bamieh and A. El Abbadi, "Convergence rates of distributed average consensus with stochastic link failures," IEEE Trans. Autom. Control, vol. 55, no. 4, pp. 880-892, 2010. [35] F. Fagnani and S. Zampieri, "Randomized consensus algorithms over large scale networks," IEEE J. on Selected Areas of Communications, vol. 26, no. 4, pp. 634-649, 2008. [36] S. Muthukrishnan, B. Ghosh, and M. Schultz, "First and second order diffusive methods for rapid, coarse, distributed load balancing," Theory of Computing Systems, vol. 31, pp. 331-354, 1998. [37] R. Diekmann, A. Frommer, and B. Monien, "Efficient schemes for nearest neighbor load balancing," Parallel Computing, vol. 25, pp. 789-812, 1999. [38] C. Intanagonwiwat, R. Govindan, and D. Estrin, "Directed diffusion: a scalable and robust communication paradigm for sensor networks," in Proceedings ACM/IEEE Conf. Mobile Computing and Networking, pp. 56-67, 2000. [39] D. Kempe, A. Dobra, and J. Gehrke, "Gossip-based computation of aggregate information," in Proc. Conf. Foundations of Computer Science, pp. 482-491, 2003. [40] A. Nedi´c, A. Olshevsky, A. Ozdaglar, and J. N. Tsitsiklis, "On distributed averaging al- gorithms and quantization effects," IEEE Trans. Automatic Control, vol. 54, no. 11, pp. 2506-2517, 2009. [41] A. Olshevsky and J. N. Tsitsiklis, "Convergence speed in distributed consensus and aver- aging," SIAM J. Control Optim., vol. 48, no.1, pp. 33-55, 2009. [42] A. Olshevsky and J. N. Tsitsiklis, "Degree fluctuations and the convergence time of con- sensus algorithms," 50th IEEE Conference on Decision and Control, Orlando, Florida, December 2011. 26 [43] J. M. Hendrickx and J. N. Tsitsiklis, "Convergence of type-symmetric and cut-balanced consensus seeking systems," submitted, available online from arXiv:1102.2361v1. [44] B. Touri and A. Nedi´c, "Alternative characterization of ergodicity for doubly stochastic chains," 50th IEEE Conference on Decision and Control, Orlando, Florida, December 2011. 27
1909.05815	1	1909	2019-09-12T17:20:15	Modeling Sensorimotor Coordination as Multi-Agent Reinforcement Learning with Differentiable Communication	[ "cs.MA", "cs.AI" ]	Multi-agent reinforcement learning has shown promise on a variety of cooperative tasks as a consequence of recent developments in differentiable inter-agent communication. However, most architectures are limited to pools of homogeneous agents, limiting their applicability. Here we propose a modular framework for learning complex tasks in which a traditional monolithic agent is framed as a collection of cooperating heterogeneous agents. We apply this approach to model sensorimotor coordination in the neocortex as a multi-agent reinforcement learning problem. Our results demonstrate proof-of-concept of the proposed architecture and open new avenues for learning complex tasks and for understanding functional localization in the brain and future intelligent systems.	cs.MA	cs	Modeling Sensorimotor Coordination as Multi-Agent Reinforcement Learning with Differentiable Communication Department of Computer Science Department of Symbolic Systems William Yin Stanford University [email protected] Bowen Jing Stanford University [email protected] Abstract Multi-agent reinforcement learning has shown promise on a variety of cooperative tasks as a consequence of recent developments in differentiable inter-agent communication. However, most architectures are limited to pools of homogeneous agents, limiting their applicability. Here we propose a modular framework for learning complex tasks in which a traditional monolithic agent is framed as a collection of cooperating heterogeneous agents. We ap- ply this approach to model sensorimotor coordination in the neocortex as a multi-agent reinforcement learning problem. Our results demonstrate proof-of-concept of the pro- posed architecture and open new avenues for learning complex tasks and for understanding functional localization in the brain and future intelligent systems. 1 Introduction Motor coordination tasks have been among the most popular applications for reinforcement learning models and significant milestones have been achieved in a diverse array of tasks in environments like MuJoCo[1] and OpenAI Gym[2]. Nearly all models of motor coordination have followed the paradigm of a single agent learning to process observations from the environment and direct all actions accordingly. In this paradigm, a single agent is analogous to a single individual who learns to perform a certain motor task. In this paper, we propose modeling the sensorimotor cortex as a collection of individual agents. We draw major inspiration from the neuro-anatomical localization of visual processing, proprio-sensory perception and motor planning to the occipital, parietal, and frontal lobes, respectively. Due to the spatial separation of these functions in the brain, we explore the perspective that each cortex performs internal computations on significantly shorter timescales and signicantly higher bandwidths than messages between cortices. Given this, it is more appropriate to model the sensorimotor cortex as a group of smaller agents which communicate and coordinate their actions to achieve a broader goal. Our approach is made possible by recent developments in multi-agent reinforcement learning which permit the learning of differentiable communication protocols amongst cooperating agents, and to our knowledge is novel in the literature. 2 Multi-Agent Reinforcement Learning Multi-agent reinforcement learning (MARL) is a promising approach for modeling the dynamics of cooperative, competitive, or predator-and-prey relationships between agents, where each agent has some unique observation of the state of the world and then proceeds to make individual decisions which contribute to the world state at the next timestep. A recent example from Liu et al. -- in which emergent cooperative behaviors were studied in agents playing a game of soccer modeled in MuJoCo -- highlights the extraordinary potential for multi-agent learning within such a model[3], as does the work of Bard et al. on Hanabi, a cooperative card game[4]. The mode of communication between agents remains an active area of research. Foerster et al.[5] recently pro- posed two means of inter-agent communication: Reinforced Inter-Agent Learning (RIAL) and Differentiable Inter-Agent Learning (DIAL). RIAL describes an approach which employs deep Q-learning and allows com- munication during action selection, while DIAL describes a means to backpropogate error derivatives through noisy communication channels. DIAL is particularly interesting because it employs centralized learning but decentralized execution; in other words, learning occurs across all agents as one unit, while action selection happens on a per-agent basis where each agent treats the other agents as elements of the environment. On the other hand, Sukhbaatar et al.[6] describe the development of a neural model as a continuous communication channel between agents (called CommNet) which facilitates the learning of a communication protocol among agents. In both of these examples, the protocol itself is not specified to the agents; rather, the agents are 'tasked' with learning the protocol themselves, and in doing so revealed interesting strategies to approaching inter-agent communication and information transfer. Despite these advancements, however, state-of-the-art MARL models still remain relatively structurally simple. In particular, agents in MARL models remain relatively homogeneous, in that agent parameters are drawn from the same distributions. The agents themselves also have the same objectives, regardless of whether the task is cooperative or competitive. In this sense, these MARL environments frequently fail to extend well to scenarios in which different agents have different roles or objectives. In this sense, we see extraordinary potential in extending MARL models to scenarios in which agents may be defined with unique roles and properties. In particular, we hypothesized that such a model would be especially appropriate in modeling motor control tasks. There already exists a wealth of research in employing policy gradient techniques in modeling motor skills and control[7], but none which aim to model each component of a motor agent as a unique agent with unique properties. The sensorimotor cortex seems to be a perfect candidate for such a model due to the separation of distinct components which interact[8]. Hence, our paper proposes the use of continuous communication protocols for message-passing across distinct, role-specific sensorimotor agents as models for components of a motor control task. 3 Architecture i i i 3.1 Model In our architecture we model the sensorimotor cortex as a set S of ns sensory agents and a set M of nm motor agents embedded in a fully deterministic Markov decision process with a set of discrete or continuous observation spaces and a set of discrete action spaces1. Each sensory agent Si ∈ S makes observations o[t] in some input domain and outputs a message mi,j to each motor agent2 (Mj, Qj) ∈ M. The sensory agents do not have access to the decision process and do not themselves take any actions. Rather, their sole effect is achieved by communicating a summary of the sensory information to the motor agents. Each motor agent consists of an M-net Mi and a Q-net Qi. The M-net takes in is all incoming messages (from all sensory agents as well as all other M-nets), in addition to the most recent action a[t−1] of the corresponding motor agent, and produces a message for each of the other motor units (Mj, Qj) ∈ M, i (cid:54)= j. The Q-net is a standard action-value function which takes as input all incoming messages from sensory agents and all motor units, as well the previous action a[t−1] and predicts a value for each possible action ai given the input messages. Messages are not sent from the M-net of each motor unit to its corresponding Q-net because doing so would give encourage the M-net to serve as intermediate information-processing agents unaffiliated with a particular Q-net. In our model of message passing, we set that each message is received by the receiving agent one time step after it is sent. This serves dual purposes: first, it captures the longer period of time it would take information to propogate from one part of the cortex to another; second, it enables the training of a message protocol via gradient descent despite a cyclic communication network. More formally, each of our model's sensory agents Si ∈ S computes: o[t] i i,1 . . . m[t] (cid:16) i,nm = Si m[t] (cid:17) (cid:16) (cid:17) 1The model can easily be adapted to continuous action spaces. 2Each motor agent is an ordered pair, as explained in the following paragraph. 2 where mi,j indicates a message vector from sensory agent Si to motor agent Mj. Each of the model's motor units (Mi, Qi) ∈ M computes (cid:16) (cid:17) (cid:16) (cid:16) . . . m (cid:48)[t] i,i−1, m (cid:48)[t] i,i+1 . . . = Mi Q[t] i (a) = Qi where 1,i m[t−1] m[t−1] 1,i . . . m[t−1] . . . m[t−1] ns,i , . . . m ns,i , . . . m (cid:48)[t−1] i,i−1 , m (cid:48)[t−1] i,i−1 , m (cid:48)[t−1] i,i+1 . . . , a[t−1] (cid:48)[t−1] i,i+1 . . . , a[t−1] i i , a (cid:17) (cid:17) i,j indicates a message vector from motor agent (Mi, Qi) to motor agent (Mj, Qj) i,i−1, m(cid:48) • m(cid:48) • . . . m(cid:48) • a is an element of Ai, the action space of agent (Mi, Qi) • Q[t] i i,i+1 . . . denotes the set of all m(cid:48) is the action-value function for motor agent (Mi, Qi) at time t i,j where i (cid:54)= j for a total of nm(nm + ns − 1) message streams. The controller then chooses the actions for each motor agent according to i ∼ soft max a[t] a∈Ai Q[t] i (a) if the controller is exploratory / learning mode and a[t] i = arg max a∈Ai if it is in greedy mode. At each timestep t, the environment provides a reward R[t] in response to the actions a[t] i τ =t γτ−tR[τ ]] by performing stochastic taken. We train the action-value functions to predict Q[t] (cid:19)2 gradient descent on the temporal-difference loss i ) − R[t] i (a[t] i (a[t] Q[t+1] (cid:18) Q[t] (a) i i − γ arg max a∈Ai Q[t] i (a) i ) = E[(cid:80)∞ . 3.2 Implementation For our training environment, we instantiate the architecture to feature three sensory agents and two motor agents. The sensory agents Sl, Se, and Sr represent the left arm propriosensory cortex, the visual cortex, and the right arm propriosensory cortex, respectively, and the motor motor agents (Ml, Ql) and (Mr, Qr) controls the left and right arm, respectively. The observation domains Sl, Sr ∈ R2 represent the (x, y) coordinates of the respective arm (we restrict the environment to two dimensions for simplicity), while Se ∈ R4 encodes the (x, y) location and x, y velocity of an object of interest -- in our case, a ball3. The action spaces Al and Ar consist of five actions: moving the respective arms a fixed amount in one of four directions, or making no movements4. We implement each functional component of the system (the sensory agents, the M-nets, and the Q-nets) as a neural network with one ten-neuron hidden layer and ReLU activations operating on a concatenation of its inputs and outputting a concatenation of its outputs. While the architecture permits each message stream to be of different length, we implement all message channels to be of fixed dimensionality dm = 5. This fixed dimensionality, the number of layers and hidden units in each neural network, the learning rate, and the discount factor 0 < γ < 1 are the hyperparameters of our model. The model is trained by regarding the o[t−1]s, m[t−2]s, and m(cid:48)[t−2]s as constant inputs to the model whose output is the set of Q[t] functions at time t. Gradient descent is then simultaneously performed on all the network weights. 3In a more sophisticated system, the sensory agent may instead be a convolutional network operating on an input image stream. 4In an ideal, more realistic model of muscle movement, the outputs would affect the acceleration of the arm. 3 4 Environment We train the model on a simplified juggling-like task, in which the model must bounce a ball using paddles in each hand and prevent it from touching the ground (Figure 1). Collisions are perfectly elastic, such that the total number of bounces is principally unbounded and the bounce frequency and heights is relatively constant5. Each bounce imparts a Gaussian-distributed random horizontal impulse to the ball; if the ball hits the wall, it is deflected. The simulation resets when the ball hits the ground. Rewards are as follows: • A penalty of 50 if the ball hits the ground and the simulation resets. • A contact award of 50 each time the agent successfully bounces the ball. We found this shortening of the time horizon to be necessary for a resonable rate of learning. • To incentivize coordination between the motor agents, a global penalty of 0.5 is applied to each movement of any arm6. Figure 1: The environment consists of a ball which the agent must keep in the air by moving paddles in each hand. See details in text. The values of these rewards, the bounce height, and simulation frequency may be adjusted as hyperparameters. In our implementation of the environment, the physical parameters of the model are analogous to a human agent with arms at height 1m, bouncing a ball dropped from a height of 3m, with paddles 30cm across, in a bounding box of width 2m, arm movements of 15cm, and a message transit time of 0.1 seconds7. 5 Experiments We train each of n = 10 instances of the model for 5000 sessions, where each session is defined as ending when the ball hits the ground. We alternate 100-session exploratory learning epochs with 100-session greedy learning epochs8. The greedy epochs revealed that the model learned to minimize movements by restricting action to one of the two agents while the other agent became increasingly dormant (Figure 2). This was an interesting and unexpected result, as it is highly analogous to the notion of hand dominance. However, no such trend was observed in the exploratory epochs, suggesting the preference was very slight (Figure 3). Taken at face value, the hand dominance in our model developed because the model found it more effective to universally suppress the action of one agent rather than trying to develop nuanced communications among agents. 5This has the same effect as the more realistic assumption that the agent imparts an compensatory upward impulse after a non-zero contact time, but that more sophisticated learning requirement may be better suited for a larger model. 6An alternative perspective is to apply a penalty whenever both agents move. 7Of course many alternative formulations exist, but we choose this distance scale for intuitive simplicity. 8The exploratory and greedy controllers are as described above. Alternatively, the "temperature" of the softmax is may be tuned. 4 Figure 2: Left arm movements as a fraction of total movements during the greedy training epochs. Notice that by the later epochs, a majority of models have a "dominant" or preferred agent to take action with. Figure 3: Left arm movements as a fraction of total movements during the exploratory training epochs. No significant deviation from an even distribution is observed. Figures 4 and 5 show the performance of the multi-agent model over the course of training. Despite performing the best among reasonable configurations explored9, the model appears to exhibit only relatively weak learning. As shown in Figure 4, the mean number of bounces per 100-session greedy learning epoch rose from 26.5 to 41.8 in the first 10 epochs (t(18) = −2.816, p = 0.011) but no significant further improvement was observed in the remaining 15 epochs (t(18) = −0.327, p = 0.747). Note that this corresponds to an average of only 0.4 bounces before the ball hits the ground, even at the end of training. In the exploratory learning epochs, performance remained at around 37 bounces per epoch throughout, with no signs of learning observed (Figure 5). This is reasonable, as the model may not be confident enough yet in the learned policy to achieve consistent performance by drawing actions from a softmax rather than an argmax. 9We were able to get much more impressive results by making the paddle comically large and the learning task almost trivial. 5 Figure 4: Number of bounces per greedy learning epoch. Figure 5: Number of bounces per exploratory learning epoch. 6 Discussion and Future Work We have demonstrated the implementation of a multi-agent model of the sensorimotor cortex trained on a simple ball-bouncing task. Although further hyperparameter tuning may improve performance to reasonable human-level expectations, we have demonstrated proof of principle that the proposed architecture is capable of learning a sensorimotor coordination task. In particular, we have demonstrated that delayed message passing from sensory agents to motor agents is sufficient to allow the learning of sensory-guided motor actions. Further work is merited to confirm the model of message-passing between motor agents as a means of coordinating actions. Such validation could be performed by ablation studies in which message-passing between such agents is removed. The hand-dominance observed in the present experiments offers no evidence of multi-agent coordination, but is instead the chief unexpected result of this paper. Because all awards and penalties are shared by both agents, there is no feature of the architecture which would encourage the unilateral, consistent suppression of a single agent. It is possible that the stochastic preferential usage of a particular agent 6 early in training reinforces differences in the learned Q-values in such a way that further biases the model towards the usage of that agent. The trajectories in Figure 2 support this possibility, as they generally converge toward the dominance which was present to a slight degree in early epochs. If so, then hand dominance would appear to be a stable configuration which naturally arises from an motion-conservative reward system and trial-and-error learning. It is worth investigating or speculating if this has any connection with the formation of hand-dominance in humans. References [1] E. Todorov, T. Erez, and Y. Tassa, "MuJoCo: A physics engine for model-based control," IEEE/RSJ International Conference on Intelligent Robots and Systems, 2012. [2] G. Brockman, V. Cheung, L. Pettersson, J. Schneider, J. Schulman, J. Tang, and W. Zaremba, "OpenAI Gym," arXiv e-prints, Jun. 2016. [3] S. Liu, G. Lever, J. Merel, S. Tunyasuvunakool, N. Heess, and T. Graepel, "Emergent Coordination Through Competition," arXiv e-prints, Feb. 2019. [4] N. Bard, J. N. Foerster, S. Chandar, N. Burch, M. Lanctot, H. F. Song, E. Parisotto, V. Dumoulin, S. Moitra, E. Hughes, I. Dunning, S. Mourad, H. Larochelle, M. G. Bellemare, and M. Bowling, "The Hanabi Challenge: A New Frontier for AI Research," arXiv e-prints, Feb. 2019. [5] J. N. Foerster, Y. M. Assael, N. de Freitas, and S. Whiteson, "Learning to Communicate with Deep Multi-Agent Reinforcement Learning," arXiv e-prints, May 2016. [6] S. Sukhbaatar, A. Szlam, and R. Fergus, "Learning Multiagent Communication with Backpropagation," arXiv e-prints, May 2016. [7] J. Peters and S. Schaal, "Reinforcement learning of motor skills with policy gradients," Neural Networks, 2008. [8] T. Grent-'t-Jong, R. Oostenveld, W. P. Medendorp, and P. Praamstra, "Separating Visual and Motor Components of Motor Cortex Activation for Multiple Reach Targets: A Visuomotor Adaptation Study," Journal of Neuroscience, 2015. 7
1011.3397	1	1011	2010-11-15T14:20:14	The Inverse Task of the Reflexive Game Theory: Theoretical Matters, Practical Applications and Relationship with Other Issues	[ "cs.MA", "cs.AI", "cs.RO" ]	The Reflexive Game Theory (RGT) has been recently proposed by Vladimir Lefebvre to model behavior of individuals in groups. The goal of this study is to introduce the Inverse task. We consider methods of solution together with practical applications. We present a brief overview of the RGT for easy understanding of the problem. We also develop the schematic representation of the RGT inference algorithms to create the basis for soft- and hardware solutions of the RGT tasks. We propose a unified hierarchy of schemas to represent humans and robots. This hierarchy is considered as a unified framework to solve the entire spectrum of the RGT tasks. We conclude by illustrating how this framework can be applied for modeling of mixed groups of humans and robots. All together this provides the exhaustive solution of the Inverse task and clearly illustrates its role and relationships with other issues considered in the RGT.	cs.MA	cs	The Inverse Task of the Reflexive Game Theory: Theoretical Matters, Practical Applications and Relationship with Other Issues Sergey Tarasenko Kyoto University, Yoshida honmachi, Kyoto 606-8501, Japan [email protected] Abstract. The Reflexive Game Theory (RGT) has been recently pro- posed by Vladimir Lefebvre to model behavior of individuals in groups. The goal of this study is to introduce the Inverse task. We consider meth- ods of solution together with practical applications. We present a brief overview of the RGT for easy understanding of the problem. We also de- velop the schematic representation of the RGT inference algorithms to create the basis for soft- and hardware solutions of the RGT tasks. We propose a unified hierarchy of schemas to represent humans and robots. This hierarchy is considered as a unified framework to solve the entire spectrum of the RGT tasks. We conclude by illustrating how this frame- work can be applied for modeling of mixed groups of humans and robots. All together this provides the exhaustive solution of the Inverse task and clearly illustrates its role and relationships with other issues considered in the RGT. Key words: Reflexive Game Theory (RGT), group behavior, society behavior, RGT Forward Task, RGT Inverse Task, Asimov's Laws of Robotics, robots in RGT, mixed groups of humans and robots, human- robot societies 1 Introduction The Reflexive Game Theory (RGT) has been entirely developed by Lefebvre [1, 2] and is based on the principles of anti-selfishness or egoism forbiddeness [1, 2] and human reflexion processes [3]. Therefore RGT is based on the human-like decision-making processes. The main goal of the theory is to model behavior of individuals in the groups. It is possible to predict choices, which are likely to be made by each individual in the group, and influence each individual's decision- making due to make this individual to make a certain choice. In particular, the RGT can be used to predict terrorists' behavior [4]. In general, the RGT is a simple tool to predict behavoir of invididuals and influence individuals' choices. Therefore it makes possible to control the individ- uals in the groups by guiding their behavoir (decision-making, choices) by means of the corresponding influences. 2 Sergey Tarasenko On the other hand, now days robots have become an essential part of our life. One of the purposes robots serve to is to substitute human beings in dangerous situations and environments, like defuse a bomb or radioactive zones etc. In contrast, human nature shows strong inclinations towards the risky be- havior, which can cause not only injuries, but even threaten the human life. The list of these reasons includes a wide range starting from irresponsible kids' behavior to necessity to find solution in a critical situation. In such a situation, a robot should full-fill a function of refraining humans from doing risky actions and perform the risky action itself, if needed. However, robots are forbidden and should not physically force people, but must convince people on the mental level to refrain from doing a risky action. This method is more effective rather than a simple physical compulsion, because humans make the decisions (choices) themselves and treat these decisions as their own. Such technique is called a reflexive control [3]. The task of finding appropriate reflexive control is closely related with the Inverse task, when we need to find suitable influence of one subject on another one or on a group of subject on the subject of interest. Therefore, it is needed to develop the framework of how to solve the Inverse task. This is the primary goal of this study. However, for better understanding of the gist of the Inverse task and its intrinsic relationships with other issues of the RGT, we introduce the entire spectrum of the tasks, which can be solved by the RGT. This forms the scope of inference algorithms used in the RGT. We present the RGT algorithms in the form of the schemas of control systems that can be instantly applied for developement of soft- or/and hardware solutions. We develop a hierarchy of control systems for abstract individual (including human subject) and robotic agent (robot) based on these control schemas. Finally, we illustrate application of the Inverse task together with other RGT inference algorithms to model robot's behavior in the mixed groups of humans and robots. 2 Brief Overview of the Reflexive Game Theory (RGT) 2.1 Representation of groups: graphs, polynomials and stratification tree The RGT deals with groups of abstract subjects (individuals, humans, au- tonomous agents etc). Each subject is assigned a unique variable (subject vari- able). Any group of subjects is represented in the shape of fully connected graph, which is called a relationship graph. Each vertex of the graph corresponds to a single subject. Therefore the number of vertices of the graph is in one-to-one correspondence with overall number of subjects in the groups. Each vertex is named after the corresponding subject variable. The RGT uses the set theory and the Boolean algebra as the basis for calcu- lus. Therefore the values of subject variables are elements of Boolean algebra. The Inverse Task 3 All the subjects in the group can have either alliance or conflict relationship. The relationships are identified as a result of group macroanalysis. It is suggested that the installed relationships can be changed. The relationships are illustrated with graph ribs. The solid-line ribs correspond to alliance, while dashed ones are considered as conflict. For mathematical analysis alliance is considered to be conjunction (multiplication) operation (·), and conflict is defined as disjunction (summation) operation (+). The graph presented in Fig. 1a or any graph containing any sub-graph isomor- phic to this graph are not decomposable. In this case, the subjects are excluded from the group one by one, until the graph becomes decomposable. The exclusion is done according to the importance of the other subjects for a particular one [1, 2]. Any other fully connected graphs are decomposable. Any decomposable graph can be presented in an analytical form of a corresponding polynomial. Any relationship graph of three subjects is decomposable (see [1, 2]). Consider three subjects a, b and c. Let subject a is in alliance with other subjects, while subjects b and c are in conflict (Fig. 1b). The polynomial corre- sponding to this graph is a(b + c). Fig. 1. The relationship graphs. Fig. 2. Polynomial Stratification Tree. Polynomials [a], [b] and [c] are elementary poly- nomials. Regarding a certain relationship, the polynomial can be stratified (decom- posed) into sub-polynomials [1, 2]. Each sub-polynomial belongs to a particular level of stratification. If the stratification regarding alliance was first built, then the stratification regarding the conflict is implemented on the next step. The stratification procedure finalizes, when the elementary polynomials, containing a single variable, are obtained after a certain stratification step. The result of stratification is the Polynomial Stratification Tree (PST). It has been proved that each non-elementary polynomial can be stratified in an unique way, i.e., each non-elementary polynomial has only one corresponding acbacbabcacbd[a(b+c)][a][b+c]⋅[b]+[c]4 Sergey Tarasenko PST (see [7] considering one-to-one correspondence between graphs and polyno- mials). Each higher level of the tree contains polynomials simpler than the ones on the lower level. For the purpose of stratification the polynomials are written in square brackets. The PST for a(b + c) polynomial is presented in Fig.2. Next, we omit the branches of the PST and from each non-elementary polyno- mial write in top right corner its sub-polynomials. The resulting tree-like struc- ture is called a diagonal form[1, 2, 5, 6]. Consider the diagonal form correspond- ing to the PST in Fig. 2: [b] + [c] [a][b + c] [a(b + c)] . Hereafter, the diagonal form is considered as a function defined on the set of all subsets of the universal set. The universal set contains the elementary actions. For example, these actions are actions α and β. By definition, the Boolean algebra of the universal set includes four elements: 1 = {α, β}, {α}, {β} and the empty set 0 = {} = Ø. These elements are all the possible subsets of universal set and considered as alternatives that each subject can choose. The alternative 0 = {} is interpreted as an inactive or idle state. In general, Boolean algebra consists of 2n alternatives, if universal set contains n actions. Accroding to definition given by Lefebvre [5], we present here exponential operation defined by formula P W = P + W , (1) where W stands for negation of W [1, 2, 4]. This exponential operation is used to fold the diagonal form. During the folding, round and square brackets are considered to be interchangeable. The following equalities are also considered to be true: x + x = 1, x + 0 = x and x + 1 = 1. Next we implement folding of diagonal form of polynomial a(b + c): [b] + [c] [a][b + c] [a]([b + c] + [b] + [c]) [a(b + c)] = [a(b + c)] = a(b + c) + a . It is considered that the levels of the PST represent different processing levels of natural or artificial cognitive system. Each level is considered as an images. The root of the tree is the input into the cognitive system and, therefore can be considered as the image of the world (environment including self and others), perceived by the subject. As it follows from the PST, there is a hierarchy of images, corresponding to a particular cognitive level. During processing along this hierarchy in the bottom-up manner, the image on the lower level undergoes an extensive process of simplification by the means of decomposition into simpler parts on the higher level. These parts are considered to be the images of the image on the previous level. Therefore, the images on the second level are different representions of the The Inverse Task 5 original image of the world. This procedure repeats until we obtain elementary part (elementary polynomials) [1, 2]. On the other hand, the PST folding procedure can be referred as top-down intergration process of simpler images from the higher levels. Therefore, the stratification procedure of original polynomial together with the folding procedure of the diagonal form illustrate the interplay of bottom-up and top-down information processes, which are widely imployed in biological [8, 9, 10, 11] and artificial [12, 13, 14] information processing systems. The idea of hierarchical structure is highly coherent with hierarchical organization of ma- jority of natural (inanimate objects) and biological (living creatures) entities. Furthermore, it has been shown that hierarchical structure is intrinsic for the relationships in societies of insects [15], animals [17, 16, 18] and human beings. Therefore hierarchical representation of the groups in the form of PST corre- spond to extraction of the hierarchical structure of the given group, while fusion of the PST and its diagonal form with diagonal form folding procedure closely resembles the way of information processing within a single independent congni- tive system as discussed above. Thus, RGT imploys the fundamental principles of hierarchical organization on both group (reflects structure of the groups) and individual (illustrates information processing within independent cognitive sys- tem of a single unit) levels. This makes RGT universal tools that mildly bridges the gap between representation and analysis. 2.2 The Decision Equation: definition and solution The goal of each subject in a group is to choose an alternative from the set of alternatives under consideration. To obtain choice of each subject, we consider the decision equations, which contain subject variable in the left-hand side and the result of diagonal form folding in the right-hand side: a = (b + c)a + a b = (b + c)a + a c = (b + c)a + a To find solution of the decision equations, we consider the following equation: x = Ax + Bx , (2) where x is the subject variable, and A and B are some sets. Eq.(2) represents the canonical form of decision equation. This equation has solution if and only if the set B is contained in set A: A ⊇ B. If this requirement is satisfied, then eq.(2) has at least one solution from the interval A ⊇ x ⊇ B [4]. Otherwise, the decision equation has no solution, and it is considered that subject cannot make a decision. In such situation, the subject is in frustration state. Therefore, to find solutions of decision equation, one should first transform it into the canonical form. Out of three presented equations only the decision 6 Sergey Tarasenko equation for subject a is in the canonical form, while other two should be trans- formed. We consider explicit transformation only of decision equation for subject b [20]: a(b + c) + a = ab + ac + a = ab + (ac + a)b + (ac + a)b = (a + a + ac)b + (ac + a)b = (1 + ac)b + (ac + a)b = b + (ac + a)b = b + (ac + ac + a)b = b + (c + a)b. Therefore, b = b + (c + a)b. (3) The transformation of equation for subject c be can be easily derived by analogy: c = c + (b + a)c. Next we consider two tasks, which can be formulated regarding the decision equation in the canonical form and provide methods to solve each task. 2.3 The Forward Task The variable in the left-hand side of the decision equation in canonical form is the variable of the equation, while other variables are considered as influences on the subject from the other subjects. The Forward task is formulated as a task to find the possible choices of a subject of interest, when the influences on him from other subjects are given. After transformation of arbitral decision equation into its canonical form, the sets A and B are functions of other subjects' influences. For example, if we consider group of subjects a, b, c, etc. togehter with the abstract representation of decision equation in canonical form for subject a, the sets A and B will be the functions of subject variables b, c, etc. : a = A(b, c, ...)a + B(b, c, ...)a . (4) In the case of only three subjects a, b and c, A(b, c, ...) = A(b, c) and B(b, c, ...) = B(b, c). All the influences are presented in influence matrix (Table 1). The main diagonal of influence matrix contains the subject variables. The rows of the matrix represent influences of the given subject on other subjects, while columns represent the influences of other subjects on the given one. The influence values are used in decision equations. Table 1. Influence Matrix a b c a a {α} {β} b {β} b {β} c {β} {β} c The Inverse Task 7 For subject a: a = ({β} + {β})a + a ⇒ a = {β}a + a. For subject b: b = b + ({α}{β} + {α})b ⇒ b = b + {β}b. For subject c: c = c + ({β}{β} + {β})c ⇒ c = c + ({β} + {α})c ⇒ c = 1. Equation for subject a does not have any solutions, since set A = A(b, c) = {β} is contained in set B = B(b, c) = 1: A ⊂ B. Thus, subject a cannot make any decision. Therefore he is considered to be in frustration state. Equation for subject b has at least one solution, since A = A(b, c) = 1 = {α, β} ⊇ B = B(b, c) = {β}. The solution belongs to the interval 1 ⊇ b ⊇ {β}. Therefore subject b can choose any alternative from Boolean algebra, which contains alternative {β}. These alternatives are 1 = {α, β} and {β}. case, when A(b, c) ≡ B(b, c). Here A = B = 1. Equation for subject c turns into equality c = 1. This is possible only in the 2.4 The Inverse Task In contrast to the Forward task, the Inverse task is formulated as a task to find all the simultaneous (or joint) influences of all the subjects together on the subject of interest that result in choice of a particular alternative or subset of alternatives. We call the subject of interest to be a controlled subject. Let subject a be a controlled subject and a∗ is a fixed value, representing an alternative or subset of alternatives, which subjects b, c, etc. want subject a to choose. We call value a∗ to be a target choice. By substituting subject variable a with fixed value a∗, we obtain the influence equation. If we substitute the subject variable a with fixed value a∗ in the canonical form of the decision equation (eq. (4)), we obtain the canonical form of the influence equation: a∗ = A(b, c, ...)a∗ + B(b, c, ...)a∗ , (5) For only three subjects a, b and c, A(b, c, ...) = A(b, c) and B(b, c, ...) = B(b, c). In contrast to the decision equation, which is equation of a single variable, the influence equation is the equation of multiple variables. However, the number of variables of influence equation is not trivial question. In fact, the number of variables in influence equation can be less then (n − 1), where n is the total number of subjects in the group. There are groups, in which sets A and B are functions of less than (n− 1) variables (see Appendix A). Therefore the variables that present in influence equation are called effective variables. The Inverse task is by definition1 formalized as to find all the joint solutions of all subjects in the group, except for the controlled one, when the target choice is represented by interval χ1 ⊇ a∗ ⊇ χ2, where χ1 and χ2 are some sets and χ1 ⊃ χ2. In such a case, to solve the Inverse task, one should solve the system of influence equations: 1 We need a system of influence equations because solutions of the influence equation a∗ = A(b, c, ...)a∗ + B(b, c, ...)a∗ itself only guaratee that the original decision equa- tion a = A(b, c, ...)a + B(b, c, ...)a turns into true equality, but it is not guaranteed that these solutions are the only ones that turn decision equation into true equality. 8 Sergey Tarasenko (cid:26)A(b, c, ...) = χ1 B(b, c, ...) = χ2 If the target choice is a single alternative, then χ1 = χ2 = a∗. The solutions of the system (6-7) are considered as reflexive control strategies. The solution of the Inverse task in particular is characterized from two points. The first point is whether it is required to find the influence of a particular single subject or joint influences of a group of subjects. The second one is whether the target choice is represented as a single alternative or as an interval of alternatives. To illustrate these points, we introduce a particular group of subjects. Let subjects a and b are in alliance with each other and in conflict with subject c. The polynomial corresponding to this graph is ab + c. The diagonal form corresponding to this polynomial and its folding is [a][b] [ab] +[c] [ab + c] = ab + c Therefore the decision equation for all the subjects in the group is x = ab + c, (8) where x can be any subject variable a, b or c. Influence of a single subject vs joint influences of a group. First we consider example, when the influence of a single subject is required. Let subject b makes influence {α} and a∗ = {α}. Then we need to find influences of a single subject c, which result in solution a∗ = {α} of decision equation a = ab + c. The canonical form of this influence equation is a∗ = (b + c)a∗ + ca∗. Since a∗ = {α}, χ1 = χ2 = {α}, we obtain a system of equations: (cid:26){α} + c = {α} c = {α} (cid:26)b + c = {α} c = {α} (6) (7) (9) (10) (11) (12) Therefore, the straight forward solution of this system is c = {α}. This simple example illustrates the very gist of the Inverse task - to find the appropriate influences, which result in target choice. Next, we consider that influence of subject b is not known. Therefore, we obtain system In this case, we need to find the values of variable b, which together with c, result in solution a∗ = {α}. In other words, we need to find all the pairs (b, c), resulting in solution a∗ = {α}. These pairs are solutions of the system (11-12). Therefore, we run all the possible values of variable b and check if the first equation of the system (11-12) turns into true equality: b = 1 : 1 + {α} = 1 ⇒ 1 (cid:54)= {α}; The Inverse Task 9 b = {α} : {α} + {α} = {α} ⇒ {α} = {α}; b = {β} : {β} + {α} = 1 ⇒ 1 (cid:54)= {α}; b = 0 : 0 + {α} = {α} ⇒ {α} = {α}. 0 are appropriate. Thus, we obtain two pairs (b, c): ({α},{α}) and ({α}, 0). Therefore, out of four possible values of variable b, only two values {α} and A single target alternative vs interval of alternatives. In the previous examples we considered a target choice to be only a single alternative. Here we illustrate the case, when a target choice is an interval. Let b = {β}, and 1 ⊇ a∗ ⊇ {α}. To find corresponding influences of subject c, we solve the system of equations: (cid:26){β} + c = 1 c = {α} (13) (14) Again, we instantly obtain the solution of this system: c = {α}. In this section, we have formulated the Inverse task in general and considered its particular formalization depending on the number of influences and what is the target choice. However, we do not have a method to solve arbitral influence equation. Therefore, we solve this problem in the next section. 3 How to Solve an Arbitral Influence Equation As an introduction for this section, we consider the fundamental proposition, which will be the conner stone to solve the influence equations. Proposition 1. Let P and Q be some abstract sets. Then P Q + P Q = 0 ⇔ P = Q. Proof. Necessity. Let P Q + P Q = 0, then P Q + P Q = 0 ⇒ P Q + P Q + P = P ⇒ P + P Q = P ⇒ P (Q + Q) + P Q = Q + P Q + P Q = P ⇒ Q = P. Therefore if P Q + P Q = 0, then P = Q. Sufficiency. Let P = Q, then P P + P P = 0. (cid:3) Now let us consider the new type of equation: This equation has solution if and only if A1 ⊇ x ⊇ B1. A1x + B1x = 0 (15) 10 Sergey Tarasenko 3.1 Solving Influence Equations There are three operations defined on the Boolean algebra. They are conjunc- tion (· or multiplication), disjunction (+ or summation) and negation (x, where x is subject variable). The negation operation is unary operation, while other two operations are binary. Using combination of these three operations, we can compose any influence equation. Since, it is obvious how to solve the equation including only unary operation, we discuss how to solve influence equations in- cluding a single binary operation. For this perpose, we consider two abstract subject variables x1 and x2 and abstract alternative χ. Lemma 1. The solution of equation (16) regarding variable xi, where i = 1, 2, is given by the interval χ ⊇ xi ⊇ (χxj + xjχ), where j = 1, 2; j (cid:54)= i. x1 + x2 = χ Proof. According to Proposition 1, P = x1 + x2, Q = χ, P = x1 + x2 = x1 x2 and Q = χ. Therefore, P Q + P Q = (x1 + x2)χ + x1 x2χ = x1χ + x2χ + x1 x2χ. Conse- quently, we obtain eq.(17): x1χ + x2χ + x2χx1 = 0 (17) We solve eq.(17) regarding variable x1. First, we transform eq.(17) into canon- ical form: χx1 + (χx2 + χx2)x1 = 0 Therefore, the solution of eq.(18) is given by the interval χ ⊇ x1 ⊇ (χx2 + x2χ). (18) (19) Since variables x1 and x2 are interchangable and it is possible to solve eq.(17) regarding variable x2 as well, the general form of solution of eq.(16) is the interval χ ⊇ xi ⊇ (χxj + xjχ). (20) where i = 1, 2 and j = 1, 2; j (cid:54)= i.(cid:3) Lemma 2. The solution of equation (21) regarding variable xi, where i = 1, 2, is given by the interval (χxj + χ xj) ⊇ xi ⊇ χ, where j = 1, 2; j (cid:54)= i. x1x2 = χ The Inverse Task 11 Proof. According to Proposition 1, P = x1x2, Q = χ, P = x1x2 = x1 + x2 and Q = χ. Therefore, P Q + P Q = (x1x2)χ + (x1 + x2)χ = x2χx1 + x1χ + x2χ. Thus, we obtain eq.(22): x2χx1 + x1χ + x2χ = 0 (22) We solve eq.(22) regarding variable x1. First, we transform eq.(22) into canon- ical form: (χx2 + χx2)x1 + χx1 = 0 (23) Since χx2 + χx2 = χx2 + χ x2, the solution of eq.(23) is given by the interval (χx2 + χ x2) ⊇ x1 ⊇ χ. (24) Since variables x1 and x2 are interchangable and it is possible to solve eq.(22) regarding variable x2 as well, the general form of solution of eq.(21) is the interval (χxj + χ xj) ⊇ xi ⊇ χ. (25) where i = 1, 2 and j = 1, 2; j (cid:54)= i.(cid:3) Since one bound of the solution intervals for eqs.(16) and (21) are functions of the second variable, we need to run all the possible values of the second variable in order to obtain all possible solutions of these equations in the form of pairs (x1, x2). Next we consider several examples, illustrating application of Lemmas 1 and 2. Example 1. For illustration, we solve equation a∗ = ba∗ + c. Consider χ = a∗, x1 = ba∗ and x2 = c, we obtain the solution interval for variable x2 = c: χ ⊇ c ⊇ (χχb + χ χb). After simplfication, we get interval (26): χ ⊇ c ⊇ χb (26) Next we consider examples with particular alternatives. Let it be alternative {α} : χ = {α}. The solution interval is then {α} ⊇ c ⊇ {α}b. Since the lower bound of this interval is a function of variable b, to find all solutions of equation a∗ = ba∗ + c, we calculate value of expression {α}b for all possible values of variable b (Table 2). To reesure that solutions are correct, we check that decision equation a = ba + c turns into true equality for the obained pairs (b, c): ({α},{α}): {α}{α} + {α} = {α} ⇒ {α} = {α} is true; ({α}, 0): {α}{α} + 0 = {α} ⇒ {α} = {α} is true; ({β},{α}): {α}{β} + {α} = {α} ⇒ {α} = {α} is true; (1,{α}): {α}1 + {α} = {α} ⇒ {α} = {α} is true; (1, 0): {α}1 + 0 = {α} ⇒ {α} = {α} is true; (0,{α}): {α}0 + {α} = {α} ⇒ {α} = {α} is true. 12 Sergey Tarasenko So far, we have illustrated how to solve the influence equation. We as well showed that the pairs (b, c) obtained by solving equation a∗ = ba∗ + c in ac- cordance with Proposition 1 and Lemmas 1 and 2 are indeed solutions of this equation. Table 2. Solutions of the influence equation a∗ = ba∗ + c Values of b {α} {β} 1 0 Pairs (b, c) ({α},{α}) ({β},{α}) (1,{α}) (0,{α}) ({α}, 0) (1, 0) Example 2. We consider influence equation for subject b obtained from eq.(3). (c + a)χ + χ = χ (27) First, we transform the left-hand side of eq.(27): (c + a)χ + χ = cχ + aχ + χ = cχ + aχ + (c + a + 1)χ = c + a + χ. Therefore, eq.(27) can be rewritten as follows: c + a + χ = χ (28) Considering, x1 = c and x2 = a + χ, we instantly obtain the solution interval of eq.(28): χ ⊇ c ⊇ (χ(a + χ) + χ(a + χ)) ⇒ χ ⊇ c ⊇ (χ a + χχa). Finally, χ ⊇ c ⊇ χ a Example 3. Next, we consider influence equation ab + χ = χ (29) (30) Considering, x1 = ab and x2 = χ, we instantly obtain the solution interval χ ⊇ ab ⊇ (χχ + χχ) or χ ⊇ ab ⊇ 0 (31) Therefore, in order to find all solutions of eq.(30), we need to solve the equa- tions ab = y (32) where y is any sub-set of set χ (y ⊇ χ). Each equation can be solved according to Lemma 2. Example 4. As a final example, we again consider influence equation a∗ = (b + c)a∗ + ca∗ and show how application of Lemma 1 essentially simplifies its solution. We get the system of influence equations: (cid:26)b + c = {α} ; c = {α} . From this system we obtain a single equation: b + {α} = {α} . The Inverse Task 13 (33) (34) (35) According to Lemma 1, we instantly obtain the solution interval of eq.(35): (36) Thus, eq.(35) has two solutions: b = {α} and b = 0. Therefore the solution of system (33-34) consists of two pairs ({α},{α}) and (0,{α}). {α} ⊇ b ⊇ 0 . To conclude this section, we provide its brief summary. We have shown how to solve the Inverse task by means of influence equations. We have proved two fundamental lemmas, which allow to solve any influence equation regardless of the number of variables. Finally, we have illustrated several examples of how apply these lemmas. 3.2 Analysis of Extreme Cases 1: Frustration In this section we analyze the situation, when subject can appear in frustration state, from the point of view of the inverse task. Let us consider the polynomial a(b + c) discussed in the section 2.1. The decision equation that corresponds to this polynomial is x = (b + c)a + a, where x can be any subject variable. Next we try to find all the pairs (b, c) such that result in selection of a particular alternative by subject a. The decision equation for subject a is a = (b + c)a + a. The solution interval of this decision equation is b + c ⊇ a ⊇ 1. We need to check which alternative subject a can be convinced to choose. To do this, we consider the system of equation for each alternative. Alternative {α}: Alternative {β}: Alternative 0 = {}: 1 = {α} (cid:26)b + c = {α} (cid:26)b + c = {β} (cid:26)b + c = 0 1 = {β} 1 = 0 (37) (38) (39) (40) (41) (42) In these systems the second equation is incorrect equality. Therefore these systems have no solution. Alternative 1 = {α, β}: 14 Sergey Tarasenko (cid:26)b + c = 1 1 = 1 (43) (44) The second equation is correct equality. Therefore this system has solution. Thus, out of four possible alternatives, subject a actually can choose only alternative 1 = {α, β}. To find solutions, resulting in selection of the alternative 1 = {α, β}, we need to solve only eq.(43), since eq.(44) turns into the true equality. According to Lemma 1, we instantly obtain the solution interval for eq.(43): 1 ⊇ b ⊇ c (45) We calculate the pairs (b, c) for all possible values of variable c (Table 3). Table 3. Solutions of the influence equation b + c = 1 Values of c Pairs (b, c) {α} {β} ({β},{α}) ({α},{β}) (1,{α}) (1,{β}) 1 0 (1, 0) (0, 1) ({α}, 1) ({β}, 1) (1, 1) Therefore, the influence analysis of the decision equation a = (b+c)a+a shows that the only alternative that subject a can choose is alternative 1 = {α, β}. The influence analysis provides us with the set (exhaustive list) of pairs (b, c) of joint influences resulting in selection of alternative 1 = {α, β}. Therefore, if the pair of influences does not match any pair from this list, the decision equation has no solution and this results in frustration state. Summarizing, this section we note that in general there are two sets. The set D contains alternatives that a controlled subject can choose. The set U is the set of altertanives of the target choice. Therefore, the need to put subject a into frustration state emerges, if the target choice of a controlled subject cannot be made by this subject. In other words, we need to put a subject into frustration state, if D ∩ U = Ø. 3.3 Analysis of Extreme Cases 2: What to do with Super-Active Groups Among all the possible groups, there are groups, in which subjects will always choose only the alternative 1 = {α, β} regardless of the influence of other sub- jects. Such groups are called super-active groups. The Inverse Task 15 Next we consider one special case of super active groups - the homogenous groups. The group is called homogenous, if all the subjects in the group are connected with the same relationship. Here we provide proof of the lemma about homogenous groups originally formulated by Lefebvre [1, 2]. Lemma 3. Any homogenous group is the super-active group. Proof. We consider the homogenous groups, where all the subjects are connected with alliance (alliance groups) and conflict (conflict groups) relationship, sepa- rately. Without loss of generallity, we suggest that there are n subjects a1, a2, ..., an. Alliance groups. The polynomial corresponding to the alliance group of n subject is a1a2...an. Next we construct the diagonal form and apply folding procedure: [a1][a2]...[an] [a1a2...an] = [a1a2...an] + [a1][a2]...[an] = 1 . Therefore the alliance groups are always super-active. Conflict groups. The polynomial corresponding to the conflict group of n subject is a1 + a2 + ... + an. Next we construct the diagonal form and apply folding procedure: [a1] + [a2] + ... + [an] [a1 + a2 + ... + an] [a1 + a2 + ... + an]+ [a1] + [a2] + ... + [an] = 1 . = Therefore the conflict groups are always super-active. Since both the alliance and the conflict groups are super-active, this lemma is proved. (cid:3) However, there are non-homogenous super-active groups as well (see Ap- pendix B). Summarizing this section, we note that subjects in the super-active groups cannot be controlled in their choices and the entire groups is uncontrolable. Therefore, once the super-active groups emerges, the only way to make it con- trollable is to change the relationships in the group. 4 The Basic Control Schema of an Abstract Subject (BCSAS) in the RGT We have presented the detailed description of the RGT including solution of the Forward and Inverse tasks. We have also considered the extream cases of decisions like putting a subject into frustration state or changing structure of a super-active group. As a final stroke, we summarize all the presented material in 16 Sergey Tarasenko Fig. 3. The Block schema for extracting sets Dh and Zh. the form of Basic Control Schema of an Abstract Subject (BCSAS) in the RGT. The input comes from the environment and is formalized in the form of exter- nal Influences on the subject, the Boolean algebra of Alternatives and Structure of a Group. Information about the Influences, Boolean algebra and Group Structure is propagated into the Decision Module. The Decision Module implements solution of the Forward task. Therefore the output set D of the Decision Module is the set of possible alternatives, which subject can choose under the given conditions. The information about Boolean algebra and Group Structure is propagated into the Influence Module. The Influence Module solves the Inverse task. The output set Dh of the Influence Module is the set of the pairs (χ,Zχ)x, where χ is the target alternative, the set Zχ is the set of all the joint influences, resulting in selection of the target choice; and x represents a subject variable. Each (χ,Zχ)x represents a reflexive control strategy. Therefore, the decision to put a subject into f rustration state is justified if it is impossible to make subject x choose the target alternative χ, i.e., if for pair (χ,Zχ)x set Zχ = {}, and subject x should not choose any other alternative except for the target one. Zχ== {}? yesnoPairs(M)M= 1Start M=<NM = M + 1End yesnoDh = Dh+ χRead Pairs (χ, Zχ)xSave DhZh= Zh+ ZχSave ZhThe Inverse Task 17 4.1 Schema for Iterative Algorithm to Obtain Output of the Influence Module The alternatives χ with corresponding non-empty sets Zχ are included into the set Dh. Here we introduce set Zh to store the non-empty sets Zχ. The schema of the algorithm for extracting sets Dh and Zh is presented in Fig. 3. First the sets Dh and Zh are empty: Dh = {} and Zh = {}. The algorithm reads the set of pairs (χ,Zχ)x and stores it in array P airs(M ), where M is a counting variable, N is the total number of pairs. Then it is checked for each pairs from array P airs whether set Zχ is empty: Zχ == {}? . If 'yes', the algorithm increments counting variable M (M = M + 1) and proceeds to the next pair from array Pairs. If 'no', then alternative χ is included into the set Dh(Dh = Dh + χ), set Dh is saved, the set Zχ is included into set Zh (Zh = Zh + Zχ) and set Zh is saved. The process is run while M ≤ N . In this iterative algorithm, we separately store the alternatives χ , which can be chosen by a certian subject, in the set Dh and the joint influences Zχ , which result in selection of alternative χ, in the set Zh. Therefore, we should modify the schema of Influence Module in BCSAS as follows. We present elaborated schema, where sub-module "Solution: Dh" is ac- companied with sub-module "Solution: Zh". Together these sub-modules are included into the "Solutions" sub-module. BCSAS is the fundamental schema of an abstract subject, which is used through out the RGT. The BCSAS is presented in Fig.4. This concludes the overview of RGT and description of tasks within the scope of the general theory. Therefore, we continue with application of the RGT to the mixed groups of humans and robots. Fig. 4. The Basic Control Schema of an Abstract Subject (BSCAS). Decision equation of a robotDecision Module Solutions : DBoolean Algebra ofAlternativesEnvironmentDecision equation of a humanInfluence ModuleSolution: DhRealization of an alternativeReflexive controlInfluencesSystem of Influence eqs.Structure of a GroupSolution: ZhSolutions18 Sergey Tarasenko 5 Defining Robots in RGT As we have noted in the Introduction section, the goal of the robots in mixed groups of humans and robots is to refrain human subject from choosing risky actions, which might result in injuries or even threaten live. It is considered by default that robot follows the program of behavior. Such program consists of at least three modules. The Module 1 implements robot's ability of human-like decision-making based on the RGT. The Module 2 contains the rules, which refrain robot from making a harm to human beings. The Module 3 predicts the choice of each human subject and suggests the possible reflexive control strategies. The Modules 1 and 3 are inhereted from the BCSAS of an Abstract Individ- ual. They correspond to Decision Module and Influence Module of the BCSAS (Fig. 4), respectively. Therefore all the properties and meaning of outputs of the Modules 1 and 3 are the same as the ones for Decision and Influence modules, respectively. The Module 2 is the new module, which is intrinsic for robotic agents studied in the context of mixed groups of humans and robots. This module is responsible for extraction of only harmless or non-risky alternatives for human subject. We suggest to apply Asimov's Three Laws of robotics [19], which formulate the basics of the Module 2: 1) a robot may not injure a human being or, through inaction, allow a human being to come to harm; 2) a robot must obey any orders given to it by human beings, except where such orders would conflict with the First Law; 3) a robot must protect its own existence as long as such protection does not conflict with the First or Second Law. We consider that these laws are intrinsic part of robots "mind", which cannot be erased or corrupted by any means. The interaction of Modules 1 and 2 is performed in the Interaction Module 1. The interaction of Modules 3 and 2 is implements in the Interaction Module 2. The Boolean algebra is filtered according to Asimov's laws in Module 2. The output of Module 2 is set U of approved alternatives. This data is then propagated into interaction modules. The output of the Module 1 is set D of alternatives, which robot has to choose under the given joint influences. In the Interaction Module 1, the conjunction of sets D and U is performed: D ∩ U = DU. If set DU is not empty set, this means that there are aproved alternatives among the alternatives that robot should choose in accordance with the joint influences. Therefore, robot can implement any alternative from the set DU. If set DU is empty, this means that under given joint influences robot cannot choose any approved alternative, therefore robot will choose an alternative from set U. This is how the Interaction Module 1 works. The output of the Module 3 contains sets Dh and Zh. The goal of the robot is to refrain human subjects from choosing risky alternative. This can be done The Inverse Task 19 Fig. 5. The Basic Control Schema of a Robotic Agent (BCSRA). by convincing human subjects to choose alternatives from the set U. First, we check whether Dh contains any approved alternative. We do so by performing conjunction of sets Dh and U: Dh ∩ U = DhU. If set DhU is not empty, then it means that it is possible to make a human subject to choose some non-risky alternative. Therefore, we should choose the corresponding reflexive control strategy from the set Zh. However, if set DhU is empty, we have to find the reflexive control strategy that will make human subject to select approved alternative from set U. For this purpose, we construct set ZU by including all the joint influences Zχ for approved alternatives: Zχ ∈ ZU ⇔ χ ∈ U. Next we check whether set ZU is empty. If set ZU is empty this means it is impossible to convince a human subject to choose non-risky alternative. Therefore, the only option of reflexive control in this case is to put this subject into frustration state. However, if set ZU is not empty, this means that there exist at least one reflexive control strategy that results in selection of alternative from the set of the approved (non-risky) ones. Therefore, the BCSRA inherits the entire structure of the BCSAS and aug- ments it with Module 2 of Asimov's Laws together with Interaction Modules 1 and 2. The original schema of robot's control system has been recently presented in [20]. The BCSRA is extended version of the original schema. The BCSRA Asimov Laws' based filterset Uof approved alternativesModule2Decision equation of a robotModule1Solutions : DDU== {}? UDUBoolean Algebra ofAlternativesyesnoDhU== {}? UDhUyesnoEnvironmentRealization of an alternativeReflexive controlDecision equation of a humanModule3Solution: DhInfluencesStructure of a GroupSolution: ZhSolutionsFrustrationyesX=DhU,for∀χ∈X get(χ, Zχ)for∀Zχ∈ZUget (χ, Zχ)ZU == {}? get set ZUof Zχ≠{}:∀χ∈UnoInteraction Module1Interaction Module220 Sergey Tarasenko provides comprehensive approach of how Forward and Inverse tasks are solved in the robot's "mind". Thus, in this section we have presented the formalization of robotic agent in the RGT. We outlined the specific features of robotic agents, which distinguish them from other subjects. Furthermore, we provided detailed explanation of how the Forward and Inverse tasks are solved in the framrework of control system (BCSRA) of robots. Next, we proceed with consideration of sample sutiations of interactions be- tween humans and robots. 6 Extended Sample Analysis of Mixed Groups Here we elaborate two examples, presented in the previous study [20], of how robots in the mixed groups can make humans refrain from risky actions. We discuss the application of the extended schema of robot's control system and provide explicit derivation of reflexive control strategies, which has been applied in these examples in the prevous study [20]. 6.1 Robots Baby-Sitters Suppose robots have to play a part of baby-sitters by looking after the kids. We consider a mixed group of two kids and two robots. Each robot is looking after a particular kid. Having finished the game, kids are considering what to do next. They choose between "to compete climbing the high tree" (action α) and "to play with a ball" (action β). Together actions α and β represent the active state 1={α, β} = {α} +{β}. Therefore the Boolean algebra of alternatives consists of four elements: 1) the alternative {α} is to climb the tree; 2) the alternative {β} is to play with a ball; 3) the alternative 1 = {α, β} means that a kid is hesitating what to do; and 4) the alternative 0 = {} means to take a rest. We consider that each kid considers his robot as ally and another kid and his robot as the competitors. The kids are subjects a and c, while robots are subjects b and d. The relationship graph is presented in Fig. 6. Fig. 6. The relationship graph for robots baby-sitters examples. Next we calculate the diagonal form and fold it in order to obtain decision equation for each subject: acbdThe Inverse Task 21 [a][b] [c][d] [ab] +[cd] [ab + cd] = ab + cd . From two actions α and β, action α is a risky action, since a kid can fall from the tree and this is real threat for his health or even life. Therefore according to Asimov's laws, robots cannot allow kids to start the competition. Thus, robots have to convince kids not to choose alternative {α}. In terms of alternatives, the Asimov's laws serve like filters which filter out the risky alternatives. The remaining alternatives are included into set U. In this case, U = {{β},{}}. Next we solve the Inverse taks, regarding alternatives {β} and {}. We conduct the analysis regarding kid a. This analysis can be further extended for kid c in the similar manner. Solution of the Inverse task for kid a with approved alternatives as target choice. The decision equation for kid a is a = ab + cd. First, we transform it into canonical form: a = (b + cd)a + cda. Next we consider system of influence equations: (cid:26)b + cd = χ (46) (47) where alternative χ ∈ U. Regarding eq.(47), eq.(46) is transformed into equation cd = χ, b + χ = χ variable d, we obtain all the pairs (c, d): (48) The solution of eq.(48) directly follows from Lemma 1: χ ⊇ b ⊇ 0. Therefore for χ = {β} and χ = {} the solutions are {β} ⊇ b ⊇ 0 and b = 0, respectively. The eq.(47) can be instantly solved according to Lemma 2: χd + χ d ⊇ c ⊇ χ. Consider χ = {β} first. Then {β}d + {α}d ⊇ c ⊇ {β}. By varying values of d = 1: {β} ⊇ c ⊇ {β} ⇒ c = {β}. Therefore the solution is pair ({β}, 1); d = 0: {α} ⊇ c ⊇ {β}. Since {α} ∩ {β} = {}, there is no solution; d ={α} : 0 ⊇ c ⊇ {β}. Since {β} ⊇ {}, there is no solution; d ={β} : 1 ⊇ c ⊇ {β}. Therefore there are two solutions (1,{β}) and Therefore equation cd = {β} has three solutions ({β}, 1), (1,{β}) and ({β},{β}). ({β},{β}). Thus, we have solved both equations from system (46-47). The solutions of this system are the triplets (b, c, d) of joint influences, which are all possible com- binations of solutions of both equations. Since there are two solution of eq.(46) and three solutions of eq.(47), there are six triplets (b, c, d) in total: (0,{β}, 1) and ({β},{β}, 1); (0, 1,{β}) and ({β}, 1,{β}); (0,{β},{β}) and ({β},{β},{β}). Now we consider the case, when χ = 0 = {}. Then d ⊇ c ⊇ 0. We obtain d = 1: 1 ⊇ c ⊇ 0 ⇒ c = 0. Thus, there is only one solution (0,1); pairs (c, d) for all values of variable d: 22 Sergey Tarasenko (1, 0); d = 0: 1 ⊇ c ⊇ 0. Thus, there are four solutions (1, 0), ({α}, 0), ({β}, 0) and d = {α}: {β} ⊇ c ⊇ 0. Thus, there are four solutions ({β},{α}) and (0,{α}); d = {β}: {α} ⊇ d ⊇ 0. Thus, there are four solutions ({α},{β}) and (0,{β}). In total, equation cd = 0 has 9 solutions. Therefore system (49-50) also has 9 solutions as triplets (b, c, d): (0, 1, 0), (0, 0, 0), (0, 0,{α}), (0, 0,{β}), (0, 0, 1), (0,{α},{β}), (0,{α}, 0), (0,{β},{α}) and (0,{β}, 0). We have considered two cases, when both upper and lower bounds of the interval of decision equation equal to the same alternative. Now we discuss a new situation, when variable a should take not a single value, but several values. In this case, we should find the joint influences (b, c, d) that result in selection of either alternative {β} or {}. Since, {β} ⊇ {}, we need to find all the triplets (b, c, d), resulting in the solution of decision equation as interval {β} ⊇ a ⊇ {}. Thus, {β} ⊇ a∗ ⊇ {}. Therefore, we need to solve the following system of equations: (cid:26)b + cd = {β} (49) cd = 0. (50) The eq.(49) turns into equality b = {β}, and we need to solve eq.(50). How- ever, this equation has been already solved in the previous example. Therefore we obtian the solutions of the system (49-50): ({β}, 1, 0), ({β}, 0, 0), ({β}, 0,{α}), ({β}, 0,{β}), ({β}, 0, 1), ({β},{α},{β}), ({β},{α}, 0), ({β},{β},{α}) and ({β},{β}, 0). Comparing solutions of all three system of influence equation, we can see that there are four remarkable solutions ({β},{β},{β}) and ({β},{},{β}); ({β}, 1,{β}) and ({β},{α},{β}). The first pair of solution results in choice of only alternative {β}, while second pair of solutions results in selection of eighter alternative {β} or alternative {}. These four solutions together illustrate that if b = d = {β}, it is guaranteed that regardless of influence of kid c, kid a will choose either of approved alternatives. By analogy, we can see that among solutions of system (46-47) with χ = {}, there are four solutions (0, 1, 0),(0, 0, 0), (0,{α}, 0) and (0,{β}, 0). Therefore, if b = d = 0, kid a will choose alternative 0 = {} regardless of influence of kid c. These two examples of binding variables b and d were considered in Scenario 1 and Scenario 2 of sample situation with robot baby-sitters, originally presented in [20]. Summarizing the results of this section, we have shown that robots can suc- cessfully control kids' behavior by refraining them from doing risky actions. The basic of this control is entirely based on the proposed schema of robot's control system. We have analyzed all the possible reflexive control strategies by solving three systems of influence equation: two systems regarding a single alternative and one system regarding the interval of alternatives. Therefore, we have shown how the Inverse task can be effectively solved by our proposed algorithm in situation similar to the real conditions. The Inverse Task 23 6.2 Mountain-Climbers and Rescue Robot We consider that there are two climbers in the mountain and rescue robot. The climbers and robot are communicating via radio. One of the climbers (subject b) got into difficult situation and needs help. Suggest, he fell into the rift because the edge of the rift was covered with ice. The rift is not too deep and there is a thick layer of snow on the bottom, therefore climber is not hurt, but he cannot get out of the rift himself. The second climber (subject a) wants to rescue his friend himself (action α), which is risky action. The second option is that robot will perform rescue mission (action β). Since inaction is inappropriate solution according to the First Law, the set U of approved alternatives for robot includes only alternative {β}. The goal of the robot is to refrain the climber a from choosing alernative {α} and perform rescue mission itself. We suggest that from the beginning all subjects are in alliance. The cor- responding graph is presented in Fig. 1c and its polynomial is abc. Therefore by definition it is homogenous group and, consequently, it is super-active group according to Lemma 3. Thus, any subject in the group is in active state. Therefore, group is un- controllable (see Section 3.3). In this case, robot makes decision to change his relationship with the climber b from alliance to conflict. Robot can do that, for instance, by not responding to climber's orders. Which reflexive control leads to frustration state? Then the polynomial corre- sponding to the new group is a(b + c). This polynomial has been already broadly discussed in the Section 3.2. Therefore, we know decision equation for subject a: a = (b+c)a+a. We have shown as well that subject a can choose only alternative 1 = {α, β}, if appropriate joint influences are applied (see Section 3.2), overwise subject a is in frustration state and cannot make any choice. Therefore, in or- der to put subject a into frustration state, the reflexive control strategy should N OT be selected from the list of solutions (Section 3.2): ({β},{α}); (1,{α}); ({α},{β}); (1,{β}); (0, 1); ({α}, 1); ({β}, 1); (1, 1) and (1, 0). Here we provide two examples of such joint influences (b, c): ({α},{α}) ⇒ ({α} + {α}) = {α} ⊂ 1 and ({β},{}) ⇒ ({β} + {}) = {β} ⊂ 1. Whether robot can complete mission regardless of joint influences of other subjects? The decision equation for robot c is c = c + (b + a)c. The corresponding solution interval is 1 ⊇ c ⊇ (b + a). Here we analyze all 16 possible reflexive control strategies (a, b) that climbers can apply to robot c. Examples with emtpy set DU. For (0, b), there will be the same situation regardless of value of variable b : 1 ⊇ c ⊇ (b + 0) ⇒ 1 ⊇ c ⊇ (b + 1) ⇒ c = 1. 1 ⊇ c ⊇ (1 + a) ⇒ c = 1. For (a, 1), there will be the same situation regardless of value of variable a : For ({α},{α}): 1 ⊇ c ⊇ ({α} + {α}) ⇒ 1 ⊇ c ⊇ ({α} + {β}) ⇒ c = 1. For ({β},{β}): 1 ⊇ c ⊇ ({β} + {β}) ⇒ 1 ⊇ c ⊇ ({β} + {α}) ⇒ c = 1. Therefore in these cases set D = {{α, β}}. Next we consider other pairs (a, b). 24 Sergey Tarasenko (1,{α}): 1 ⊇ c ⊇ ({α} + 1) ⇒ 1 ⊇ c ⊇ {α}. Here set D = {{α, β},{α}}. ({β},{α}): 1 ⊇ c ⊇ ({α} +{β}) ⇒ 1 ⊇ c ⊇ {α}. Here set D = {{α, β},{α}}. ({β}, 0): 1 ⊇ c ⊇ (0 +{β}) ⇒ 1 ⊇ c ⊇ {α}. Therefore, set D = {{α, β},{α}}. Since U = {{β}}, DU = {} for all the cases considered above, robot will choose alternative {β} from the set U. Examples with non-empty set DU. Consider the following pairs (a, b): (1,{β}): 1 ⊇ c ⊇ ({β} + 1) ⇒ 1 ⊇ c ⊇ {β}. Therefore, set D = {{α, β},{β}}. (1, 0): 1 ⊇ c ⊇ (0 + 1) ⇒ 1 ⊇ c ⊇ 0. Thus, set D = {{α, β},{α},{β},{}}. ({α},{β}): 1 ⊇ c ⊇ ({β}+{α}) ⇒ 1 ⊇ c ⊇ {β}. Thus, set D = {{α, β},{β}}. ({α},{β}): 1 ⊇ c ⊇ ({β}+{α}) ⇒ 1 ⊇ c ⊇ {β}. Thus, set D = {{α, β},{β}}. ({α}, 0): 1 ⊇ c ⊇ (0 + {α}) ⇒ 1 ⊇ c ⊇ {β}. Thus, set D = {{α, β},{β}}. Since U = {{β}}, DU = {{β}} for all the cases considered above, robot will choose alternative {β} from the set DU. Thus, we have shown that under all 16 reflexive control strategies (a, b), robot c can choose the alternative {β}, which is to perform the rescue mission itself. Therefore robot will choose alternative {β} regardless of the joint influences (a, b) of the climbers. The discussed example illustrates how robot can transform uncontrollable group into controllable one by manipulating the relationships in the group. In the controllable group by its influence on the human subjects, robot can refrain the climber a from risky action to rescue climber b. Robot achieves its goal by putting climber a into frustration state, in which climber a cannot make any decision. On the other hand, set U of approved alternatives guarantees that robot itself will choose the option with no risk for humans and implement it regardless of climber's influence. Therefore, in this section we have illustrated robot's ability to refrain human being from risky actions and to perform these risky actions itself. This proves that our approach achieves both goals of robotic agent: 1) to refrain people from risky actions and 2) to perform risky actions itself regardless of human's influences. 7 Discussion and Conclusion Summarizing, the results of this paper, we outline the most important of them. First of all, we have introduced the Inverse task and developed the ultimate methods to solve it. We have provided a comprehensive tutorial to the brand new Reflexive Game Theory recently formulated and proposed by Vladimir Lefebvre [1, 2, 3, 4]. The tutoral contains the detailed description of the Forward and Inverse tasks together with methods to solve them. We propose control schemas for both abstract subject (BCSAS) and robotic agent (BCSRA). These schemas were specially designed to incorporate solution of the Forward and Inverse tasks, thus providing us with autonomous units The Inverse Task 25 (individuals, subjects, agents) capable of making decisions in the human-like manner. We have shown that robotic agents based on BCSRA can be easily included into the mixed groups of humans and robots and effectively serve their fundamental goals (refraining humans from risky actions and, if needed, perform the risky acions itself). Therefore, we consider that present study provides the comprehensive overview of the classic RGT proposed by Vladimir Lefebvre [1, 2, 3, 4] and newly developed self-consistent framework for analysis of different kinds of groups and societies, including human social groups and mixed groups of humans and robots together with application tutorial of this new framework. This framework is entirely based on the principles of the RGT and brings together all its elements. The solution of the Inverse task, presented in this paper, plays a crutial role in formation of this framework. Therefore, by having the Inverse task as one of its fundamentals, this framework illustrates the role of the Inverse task and its relationship with other issues considered in the RGT. References 1. Lefebvre, V.A.: Lectures on Reflexive Game Theory. Leaf & Oaks, Los Angeles (2010). 2. Lefebvre, V.A.: Lectures on Reflexive Game Theory. Cogito-Center, Moscow (2009) [in Russian]. 3. Lefebvre, V.A.: The basic ideas of reflexive game's logic. Problems of research of systems and structures. pp. 73-79 (1965) [in Russian]. 4. Lefebvre, V.A.: Reflexive analysis of groups. In: Argamon, S. and Howard, N. (eds.) Computational models for counterterrorism. pp. 173-210. Springer, Heidel- berg (2009). 5. Lefebvre, V.A.: Algebra of Conscience. D. Reidel, Holland (1982). 6. Lefebvre, V.A.: Algebra of Conscience. 2nd Edition. Holland: Kluwer (2001). 7. Batchelder, W.H., Lefebvre, V.A.: A mathematical analysis of a natural class of partitions of a graph. J. Math. Psy. 26, pp. 124-148 (1982). 8. Kobatake, E., and Tanaka, K.: Neuronal Selectivities to Complex Object Features in the Ventral Pathway of the Macaque Monkey. Journal of Neurophysiology, 71, 3, pp. 856-867 (1994). 9. Koerner, E., Gewaltig, M.-O., Koerner, U., Richter, A., and Rodemann, T.: A model of computation in neocortical architecture. Neural Networks, 12, pp. 989- 1005 (1999). 10. Lucke, J., and von der Malsburg, C.: Rapid processing and unsupervised learning in a model of the cortical macrocolumn. Neural Computation, 16, pp. 501-533 (2003). 11. Schrander, S., Gewaltig, M.-O., Korner, U. and Korner, E.: Cortext: A columnar- model of bottom-up and top-down processing in the neocortex. Neural Networks, 22, pp. 1055-1070 (2009). 12. Fukushima, K.: Neocognitron: a self-organizing neural network model for a mech- anism of pattern recognitition unaffected by shift and position, Biological Cyber- natics, 36, pp. 193-201 (1980). 13. Riesenhuber, M. and Poggio, T.: Hierarchical models of object recognition in cor- tex. Nature Neuroscience, 2, 11, pp. 109-125 (1999). 26 Sergey Tarasenko 14. T. Serre, L. Wolf, S. Bileschi, M. Riesenhuber, and T. Poggio.: Robust Object Recognition with Cortex-like Mechanisms, IEEE Transactions on pattern analysis and machine intelligence, 29, 3, pp. 411-426 (2007). 15. Hienze, J.: Hierarchy length in orphaned colonies of the ant Temnothorax nylanderi Naturwissenschaften, 95, 8, pp. 757-760 (2008). 16. Chase, I., D.: Models of hierarchy formation in animal societies. Behavioral Science, 19, 6, pp. 374-382 (2007). 17. Chase I., Tovey C., Spangler-Martin D., Manfredonia M.: Individual differences versus social dynamics in the formation of animal dominance hierarchies. PNAS, 99, 9, pp. 5744-5749 (2002). 18. Buston P.: Social hierarchies: size and growth modification in clownfish. Nature, 424, pp. 145-146 (2003). 19. Asimov, I.: Runaround. Astounding Science Fiction, March, pp. 94-103 (1942). 20. Tarasenko, S.: Modeling mixed groups of humans and robots with Reflexive Game Theory. In Lamers, M.H., and Verbeek, F.J. (eds.): HRPR 2010, LINCST 59, pp. 108-117 (2011). Appendix A When sets A and B are functions of less than total number of subject minus one variables Consider groups of four subjects a, b, c and d. Suggest the polynomial corre- sponding to this group is b(a + d) + c. Next we construct diagonal form and perform folding operation: [a] + [d] [b][a + d] [b(a + d)] [b(a + d) + c] +[c] = [b]([a + d] + [a] + [d]) [b(a + d)] [b(a + d) + c] +[c] = [b] [b(a + d)] +[c] [b(a + d) + c] = = b(a + d) + c + b(a + d) + b + c Next we simplify the resultant expression of diagonal form folding: b(a + d) + c + b(a + d) + b + c = b(a + d) + c + b(a + d)cb = b(a + d) + cb + cb + b(a + d)cb = b((a + d) + c + b(a + d)c) + cb = b((a + d)c + (a + d)c + c + (b + (a + d))c) + cb = b((a + d)c + (a + d)c + c + bc + (a + d)c) + cb = b(c + (a + d)c + ((a + d) + (a + d))c) + cb = b((a + d)c + c + c) + cb = b((a + d)c + 1) + cb = b + cb = b + c The Inverse Task 27 Consequently, [b(a + d)] + [b] + [c] [b(a + d) + c] = b + c Therefore, the decision equation includes only two subject variables instead of four. Consequenly, for subjects a and d the decision equations in canonical forms are a = (b + c)a + (b + c)a d = (b + c)d + (b + c)d (51) (52) Thus, the sets A and B for subjects a and d are equal. The sets A and B are functions of only variables b and c: A = A(b, c) = b + cb and B = B(b, c) = b + cb. The canonical forms of decision equations for subjects b and c are: b = b + cb c = c + bc (53) (54) Therefore, set A = 1 for both subjects. Set B is a functions of a single variable: B(c) = c and B(b) = b for subjects b and c, respectively. B Example of non-homogenous super-active groups Here we provide an example of non-homogenous super-active group. Consider the group of four subject a, b, c and d, which is described by poly- nomial c(ab + b). Let us build the diagonal form and perform its folding: [a][b] [ab] +[d] [c][ab + d] [c(ab + d)] = ([ab] + [a][b]) + [d] [c][ab + d] = [c(ab + d)] = [c][ab + d] 1 = = [c(ab + d)] = [c(ab + d)] + [c][ab + d] = 1 (cid:3)
1911.03743	1	1911	2019-11-09T17:56:47	A perspective on multi-agent communication for information fusion	[ "cs.MA", "cs.AI", "cs.CL", "cs.LG" ]	Collaborative decision making in multi-agent systems typically requires a predefined communication protocol among agents. Usually, agent-level observations are locally processed and information is exchanged using the predefined protocol, enabling the team to perform more efficiently than each agent operating in isolation. In this work, we consider the situation where agents, with complementary sensing modalities must co-operate to achieve a common goal/task by learning an efficient communication protocol. We frame the problem within an actor-critic scheme, where the agents learn optimal policies in a centralized fashion, while taking action in a distributed manner. We provide an interpretation of the emergent communication between the agents. We observe that the information exchanged is not just an encoding of the raw sensor data but is, rather, a specific set of directive actions that depend on the overall task. Simulation results demonstrate the interpretability of the learnt communication in a variety of tasks.	cs.MA	cs	A perspective on multi-agent communication for information fusion Homagni Saha Department of Mechanical Engineering Iowa State University Ames, IA 50011 [email protected] Alberto Speranzon Honeywell Aerospace Plymouth, MN 55441 [email protected] Vijay Venkataraman Honeywell Aerospace Plymouth, MN 55441 [email protected] Soumik Sarkar Department of Mechanical Engineering Iowa State University Ames, IA 50011 [email protected] Abstract Collaborative decision making in multi-agent systems typically requires a prede- fined communication protocol among agents. Usually, agent-level observations are locally processed and information is exchanged using the predefined protocol, enabling the team to perform more efficiently than each agent operating in isola- tion. In this work, we consider the situation where agents, with complementary sensing modalities must co-operate to achieve a common goal/task by learning an efficient communication protocol. We frame the problem within an actor-critic scheme, where the agents learn optimal policies in a centralized fashion, while tak- ing action in a distributed manner. We provide an interpretation of the emergent communication between the agents. We observe that the information exchanged is not just an encoding of the raw sensor data but is, rather, a specific set of di- rective actions that depend on the overall task. Simulation results demonstrate the interpretability of the learnt communication in a variety of tasks. 1 Introduction In this paper, we analyze communication protocols learnt by a team of agents equipped with com- plementary sensor modalities and tasked with a common goal. We call this "task based multi-modal decision making", wherein agents learn to map their sensor measurements and the information com- municated by other agents, directly into actions based on the common goal. In this setting, each agent has access only to its own sensor data but needs to rely on the communication with other agents to obtain task relevant information from that agent's sensor modality. We present a way to interpret the emergent communication by visualizing this mapping into the agent's action space. We find the communication that is learnt, within a reinforcement learning paradigm, is not only emergent, but is task dependent and adaptive to the size of the communication channel. In relation to existing literature on learning for multi-agent systems our work borrows from the gen- eral framework of "Markov-games", proposed in Lowe et al. [2017], Mordatch and Abbeel [2018], Foerster et al. [2016] and references therein. In particular, we consider ( Lowe et al. [2017]) for our learning problem. Related to the emergent communication aspect, central to this work, we consider ideas from the literature on "multi-agent referential games" ( Golland et al. [2010], Andreas and Klein [2016], Evtimova et al. [2017], Lazaridou et al. [2018]), where a (sender) agent communi- cates highly structured information (images and text) to a (receiver) agent which has to interpret what the other agent saw. However, here we are interested in the evolution of communication for Preprint. Under review. unstructured data under joint interactions using an actor-critic algorithm. Emergent communication was also studied in ( Kottur et al. [2017], Cao et al. [2018]), however, we focus our analysis of emergent communication on the action space. Specifically, we project the learnt communication on the action space and visually analyze the results. This enables us to more clearly interpret the learnt communication. It is shown that powerful joint representations of the world can be encoded through task dependent communication which is easy to interpret under complementary sensing modality constraints. 2 Environment and tasks For our experimental study, we consider a two-dimensional world with two agents and L landmarks. Each landmark has a color {red, green, blue} and shape {triangle, circle, square} property. Our agents have complementary sensing modalities: one of the agents, denoted as color-agent, can only observe the color of the landmarks and the other, denoted as shape-agent, can only observe the shape of the landmarks. We assume that both agents can measure their (relative) distance from all land- marks but cannot measure their distance from each other. At every discrete time step the agents take both physical movement actions (a unit movement in one of the four directions or stand still) and communicative actions, namely, broadcast a k-bit message. The communication message sent by one agent is received by the other in the next time step. Each agent's observation is a vector oi = [xi1, yi1, mi1, . . . , xiL, yiL, miLci1, . . . , cik gi1, . . . , giL], where xij and yij denote the horizontal and vertical distances, respectively, between the ith agent and the jth landmark; mij denotes a one hot encoding of the jth landmark's property as sensed by the ith agent. For example, m1j ∈ {[1, 0, 0], [0, 1, 0], [0, 0, 1]} denotes the encoding of the jth landmark's color for the color agent or shape for the shape agent. The vector [ci1, . . . , cik], denotes the k bit word received bythe ith agent. Finally, [gi1, . . . , giL] denotes one hot encoding of the target landmark properties provided to agent i. We base our study on the three collaborative tasks, described below. Task 1: Cross modal information exchange: In this task, the map contains three landmarks. No two landmarks have the same shape or color. During each episode, agents and landmarks are placed randomly in the map. One of the landmarks is designated as "target" and the goal, for both agents, is to reach the designated target landmark. This target landmark's property is indicated to the agents using [gi1, . . . , giL] as described previously. Consider as example where the target landmark is a blue circle, if we were to provide to the color agent the color properties, it can trivially navigate to the target given that it has full knowledge of where the different colored landmarks are with respect to itself. To avoid this and to encourage communication, we pass the encoding corresponding to the shape of the target landmark to the color-agent (circle in this example) and vice versa for the shape agent. This creates a situation where the agents need to exchange information in order to success- fully navigate to the right landmark. Task 2: Multi target consensus: For this task, the map contains six landmarks each with shape and color properties and no two landmarks have the same set of properties. However, there can be two landmarks with the same color or shape and the target landmark is unique when both properties are considered. The goal is for both agents to move to the target landmark. The agent observation and action spaces are similar to the previous scenario, but here the property of the target landmark is specified in the agent's own modality. Specifying the encoding for circle, as the target, to the shape agent does not trivially solve the problem as there can be two circles and co-ordination with the color agent is necessary to figure out which circle is blue and then move towards it. In this task, the agents need to reach a consensus on which is the target landmark by learning to reasoning over their observation spaces. Task 3: Collaborative localization: Here the setup is similar to the information exchange task. However, no target landmark is specified as the goal is for the agents to meet with each other in the shortest possible time. Here a constant negative reward Rt supplied to the agents at each time-step to encourage meeting up fast. Here the agents must learn to estimate their relative position with respect to each other and then take actions to move closer. Summarizing, all the above tasks share the following key challenging characteristics. Agents have (i) different sensing modalities; (ii) No knowledge about other agent's sensor or position; (iii) No common world coordinate frame in their state space; (iv) a finite communication bandwidth. Reward structure and learning framework: We primarily used three types of rewards: Rd = iT , where i is the agent number, xiT and yiT are the horizontal and vertical dis- tances of ith agent from the target landmark, and n is the number of agents. We define an instanta- (cid:80)i=n (cid:112)x2 i=1 iT + y2 2 Task Information exchange Multi target consensus Collaborative localization k Reward Metric (%) M1 M2 80 99 2 89.5 100 3 98.9 100 4 4 15.1 0.5 61.6 91.8 4 Rd + Ri 80.5 100 3 Rd + Rt Rd Rd Rd Rd Table 1: Performance metrics for the three test tasks with variations in reward structure and communi- cation channels. neous reward Ri = H, where H is a large number if at least one agent is touching the target at the current time step and 0 otherwise. For collaborative localization task, Rt is a constant penalty per time step and Rd is the inter agent distance. For all tasks, our reinforcement learning (RL) frame- work is based on the MADDPG algorithm (Lowe et al. [2017]), which relies on centralized training and decentralized execution, making it suitable for multi-agent problems. The core of MADDPG is an actor-critic scheme (Grondman et al. [2012]) that maintains a critic for each agent and the critics have access to actions (movement and communications) and rewards of all the agents. This helps with the problem of non-stationarity in multi-agent environments. In all experiments, we param- eterize the output of both actors and the critics with a three layered fully connected network with ReLU activations. It must be noted that, although the agent state space allows for real numbers in the communication stream, the use of Gumbel-Softmax estimator (Jang et al. [2016]) transforms these into discrete valued messages. While 2k word variations are possible, we observe that the agents limit their vocabulary use to k + 1 words. For 3 channels of communication, the word vocabulary was limited to w0 = [0, 0, 0], w1 = [1, 0, 0], w2 = [0, 1, 0], w3 = [0, 0, 1]. Details of the learning framework, training hyperparameters and reward curves are provided in Appendix A. 3 Results In the following we evaluate agent performance using simple metrics and then provide interpreta- tions of the emerged communication between the agents. Performance metrics: We use two metrics to evaluate the performance of the agents in achieving the common goal. Let m1 denote the number of episodes in which at least one of the agent reaches the target landmark, m2 denote the number of episodes where both agents reach the target land- mark, and N denote the total number of test episodes. We use N = 1000, let M1 = m1/N and M2 = m2/N. For the collaborative localization task m2 denotes the number of times the agents meet with each other, and m1 is the starting distance between agents divided by total distance trav- elled by both agents in the episode. As we mentioned, k is the number of communication bits available to the agent. Table 3 shows these metrics for the three different tasks. Emerged communication: Table 3 shows that the agents are able to successfully complete the information exchange and collaborative localization tasks almost every time. In the more complex multi target consensus task the agents achieve a 61.6% success rate which is better than random chance of 25%. In order to better understand how the communication aid the agents, we devise a way to visualize this relation as follows. At every time step the ith agent decides its actions based on its observations oi comprising of relative position to the landmarks, the word received from the other agent and target landmark (if applicable). For a given test case the target landmark is fixed. Then for every possible word that can be received, we can place the agent in a fixed position of the environment and query the learnt policy to find in which direction the agent would move. We can then repeat this for all possible agent positions and color code its preferential direction of motion at each location, obtaining a picture as shown in figure 1. As expected, we observe random motion in the beginning of the training. Over time the color agent learns to move to the blue triangle if the word uttered by the shape agent is [1, 0, 0]. Similarly the shape agent learns to move towards the blue circle for the same utterance by the color agent. Visualization for other word utterances are given in the Appendix, see figure 6. Note that both agents are either focusing on a blue or circular object for all word utterances as the given target in this example is a blue circle. It is remarkable that the agents are able to solve a complex map alignment and reasoning problem by directing each others actions through communication. Examples of the final learnt policy for the information ex- change and collaborative localization tasks are shown in figure 2 left and right respectively. In the information exchange task, we observe that each unique word uttered by the color agent causes the shape-agent to move close to a specific shape of landmark irrespective of its current position in the map (e.g [1, 0, 0, 0] causes movement towards the circle). In the collaborative localization task each unique word uttered by the color agent causes the shape agent to move towards a focus point in the 3 Figure 1: Multi target con- sensus task. Evolution of policies, for the word utter- ance [1, 0, 0, 0], for both color agent (top row) and shape- agent (bottom row). Policy re- lation to color: Green - Go down, Turquoise - Go left, Blue - Go up, Yellow - Go right, Grey - No movement. The meeting point of the dif- ferent colors (vertex) is the equilibrium point at which the agent may come to rest. Figure 2: Left-Information exchange task. From left to right, the color-agent utters the words [1, 0, 0, 0], [0, 1, 0, 0], [0, 0, 1, 0] to the shape agent. Top and bottom rows represent different land- mark configurations. Right-Collaborative localization task. From left to right, policy for the words [1, 0, 0], [0, 1, 0], [0, 0, 1] are visualized for both color (top row) and shape agents (bottom row). map and vice-versa. During the episode, both agents continuously change their utterance in order to force their partner to travel towards each other and meet in the shortest time possible. Effect of communication channels: Changing the number of channels affects the learning capac- ity. In the information exchange task when 3 channels are provided agents associate their target landmarks as "topmost","leftmost", or "bottom most" landmark in the map, depending on received communication. While this is a clever reference, as the "topmost" etc. landmark can easily be the same for the two agents as their reference axes are only translated from each other. However, this leads to a failure when the target landmark is located in the middle or is the "rightmost". When using 4 channels agents can directly associate their targets to the property communicated to them by the other agent. In figure 2 (left), shape agent interprets [1, 0, 0, 0] from the color agent as a signal to go to a circle, [0, 1, 0, 0] to a triangle and [0, 0, 1, 0] to a square. This improves performance greatly. 4 Conclusion We studied the application of multi agent reinforcement learning for task driven multimodal decision making. We analyzed the emergence of interpretable communication between agents and found that adaptive and non trivial communication protocols can be learned based on number of available communication channels and imposed reward structures. The size of the communication channel can be crucial in deciding the amount of information that is required to reconcile various modalities with each other and reward structures affect the nature of the learned communication. Visualizing emergent policies in the agent's action spaces confirms that powerful joint representation of the world can be encoded through communication. 5 Acknowledgement We are thankful to Shashank Shivkumar at Honeywell Aerospace for discussions related to this paper ranging from initial idea for the research through algorithm implementation and testing. 4 References Jacob Andreas and Dan Klein. Reasoning about pragmatics with neural listeners and speakers. arXiv preprint arXiv:1604.00562, 2016. Kris Cao, Angeliki Lazaridou, Marc Lanctot, Joel Z Leibo, Karl Tuyls, and Stephen Clark. Emergent communication through negotiation. arXiv preprint arXiv:1804.03980, 2018. Katrina Evtimova, Andrew Drozdov, Douwe Kiela, and Kyunghyun Cho. Emergent communication in a multi-modal, multi-step referential game. arXiv preprint arXiv:1705.10369, 2017. 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. Dave Golland, Percy Liang, and Dan Klein. A game-theoretic approach to generating spatial de- In Proceedings of the 2010 conference on empirical methods in natural language scriptions. processing, pages 410 -- 419. Association for Computational Linguistics, 2010. Ivo Grondman, Lucian Busoniu, Gabriel AD Lopes, and Robert Babuska. A survey of actor-critic reinforcement learning: Standard and natural policy gradients. IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews), 42(6):1291 -- 1307, 2012. Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. Satwik Kottur, Jos´e MF Moura, Stefan Lee, and Dhruv Batra. Natural language does not emerge'naturally'in multi-agent dialog. arXiv preprint arXiv:1706.08502, 2017. Angeliki Lazaridou, Karl Moritz Hermann, Karl Tuyls, and Stephen Clark. Emergence of lin- guistic communication from referential games with symbolic and pixel input. arXiv preprint arXiv:1804.03984, 2018. Ryan Lowe, Yi Wu, Aviv Tamar, Jean Harb, Pieter Abbeel, and Igor Mordatch. Multi-agent actor- critic for mixed cooperative-competitive environments. In Advances in Neural Information Pro- cessing Systems, pages 6379 -- 6390, 2017. Igor Mordatch and Pieter Abbeel. Emergence of grounded compositional language in multi-agent populations. In Thirty-Second AAAI Conference on Artificial Intelligence, 2018. 5 Appendix: A perspective on multi-agent communication for information fusion A : Framework and training An overview of the learning framework we used for the three different tasks is shown in Figure 3. In the information exchange and collaborative localization tasks we train with a maximum episode length of 60. In multi-target consensus task, agents are trained for 120000 episodes with a maximum episode length of 80 steps. We use Adam optimizer with learning rate of 0.01, discount factor of 0.001 for the critics, and a batch size of 1024. In general, convergence was observed within 5000 episodes. Figure 4 shows the training progress over number of episodes for the three different Figure 3: Learning framework used in this work. tasks. In the information exchange and collaborative localization tasks, the majority of the policy improvement takes place in the initial episodes and the agent maintains the word-action associations it learns over the following episodes. In the more complex multi target consensus task, the agents take the longest to learn meaningful policies that maximize reward. Reward curve has a sudden peak near 5000 episodes, however the policies still gradually keep improving over time till 80000 episodes. B : Effect of reward structure on communication For the complex multi-target consensus task, we found that just using a continuous average distance penalty (Rd) for each time step in the episode is detrimental to learning. As the training progresses, the agents learn to ignore communications and stay in the same place where they started: the whole Figure 4: Plots of average total cumulative rewards vs training episode number Left- multi target consensus task. Middle- information exchange task. Right - collaborative localization task. 1 Figure 5: Evolution of policy in multi target consensus when only average distance penalty is used. Two scenarios are shown. Identical plots are obtained for shape agent. policy map changes to grey (no movement) in the final epochs as shown in figure 5. To encourage more exploration, we introduced an instantaneous touching reward Ri in addition to the constant average distance penalty and observe improved performance. We also experimented with just using Ri alone and the policies learned are sub optimal. So it appears that both the reward types are necessary for learning in complex multimodal scenarios. C : Number of words learned While k bit communication channel is used, we experimentally verified that 2k different words are valid. However the agents mainly chose to use only k + 1 words while navigating. An example policy evolution corresponding to all the words in multi-target consensus is shown in figure 6. To understand what the other unused words meant we visualized the policy corresponding to two such words as shown in figure 7. We find that the (combination) word utterance [1, 1, 0] will produce a new focus / equilibrium point different from that of the individual words [1, 0, 0] and [0, 1, 0]. While it can be useful to use another focus point to guide the movement of the other agent in the collaborative localization environment, the agents prefer to just utter different simple words sequentially, rather than a more complex word once. 2 Figure 6: Evolution of policy in multi target consensus for shape-agent (top) and color-agent (bot- tom) for different words uttered (left to right). Figure 7: Policy visualizations corresponding to simple [1, 0, 0], [0, 1, 0], [0, 0, 1] and complex words [1, 1, 0], [1, 0, 1] in collaborative localization task. 3
1902.07497	3	1902	2019-04-10T13:46:37	The Representational Capacity of Action-Value Networks for Multi-Agent Reinforcement Learning	[ "cs.MA" ]	Recent years have seen the application of deep reinforcement learning techniques to cooperative multi-agent systems, with great empirical success. However, given the lack of theoretical insight, it remains unclear what the employed neural networks are learning, or how we should enhance their representational power to address the problems on which they fail. In this work, we empirically investigate the representational power of various network architectures on a series of one-shot games. Despite their simplicity, these games capture many of the crucial problems that arise in the multi-agent setting, such as an exponential number of joint actions or the lack of an explicit coordination mechanism. Our results quantify how well various approaches can represent the requisite value functions, and help us identify issues that can impede good performance.	cs.MA	cs	THE REPRESENTATIONAL CAPACITY OF ACTION-VALUE NETWORKS FOR MULTI-AGENT REINFORCEMENT LEARNING Jacopo Castellini Dept. of Computer Science University of Liverpool [email protected] A PREPRINT Frans A. Oliehoek Interactive Intelligence Group Delft University of Technology [email protected] Rahul Savani Dept. of Computer Science University of Liverpool [email protected] Shimon Whiteson Dept. of Computer Science University of Oxford [email protected] ABSTRACT Recent years have seen the application of deep reinforcement learning techniques to cooperative multi-agent systems, with great empirical success. However, given the lack of theoretical insight, it remains unclear what the employed neural networks are learning, or how we should enhance their representational power to address the problems on which they fail. In this work, we empirically investigate the representational power of various network architectures on a series of one-shot games. Despite their simplicity, these games capture many of the crucial problems that arise in the multi- agent setting, such as an exponential number of joint actions or the lack of an explicit coordination mechanism. Our results quantify how well various approaches can represent the requisite value functions, and help us identify issues that can impede good performance. Keywords multi-agent systems · neural networks · decision-making · action-value representation · one-shot games 1 Introduction In future applications, intelligent agents will cooperate and/or compete as part of multi-agent systems (MASs) [30, 15, 33, 7]. Multi-agent reinforcement learning (MARL) uses RL to solve such problems and can lead to flexible and robust solutions [2], and recently, a variety of work [28, 24, 4, 17] has successfully applied deep MARL techniques. These approaches have shown good results, but given the lack of theoretical insight, it remains unclear what these neural networks are learning, or how we should enhance their representational power to address the problems on which they fail. In this paper, we focus on value-based MARL approaches for cooperative MASs. Value-based single-agent RL meth- ods use (deep) neural networks to represent the action-value function Q(s,a; θ) to select actions directly [19] or as a 'critic' in an actor-critic scheme [20, 16]. A straightforward way to extend such methods to the multi-agent setting is by simply replacing the action by the joint action a = (cid:104)a1, . . . ,an(cid:105) of all agents Q(s,(cid:104)a1, . . . ,an(cid:105); θ). However, this approach heavily relies on the function approximation abilities of the neural network, since it will need to generalize across a discrete action space whose size is exponential in the number of agents. Moreover, selecting a joint action a that maximizes the Q-function will require that, as in deep Q-networks [19], the (now joint) actions need to be output nodes. As a result, the computational and sample costs scale poorly in the number of agents. Another approach to extending single-agent RL methods to MASs is to apply them to each agent independently. This improves scalability at the expense of quality, e.g., individual deep Q-learners may not be able to accurately represent This work as been accepted as an Extended Abstract in 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 2019, Montreal, Canada. The Representational Capacity of Action-Value Networks for Multi-Agent Reinforcement LearningA PREPRINT the value of coordination. Furthermore, the environment becomes non-stationary from the perspective of a single agent and thus, unsurprisingly, their learning process may not converge [3, 29, 32]. A middle ground is to learn factored Q-value functions [9, 11], which represent the joint value but decompose it as the sum of a number of local components, each involving only a subset of the agents. Compared to independent learning, a factored approach can better represent the value of coordination and does not introduce non-stationarity. Compared to a naive joint approach, it has better scalability in the number of agents. Recently, factored approaches have shown success in deep MARL [27, 24]. This paper examines the representational capacity of these various approaches by studying the accuracy of the learned Q-function approximations Q. We consider the optimality of the greedy joint action, which is important when using Q to select actions. We also consider distance to optimal value ∆Q = Q − Q, as verifying the optimality of the greedy action requires bounding ∆Q. Furthermore, minimising ∆Q is important for deriving good policy gradients in actor-critic architectures and for sequential value estimation in any approach (such as Q-learning) that relies on bootstrapping. However, to minimise confounding factors, we focus on one-shot (i.e., non-sequential) problems. Specifically, we investigate the representational power of various network architectures on a series of one-shot games that require a high level of coordination. Despite their simplicity, these games capture many of the crucial problems that arise in the multi-agent setting, such as an exponential number of joint actions. While good performance in such one-shot settings does not necessarily imply good performance in the sequential setting, the converse is certainly true: any limitations we find in one shot settings would imply even greater limitations in corresponding sequential settings. Thus, assessing the accuracy of various representations in one-shot problems is a key step towards understanding and improving deep MARL techniques. 2 Background 2.1 One-Shot Games i=1,{Qi}D In this work, we focus on one-shot games, which do not have a notion of environment state. The model consists of the tuple M = (cid:104)D,{Ai}D i=1(cid:105), where D = {1, . . . ,n} is the set of agents, Ai is the set of individual actions for agent i (A = ×D i=1Ai is the joint action set) and the set of reward functions Qi(a)1, depending only on the joint action a ∈ A performed by the team of agents, express how much reward agent i gets from the overall team decision. A cooperative one-shot game is a game in which all agents share the same reward function Q(a), so that the goal of the team is to maximize this shared reward by finding the optimal joint action a ∈ A to perform. In this work, we focus on cooperative games. Our work aims at investigating the representations of the action-value function obtained with various neural network approaches and how close these are to the original one. We do not investigate the learning of an equilibrium strategy for the agent to exploit, as is typically considered in works on repeated games. 2.2 Coordination Graphs In many problems, the decision of an agent is influenced by those of only a small subset of other agents [9]. This locality of interaction means the joint action-value function Q(a) can be represented as the sum of smaller reward functions, one for each factor e: C(cid:88) Q(a) = Qe(ae), (1) e=1 where C is the number of factors and ae = (cid:104)ai(cid:105)i∈e is the local joint action of the agents that participate in factor e. The structure of the interactions between the agents can be represented with a (hyper-) graph called a coordination graph [14]. A coordination graph has a node for each agent in the team and (hyper) edges connecting agents in the same factor. Figure 1 shows some examples coordination graphs. Coordination graphs are a useful instrument to represent interactions between agents and there are many algorithms that exploit such structures and require good approximations of the action-value function in order to efficiently select a maximizing joint action, e.g., variable elimination [9] or max-sum [25, 14]. 1We write Q(a) for the reward function in the one-shot problem to make the link with sequential MARL more apparent. 2 The Representational Capacity of Action-Value Networks for Multi-Agent Reinforcement LearningA PREPRINT There are many cases in which the problem itself is not perfectly factored according to a coordination graph that can be exploited. In these cases, however, it can still be useful to resort to an approximate factorization [11]: Q(a) ≈ Q(a) = Qe(ae), (2) (cid:88) obtained by considering a decomposition of the original function in a desired number of local approximate terms Qe(ae), thus forming an approximation Q of the original action-value function Q(a). e 3 Action-Value Functions for MARL Current deep MARL approaches are either based on the assumption that the joint-action value function Q(s,a) can be represented efficiently by neural networks (when, in fact, the exponential number of joint actions usually makes a good approximation hard to learn), or that it suffices to represent (approximated) individual action values Qi(si,ai) [18]. Our aim is to investigate to what degree these assumptions are valid by exploring them in the one-shot case, as well as assessing if higher order factorizations result in improved representations of such functions, while making learning easier (as there are only small factors to be learned). We use neural networks as function approximators to represent the various components of these factorizations. Our work explores a series of directions, both in terms of agent models (how to use these networks to represent the various agents/factors) and in terms of learning algorithms, combining them to independently assess the influence of these various aspects on the represented action-value functions. We couple each of the factorizations introduced later with the following two learning approaches. • Mixture of experts (MoE) [1]: each factor network optimizes its own output Qe to predict the global reward, thus becoming an "expert" on its own field of action. The loss for the network representing factor e at training sample t is defined as: Le t (ae t ) = 1 2 Q(at) − Qe(ae t ) , (3) where Q(at) is the common reward signal received after selecting joint action at and Qe(ae the network for local joint action ae is, after training, we compute it from the C factors as the mean over the appropriate local Q-values: t ) is the output of t . As we aim to assess how good the reconstructed action-value function Q(a) = 1 n Qe(ae) ∀a ∈ A. (4) • Factored Q-function (FQ) [10]: we jointly optimize the factor networks to predict the global reward as a sum of their local Q-values Qe. The loss for the sample at time t is identical for all networks: (cid:17)2 (cid:33)2 (cid:16) C(cid:88) e=1 (cid:32) Q(at) − C(cid:88) C(cid:88) e=1 Lt(at) = 1 2 Qe(ae t ) . (5) After training, the joint action-value function is reconstructed as the sum of the appropriate local Q-values: Q(a) = Qe(ae) ∀a ∈ A. (6) We investigate four different factorizations: e=1 • Single agent decomposition: each agent i is represented by an individual neural network and computes its own individual action-values Qi(ai), one for each output unit, based on its local action ai. Under the mixture of experts, this corresponds to the standard independent learning approach to MARL, in which we learn local agent-wise components, while under the factored Q-function approach this corresponds to the value decomposition networks from [27]. • Random partition: agents are randomly partitioned to form factors of size f, with each agent i involved in D f factors is represented by a different neural network that represents local only one factor. Each of the C = action-values Qe(ae), one for each of its output units, for a certain factor e, where ae is the local joint action of agents in factor e. We consider factors of size f ∈ {2,3}. 3 The Representational Capacity of Action-Value Networks for Multi-Agent Reinforcement LearningA PREPRINT e0 1 2 6 e2 (a) e1 4 3 5 2 e3 3 e5 1 e4 e0 4 1 2 e5 3 e0 e1 e8 e6 e11 e7 e2 e9 e10 4 e2 e1 6 5 (b) e4 e3 e13 e12 6 5 e14 (c) Figure 1: Example coordination graphs for: (a) random partition, (b) overlapping factors, (c) complete factorization. • Overlapping factors: a fixed number C of factors is picked at random from the set of all possible factors of size f. We require the sampled set to not include duplicate factors (we use C distinct ones) and that every agent i appears in at least one factor. Every factor e is represented by a different neural network learning local action-values Qe(ae) for the local joint actions ae, one for each output unit. In our experiments we choose C = 6, again with factors of size f ∈ {2,3}. • Complete factorization: each agent i is grouped with every possible combination of the other agents in the team D \ i to form factors of size f, resulting in C = f !(D−f )! factors, each represented by a network. Each of these networks learns local action values Qe(ae), one for each output unit of the network, conditioned on local joint actions ae of component e. As for the other factorizations, we consider factors of size f ∈ {2,3}. D! This results in the following combinations: Single agent Random p. (f = 2,3) Overlapping (f = 2,3) Complete (f = 2,3) Table 1: Combinations of factorizations and learning rules. F2R, F3R F2O, F3O F2C, F3C Mix. of Experts M1(=IL [29]) M2R, M3R M2O, M3O M2C, M3C Factored Q F1(=VDN [27]) 4 Experiments We investigate the representations obtained with the proposed combinations of factorization and learning approach on a series of challenging one-shot coordination games that do not present an explicit decomposition of the reward function Q(a) (non-factored games), and then on two factored games. The following Table summarizes the investigated games and associated parameters. Game Dispersion/Platonia Climb/Penalty Generalized FF Aloha n 6 6 6 6 Ai 2 3 2 A 64 729 64 Factored No No Yes Yes 2 (per type) 64 (8192 total) Table 2: Details of the investigated games. 4 The Representational Capacity of Action-Value Networks for Multi-Agent Reinforcement LearningA PREPRINT 4.1 Experimental Setup Every neural network has a single hidden layer with 16 hidden units using leaky ReLU activation functions, while all output units are linear and output local action-values Qe(ae) 2. Given the absence of an environment state to feed to the networks as an input, at every time step they just receive a constant scalar value. We used the mean square error as the loss function and the RMSprop training algorithm with a learning rate of η = 10−5. For every game, we trained the networks with 100,000 examples by sampling a joint action at uniformly at random3. Then, we propagate the gradient update through each network e from the output unit Qe(ae t ). The loss function minimizes the squared difference between the collected reward Q(at) at each training step and the approximation computed by the networks. After training, the learned action-value function Q is compared to the original Q one. We also consider a baseline joint learner (a single neural network with an exponential number A = Ain of output units). Every experiment was repeated 10 times with random initialization of weights, each time sampling different factors for the random partitions and the overlapping factors; we report the averages of these. 4.2 Non-Factored Games 4.2.1 Dispersion Games In the Dispersion Game, also known as Anti-Coordination Game, the team of agents must divide as evenly as possible between the two local actions that each agent can perform [8]. This game requires explicit coordination, as none of the local actions is good per se, but the obtained reward depends on the decision of the whole team. We investigate two versions of this game: in the first one the agents obtain reward proportional to their dispersion coefficient (i.e., how split the agents are in performing one of their two local actions). The reward function Q(a) for this game with n agents, each with a local action set Ai = {a0,a1} is: Q(a) = n − max{#a0, #a1}. (7) In the second version, which we call Sparse Dispersion Game, the agents receive a reward (which we set to the maximum dispersion coefficient with n agents: n 2 ) only if they are perfectly split: (cid:26) n 2 0 Q(a) = if #a0 = #a1, otherwise. (8) Figure 2 shows the Q-function reconstructed by the proposed factorizations and learning approaches for these two games. In these plots (and those that follow), the x-axis enumerates the joint actions a ∈ A and the y-axis shows the corresponding values Q(a) for the reconstructed function, with the color of the bars as an encoding of the action-value. We analyse the accuracy of the computed reconstructions considering two aspects: the total reconstruction error of Q(a) with respect to the original reward function Q(a) ∀a ∈ A, and whether a reconstruction is able to produce a correct ranking of the joint actions. For a good reconstruction, the bars have the same relative heights, indicating that the factorization correctly ranks the joint actions with respect to their value, and to be of a similar value to those in the original one (the factorization can reconstruct a correct value for that reward component). However, reconstruction error alone is not a good accuracy measure because lower reconstruction error does not imply better decision-making, as a model could lower the total error by over- or underestimating the value of certain joint actions. Figure 2(a) shows that the proposed complete factorizations are able to almost perfectly reconstruct the relative ranking between the joint actions, meaning that these architectures can be reliably used for decision making. Moreover, the ones using the factored Q-function (F2C and F3C in the plot) are also able to produce a generally good approximation of the various reward components (expressed by the value of the bars), while those based on the mixture of experts produce a less precise reconstruction: the joint optimization of the former gives an advantage in this kind of extremely coordinated problems. Smaller factorizations, like the random pairings, are not sufficient to correctly represent this function, probably because a higher degree of connection is required to represent coordination. Figure 2(b) is similar but in this case the reconstruction is less accurate and the values of the bars are quite different from those of the original one. This is possibly due to the sparsity of the reward function, requiring the networks to correctly approximate quite different values with the same output components. In this case, the sparseness of the reward function fools the representations into being similar to those of the non-sparse version. 2We also investigated deeper networks with 2 and 3 hidden layers, but did not find improvements for the considered problems. 3We do not use -greedy because we are interested in representing the whole value function and not just the best performing action at every training step and collecting the reward Q(at). 5 The Representational Capacity of Action-Value Networks for Multi-Agent Reinforcement LearningA PREPRINT (a) (b) Figure 2: Reconstructed Q(a) for (a) the Dispersion Game, and (b) its sparse variant. 6 0.02.5Original0.02.5Joint2.02.1F12.02.1M11.52.02.5F2R2.002.25M2R12F3R1.52.0M3R0.02.5F2C1.82.02.2M2C0.02.5F3C1.52.0M3C12F2O1.82.02.2M2O01020304050600.02.5F3O01020304050601.52.0M3O0.02.5Original0.02.5Joint0.91.0F10.91.0M10.51.0F2R0.751.00M2R01F3R0.51.0M3R2101F2C0.751.00M2C2101F3C0.51.0M3C01F2O0.751.00M2O0102030405060101F3O01020304050600.51.0M3OThe Representational Capacity of Action-Value Networks for Multi-Agent Reinforcement LearningA PREPRINT 4.2.2 Platonia Dilemma In the Platonia Dilemma, an eccentric trillionaire gathers 20 people together and tells them that if one and only one of them sends him a telegram by noon the next day, that person will receive a billion dollars. In our cooperative version the reward is set to the number of agents n and is received by the whole team, not just a single agent. Thus, the reward function for n agents with local action sets Ai = {send,idle} is: (cid:26)n if #send = 1, 0 otherwise. Q(a) = (9) Figure 3: Reconstructed Q(a) for the Platonia Dilemma. Figure 3 shows the reconstructed action-value functions for the Platonia Dilemma. For this problem, none of the proposed factorizations can correctly represent the action-value function. In fact, while they are perfectly able to correctly rank all the optimal actions (the ones in which only a single agent sends the telegram) at the same level, they all fail to correctly rank and reconstruct the same joint action (the one in which none of the agents sends the telegram). In fact, the unique symmetric equilibrium for the team in this game is that each of them sends the telegram with probability 1 n, so the agents usually gather more reward by not sending it themselves, but relying on someone else to do so. This results in an 'imbalanced' reward function in which the high reward is more often obtained, from an agent perspective, by choosing a certain action instead of the other, thus resulting in overestimating one of the actions (the one in which all the agents perform the same action, i.e., not sending the telegram). This imbalance in the reward given by the two actions is probably the cause of the poor reconstruction. Thus, for this kind of tightly coupled coordination problem, none of the techniques to approximate action-values currently employed in deep MARL suffice to guarantee a good action is taken, even if the coordination problem is conceptually simple. 4.2.3 Climb Game In the Climb Game [31], each agent has three local actions Ai = {a0,a1,a2}. Action a0 yields a high reward if all the agents choose it, but no reward if only some do. The other two are suboptimal actions that give lower reward but do not require precise coordination. This game enforces a phenomenon called relative overgeneralization, that pushes the agents to underestimate a certain action (in our example, a0) because of the low rewards they receive, while they could get a higher reward by perfectly coordinating on it. The reward function Q(a) is: (10) n if #a0 = n, if 0 < #a0 < n, otherwise. 0 n 2 Q(a) = 7 05Original05Joint0.02.5F10.51.0M10.02.5F2R01M2R0.02.5F3R02M3R05F2C01M2C05F3C02M3C0.02.5F2O01M2O010203040506005F3O010203040506002M3OThe Representational Capacity of Action-Value Networks for Multi-Agent Reinforcement LearningA PREPRINT (a) (b) Figure 4: Reconstructed Q(a) for the Climb Game (a) factored Q function learning approach, and (b) the mixture of experts learning approach. 8 05Original02Joint101F101F2R01F3R0.02.5F2C2.50.02.5F3C01F2O010020030040050060070002F3O05Original02Joint0.000.25M10.00.5M2R0.00.5M3R0.00.5M2C0.00.5M3C0.00.5M2O01002003004005006007000.00.5M3OThe Representational Capacity of Action-Value Networks for Multi-Agent Reinforcement LearningA PREPRINT Figure 4 shows the results obtained on the proposed Climb Game. The joint network is not able to learn the correct action-value function in the given training time, due to the large number of joint actions. This highlights again how joint learners are not suited for this kind of even moderately large multi-agent system. By contrast, all the other architectures are able to correctly rank the suboptimal actions. The ones using the factored Q-function and a complete factorization are also able to correctly reconstruct the values for those actions, as can be seen from the bars. However, only F2C can correctly rank and reconstruct the optimal action (the coordinated one), while even F3C fails to do so and gives it a large negative value. A likely cause for this effect is that, when optimizing the loss function, assigning negative values to the components forming that joint action reduces the overall error, even if one of the reconstructed reward value is totally wrong. We can also observe how the mixture of experts plot looks somewhat comparable to the one for factored Q-functions, but more 'compressed' and noisy. 4.2.4 Penalty Game Similarly to the Climb Game, in the Penalty Game [31] each agent has three local actions Ai = {a0,a1,a2}. In this game, two local actions (for example, action a0 and a2) give a high reward if the agents perfectly coordinate on one of them, but also give a negative penalty if they mix them together. The third action a1 is suboptimal and gives a lower reward when the team coordinates on it, but also no penalty if at least one of the agents uses it. This game could also lead to relative overgeneralization, as the suboptimal action gives a higher reward on average. We use the following reward function: (11)  Q(a) = if #a0 = n, or #a2 = n, n if #a1 = n, n 2 if 0 < #a1 < n, 0 −n otherwise. Figure 5 presents the representations obtained by the investigated architectures. Given the high level of coordination required, all of the architectures using the mixture of experts learn a totally incorrect approximation, biased by the larger number of joint actions that yield a penalty rather than a positive reward. For this game, none of the architec- tures can correctly reconstruct the whole structure of the action-value function, but they all fail at the two optimal joint actions (at the two sides of the bar plots). This is probably due to the large gap in the reward values that the agents can receive when choosing one of their local optimal actions: they can get a high reward if all the agents perfectly coordinate, but it is more common for them to miscoordinate and receive a negative penalty, resulting in an approx- imation that ranks those two joint actions as bad in order to correctly reconstruct the other cases. Furthermore, the suboptimal action is hard to correctly approximate because, similarly to the optimal ones, it also usually results in a smaller reward than the one it gives when all the agents coordinate on it. Only F1 and F3C rank it as better than the other, but surprisingly only F1 is also able to reconstruct the correct value. 4.3 Factored Games 4.3.1 Generalized Firefighting The Generalized Firefighting problem [22] is an extension of the standard two-agent firefighting problem with n agents. This is a cooperative graphical Bayesian game, so each agent i has some private information, called its local type θi, on which it can condition its decision. The combination of the various agents types θ = (cid:104)θ1, . . . ,θn(cid:105) determines the values of the reward function Q(a,θ). We have a team of n firefighters that have to fight possible fires at Nh different houses. Each house j can be burning, Fj, or not, Nj. Each agent i has a limited observation and action field: it can observe only No houses (so its local type is θi ∈ {Fj,Nj}No) and can fight the fire only at Na houses (the sets of the observed and reachable houses are fixed beforehand and are part of the problem specification, with No and Na being their cardinality respectively). Each house yields a reward component: if an agent fights the fire at a burning house, that house gives a positive reward q1 = 2; if the house is not burning (or if it is burning but no-one is fighting the fire at it) it does not provide any reward. The reward function is sub-additive: if two agents fight the fire at the same burning house, this gives a reward q2 = 3 < 2q1. The overall value of the reward function Q(a,θ) experienced by agents for a given joint type θ and joint action a is the sum of the rewards given by each house. In our experiments, a team of n = 6 agents have to fight fire at Nh = 7 houses. Each agent can observe No = 2 houses and can fight fires at the same set of locations (Na = 2). Figure 6 shows the representations learned for the joint type θ = {N1,F2,N3,F4,N5,F6,N7}. This game requires less coordination than those studied earlier (agents only have to coordinate with other agents that can fight fire at the same locations), and every investigated architecture correctly ranks all the joint actions, even the single agent factorizations F1 and M1 (this is true also for any other joint type, for which we are not reporting the 9 The Representational Capacity of Action-Value Networks for Multi-Agent Reinforcement LearningA PREPRINT (a) (b) Figure 5: Reconstructed Q(a) for the Penalty Game (a) factored Q function learning approach, and (b) the mixture of experts learning approach. 10 505Original20Joint0.02.5F12.50.0F2R2.50.0F3R5.02.50.0F2C505F3C2.50.0F2O01002003004005006007002.50.0F3O505Original20Joint0.50.0M110M2R10M3R10M2C10M3C10M2O010020030040050060070010M3OThe Representational Capacity of Action-Value Networks for Multi-Agent Reinforcement LearningA PREPRINT Figure 6: Reconstructed Q(a) for a single joint type of the Generalized Firefighting game. 11 05Original05Joint2.55.07.5F14.55.05.5M12.55.07.5F2R46M2R2.55.07.5F3R2.55.0M3R05F2C46M2C05F3C2.55.0M3C2.55.07.5F2O46M2O01020304050600.02.55.07.5F3O01020304050602.55.0M3OThe Representational Capacity of Action-Value Networks for Multi-Agent Reinforcement LearningA PREPRINT plots, with just the single agent factorizations reporting some isolated error). However, while those using the factored Q-function can also correctly reconstruct the reward value of each action, those using the mixture of experts are less precise in their reconstruction. Overall, this experiment demonstrates that there exist non-trivial coordination problems that can effectively be tackled using small factors, including even individual learning approaches. 4.3.2 Aloha Aloha [21] is a partially observable game in which there is a set of islands, each provided with a radio station, trying to send a message to their inhabitants. We present a slightly altered one-shot version in which the ruler of each island wants to send a radio message to its inhabitants, but, given that some of the islands are near one to another, if they all send the message the radio frequencies interfere and the messages are not correctly received by their populations. Given that all the rulers are living in peace and they want to maximize the number of received messages by their populations, the reward signal is shared and thus the game is cooperative. It is a graphical game, as the result of each island transmission is affected only by the transmissions of nearby islands. Every ruler has two possible actions: send a message or not. If they do not send a message, they do not contribute to the total reward. If they send one and the message is correctly received by the population (no interference occurs) they get a reward q1 = 2, but if they are interfering with someone else, they get a penalty of q2 = −1. The common reward that all the rulers receive at the end is the sum of their local contributions. Our experiment uses a set of n = 6 islands disposed in a 2 × 3 grid, with each island being affected only by the transmissions of the islands on their sides and in front of them (islands on the corner of the grid miss one of their side neighbours). Representations learned for this game are reported in Figure 7. Figure 7: Reconstructed Q(a) for Aloha. The plot shows clearly how this game is challenging for the proposed factorizations to learn, with only three of them (plus the joint learner) able to correctly represent the reward function. The structure of the game is similar to that of Generalized Firefighting, with an agent depending directly only on a small subset of the others, but the different properties of the reward function make it more challenging to correctly represent. This is possibly due to the large difference between the two rewards an agent can get when transmitting the radio message, depending on an eventual interference. Observing only the total reward, this action looks neutral per se, similarly to what happens for the two actions in the dispersion game, its outcome depending on the action of the neighbouring agents, thus possibly fooling many of the proposed factorizations, especially those using the mixture of experts approach. 4.4 Summary of Results Table 3 presents the accuracy using various measures of the investigated representations, both in terms of reconstruc- tion error and action ranking, as well as evaluating the action selection that these representations result in. To evaluate 12 505Original505Joint50F121M150F2R2.50.0M2R50F3R2.50.0M3R505F2C2.50.0M2C505F3C2.50.0M3C50F2O2.50.0M2O010203040506050F3O01020304050602.50.0M3OThe Representational Capacity of Action-Value Networks for Multi-Agent Reinforcement LearningA PREPRINT Model Mean square error MSE on optimal actions Optimal actions found Value loss Boltzmann value loss Correctly ranked Kendall τ Dispersion game (64 joint actions, 20 optimal) 0.00± 0.0 1.70± 1.0 0.00± 0.0 0.60± 0.5 0.00± 0.0 0.00± 0.0 0.60± 0.5 0.20± 0.4 1.30± 0.8 0.00± 0.0 0.40± 0.5 0.00± 0.0 0.00± 0.0 0.50± 0.5 0.30± 0.5 0.00± 0.0 6.00± 0.0 6.00± 0.0 6.00± 0.0 6.00± 0.0 6.00± 0.0 6.00± 0.0 6.00± 0.0 6.00± 0.0 6.00± 0.0 6.00± 0.0 6.00± 0.0 6.00± 0.0 6.00± 0.0 6.00± 0.0 13 64± 0 21± 2 24± 1 31± 3 64± 0 64± 0 36± 2 47± 3 24± 2 24± 1 31± 2 64± 0 64± 0 36± 4 40± 3 64± 0 62± 0 62± 0 62± 0 62± 0 62± 0 62± 1 62± 0 62± 0 62± 0 62± 0 62± 0 62± 0 61± 1 61± 1 727± 1 726± 0 726± 0 726± 0 729± 0 726± 0 726± 0 726± 0 726± 0 726± 0 726± 0 726± 0 726± 0 726± 0 726± 0 727± 1 722± 0 723± 0 723± 0 722± 0 724± 0 723± 0 723± 0 1.00± 0.0 0.00± 0.0 0.27± 0.0 0.39± 0.0 1.00± 0.0 1.00± 0.0 0.44± 0.0 0.70± 0.0 0.02± 0.0 0.27± 0.0 0.39± 0.0 1.00± 0.0 1.00± 0.0 0.46± 0.0 0.58± 0.1 1.00± 0.0 0.82± 0.0 0.82± 0.0 0.82± 0.0 0.82± 0.0 0.82± 0.0 0.78± 0.1 0.82± 0.0 0.82± 0.0 0.82± 0.0 0.82± 0.0 0.82± 0.0 0.82± 0.0 0.69± 0.1 0.71± 0.1 1.00± 0.0 0.98± 0.0 0.98± 0.0 0.98± 0.0 1.00± 0.0 0.98± 0.0 0.98± 0.0 0.98± 0.0 0.98± 0.0 0.98± 0.0 0.98± 0.0 0.98± 0.0 0.98± 0.0 0.98± 0.0 0.98± 0.0 1.00± 0.0 0.91± 0.0 0.92± 0.0 0.94± 0.0 0.92± 0.0 0.97± 0.0 0.95± 0.0 0.95± 0.0 Platonia dilemma (64 joint actions, 6 optimal) 0.00± 0.0 0.49± 0.0 0.38± 0.0 0.31± 0.0 0.16± 0.0 0.16± 0.0 0.32± 0.0 0.20± 0.0 0.49± 0.0 0.46± 0.0 0.39± 0.0 0.46± 0.0 0.40± 0.0 0.46± 0.0 0.40± 0.0 0.00± 0.0 0.00± 0.0 20± 0 Joint 0.62± 0.0 0.88± 0.0 5± 2 F1 0.52± 0.0 0.82± 0.0 8± 0 F2R 0.41± 0.0 0.72± 0.0 10± 1 F3R 0.09± 0.0 0.14± 0.0 20± 0 F2C 0.09± 0.0 0.14± 0.0 20± 0 F3C F2O 0.41± 0.0 0.64± 0.0 10± 1 F3O 0.19± 0.0 0.30± 0.0 13± 1 0.62± 0.0 0.88± 0.0 6± 1 M1 M2R 0.56± 0.0 0.82± 0.0 8± 0 M3R 0.46± 0.0 0.73± 0.0 10± 1 M2C 0.55± 0.0 0.81± 0.0 20± 0 M3C 0.43± 0.0 0.68± 0.0 20± 0 M2O 0.56± 0.0 0.81± 0.0 10± 2 M3O 0.44± 0.0 0.69± 0.0 11± 1 0.00± 0.0 0.03± 0.1 6± 0 Joint 2.22± 0.0 15.55± 0.1 5± 0 F1 2.11± 0.0 14.17± 0.1 5± 0 F2R 2.00± 0.0 12.90± 0.1 5± 0 F3R 1.69± 0.0 8.92± 0.1 5± 0 F2C 1.69± 0.0 8.97± 0.1 5± 0 F3C F2O 2.00± 0.0 12.83± 0.1 5± 0 F3O 1.78± 0.0 10.10± 0.4 5± 0 2.80± 0.0 27.01± 0.0 5± 0 M1 M2R 2.53± 0.0 23.93± 0.0 5± 0 M3R 2.28± 0.0 20.51± 0.1 5± 0 M2C 2.52± 0.0 23.90± 0.0 5± 0 M3C 2.25± 0.0 20.54± 0.0 5± 0 M2O 2.54± 0.0 23.93± 0.1 4± 0 M3O 2.28± 0.0 20.60± 0.1 4± 0 0.01± 0.0 4.19± 0.0 4.09± 0.0 4.04± 0.0 4.39± 0.0 4.52± 0.0 3.94± 0.0 4.15± 0.1 5.15± 0.0 4.93± 0.0 4.64± 0.0 4.92± 0.0 4.62± 0.0 4.92± 0.0 4.61± 0.0 Climb game (729 joint actions, 1 optimal) 1.52± 0.3 2.16± 0.0 2.06± 0.0 1.92± 0.0 1.40± 0.0 0.96± 0.0 1.94± 0.0 1.54± 0.0 2.36± 0.0 2.30± 0.0 2.20± 0.0 2.30± 0.0 2.20± 0.0 2.30± 0.0 2.20± 0.0 Penalty game (729 joint actions, 2 optimal) 3.24± 0.1 3.31± 0.0 3.33± 0.0 3.31± 0.0 3.30± 0.0 2.06± 0.0 3.32± 0.0 3.29± 0.0 2.70± 0.9 3.00± 0.0 3.00± 0.0 3.00± 0.0 0.00± 0.0 3.00± 0.0 3.00± 0.0 3.00± 0.0 3.00± 0.0 3.00± 0.0 3.00± 0.0 3.00± 0.0 3.00± 0.0 3.00± 0.0 3.00± 0.0 0.90± 1.4 3.00± 0.0 3.00± 0.0 3.00± 0.0 6.00± 0.0 3.00± 0.0 6.00± 0.0 6.00± 0.0 0.17± 0.1 18.45± 4.9 0± 0 Joint 0.58± 0.0 52.29± 0.1 0± 0 F1 0.52± 0.0 40.95± 0.0 0± 0 F2R 0.44± 0.0 36.51± 0.2 0± 0 F3R 0.25± 0.0 7.86± 0.1 1± 0 F2C 0.17± 0.0 70.77± 0.7 0± 0 F3C F2O 0.45± 0.0 30.83± 0.1 0± 0 F3O 0.30± 0.0 28.89± 1.9 0± 0 0.71± 0.0 35.91± 0.0 0± 0 M1 M2R 0.63± 0.0 35.77± 0.0 0± 0 M3R 0.53± 0.0 35.34± 0.1 0± 0 M2C 0.62± 0.0 35.77± 0.0 0± 0 M3C 0.51± 0.0 35.30± 0.1 0± 0 M2O 0.63± 0.0 35.74± 0.0 0± 0 M3O 0.52± 0.0 35.31± 0.1 0± 0 1.60± 0.4 18.21± 4.3 1± 0 Joint 2.18± 0.0 63.71± 0.1 0± 0 F1 2.00± 0.0 65.95± 0.3 0± 0 F2R 1.75± 0.0 66.27± 1.0 0± 0 F3R 1.29± 0.0 79.66± 0.0 0± 0 F2C 0.54± 0.0 82.77± 0.0 0± 0 F3C F2O 1.82± 0.0 68.81± 0.3 0± 0 F3O 1.27± 0.0 73.48± 1.4 0± 0 The Representational Capacity of Action-Value Networks for Multi-Agent Reinforcement LearningA PREPRINT 2.71± 0.0 45.27± 0.1 0± 0 M1 M2R 2.43± 0.0 49.11± 0.2 0± 0 M3R 2.09± 0.0 52.23± 0.8 0± 0 M2C 2.41± 0.0 49.12± 0.1 0± 0 M3C 2.02± 0.0 52.45± 0.2 0± 0 M2O 2.43± 0.0 49.11± 0.1 0± 0 M3O 2.06± 0.0 52.56± 0.5 0± 0 3.00± 0.0 3.00± 0.0 3.00± 0.0 3.00± 0.0 3.00± 0.0 3.00± 0.0 3.00± 0.0 3.66± 0.0 3.54± 0.0 3.43± 0.0 3.54± 0.0 3.43± 0.0 3.54± 0.0 3.43± 0.0 722± 0 723± 0 723± 0 724± 0 724± 0 724± 0 723± 2 0.91± 0.0 0.93± 0.0 0.94± 0.0 0.97± 0.0 0.97± 0.0 0.97± 0.0 0.96± 0.0 Generalized firefighting (8192 joint actions, 779 optimal) 1.29± 2.5 4.96± 7.2 656± 12361.60± 68.2 46.42± 48.4 6,893± 1,475 0.85± 0.2 Joint 6,236± 38 6.38± 0.1 0.16± 0.0 0.20± 0.0 700± 7 26.20± 7.3 0.88± 0.0 F1 6,777± 383 0.91± 0.0 4.78± 1.4 0.12± 0.0 0.15± 0.0 722± 19 16.80± 8.5 F2R 7,288± 558 0.94± 0.0 0.09± 0.0 0.11± 0.1 743± 25 11.10± 10.2 3.42± 1.7 F3R 1.00± 0.0 8,192± 0 0.00± 0.0 0.00± 0.0 0.00± 0.0 0.00± 0.0 779± 0 F2C 8,192± 0 0.00± 0.0 0.00± 0.0 779± 0 0.00± 0.0 1.00± 0.0 0.00± 0.0 F3C 7,333± 382 F2O 0.09± 0.0 0.10± 0.0 747± 18 8.00± 7.1 3.75± 1.1 0.95± 0.0 1.00± 0.0 1.14± 0.9 F3O 0.03± 0.0 0.03± 0.0 778± 4 8,149± 130 0.20± 0.6 0.88± 0.0 163.84± 0.1 6,220± 30 3.55± 0.0 8.35± 0.0 700± 6 27.80± 6.4 M1 0.90± 0.0 124.61± 0.6 6,602± 301 M2R 1.85± 0.1 4.92± 0.2 718± 12 20.60± 5.4 M3R 1.09± 0.2 2.81± 0.2 739± 33 14.60± 13.8 88.58± 2.3 7,097± 703 0.93± 0.0 M2C 1.82± 0.0 4.90± 0.0 777± 0 0.00± 0.0 124.39± 0.1 7,826± 0 0.97± 0.0 1.00± 0.0 8,151± 4 88.09± 0.1 0.00± 0.0 M3C 0.85± 0.0 2.58± 0.0 778± 1 M2O 1.97± 0.1 5.08± 0.1 741± 13 11.70± 5.0 127.21± 1.9 5,628± 249 0.84± 0.0 M3O 1.03± 0.1 2.76± 0.2 774± 4 0.50± 1.0 90.31± 3.8 6,220± 382 0.88± 0.0 Aloha (64 joint actions, 2 optimal) 1.13± 0.0 0.00± 0.0 2± 0 Joint 4.78± 0.0 50.93± 0.1 0± 0 F1 4.05± 0.4 35.00± 7.0 0± 0 F2R 3.16± 0.5 20.64± 4.6 0± 0 F3R 0.91± 0.0 0.14± 0.0 2± 0 F2C 2± 0 0.07± 0.0 0.14± 0.0 F3C F2O 3.27± 0.3 20.63± 3.0 0± 0 F3O 1.46± 0.3 3.55± 1.3 1± 1 8.26± 0.0 50.84± 0.1 0± 0 M1 M2R 6.52± 0.2 44.17± 1.5 0± 0 M3R 4.53± 0.5 31.35± 3.8 0± 0 M2C 6.51± 0.0 44.65± 0.1 0± 0 M3C 4.56± 0.0 33.45± 0.1 0± 0 M2O 6.63± 0.4 44.51± 1.1 0± 0 M3O 4.71± 0.3 33.36± 1.2 0± 0 0.08± 0.0 0.00± 0.0 4.04± 0.0 6.00± 0.0 5.00± 1.3 3.69± 0.4 3.23± 0.9 4.20± 1.4 0.00± 0.0 −0.04± 0.0 0.22± 0.0 0.00± 0.0 3.24± 0.5 4.40± 1.2 1.19± 0.4 0.80± 1.3 6.00± 0.0 5.47± 0.0 5.00± 1.3 4.57± 0.1 3.41± 0.6 4.00± 2.2 4.59± 0.0 6.00± 0.0 3.62± 0.0 6.00± 0.0 5.20± 1.0 4.62± 0.2 5.20± 1.0 3.65± 0.2 51± 1 27± 1 22± 4 26± 4 42± 0 64± 0 23± 4 29± 5 27± 1 25± 4 28± 5 28± 0 36± 1 22± 5 25± 4 0.88± 0.0 0.67± 0.0 0.70± 0.0 0.74± 0.0 0.89± 0.0 1.00± 0.0 0.74± 0.0 0.83± 0.0 0.67± 0.0 0.70± 0.0 0.77± 0.1 0.76± 0.0 0.86± 0.0 0.66± 0.0 0.74± 0.0 Table 3: Accuracy results for the different games. 14 The Representational Capacity of Action-Value Networks for Multi-Agent Reinforcement LearningA PREPRINT the reconstruction error, we compute the mean square error over all the joint actions and the same measure restricted only to those actions that are optimal in the original reward function. We also assess how many of the optimal actions are considered optimal also by the reconstructions, and compute the value loss (regret) obtained by following the represented value functions. We also provide a different version, that we call Boltzmann value loss, which expresses the value loss obtained by the expected reward accrued by defining a softmax distribution over all the joint actions (this gives an indication of value loss amongst all good actions). Finally, we compute the number of correctly ranked actions (accounting for ties were needed) and the corresponding Kendall τ-b coefficient [13] between the computed ranking and the original one. For every method, mean values and standard errors across 10 runs are reported. While many aspects can influence the learning outcome, our results have four main takeaways: • There are pathological examples, like the Platonia Dilemma, where all types of factorization result in selecting the worst possible joint action. Given that only joint learners seem to be able to address such problems, currently no scalable deep RL methods for for dealing with such problems seem to exist. • Beyond those pathological examples, 'complete factorizations' of modest factor size coupled with the fac- tored Q-function learning approach yield near-perfect reconstructions and rankings of the actions, also for non-factored action-value functions. Moreover, these methods scale much better than joint learners: for a given training time, we say that these complete factorizations already outperform fully joint learners on mod- estly sized problems (like the Climb Game or the Generalized Firefighting with 6 agent), as can be seen from the training curves in Figure 8 for two of the proposed games. (a) (b) Figure 8: Training curves for the investigated models for (a) the Dispersion Game, and (b) the Generalized Firefighting game. • For these more benign problems, random overlapping factors also achieve excellent performance, especially in terms of value loss, comparable to those of more computationally complex methods like joint learners and complete factorizations. This suggests that such approaches are a promising direction forward for scalable deep MARL in many problem settings. • Factorizations with the mixture of experts approach usually perform somewhat worse than the corresponding factored Q-function approaches. However, in some cases they perform better (e.g. in terms of value loss and MSE on optimal actions in the Penalty Game) or comparable (Dispersion, Generalized Firefighting), in which M2R and M3R still outperform F1 (i.e., VDNs) in terms of value loss. This is promising, because the mixture of experts learning approach does not require any exchange of information between the neural networks, thus potentially facilitating learning in settings with communication constraints, and making it easier to parallelize across on multiple CPUs/GPUs. These observations shed light on the performance of independent learners in MARL: while they can outperform joint learners on large problems, the degree of independence and the final outcome is hard to predict and is affected by different factors. Designing algorithms that are able to overcome these difficulties should be a primary focus of MARL research. 5 Related Work Recently, many works have applied deep RL techniques to MASs, achieving great performance. Gupta et al. [12] com- pare the performance of many standard deep reinforcement learning algorithms (like DQN, DDPG and TRPO) using 15 The Representational Capacity of Action-Value Networks for Multi-Agent Reinforcement LearningA PREPRINT a variety of learning schemes (joint learners, fully independent learners, etc.) on both discrete and continuous tasks, assessing and comparing their performance. Tampuu et al. [28] present a variation on DQN capable of dealing with both competitive and cooperative settings in Pong. Applications of techniques to enhance the learning process have also been investigated: Palmer et al. [23] apply leniency to independent double DQN learners in a coordinated grid world problem, while Foerster et al. [5] propose a novel approach to stabilize the experience replay buffer in DQNs by conditioning on when the samples were collected, thereby easing the non-stationarity of independent learning. Communication between agents has also been explored: Sukhbaatar et al. [26] investigate the emergence of a com- munication mechanism that can be directly learned through backpropagation. However, none of these works compare alternate representations of Q-values for such networks. Many works address the problem of the exponentially large joint action space by exploiting centralized learning. Foerster et al. [6] present an architecture based on the actor-critic framework with multiple independent actors but a single centralized critic used to efficiently estimate both Q-values and a counterfactual baseline to tackle the credit assignment problem and guide the agents through the learning process under partial observability. However, this work still represents Q(s,a) monolithically, and thus can experience scalability issues. On the other hand, Lowe et al. [17] maintain a different critic network for every actor, together with an inferred policy for the other agents in the environment, and apply the approach both on cooperative and competitive task. Sunehag et al. [27] address the problem by training the agents independently by using a value decomposition method that represents the original Q- function as the sum of local terms depending only on agent-wise information, while Rashid et al. extend this idea by representing the Q-function using a monotonic nonlinear combination with a mixing network on top of such individual terms, so that the maximization step can still be performed in an efficient way. While such mixing networks may lead to more accurate Q-values, our investigation shows that for many coordination problems, individual Q-components may not suffice. 6 Conclusions In this work, we investigated how well neural networks can represent action-value functions arising in multi-agent systems. This is an important question since accurate representations can enable taking (near-) optimal actions in value-based approaches, and computing good gradient estimates in actor-critic methods. In this paper, we focused on one-shot games as the simplest setting that captures the exponentially large joint action space of MASs. We compared a number of existing and new action-value network factorizations and learning approaches. Our results highlight the difficulty of compactly representing action values in problems that require tight coordination, but indicate that using higher-order factorizations with multiple agents in each factor can improve the accuracy of these representations substantially. We also demonstrate that there are non-trivial coordination problems - some without a factored structure - that can be tackled quite well with simpler factorizations. Intriguingly, incomplete, overlapping factors perform very well in several settings. There are also settings where the mixtures of experts approach, with its low communication requirements and amenability to parallelization, is competitive in terms of the reconstructions. While our results emphasize the dependence of appropriate architectural choices on the problem at hand, our analysis also shows general trends that can help in the design of novel algorithms and improve general performance, highlight- ing how the use of factored action-value function can be a viable way to obtain good representations without incurring in excessive costs. Acknowledgements This research made use of a GPU donated by NVIDIA. F.A.O. is funded by EPSRC First Grant EP/R001227/1. This project had received funding from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme (grant agreement No. 758824 -- INFLUENCE). References [1] Christopher Amato and Frans A. Oliehoek. Scalable planning and learning for multiagent pomdps. In Proceed- ings of the 29th AAAI Conference on Artificial Intelligence, AAAI'15, pages 1995 -- 2002. American Association for Artificial Intelligence, 2015. 16 The Representational Capacity of Action-Value Networks for Multi-Agent Reinforcement LearningA PREPRINT [2] Lucian Busoniu, Robert Babuska, and Bart De Schutter. A comprehensive survey of multiagent reinforcement learning. IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews), 38:156 -- 172, 2008. [3] Caroline Claus and Craig Boutilier. The dynamics of reinforcement learning in cooperative multiagent systems. In Proceedings of the 15th/10th AAAI Conference on Artificial Intelligence/Innovative Applications of Artificial Intelligence, AAAI'98/IAAI'98, pages 746 -- 752. American Association for Artificial Intelligence, 1998. [4] Jakob Foerster, Ioannis Alexandros Assael, Nando de Freitas, and Shimon Whiteson. Learning to communi- cate with deep multi-agent reinforcement learning. In Advances in Neural Information Processing Systems 29, NIPS'16, pages 2137 -- 2145. Curran Associates, Inc., 2016. [5] Jakob N. Foerster, Nantas Nardelli, Gregory Farquhar, Philip H. S. Torr, Pushmeet Kohli, and Shimon Whiteson. Stabilising experience replay for deep multi-agent reinforcement learning. In Proceedings of the 34th Interna- tional Conference on Machine Learning, ICML'17, pages 1146 -- 1155. PMLR, 2017. [6] Jakob N. Foerster, Gregory Farquhar, Triantafyllos Afouras, Nantas Nardelli, and Shimon Whiteson. Coun- terfactual multi-agent policy gradients. In Proceedings of the 32th AAAI Conference on Artificial Intelligence, AAAI'18, pages 2974 -- 2982. American Association for Artificial Intelligence, 2018. [7] Mohammad Ghavamzadeh, Sridhar Mahadevan, and Rajbala Makar. Hierarchical multi-agent reinforcement learning. Autonomous Agents and Multi-Agent Systems, 13(2):197 -- 229, 2006. [8] Trond Grenager, Rob A. Powers, and Yoav Shoham. Dispersion games: General definitions and some specific learning results. In Proceedings of the 18th AAAI Conference on Artificial Intelligence, AAAI'02, pages 398 -- 403. American Association for Artificial Intelligence, 2002. [9] Carlos Guestrin, Daphne Koller, and Ronald Parr. Multiagent planning with factored mdps. In Advances in Neural Information Processing Systems 14, NIPS'02, pages 1523 -- 1530. Morgan Kaufmann Publishers Inc., 2002. [10] Carlos Guestrin, Michail G. Lagoudakis, and Ronald Parr. Coordinated reinforcement learning. In Proceed- ings of the 19th International Conference on Machine Learning, ICML'02, pages 227 -- 234. Morgan Kaufmann Publishers Inc., 2002. [11] Carlos Guestrin, Daphne Koller, Ronald Parr, and Shobha Venkataraman. Efficient solution algorithms for fac- tored mdps. Journal of Artificial Intelligence Research, 19(1):399 -- 468, 2003. [12] Jayesh K. Gupta, Maxim Egorov, and Mykel Kochenderfer. Cooperative multi-agent control using deep re- In Autonomous Agents and Multi-Agent Systems, pages 66 -- 83. Springer International inforcement learning. Publishing, 2017. [13] Maurice Kendall and Jean D. Gibbons. Rank Correlation Methods. A Charles Griffin Title, 5 edition, 1990. [14] Jelle R. Kok and Nikos Vlassis. Collaborative multiagent reinforcement learning by payoff propagation. Journal of Machine Learning Research, 7:1789 -- 1828, 2006. [15] Joel Z. Leibo, Vinicius Zambaldi, Marc Lanctot, Janusz Marecki, and Thore Graepel. Multi-agent reinforcement In Proceedings of the 16th International Conference on Autonomous learning in sequential social dilemmas. Agents and MultiAgent Systems, AAMAS'17, pages 464 -- 473. International Foundation for Autonomous Agents and Multiagent Systems, 2017. [16] Timothy P. Lillicrap, Jonathan J. Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. CoRR, abs/1509.02971, 2015. [17] Ryan Lowe, Yi Wu, Aviv Tamar, Jean Harb, Pieter Abbeel, and Igor Mordatch. Multi-agent actor-critic for mixed In Advances in Neural Information Processing Systems 30, NIPS'17, cooperative-competitive environments. pages 6379 -- 6390. Curran Associates, Inc., 2017. [18] Laetitia Matignon, Guillaume J. Laurent, and Nadine Le Fort-Piat. Independent reinforcement learners in coop- erative markov games: a survey regarding coordination problems. Knowledge Engineering Review, 27(1):1 -- 31, 2012. [19] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Bellemare, Alex Graves, Martin Riedmiller, Andreas K. Fidjeland, Georg Ostrovski, Stig Petersen, Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Kumaran, Daan Wierstra, Shane Legg, and Demis Hassabis. Human- level control through deep reinforcement learning. Nature, 518(7540):529 -- 533, 2015. [20] Volodymyr Mnih, Adri`a Puigdom`enech Badia, Mehdi Mirza, Alex Graves, Tim Harley, Timothy P. Lillicrap, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In Proceedings of the 33rd International Conference on Machine Learning, volume 48 of ICML'16, pages 1928 -- 1937. PMLR, 2016. 17 The Representational Capacity of Action-Value Networks for Multi-Agent Reinforcement LearningA PREPRINT [21] Frans A. Oliehoek. Value-Based Planning for Teams of Agents in Stochastic Partially Observable Environments. PhD thesis, Informatics Institute, University of Amsterdam, 2010. [22] Frans A. Oliehoek, Shimon Whiteson, and Matthijs T. J. Spaan. Exploiting agent and type independence in collaborative graphical bayesian games, 2011. [23] Gregory Palmer, Karl Tuyls, Daan Bloembergen, and Rahul Savani. Lenient multi-agent deep reinforcement learning. In Proceedings of the 17th International Conference on Autonomous Agents and MultiAgent Systems, AAMAS'18, pages 443 -- 451. International Foundation for Autonomous Agents and Multiagent Systems, 2018. [24] Tabish Rashid, Mikayel Samvelyan, Christian Schroder de Witt, Gregory Farquhar, Jakob N. Foerster, and Shi- mon Whiteson. Qmix: Monotonic value function factorisation for deep multi-agent reinforcement learning. In Proceedings of the 35th International Conference on Machine Learning, ICML'18, pages 4292 -- 4301. JMLR.org, 2018. [25] A. Rogers, A. Farinelli, R. Stranders, and N. R. Jennings. Bounded approximate decentralised coordination via the max-sum algorithm. Artificial Intelligence, 175(2):730 -- 759, 2011. [26] Sainbayar Sukhbaatar, Arthur Szlam, and Rob Fergus. Learning multiagent communication with backpropaga- tion. In Proceedings of the 30th International Conference on Neural Information Processing Systems, NIPS'16, pages 2252 -- 2260. Curran Associates, Inc., 2016. [27] Peter Sunehag, Guy Lever, Audrunas Gruslys, Wojciech Marian Czarnecki, Vinicius Zambaldi, Max Jaderberg, Marc Lanctot, Nicolas Sonnerat, Joel Z. Leibo, Karl Tuyls, and Thore Graepel. Value-decomposition networks for cooperative multi-agent learning based on team reward. In Proceedings of the 17th International Conference on Autonomous Agents and MultiAgent Systems, AAMAS'18, pages 2085 -- 2087. International Foundation for Autonomous Agents and Multiagent Systems, 2018. [28] Ardi Tampuu, Tambet Matiisen, Dorian Kodelja, Ilya Kuzovkin, Kristjan Korjus, Juhan Aru, Jaan Aru, and Raul Vicente. Multiagent cooperation and competition with deep reinforcement learning. PLoS ONE, 12(4):1 -- 15, 2017. [29] Ming Tan. Multi-agent reinforcement learning: Independent vs. cooperative agents. In Proceedings of the 10th International Conference on Machine Learning, ICML'93, pages 330 -- 337. Morgan Kaufmann Publishers Inc., 1993. [30] Elise Van der Pol and Frans A. Oliehoek. Coordinated deep reinforcement learners for traffic light control. In NIPS'16 Workshop on Learning, Inference and Control of Multi-Agent Systems, 2016. [31] Ermo Wei and Sean Luke. Lenient learning in independent-learner stochastic cooperative games. Journal of Machine Learning Research, 17(84):1 -- 42, 2016. [32] Michael Wunder, Michael L. Littman, and Monica Babes. Classes of multiagent q-learning dynamics with In Proceedings of the 27th International Conference on Machine Learning, epsilon-greedy exploration. ICML'10, pages 1167 -- 1174. Omnipress, 2010. [33] Dayong Ye, Minjie Zhang, and Yun Yang. A multi-agent framework for packet routing in wireless sensor net- works. Sensors, 15(5):10026 -- 10047, 2015. 18
0911.2902	1	0911	2009-11-15T18:31:21	Simulation of Pedestrians Crossing a Street	[ "cs.MA" ]	The simulation of vehicular traffic as well as pedestrian dynamics meanwhile both have a decades long history. The success of this conference series, PED and others show that the interest in these topics is still strongly increasing. This contribution deals with a combination of both systems: pedestrians crossing a street. In a VISSIM simulation for varying demand jam sizes of vehicles as well as pedestrians and the travel times of the pedestrians are measured and compared. The study is considered as a study of VISSIM's con ict area functionality as such, as there is no empirical data available to use for calibration issues. Above a vehicle demand threshold the results show a non-monotonic dependence of pedestrians' travel time on pedestrian demand.	cs.MA	cs	Simulation of Pedestrians Crossing a Street Cornelia Bonisch and Tobias Kretz PTV Planung Transport Verkehr AG Stumpfstrasse 1 D-76131 Karlsruhe {Cornelia.Boenisch,Tobias.Kretz}@ptv.de October 28, 2018 Abstract The simulation of vehicular traffic as well as pedestrian dynamics meanwhile both have a decades long history. The success of this con- ference series, PED and others show that the interest in these topics is still strongly increasing. This contribution deals with a combination of both systems: pedestrians crossing a street. In a VISSIM simulation for varying demand jam sizes of vehicles as well as pedestrians and the travel times of the pedestrians are measured and compared. The study is considered as a study of VISSIM's conflict area functionality as such, as there is no empirical data available to use for calibration issues. Above a vehicle demand threshold the results show a non-monotonic dependence of pedestrians' travel time on pedestrian demand. 1 Introduction For vehicles and pedestrians alike the single mode systems have attracted first and much interest: highway traffic for vehicles [1 -- 3] and evacuations for pedestrians [4 -- 6]. These systems respectively situations are comparatively easy to handle in analytical or numerical terms and of special interest as they are most present in public awareness. During the post-war re-building period of European cities the focus of city planning was strongly set to vehicular traffic. But recognizing economic, ecologic and aesthetic needs, the needs of pedestrians and cyclists were gain- ing ground. This shift in paradigms becomes most visible in the idea of and discussion on "shared space" and cities like Copenhagen that have explicitly put pedestrians into the top priority position. In parallel to this development the rapid progress in computation hard- ware within just three decades made it possible to advance from simplified 1 vehicular traffic simulations with limited extend to large scale combined simulations of vehicular traffic, cyclists, and pedestrians. Recently VISSIM was the first professional tool to incorporate the sim- ulation of vehicles and pedestrians as well as zones of interaction between these modes of traffic [7]. The amount of work done on "interaction issues" -- be it on signalized or non-signalized crossings -- is still marginal compared to the work done in the separate fields [8 -- 15]. In this work pedestrians crossing a street with a lane for each direction are simulated in VISSIM. The pedestrians only walk one-way, there is no counterflow. For the simulation of vehicles and pedestrians VISSIM's stan- dard models are applied [5, 7, 16 -- 19]. One can expect that the underlying operational models only have a marginal influence on the results1. In each simulation 10 hours were simulated. Demand was kept constant within a simulation and varied between simulations. For details of the geometry see figure 1. Figure 1: Pedestrians are inserted with a given average frequency to the simulation on the yellow area. Travel time measurement starts as soon as they are on the magenta area. They are counted as "jam" while they are either on the blue or the magenta area. The conflict areas are depicted red and each have an area of 3.5 x 3.5 sqm. Once a pedestrian has reached the green area, he's taken out of the simulation. The distance for the travel time measurement is 26 m. The vehicle lanes stretch 500 m to both sides. (They are not shown in their entirety here.) A note on demand and input areas: the maximum density on input areas at which still pedestrians are inserted to the simulation is 5 pedestrians per square meter. Even without any vehicles at all, this value is reached for input values of slightly less than 12,000 pedestrians per hour (in this case about 1,000 pedestrians in ten hours are skipped at input). Increasing the 1It has to be considered in the construction of the conflict area that the acceleration in the Social Force Model is finite, if τ is larger than the simulation time step. 2 input further only has statistical effects on the simulation result. 2 Conflict Areas The interaction zone is modeled as a "conflict area" similar to the way vehicle-vehicle conflicts were modeled earlier in VISSIM [20]. A conflict area for vehicle-vehicle conflicts is an area, because vehicles have a width. In an abstract representation, it would be a conflict point, as two basically one- dimensional objects -- the links -- intersect. A conflict area for pedestrian- vehicle interaction is an area, because the pedestrians -- although compared to vehicles almost point-like -- can also have a transversal component in their movement compared to the main direction of motion, and the available passage width pedestrians have at crossings is usually a multiple of their body extension, which is different for vehicles, which on one lane have almost no transversal freedom. Once a conflict area is defined, the priority is given either to vehicles or to pedestrians. This study deals exclusively with conflict areas with vehicle priority, i.e. a normal part of a street, no pedestrian crossing or even signalisation. When approaching a conflict area, pedestrians calculate, if there is enough time to cross the street in time before the next vehicle arrives. But if the density of pedestrians is sufficiently high, pedestrians may be forced to slow down, or evade other pedestrians (move transversally and by this reduce the lateral speed). In this case, it may happen that they do not make it in time to the other side of the street. The information of pedestrians being on the conflict area is then given to the approaching vehicles, which in turn slow down, notwithstanding their right of way (an animation of this can be found online [7]). Pedestrians approaching from behind to the conflict area do not always base their decision to walk or to keep standing at the edge of the road on the vehicle's speed, but -- if there is at least one pedestrian on the conflict area -- they estimate, if they could overtake pedestrians on the conflict area with their desired speed. VISSIM's conflict areas have five parameters, whose values can have an effect in the situation discussed in this paper. They were not part of the investigation and thus set to equal values for all simulations. The visibility of both links was 100 m, the front and rear gap 0.5 s and the safety distance factor 1.5. The usage of these parameters for conflicts between vehicles is fully de- scribed elsewhere [7], discussing it here, would exceed the size of this con- tribution. For the investigated situation where pedestrians are crossing a vehicle link the parameter for the rear gap is the most important one. The rear gap for pedestrians has the same meaning like for vehicles, i.e. it is the 3 time gap in seconds after the pedestrian left the conflict area and before the next vehicle enters. Figure 2: A screenshot from a 3d animation of the simulation (animation online [7]). 3 Results Figure 3 shows the dependency of pedestrians' travel times on pedestrian demand for various vehicle demands. Above a vehicle demand threshold of about 700 to 800 vehicles per hour the dependencies show a local maximum. Above the maximum the flow of pedestrians is large enough that subsequent pedestrians can make a profit of vehicles having to slow down, when pre- ceding pedestrians could not cross the street in time. Figure 4 shows that the dependencies for no and large vehicle demand converge for pedestrian demand toward capacity. At pedestrian demands just below convergence the average jam length appears to be unstable, as the total simulation time is only a small multiple of the typical oscillation period (see figures 5 and 6). References [1] K. Nagel, "Particle Hopping Models and Traffic Flow Theory", Phys. Rev. E 53 (1996) 4655 -- 4672, cond-mat/9509075v1. [2] D. Chowdhury, L. Santen, and A. Schadschneider, "Statistical Physics of Vehicular Traffic and Some Related Systems", Phys. Rep. 329 (2000) 199 -- 329, cond-mat/0007053v1. [3] D. Helbing, "Traffic and related self-driven many-particle systems", Rev. Mod. Phys. 73 (2001) no. 4, 1067 -- 1141, cond-mat/0012229v2. 4 Figure 3: Dependency of pedestrians' travel times on pedestrian demand well below capacity and for different vehicle demands. Figure 4: Dependency of pedestrians' travel times on pedestrian demand well below capacity and for the two extreme vehicle demands (none and 1,800/h). Note that in effect the simulations with a demand of 12,000 and 15,000 pedestrians (and above) are the same in all relevant aspects, as the density on the input area is quickly too large (above 5 pedestrians per sqm) to insert further pedestrians. 5 Figure 5: Vehicle jam size over time for a demand of 1,800 vehicles per hour and lane and 900 (upper) or 6,000 pedestrians per hour (lower). 6 Figure 6: Upper: Vehicle jam size over time for a demand of 1,800 vehicles per hour and 12,000 pedestrians per hour. Lower: vehicle jam size time average in dependence of pedestrian demand. 7 [4] A. Schadschneider, W.W.F. Klingsch, H. Klupfel, T. Kretz, C. Rogsch, and A. Seyfried, "Evacuation Dynamics: Empirical Results, Modeling and Applications", in Meyers [22], p. 3142. arXiv:0802.1620 [physics.soc-ph]. ISBN:978-0-387-75888-6. [5] D. Helbing and A. Johansson, "Pedestrian, Crowd and Evacuation Dynamics", in Meyers [22], p. 6476. ISBN:978-0-387-75888-6. [6] A. Schadschneider, H. Klupfel, T. Kretz, C. Rogsch, and A. Seyfried, "Fundamentals of Pedestrian and Evacuation Dynamics", in Bazzan and Klugl [21], ch. VI, pp. 124 -- 154. ISBN:978-1-60566-226-8. [7] PTV, VISSIM 5.10 User Manual. PTV Planung Transport Verkehr AG, July, 2008. http://www.vissim.de/. The simulations were done using version 5.10-06. For other purposes until 5.20-04 slight modifications were applied to the conflict areas, which might change the results of a simulation of the model of this contribution. [8] J.C. Tanner, "The delay to pedestrians crossing a road", Biometrika (1951) 383 -- 392. [9] J. Hunt and J. Abduljabbar, "Crossing the road: a method of assessing pedestrian crossing difficulty", Traffic Engineering + Control 34 (1993) no. 11, 526 -- 531. [10] G. Vogts, " Uberprufung der Verkehrsqualitat von Fussgangern an signalisierten Fussgangeruberwegen", Master's thesis, Universitat Hannover, 2001. (in German). [11] R. Jiang, Q. Wu, and X. Li, "Capacity drop due to the traverse of pedestrians", Phys. Rev. E 65 (2002) no. 3 (2), 36120 -- 36120. [12] N. Lehnhoff, "Fussgangerquerverkehr an Querungsstellen." Seminar der Vereinigung der Strassenbau- und Verkehrsingenieure, 2004. (in German). [13] D. Helbing, R. Jiang, and M. Treiber, "Analytical investigation of oscillations in intersecting flows of pedestrian and vehicle traffic", Phys Rev E 72 (2005) 046130, physics/0507178. [14] M.M. Ishaque and R.B. Noland, "Trade-offs between vehicular and pedestrian traffic using micro-simulation methods", Transport Policy 14 (2007) no. 2, 124 -- 138. [15] M.M. Ishaque and R.B. Noland, "Behavioural Issues in Pedestrian Speed Choice and Street Crossing Behaviour: A Review", Transport Reviews 28 (2008) no. 1, 61 -- 85. 8 [16] R. Wiedemann, "Simulation des Strassenverkehrsflusses", Schriftenreihe des IfV 8 (1974) . Institut fur Verkehrswesen. Universitat Karlsruhe. [17] D. Helbing and P. Molnar, "Social force model for pedestrian dynamics", Phys. Rev. E 51 (1995) 4282 -- 4286, cond-mat/9805244. [18] D. Helbing, I. Farkas, and T. Vicsek, "Simulating dynamical features of escape panic", Nature 407 (2000) 487 -- 490, cond-mat/0009448. [19] A. Johansson, D. Helbing, and P.K. Shukla, "Specification of the Social Force Pedestrian Model by Evolutionary Adjustment to Video Tracking Data", Advances in Complex Systems 10 (2007) no. 4, 271 -- 288, arXiv:0810.4587 [physics.soc-ph]. [20] PTV, VISSIM 5.00 User Manual. PTV Planung Transport Verkehr AG, July, 2007. http://www.vissim.de/. [21] 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. [22] R.A. Meyers, ed., Encyclopedia of Complexity and Systems Science. Springer, Berlin Heidelberg New York, 2009. ISBN:978-0-387-75888-6. 9
1105.1564	1	1105	2011-05-09T01:00:04	Complex Adaptive Digital EcoSystems	[ "cs.MA" ]	We investigate an abstract conceptualisation of DigitalEcosystems from a computer science perspective. We then provide a conceptual framework for the cross pollination of ideas, concepts and understanding between different classes of ecosystems through the universally applicable principles of Complex Adaptive Systems (CAS) modelling. A framework to assist the cross-disciplinary collaboration of research into Digital Ecosystems, including Digital BusinessEcosystems (DBEs) and Digital Knowledge Ecosystems (DKEs). So, we have defined the key steps towards a theoretical framework for Digital Ecosystems, that is compatible with the diverse theoretical views prevalent. Therefore, a theoretical edifice that can unify the diverse efforts within Digital Ecosystems research.	cs.MA	cs	Complex Adaptive Digital EcoSystems Gerard Briscoe Intelligent Systems Lab Department of Computer Science Heriot Watt University United Kingdom [email protected] ABSTRACT We investigate an abstract conceptualisation of Digital Ecosystems from a computer science perspective. We then provide a conceptual framework for the cross pollination of ideas, concepts and understanding between different classes of ecosystems through the universally applicable principles of Complex Adaptive Systems (CAS) modelling. A framework to assist the cross-disciplinary collaboration of research into Digital Ecosystems, including Digital Business Ecosystems (DBEs) and Digital Knowledge Ecosystems (DKEs). So, we have defined the key steps towards a theoretical framework for Digital Ecosystems, that is compatible with the diverse theoretical views prevalent. Therefore, a theoretical edifice that can unify the diverse efforts within Digital Ecosystems research. Keywords Complex Adaptive Systemss, Multi-Agent Systems, Busi- ness Ecosystems, Knowledge Ecosystems 1. INTRODUCTION Conceptualising ecosystems has been an inherent part of efforts, which presents us with an opportunity to for- malise our current and future efforts to improve the cross- disciplinary knowledge transfer required. In the creation of Digital Ecosystems [5, 9, 4, 6] we con- sidered aspects of biological ecosystems, including Agent- Based Modelling (ABM) [20] and CAS [25], and then constructed their counterparts in Digital Ecosystems. After which we considered the possibility of a Generic Ecosystem definition [4], because we made use of a direct unidirectional flow of information and models from biological ecosystems to Digital Ecosystems as shown in Figure 1. Without the Generic Ecosystem concept some of the counterparts of biological ecosystems that we constructed in Digital Ecosystems appeared to be compromised, when they were actually the realisation of generic abstract concepts. Most notably the network structure, which is energy-centric in 1 1 0 2 y a M 9 ] A M . s c [ 1 v 4 6 5 1 . 5 0 1 1 : v i X r a Figure 1: Creation of Digital Ecosystems: We considered aspects of biological ecosystems, using a direct unidirectional flow of information and models from biological ecosystems to Digital Ecosystems. Figure 2: Hypothetical Abstract Ecosystem Definition: Given an abstract ecosystem class in the UML, then the Digital Ecosystem and biological ecosystem classes would both inherit from it, but implement its attributes differently. biological ecosystems [1], while information-centric in Digi- tal Ecosystems, as shown in Figure 2. So, given an abstract ecosystem class in the UML, then the Digital Ecosystem and biological ecosystem classes would both inherit from it, but implement its attributes differently. So, we argued that the apparent compromises in mimicking biological ecosystems were actually features unique to Digital Ecosystems [4]. Therefore, there is potential to create a Generic Ecosystem definition, using a suitable modelling technique such as CAS [46], which would abstractly define the key properties of an ecosystem, and would theoretically be applicable to any domain where the modelling technique has been applied. We can create a definition of a Generic Ecosystem based on CAS, making use of the ABM and Multi-Agent Systems (MASs),1 which will define the key properties of an ecosystem, in any domain, in an abstract extensible form. 1An ABM is a class of computational models for simulating the actions and interactions of autonomous agents with a view to assessing their effects on the system as a whole. It combines elements of game theory, complex systems, emergence, computational sociology, MASs, and evolutionary programming. Biological EcosystemDigital EcosystemGenericEcosystemNetworkEcosystemNetwork: EnergyCentricBiological EcosystemNetwork: InformationCentricDigital EcosystemTherefore, the Generic Ecosystem definition will provide a framework for the application of ideas, concepts, and models from one class of ecosystem to another, including Digital, Business and Knowledge Ecosystems. Naturally, biological ecosystems will be the main source of information for the conceptualisation of the Generic Ecosystem. 2. BIOLOGICAL ECOSYSTEMS In order to create a Generic Ecosystem we will consider biological ecosystems in terms of the key properties, be- haviours and structures. These were considered extensively [4, 9, 8], and so we will summarise the main findings. Ecosystems are often described as CAS, because like them, they are systems made from diverse, locally interact- ing components that are subject to selection. Other CAS include brains, individuals, economies, and the biosphere. All are characterised by hierarchical organisation, continual adaptation and novelty, and non-equilibrium dynamics. These properties lead to behaviour that is non-linear, historically contingent, subject to thresholds, and contains multiple basins of attraction [25]. The features of these systems, especially non-linearity and non-equilibrium dy- namics, offer both advantages and hazards for adaptive problem-solving. The major hazard is that the dynamics of CAS are intrinsically hard to predict because of the non- linear emergent dynamics [26]. The occurrence of multiple basins of attraction in CAS suggests that even a system that functions well for a long period may suddenly at some point transition to a less desirable state [18]. Non-linear behaviour provides the opportunity for scalable organisation and the evolution of complex hierarchical solutions, while rapid state transitions potentially allow the system to adapt to sudden environmental changes with minimal loss of functionality [25]. In creating Digital Ecosystems, the digital counterpart of biological ecosystems, we naturally asked their likeness to the biological ecosystems from which they came [4]. Further to this, we could consider the applicability of other aspects of ecosystems theory in understanding and analysing the dynamics of Digital Ecosystems. For example, energy pyramids2 of biological ecosystems, what is their equivalent 2Energy pyramids show the dissipation of energy at trophic levels, positions that organisms occupy in a food chain, e.g. producers or consumers [40]. in Digital Ecosystems? Given that Digital Ecosystems are information-centric, whereas biological ecosystems are energy-centric [1], they would undoubtedly be information pyramids, but further definition would naturally require more research. So, we can define a framework for understanding biolog- ical ecosystems, with the aim of applying that understand- ing to Digital Ecosystems, through a Generic Ecosystem definition. Our understanding of a biological ecosystem is summarised as a mind-map in Figure 3. This mind- map will allow for the easy transition of understanding from biological ecosystems to the Generic Ecosystem. It can also be easily extended if and when we find new and relevant understanding, given that research into biological ecosystems is ongoing. 3. GENERIC ECOSYSTEM The Generic Ecosystem will provide a framework for the application of ideas, concepts of models from one class of ecosystem to another, which will be fundamental when combining different classes of ecosystems to create and define applied Digital Ecosystems. Biological ecosystems can be considered in terms of ABM [20] and CAS, leading us to define a Generic Ecosystem in terms of MAS and CAS, with agents to represent organisms and a network to represent the geographical landscape, as shown in Figure 4. The evolution, change over time, is biological (Darwinian) [13], and not the more general mathematical interpretation often associated with CAS. The instantiation of the Generic Ecosystem within a specific domain will create a class of that type of system with ecological properties. While some properties, behaviours, and structures will transition easily between domains, as counterparts already exist or can be easily constructed (e.g. evolution), others will prove more challenging (e.g. ecological dynamics). Assuming the motivation for engineering an applied Digital Ecosystem is the development of scalable, adaptive solutions to complex dynamic problems, certain generalisa- tions can be made from biological ecosystems. Sustained diversity [18], is a key requirement for dynamic adapta- tion. In any applied Digital Ecosystem, diversity must be Figure 3: Biological Ecosystem: Key properties, behaviours and structures based on our understand- ing from [4, 9, 8]. This mind-map will allow for the easy transition of understanding from biological ecosystems to the Generic Ecosystems. Figure 4: Generic Ecosystem: Key properties, be- haviours and structures based on our understanding from biological ecosystems. The items in bold are the ones that have changed to more generic concepts from biological ecosystems, as defined in Figure 3. Biological Ecosystem7. Evolution6. Ecology2. Organism8. Geography4. Community1. Environment3. Population5. Multi-Agent SystemGeneric Ecosystem7. Evolution6. Dynamics2. Agent8. Network4. Community1. Environment3. Population5. Multi-Agent Systembalanced against adaptive efficiency because maintaining large numbers of poorly-adapted solutions is costly. The exact form of this trade off will be guided by the specific requirements of the system in question. Stability [25], is likewise, a trade-off: we want the system to respond to environmental change with rapid adaptation, but not to be so responsive that mass extinctions deplete diversity or sudden state changes prevent control. This is an example of the kind of cross ecosystem knowledge transfer to be facilitated, which could be achieved through Biological Design Patterns (BDPs). A design pattern is a general reusable solution to a commonly occurring problem in software design [19]. It is not a finished design that can be transformed directly into code, but a description or template for how to solve a problem that can be used in many different situations [19]. For example, object-oriented design patterns typically show relationships and interactions between classes or objects, without specifying the final application classes or objects that are involved [19]. A BDP would extend this concept to catalogue common interac- tions between biological structures using a pattern-oriented modelling approach [21], which when applied would endow systems with the desirable properties of biological systems, such as self-organisation, self-management, scalability and sustainability. 3.1 Digital Ecosystem Figure 5 shows the key properties, behaviours and struc- tures of a Digital Ecosystem, based on our understanding of a Generic Ecosystem, with the concepts in bold having changed to more domain specific ones. The concept of agent from the Generic Ecosystem naturally maps to software agents [5, 4]. The agents of the Digital Ecosystem are functionally analogous to the organisms of biological ecosystems, including the behaviour of migration and the ability to be evolved [1], and will be achieved through using a hybrid of different technologies. The ability to migrate is provided by using the paradigm of agent mobility from mobile agent systems [42], with the habitats of the Digital Ecosystem provided by the facilities of agent stations from mobile agent systems [30], i.e. a distributed network of locations to migrate to and from. The concept of evolution easily maps to evolutionary computing, and therefore so does the concept of population, to which the process of evolution applies [13]. However, the specifics of applying evolutionary computing in Digital Ecosystems required some consideration. To evolve high- level software components in Digital Ecosystems, we pro- posed taking advantage of the native method of software advancement, human developers, and making use of evolu- tionary computing [17] for combinatorial optimisation [41] of the available software agents (which represent services). This involves treating developer-produced software services as the functional building blocks, as the base unit in a genetic-algorithms-based process. Furthermore, such an approach requires a modular reusable paradigm to soft- ware development, such as Service-Oriented Architectures (SOAs) [39]. Mapping the concept of network required more effort. Specifically, a distributed information-centric dynamically re-configurable network topology to support the constantly changing multi-objective information-centric selection pres- sures of a user base. This would allow for the connectivity of its habitats to adapt to the connectivity within a user base, with a cluster of habitats representing a community within the user base. So, a network topology that will be discovered with time, and which reflects the connectivity within a user base [4, 5]. The mapping of dynamics (ecology) lead to novel form of distributed evolutionary computation for our Ecosystem- Oriented Architecture (EOA) [4, 5]. This novelty came from the creation of multiple evolving populations in response to similar requests, whereas in the island-models of distributed evolutionary computing there are multiple evolving popu- lations in response to only one request [27]. So, different requests are evaluated on separate islands (populations), with their evolution accelerated by the sharing of solutions between the evolving populations (islands), because they are working to solve similar requests (problems). 3.2 Social Ecosystem According to Social Ecosystem theory, populations adapt to their environment in order to survive, since it is in the environment where they find the sustenance resources needed for survival, but human populations are the only ones to adapt to their environment through culture [16]. Therefore, culture may be considered an instrumental re- sponse on the part of human populations in order to achieve a better adaptation to their environment [22, 15]. Different Figure 5: Digital Ecosystem: Key properties, be- haviours and structures based on our understanding of a Generic Ecosystem, with the concepts in bold having changed to more domain specific ones, e.g. network to topology. Figure 6: Social Ecosystem: Key properties, behaviours and structures based on our under- standing of a Generic Ecosystem. The concept of environment from the Generic Ecosystem maps to that of society, while the concept of agent maps to that of person. DigitalEcosystem7. Evolutionary Computing6. Ecology2. Software Agent4. Community1. Environment3. Population5. Multi-Agent System8. TopologySocial Ecosystem7. Co-Evolution6. Dynamics2. Person8. Network4. Community3. Population5. Multi-Agent System1. Societyforms of social organisation constitute cultural responses to the problem of adaptation faced by any population that must survive with the resources which it finds in its environment. So, naturally the concept of agent from the Generic Ecosystems maps to that of person, and the concept of environment maps to that of society, as shown in Figure 6. In biological ecosystems a community is a group of interacting populations sharing an environment [1]. While the concept of community is well established within Social Ecosystems, there is no single agreed definition [31]. How- ever, in summary, in communities, intent, belief, resources, preferences, needs, risks, and a number of other conditions may be present and common, affecting the identity of the participants and their degree of cohesiveness [31]. The word is often used to refer to a group that is organised around common values and is attributed with social cohesion within a shared geographical location, generally in social units larger than a household. The word can also refer to the national community or global community. Since the advent of the Internet, the concept of community no longer has geographical limitations, as people can now virtually gather in an online community and share common interests regardless of physical location. So, we can map the abstract concept of community from the Generic Ecosystem to its more domain specific variant. The concept of evolution is mapped to that of co- evolution, because the key differentiating point of a Social Ecosystem from a social system is the interdependence among the entities within it, which occurs through the phe- nomenon of co-evolution [32]. The is consistent biologically, because the environment is society, such that for any person their environment is other people. In a biological ecosystems, co-evolution is the evolution- ary change of an organism triggered by the change of a related organism [24]. Each party in a co-evolutionary relationship exerts selective pressures on the other, thereby affecting each others evolution. Co-evolution may occur in a one-on-one interaction, such as that between predator and prey, host-symbiont or host-parasitic pair, but many cases are less clear-cut; a species may evolve in response to a number of other species, each of which is also evolving in response to a set of species. This situation has been referred to as diffuse co-evolution [45], and for many organisms the biotic (living) environment is the most prominent selective pressure resulting in evolutionary change. We would suggest that the same is true for Social Ecosystems, such that the majority of its co-evolution is also diffuse. 3.3 Business Ecosystem The concept of a business ecosystem [33] is well-defined and is focused on the micro-economic view of business networks, whereas the Business Ecosystem has a macro- economic perspective [37]. However, it should not be con- fused with ecological economics, which is a transdisciplinary field that aims to address the interdependence of human economies and biological ecosystems [12]. Therefore, the concept of environment from the Generic Ecosystem maps to that of the economy, as shown in Figure 7. While the concept of an agent from the Generic Ecosystem naturally maps to that of a business in the Business Ecosystem. Figure 7: Business Ecosystem: Key properties, behaviours and structures based on our under- standing of a Generic Ecosystem. The concept of environment from the Generic Ecosystem maps to that of economy. Each agent (business) is a participant which both influences and is influenced by the environment (economy) of the Business Ecosystem, which is made up of all the businesses, consumers, and suppliers, as well as economic and legal institutions [32]. Evolutionary theory is well understood within economics [38], so the concept of evolution from the Generic Ecosys- tems can be mapped to its more domain specific variant, as can the concept of population to which the process of evolution occurs [13]. However, ecosystems theory, including ecological dynamics, is not well understood within economics. We could use our efforts with Digital Ecosys- tems as a case study, following the same process to define Business Ecosystems. Alternatively, we could instead make use of the Generic Ecosystem definition, because there is extensive work on the ABM of economic systems [44], which we can take advantage of in defining a CAS/MAS-based definition for an Business Ecosystem. 3.4 Knowledge Ecosystem An extension of knowledge management ideas, a Knowl- edge Ecosystem fosters the dynamic evolution of knowledge interactions between entities to improve decision-making and innovation. This bottom-up approach seeks to be more resilient [28]. In contrast to directive management efforts that attempt either to manage or direct outcomes, Knowledge Ecosystems espouse that knowledge strategies should focus more on enabling self-organisation in response Figure 8: Knowledge Ecosystem: Key properties, behaviours and structures based on our understand- ing of a Generic Ecosystem. The concept of agent from the Generic Ecosystems maps to that of meme. BusinessEcosystem7. Evolution6. Dynamics2. Business8. Network4. Community1. Economy3. Population5. Multi-Agent SystemKnowledge Ecosystem6. Dynamics2. Meme8. Network4. Community1. Environment3. Population5. Multi-Agent System7. Evolutionto changing environments [11]. Articles discussing these ecologically-oriented approaches typically incorporate ele- ments of CAS [3]. There is no single agreed definition of knowledge, but instead numerous competing theories. Still, one way to consider knowledge constructs is as memes. A meme, as defined within memetic theory, comprises a unit of cultural information, the building block of cultural evolution or diffusion that propagates from one mind to another analogously to the way in which a gene propagates from one organism to another as a unit of genetic information and of biological evolution [14]. So, the concept of agent from the Generic Ecosystems would map to that of meme. Therefore, with memes, some knowledge will propagate less successfully and become extinct, while others will survive, spread, and, for better or for worse, mutate [14]. Meme theorists contend that memes evolve by natural selection similarly to Darwinian biological evolution through the processes of variation, mutation, competition, and inheritance. So we can map the concept of evolution from the Generic Ecosystem to the more domain specific variant, as a well as the concept of population to which the process of evolution occurs [13]. 4. APPLIED DIGITAL ECOSYSTEMS As Figure 9 shows, with this conceptual framework the majority of information flow for defining a Generic Ecosystem comes, unsurprisingly, from biological ecosys- tems. However, it also allows for the transfer of realised abstract concepts, through the Generic Ecosystem, from one class of ecosystem to another. We can now define an applied Digital Ecosystem as its combination with a Social Ecosystem; therefore, any dis- tributed adaptive open socio-technical system, with prop- erties of self-organisation, scalability and sustainability, inspired by biological ecosystems, as shown in Figure 10. The items in bold are the ones that have changed to more domain specific concepts, with the background colours indicating the class of ecosystem from which the concepts originate. So, an applied Digital Ecosystem adopts the Figure 9: Ecosystems: The arrows represent information flow between conceptual models of understanding, with the majority coming from So, an applied Digital biological ecosystems. Ecosystem as combination with a Social Ecosystem. its Figure 10: Applied Digital Ecosystem: Key properties, behaviours and structures based on combining concepts and understanding from Social Ecosystems, Digital Ecosystems, and biological ecosystems through our Generic Ecosystem. concept of ecology from biological ecosystems, the concept of society and community from Social Ecosystems, and the concept of topology from Digital Ecosystems (as we will have a digital information-centric network, rather than a biological energy-centric or a sociological geographically- centric one). The other concepts will depend on the application of the Digital Ecosystem, for example evolution from the Generic Ecosystem could map to the evolution of biological ecosystems, evolutionary computing of Digital Ecosystems, the co-evolution of Social Ecosystems, or a domain specific variant of evolution from the application space, or a combination of these. The same therefore also applies to the concept of agent and population, to which the process of evolution will occur. Also, the concept of society will become more specific depending upon the application to which the Digital Ecosystem is applied. This will be further explained as we consider applied Digital Ecosystems which make use of the Business and Knowledge Ecosystems. Furthermore, all these classes of ecosystems can be modelled through ABM as MASs, allowing us to reasonably combine concepts from these different ecosystems. 4.1 Digital Business Ecosystem The DBE is a proposed methodology for economic and technological innovation. Specifically, the DBE is a software infrastructure for supporting large numbers of interacting business users and services [36]. It aims to be a next generation Information and Communications Technology (ICT) that will extend the SOA concept with the automatic combining of available and applicable services in a scalable architecture, to meet business user requests for applications that facilitate business processes. In essence, the DBE will be an Internet-based environment in which businesses will be able to interact with each other in very effective and efficient ways [37]. The synthesis of the concept of Digital Business Ecosystems emerged by adding [35] digital in front of business ecosystem [33]. The term Digital Business Ecosystem was used earlier, but with a focus exclusively on developing countries [34]. The generalisation of the term to refer to a new interpretation of what socio-economic development catalysed by ICT means was new, emphasising the co-evolution between the business ecosystem and its partial digital representation: the digital ecosystem. The term Digital Business Ecosystem came to represent the combination of these two ecosystems [36]. Biological EcosystemSocial Ecosystem Digital EcosystemBusiness EcosystemGeneric Ecosystem(applied)Knowledge EcosystemDigital Ecosystem(applied)7. Evolution6. Ecology2. Agent4. Community1. Society3. Population5. Multi-Agent Ecosystem8. TopologyFigure 11: Digital Business Ecosystem: Key properties, behaviours and structures based on combining concepts and understanding from Social Ecosystems, Digital Ecosystems, and biological ecosystems through our Generic Ecosystem. Figure 12: Digital Knowledge Ecosystem: The key properties, behaviours and structures of a DKE, based on combining concepts from Social Ecosys- tems, Digital Ecosystems, Knowledge Ecosystems and biological ecosystems. However, we can now define a DBE as a combination of Digital, Social, and Business Ecosystems; therefore, any distributed adaptive open socio-technical system for business, with properties of self-organisation, scalability and sustainability, inspired by biological ecosystems. The concept of environment from the Generic Ecosystem maps to the economy of society. In addition to our definition of applied Digital Ecosystems in the previous section, we have mapped the abstract concept of agent, population and evolution from the Generic Ecosystem to the domain specific concept of business and variants of population and evolution. The mapping proposed not only depends on how the Digital Ecosystem would be implemented, but also how Business Ecosystems are interpreted. For example, some efforts in defining DBEs are more biased towards to the Digital Ecosystem [5, 4], while more recent efforts [43] are more aligned with this definition. 4.2 Digital Knowledge Ecosystem We can now define a DKE as a combination of Digital, So- cial, and Knowledge Ecosystems; therefore, any distributed adaptive open socio-technical system for knowledge sharing and management, with properties of self-organisation, scal- ability and sustainability, inspired by biological ecosystems. The concept of environment from the Generic Ecosystem maps to that of society in a DKE. In addition to our definition of applied Digital Ecosystems in the previous section, we have mapped the abstract concept of agent, population and evolution from the Generic Ecosystem to the domain specific concept of meme and variants of population and evolution. The mapping proposed not only depends on how the Digital Ecosystem would be implemented, but also how the Knowledge Ecosystem is interpreted. Wikipedia and Arxiv.org could be considered DKEs, because they have many of the necessary properties, ex- cept for the topology (distributed technical infrastructure). The Digital Ecosystem for Agriculture & Rural Livelihood (DEAL) [10] can also be considered to be a DKE, where the knowledge sharing and management is for the benefit of the society of rural agriculture. However, the topology (distributed technical infrastructure) of Digital Ecosystems is still lacking. In addition to the required distributed technical infrastructure, the necessary legal framework and political support are required for the development and deployment of Digital Ecosystems [43]. 5. CONCLUSIONS We have provided a conceptual and theoretical discussion, from a computer science epistemology, of the characteristics of Digital Ecosystems. Including a discussion of the relevant interfaces between complexity science, sociology, economics and biology to define the nature of Digital Ecosystems and their application. We have provided a conceptual framework for the cross pollination of ideas, concepts and understanding between different classes of ecosystems, based on the universally applicable principles of ABM and CAS. Using ABM to interpret the different classes of ecosystems as MASs and therefore facilitate cross- disciplinary understanding between them. Furthermore, we have used this approach to robustly define Digital Ecosystems, including different classes of applied Digital Ecosystems. Therefore, providing a framework to assist cross-disciplinary collaboration in Digital Ecosystems re- search. There are of course other dimensions to Digital Ecosys- tems to be considered, such as the necessary technical infrastructure (i.e. access to the Internet), legal frameworks and political support required for the development and effective deployment of Digital Ecosystems at all levels (economic, social, technical and political). For example, DBEs to produce real impacts in the economic activities of regions through the improvement of their Small and Medium sized Enterprise (SME) business environments [43]. Also the realisation of Digital Ecosystems in the context of emerging computational paradigms, such as Cloud Comput- ing and Sustainable Computing, which have the potential to radically change the landscape of computational resource provisioning [7]. For example, Digital Ecosystems risk being subsumed into Cloud Computing at the infrastructure level, while striving for decentralisation at the service level, which would clearly be incompatible with its principles. So, the realisation of the Digital Ecosystems vision requires a form of Cloud Computing, but within the principle of distributed community-based infrastructure, where individual users share ownership [29]. Sustainable Computing is concerned with achieving environmental sustainability, while abiding by social and ethical responsibilities. So, while Digital Ecosystems would be socially sustainable, there is a lack of a position on environmental sustainability, which is becoming of ever greater importance. We believe that a framework for understanding, such as we propose here, will be required to Digital BusinessEcosystem7. Evolution6. Ecology2. Business4. Community1. Economy3. Population5. Multi-Agent Ecosystem8. TopologyDigital Knowlege Ecosystem7. Evolution6. Ecology2. Meme4. Community1. Society3. Population5. Multi-Agent Ecosystem8. Topologyaffectively address these and other, issues and dimensions of Digital Ecosystems. 6. ACKNOWLEDGMENTS This work was supported by the EU-funded Open Philoso- phies for Associative Autopoietic Digital Ecosystems (OPAALS) Network of Excellence (NoE), Contract No. FP6/IST-034824. 7. REFERENCES [1] M. Begon, J. Harper, and C. Townsend. Ecology: Individuals, Populations and Communities. Blackwell Publishing, 1996. [2] H. Boley and E. Chang. Digital ecosystems: Principles and semantics. In Inaugural IEEE International Conference on Digital Ecosystems and Technologies, pages 398 -- 403. Citeseer, 2007. [3] B. Bowonder and T. Miyake. Technology management: a knowledge ecology perspective. International Journal of Technology Management, 19(7):662 -- 684, 2000. [18] C. Folke, S. Carpenter, B. Walker, M. Scheffer, T. Elmqvist, L. Gunderson, and C. Holling. Regime shifts, resilience, and biodiversity in ecosystem management. Annual Review of Ecology, Evolution, and Systematics, 35:557 -- 581, 2004. [19] E. Gamma, R. Helm, R. Johnson, and J. Vlissides. Design patterns: elements of reusable object-oriented software. Addison-Wesley, 1995. [20] D. Green, N. Klomp, G. Rimmington, and S. Sadedin. Complexity in Landscape Ecology. Springer, 2006. [21] V. Grimm, E. Revilla, U. Berger, F. Jeltsch, W. Mooij, S. Railsback, H. Thulke, J. Weiner, T. Wiegand, and D. DeAngelis. Pattern-oriented modeling of agent-based complex systems: Lessons from ecology. Science, 310:987 -- 991, 2005. [22] A. Hawley. Human ecology: A theoretical essay. University of Chicago Press, 1986. [23] T. Heistracher, T. Kurz, C. Masuch, P. Ferronato, M. Vidal, A. Corallo, P. Dini, and G. Briscoe. Pervasive service architecture for a digital business ecosystem. In Springer European Conference on Object-Oriented Programming, 2004. [24] E. Lawrence. Henderson's dictionary of biological [4] G. Briscoe. Digital Ecosystems. PhD thesis, Imperial terms. Pearson Education, 2005. College London, 2009. [5] G. Briscoe and P. De Wilde. Digital Ecosystems: Evolving service-oriented architectures. In IEEE Bio Inspired Models of Network, Information and Computing Systems Conference, 2006. [6] G. Briscoe and P. De Wilde. Computing of applied digital ecosystems. In ACM Management of Emergent Digital Ecosystems Conference, 2009. [7] G. Briscoe and A. Marinos. Digital ecosystems in the clouds: Towards community cloud computing. In IEEE Digital Ecosystems and Technologies Conference, 2009. [8] G. Briscoe and S. Sadedin. Natural science paradigms. In Digital Business Ecosystems, pages 48 -- 55. European Commission, 2007. [9] G. Briscoe, S. Sadedin, and G. Paperin. Biology of applied digital ecosystems. In IEEE Digital Ecosystems and Technologies Conference, pages 458 -- 463, 2007. [10] J. Chatterjee and T. Prabhakar. On to action- Building a digital ecosystem for knowledge diffusion in rural India. In International Conference on Knowledge Management, 2005. [11] J. Clippinger. The biology of business: Decoding the natural laws of enterprise. Jossey-Bass, 1999. [12] R. Costanza. An introduction to ecological economics. CRC, 1997. [13] C. Darwin. On the Origin of Species by Means of Natural Selection. John Murray, 1859. [14] R. Dawkins. The Selfish Gene. Oxford University Press, 2006. [15] J. Diez Nicol´as. Ecolog´ıa humana y ecosistema social. Sociolog´ıa y Medio Ambiente, 12:19 -- 31, 1983. [25] S. Levin. Ecosystems and the biosphere as complex adaptive systems. Ecosystems, 1:431 -- 436, 1998. [26] S. Levin. Fragile dominion: complexity and the commons. Perseus Books Group, 1999. [27] S. Lin, W. Punch III, and E. Goodman. Coarse-grain parallel genetic algorithms: categorization and new approach. In Symposium on Parallel and Distributed Processing, pages 28 -- 37. IEEE Press, 1994. [28] J. March. The pursuit of organizational intelligence. Wiley-Blackwell, 1999. [29] A. Marinos and G. Briscoe. Community cloud computing. In Springer Cloud Computing Conference, 2009. [30] F. McCabe and K. Clark. April-agent process interaction language. In M. Wooldridge and N. Jennings, editors, Intelligent Agents: Workshop on Agent Theories, Architectures, and Languages, pages 324 -- 340. Springer, 1994. [31] D. McMillan and D. Chavis. Sense of community: A definition and theory. Journal of Community Psychology, 14(1):6 -- 23, 1986. [32] E. Mitleton-Kelly. Ten principles of complexity and enabling infrastructures. Complex systems and evolutionary perspectives on organisations: The application of complexity theory to organisations, pages 23 -- 50, 2003. [33] J. Moore. The Death of Competition: Leadership and Strategy in the Age of Business Ecosystems. Harvard Business School Press, 1996. [34] J. Moore. Digital business ecosystems in developing countries: An introduction. Technical report, Berkman Center for Internet and Society, Harvard Law School, 2003. [16] J. Diez Nicol´as. Postmaterialism and the social [35] F. Nachira. Towards a network of digital business ecosystem. Culture Within Nature/Culture dans la Nature, 1995. [17] A. Eiben and J. Smith. Introduction to Evolutionary Computing. Springer, 2003. ecosystems fostering the local development. Technical report, Directorate General Information Society and Media, European Commission, 2002. [36] F. Nachira, P. Dini, and A. Nicolai. A network of digital business ecosystems for europe: Roots, processes and perspectives. In Nachira et al. [37]. [37] F. Nachira, A. Nicolai, P. Dini, M. Le Louarn, and L. Rivera Le´on, editors. Digital Business Ecosystems. European Commission, 2007. [38] R. Nelson and S. Winter. An Evolutionary Theory of Economic Change. Harvard University Press, 1982. [39] E. Newcomer and G. Lomow. Understanding SOA with web services. Addison-Wesley, 2005. [40] E. Odum. Energy flow in ecosystems: a historical review. Integrative and Comparative Biology, 8(1):11, 1968. [41] C. Papadimitriou and K. Steiglitz. Combinatorial Optimization: Algorithms and Complexity. Dover Publications, 1998. [42] V. Pham and A. Karmouch. Mobile software agents: an overview. IEEE Communications Magazine, 36:26 -- 37, 1998. [43] J. Stanley and G. Briscoe. The ABC of digital business ecosystems. Communications Law - Journal of Computer, Media and Telecommunications Law, 15(1), 2010. [44] L. Tesfatsion. Agent-based computational economics: Growing economies from the bottom up. Artificial life, 8(1):55 -- 82, 2002. [45] J. Thompson. The coevolutionary process. University of Chicago Press, 1994. [46] M. Waldrop. Complexity: The Emerging Science at the Edge of Order and Chaos. Simon & Schuster, 1992.
1712.05990	1	1712	2017-12-16T17:01:40	Using Machine Learning to Enhance Vehicles Traffic in ATN (PRT) Systems	[ "cs.MA" ]	This paper discusses new techniques to enhance Automated Transit Networks (ATN, previously called Personal Rapid Transit - PRT) based on Artificial Intelligence tools. The main direction is improvement of the cooperation of autonomous modules that use negotiation protocols, following the IoT paradigm. One of the goals is to increase ATN system throughput by tuning up autonomous vehicles cooperation. Machine learning (ML) was used to improve algorithms designed by human programmers. We used "existing controls" corresponding to near-optimal solutions and built refinement models to more accurately relate a system's dynamics to its performance. A mechanism that mostly influences ATN performance is Empty Vehicle Management (EVM). The algorithms designed by human programmers was used: calls to empty vehicles for waiting passengers and balancing based on reallocation of empty vehicles to achieve better regularity of their settlement. In this paper we discuss how we can improve these algorithms (and tune them to current conditions) by using ML to tailor individual behavioral policies. Using ML techniques was possible because our algorithm is based on a set of parameters. A number of weights and thresholds could be tuned up to give better decisions on moving empty vehicles across the track.	cs.MA	cs	Bogdan Czejdo, Wiktor B. Daszczuk, Mikołaj Baszun I USING MACHINE LEARNING TO ENHANCE VEHICLES TRAFFIC IN ATN (PRT) SYSTEMS This paper discusses new techniques to enhance Automated Transit Networks (ATN, previously called Personal Rapid Transit - PRT) based on Artificial Intelligence tools. The main direction is improvement of the cooperation of autonomous modules that use negotiation protocols, following the IoT paradigm. One of the goals is to increase ATN system throughput by tuning up autonomous vehicles cooperation. Machine learning (ML) was used to improve algorithms designed by human programmers. We used "existing controls" corresponding to near-optimal solutions and built refinement models to more accurately relate a system's dynamics to its performance. A mechanism that mostly influences ATN performance is Empty Vehicle Management (EVM). The algorithms designed by human programmers was used: calls to empty vehicles for waiting passengers and balancing based on reallocation of empty vehicles to achieve better regularity of their settlement. In this paper we discuss how we can improve these algorithms (and tune them to current conditions) by using ML to tailor individual behavioral policies. Using ML techniques was possible because our algorithm is based on a set of parameters. A number of weights and thresholds could be tuned up to give better decisions on moving empty vehicles across the track. Keywords - Automated Transit Networks, ATN traffic, Personal Rapid Transit, IoT, negotiation protocols, Machine Learning, Neural networks INTRODUCTION New automatization techniques in urban transport are under development. Many of them concern autonomous vehicles moving in urban traffic [1] [2] [3]. Also, automatization techniques are applied for driver support [4]. In Automated Transit Networks (ATN, also called Personal Rapid Transit - PRT [5]), the movement takes place on a separate track, typically raised above the ground level. This approach liberates the designers from most problems of recognition of the environment. Main targets are keeping up the track [6], routing [7] and empty vehicles management [8]. Some modern ATN systems are decentralized, where decisions are made using simple communication protocols between autonomous modules. For example, a delivery of empty vehicles may follow this principle [8] [9]. The distributed cooperation of autonomous modules on Internet of Things paradigm (IoT [10]), using simple negotiation protocols [11] may be used in ATN station maneuvering, where track segment controllers guide vehicles between charging lot, parking lots and boarding/alighting lots [12] [13]. Such a maneuvering may cause traffic conflicts, if more than one vehicle take part in a change of places. An example of an automatic vehicle guidance system based on cooperating track segment controllers is described in [14]. The other subject of autonomous vehicles cooperation is Empty Vehicles Management, in which vehicles reallocate in ATN, anticipating future needs and compared with current state of the network [15] [16]. The network has some ridership, counted as a number of trips that may be executed in a time unit, for a given Origin-Destination matrix and a given number of vehicles. Various methods are used to check the ridership [8] [17] but they give similar results. Our approach is simulation-based, in which there are infinite queues of passengers on every station and therefore every vehicle finishing its trip is taken by a waiting passenger group. Having real demand, the network performs a number of trips lower than the maximum ridership. We call throughput a number of full trips (with passengers) preformed in a time unit in real traffic conditions. In addition to a number of vehicles and an Origin- Destination matrix, a demand is a parameter of the throughput. Empty the throughput. Higher throughput shortens the average passenger waiting time (at a station in a queue). the system allow trips organized increase in to traffic is one of in ATN systems Machine learning (ML) in some problem domains can generate solutions better than algorithms designed by human programmers [18] [19]. There are also domains where using machine solutions improve existing hardware/software system. Controlling can vehicles these domains. Theoretically it is possible to collect enough data either through a simulation or on-line monitoring system to completely "control" ATN Systems by a trained model. Practically, however, it is much more realistic to use "existing controls" based on near-optimal solutions and build refinement models to more accurately relate a system's dynamics to its performance. In Automated Transit Networks we can control in a relatively robust and reliable fashion as shown [8]. A mechanism that mostly influences ATN performance is Empty Vehicle Management (EVM) [8] [16]. The two algorithms are used: calling empty vehicles for waiting passengers and balancing - reallocation of empty vehicles to achieve better regularity of their settlement [8] (or, on the contrary, irregular allocation of vehicles for special events [20]). The goal of these algorithms is better operation of ATN system, expressed as shortening of passenger waiting time at the cost (possibly smallest) of some number of empty trips performed. We will discuss how we can improve Automated Transit Networks performance on a cooperative transportation task by using ML to adjust individual behavior selection polices from models related to global performance. Unlike some EVM algorithms proposed [15] [21] [22], our algorithm is based on a set of parameters measuring physical features of the net (like distances between stations and their capacity), historical data 1 I (past demand used to predict future demand) and current traffic conditions (current vehicles allocation and a quantity of passenger groups waiting in queues at the stations). A number of weights and thresholds applied to these features give the decisions on moving empty vehicles across the track. Choosing of proper set of parameter values is a difficult process because of large number of parameters and long simulation time of every experiment. Machine learning techniques [23] [24] may be applied to obtain best values of a set of parameters governing EVM. The control system may be automatically learned to react properly in typical situations, in which current demand fits historical data, and in a case of unusual situation when a social event or an accident changes the demand rapidly. Also, learning may be performed during every day operation to adopt the control to changes in number of vehicles available, Origin-Destination matrix structure, physical track changes due to vehicle/track malfunction, conservation conditions, change of maximum velocity allowed (atmospheric conditions, bending of the track, etc.). There are many enhancements for the vehicle traffic in ATN systems based on machine learning. In this paper we classify the efforts based on area of enhancement, data used and machine learning technique. In this paper we discuss efforts related to applying the following techniques: Supervised Learning, Clustering and Reinforcement Learning. Section 1 presents general analysis of machine learning use in ATN systems. In Section 2, an application of neural networks to implement ATN throughput enhancements is discussed. The conclusions are presented in Section 3. 1. MACHINE LEARNING FOR ATN SYSTEMS One of the enhancement of the current control is to generate adjustments for EVM better performance. Due to huge collection of parameter values sets, which cannot be searched completely, two modes of learning should be applied. In training phase, many parameter sets and many working condition parameters should be applied to obtain best tuning of the algorithm for given net structure. Then, during normal execution of the system, algorithm parameters should be changed slightly to learn how to fine-tune the algorithm and to adopt it to changing operation conditions. Many algorithms for EVM are described in the literature, for example [15] [21] [22]. Most algorithms are focused on the optimal reallocation of empty vehicles, and some of them on delivering empty vehicles for waiting passengers. Both of them may reduce the passenger waiting time significantly. Reallocation algorithms are usually based on past demand estimates and future forecast. All mentioned approaches use a form of central data base in which historical demand and actual positions of empty vehicles are stored. Our algorithm, described in [8], uses decentralized data on demand (passenger queues), past demand, vehicles staying at the stations and on the move (performing full and empty trips). Any station retrieves data from its neighboring stations, which allows for distributed implementation. If the demand follows historical trends, the algorithm acts like typical predictive ones. Yet, in a case of non- typical situation (for example a social event organized in a given quarter of a town), other parameters cause the change in EVM to cover this unusual demand [20]. In fact, two EVM algorithms with identical structure but separate sets of parameters are used: calling for vehicle delivery and balancing for vehicle redistribution. Both of them cooperate to enhance the network throughput, which is measured by passenger waiting time. Based on our model we can generate data for machine learning processing. For the calculation of the values of the two functions (for calling and balancing algorithms), a vector of weighting factors and threshold values has been defined. The weighting factors determine how strongly given parameters influence the decision to move a given vehicle. The thresholds define minimum values of measured features, which allow a vehicle to move. – FQ - passenger queue factor - determines the impact of the passenger queue length in target station; – FEB - empty berths factor - the impact of a number of empty berths in target station or capacitor; – FND - normalized inverse distance factor - the impact of normalized inverse distance between nodes (NDij=Dav/Dij, Where Dij - shortest distance from station si to station sj, Dav - average distance between a pair of distinct stations; Notice is a denominator: the shorter the actual distance is, the greater the distance between stations Dij that Fig. 1. Training Mode 2 ATN Throughput Environment Behavior ATN Control ATN System Parameters Machine Learning – Training Mode I Fig. 2. Model Execution Mode is the value of NDij; ND=1 for mean distance). – FAI - historical demand factor - the impact of mean value of passenger groups inter-arrival time distribution at target station during previous days (a measure of predicted demand); the mean value is a denominator, because the shorter is the time between occurrences of two consecutive groups, the stronger the impact is; – TQ - passenger queue threshold - if in a queue there are less passenger groups than TQ, then a vehicle is not moved; – TEB - empty berths threshold - if there are less empty berths than TEB, then a vehicle is not moved; – TEV - empty vehicles threshold - if there are less empty vehicles in berths than TEV, then a vehicle is not moved; – TND - normalized inverse distance threshold (inverse of the horizon) - if the distance between nodes is greater than TND (note that the actual distance between nodes is a denominator), then a vehicle is not moved; – T - total function threshold - if the value calculated as the sum of products of individual factors by corresponding static or dynamic parameter values is less than T, then a vehicle is not moved. Each function has its separate set of the above weighting factors and thresholds. The factors and thresholds for the balancing function have B prefix (i.e., BFQ, BFEB, etc.), and for the calling function they have C prefix (i.e., CFQ, CFEB, etc.). The details of the algorithm may be found in [8]. In addition to the algorithm parameters, some values describe the operating conditions: – Number of vehicles; – Maximum velocity allowed; – Actual demand: total and for selected stations (for example 4); – Several structures of Origin-Destination matrix (for example 4). This gives a set of 11 additional parameters. The simulation methods can provide verification of some solutions satisfying the main constrains but simulation methods alone cannot guarantee an optimal solution in all circumstances. Various approaches have been proposed that heavily sample the search space in search for the global optimal solution but most of them are very expensive and too slow for online applications especially for significantly changing environment requirements. There is no global optimization method that can solve this problem with any theoretical guarantee of success. Combining simulation with machine learning can improve this process in many aspects New applications of Machine Learning (ML) [25] are identified very rapidly. These changes are facilitated by new approaches and breakthroughs in parallelization of artificial neural networks algorithms that takes advantage of fast and parallel hardware [24]. Each new application, however, requires careful determination of training data. When using Deep Learning approach the ML model can be created from raw data and labels assuming that we have enough labeled data. In a more traditional approach the raw data is first transformed to, so called, feature vector and then the training is performed on such labeled data. In our case the use of machine learning has a more complex architecture, but is shown in a simplified form in Fig. 1. We assume an ATN system described by distributed agents' behavior. ATN system behavior, or shortly ATN throughput, has many parameters and acts within an environment i.e. ATN throughput responds to environmental requirements. These environmental requirements can change with time, therefore they are referred to as Environment Behavior. The ATN system performance (throughput) should be measured with respect to Environment Behavior. Therefore, as shown in Fig. 1, our training data includes Environment Behavior, ATN throughput and some measures of resulting performance. In case of Empty Vehicle Management the best measures of the performance are related to throughput. The data describing ATN throughput can be collected during actual ATN runs and/or by simulations. Different types of approaches can be used to determine ATN throughput. The typical approach is to develop a general evaluation functions, e.g. number of passengers transported per hour, the total distance for all passengers transported per hour, the average passenger waiting time, minimum energy, etc. These functions have a tendency to emphasize group interests rather than the individuals. As in typical ML approach several modes are defined. In training mode, many working condition parameters (Environment Behavior) and many ATN system parameter sets are generated to obtain best ATN throughput tuning for the given Environment Behavior. Then, during testing mode and normal execution mode of the system, ML model will compute the parameters that will be used to control the individual vehicles as is shown in Fig. 2. 3 ATN Throughput ATN Control ATN System Parameters Machine Learning – Model Execution Environment Behavior I There are also possible improvement for this approach by applying additionally reinforcement learning. The architecture of such system will be very similar to Fig. 2, but with the additional arrow indicating that the throughput will be provided for ML system to describe the feedback to change ML model slightly by learning how to fine-tune parameters and to better adopt it to changing operation conditions. 2. ML USING NEURAL NETWORKS We considered use of various machine (ML) algorithms including neural networks [26] [27] [28] [29] [30] [31], Support Vector Machines (SVM) [23], and random forest [32], to combine measurable features into a single number estimate of ATN throughput [15] [16] . learning Neural networks were selected as ML solution since they support both traditional ML using feature vectors and Deep Learning using raw data. A neuron is a cell in the brain that is responsible for collecting processing and dissemination of electrical signals [24]. The brain's information processing capability is associated with the network of such neurons [33]. This is one of the reason that artificial neural network always were important part of Artificial Intelligence area and were of great interest to students. Each neuron has n inputs with activation signals x0xn. For each input a weight is assigned w0wn. Positive value of the weight corresponds the weight deactivate. Output signal y, i.e. activation of the neuron, can be described as: to activation and negative value of (1) Where function f() – is an activation function. The activation function can be linear: (2) Very often the step functions are used as activation function or in case of more complex tasks the function of hyperbolic tangent or sigmoidal functions are applied. Neurons can be combined into layers. There can be any number of neurons in each layer. Neural network is composed from many layers [34]. There is an input layer, the output layer and possibly a hidden layer. There can be any number of hidden layers within a neural network but one is usually enough to support typical problems. The design of the neural network requires specification of number of neuron in each layer and the learning algorithm which is associated with the activation function. Let us discuss the design a neural net that will recognize data represented by Environment Behavior parameters. Clustering Environment Behavior values and assuming small randomly distributed changes around the basic cluster can help to simplify ML training. Absence or presence of clustering of Environment Behavior defines different types of Neural Network. The first case is to use raw Environment Behavior data and to train ML models on such data. It means that the neural network has to accept the state of Environment Behavior vector as an input. It would result in the neural network number of neuron equal to 11 in Clusters Fig. 3. Success rate for randomly generated data for cluster number six 4 )(0niiiwxfyRaforassf)(I the input layer [23]. The second NN would use values that label the clusters. That would make much easier to train ML models, but they might fail to pick best ATN throughput, especially when applied to Environment Behavior that is not clearly within a cluster. Our experiments suggested using richer environment data. Although this method requires larger training set, it provides a more precise response to the environment variable. Yet another NN can be constructed term tendencies and cyclic changes of the Environment. Such a NN would require additional neurons in the input layer to accommodate additional variable describing explicitly typical dynamics of Environment Behavior. improve decoding short term and long to There can be a similar discussion for clustering ATN System parameters values Absence or presence of clustering of ATN System parameters defines again different types of Neural Network. The first case is to use raw ATN system parameters and to train ML models to predict such data. The second NN would use values that label the clusters. That would make much easier to train ML models. Our experiments here, however, did not suggested using richer control space because of instances of overfitting. In order to provide a resilience to changes of Environment Behavior the cluster approach was preferred. In our neural networks output need to be related to ATN Behavior parameters or their clusters. For example when we have 10 clusters then we might need 10 classifying states resulting in 10 neurons in the output layer [32]. Each neuron will have output either a 1 or zero. The value 1 indicates that it neural network recognized the given category. This happens if there is 100% of success in recognition of the ATN parameters. In the ambiguous situation that success rate is much lower, especially when the noise is present. The number of neurons in the hidden layer needs typically to be identified through experiments. Once the neural network has been designed it needs to be trained. We mostly used supervised learning as shown in Figs. 1 and 2, which means that we feed the neural network with data set called a training set. After that the testing takes place that tries to identify an error related to applying the trained neural network to a data set called a testing set. Since typically the data for training set and testing set is obtained randomly from the provided set, the additional evaluation of the network is necessary. Thus each application requires extensive study to guarantee the proper performance of neural network. Let us discuss the specific implementation of NN for ATN system. An important part of our efforts is to set up the proper environment consisting on appropriate ML library. One possibility is NN Toolbox [26]. We used another popular environment based on theano. Our environment and libraries included: 1. theano - provides fast multi-dimensional array calculation, allows GPU usage Code: import theano 2. pandas - converts csv files to DataFrames used for training. Code: import pandas as pd 3. numpy - numerical library for python Code: import numpy as np 4. Keras - Deep learning library built on top of Theano that allows for fast architecture iteration. Sequential - an object that encompasses deep learning operations Code: from keras.models import Sequential 5. Add-ons: Dense - provides access to the classic neural network layer. Dropout - gives access to the ability to randomly drop certain neurons during test time, which has been shown to dramatically improve performance. Activation - gives access to varying types of neural activators, for example relu and tanh Code: from keras.layers.core import Dense, Dropout, Activation 6. Optimizers - give access to different types of gradient descent algorithms for model training from keras.optimizers Code: RMSprop import SGD, Adam, 7. Keras utilities - give access to evalutation metrics Code: from keras.utils import np_utils 8. sklearn preprocessing - gives access to min max scaler Code: from sklearn import preprocessing 9. Custom library built for custom metrics Code: import Throughput as Th The example execution of the ML model is shown in Fig. 3. The visualization tools displays discovered association of a given Environment Behavior with one of the 10 clusters constructed for ATN System parameters. The specific noise level was applied that modified Environment Behavior. More study are needed to formally determine the success rate for other clusters and for other space reduction methods with various controlled noise level. We investigate the two algorithms: calling the vehicles for waiting passengers and empty vehicles redistribution. They both must be active in ATN operation for its best throughput. A quality measure for every algorithm is passenger waiting time, therefore every algorithm should be tuned separately, with fixed set of parameters of the other one. This procedure should be repeated for several times, switching the tuning between the two algorithms. 3. CONCLUSIONS to produce simple but Our main focus was to facilitate "intelligent" solutions to be applied for ATN systems. We showed how we can include in ATN systems soft computing interesting "intelligent" behaviors. There are many benefits of using our testbed system with NN as the tool for improving performance of ATN systems including optimization of throughput as emphasized in this paper. More study are needed to compare ML results using raw data and various space reduction algorithms including clustering. Especially important is to formally determine the resilience of ATN systems to unexpectedly changing Environment Behavior. So far, we have been exploring distributed protocols using simulation methods, with parameter sets assigned arbitrary. This is our first article in which we describe our experience of using ML for ATN. We considered several classification algorithms but we chose neural networks mainly for practical reasons because of the library keras which not only allows for "feature vector based" ML but also "Deep Learning". With the production of more simulation data we plan to use the same tool (keras) for "deeper" analysis reducing the role of "feature vector". BIBLIOGRAPHY [1] S. Parkinson, P. Ward, K. Wilson, J. Miller, Cyber Threats Facing Autonomous and Connected Vehicles: Future Challenges, IEEE Trans. Intell. Transp. Syst. PP(99), 2017 1– 18. doi:10.1109/TITS.2017.2665968. [2] T.U. Kim, J.W. Lee, S. Yang, Study on development of autonomous vehicle using embedded control board, in: 2016 11th Int. Forum Strateg. Technol. (IFOST), Novosib. Russ. 1-3 June 599–603. doi:10.1109/IFOST.2016.7884331. IEEE, 2016, 2016: pp. [3] S. Zhang, I.-L. Yen, F. Bastani, H. Moeini, D. Moore, A Semantic Model for Information Sharing in Autonomous Vehicle Systems, in: 2017 IEEE 11th Int. Conf. Semant. Comput. (ICSC), San Diego, CA, 30 Jan - 1 Feb 2017, IEEE, 5 I 2017: pp. 32–39. doi:10.1109/ICSC.2017.93. [4] W. Choromański, I. Grabarek, M. Kozłowski, Simulation and Experimental Study of Selected Parameters of the Multifunction Steering Wheel in the View of Users' Abilities and Accuracy of Vehicle Maneuvers, Procedia Manuf. 3 (2015) 3085–3091. doi:10.1016/j.promfg.2015.07.855. [5] S.S. McDonald, Personal Rapid Transit (PRT) system and its Development, in: Encycl. Sustain. Sci. Technol., Springer New York, New York, NY, 2012: pp. 7777–7797. doi:10.1007/978- 1-4419-0851-3_671. [6] M. Kozłowski, W. Choromański, J. Kowara, Parametric sensitivity analysis of ATN-PRT vehicle (Automated transit network – personal rapid transit), J. Vibroengineering. 17 (2015) 1436–1451. https://www.scopus.com/inward/record.uri?eid=2-s2.0- 84946555953&partnerID=40&md5=1b3eb293ddc7bca0d8e9c cfba1974cdb. [7] E. Fatnassi, O. Chebbi, J.C. Siala, Bee colony algorithm for the routing of guided automated battery-operated electric vehicles in personal rapid transit systems, in: 2014 IEEE Congr. Evol. Comput. (CEC), Beijing, China, 6-11 July 2014, IEEE, 2014: pp. 536–543. doi:10.1109/CEC.2014.6900445. [8] W.B. Daszczuk, J. Mieścicki, W. Grabski, Distributed algorithm for empty vehicles management in personal rapid transit (PRT) network, J. Adv. Transp. 50 (2016) 608–629. doi:10.1002/atr.1365. [9] O. Chebbi, J. Chaouachi, A Decentralized Management Approach for On-Demand Transit Transportation System, in: Second Int. Afro-European Conf. Ind. Adv. AECIA 2015, Paris - Villejuif, Fr. 9-11 Sep 2015, AISC 427, Springer International Publishing, Cham, Switzerland, 2015: pp. 175–184. doi:10.1007/978-3-319-29504-6_18. [10] G.M. Lee, N. Crespi, J.K. Choi, M. Boussard, Internet of Things, in: Evol. Telecommun. Serv. LNCS 7768, Springer- Verlag, 257–282. doi:10.1007/978-3-642-41569-2_13. Heidelberg, Berlin 2013: pp. [11] Z. Garofalaki, D. Kallergis, G. Katsikogiannis, I. Ellinas, C. Douligeris, Transport services within the IoT ecosystem using localisation parameters, in: 2016 IEEE Int. Symp. Signal Process. Inf. Technol. (ISSPIT), Limassol, Cyprus, 12-14 December 87–92. doi:10.1109/ISSPIT.2016.7886014. IEEE, 2016, 2016: pp. [12] J.-M. Won, H. Choe, F. Karray, Optimal design of personal rapid transit, in: 2006 IEEE Intell. Transp. Syst. Conf. Toronto, Canada, 17-20 Sept 2006, IEEE, 2006: pp. 1489–1494. doi:10.1109/ITSC.2006.1707434. [13] E. Fatnassi, J. Chaouachi, Discrete Event Simulation of Loading Unloading Operations Intermodal Transportation Context, in: 5th Comput. Sci. On-Line Conf. 2016 (CSOC2016), Softw. Eng. Perspect. Appl. Intell. Syst. Vol 2, AISC 465, Springer International Publishing, Cham, Switzerland, 2016: pp. 435–444. doi:10.1007/978-3-319- 33622-0_39. in a Specific [14] B. Czejdo, S. Bhattacharya, M. Baszun, W.B. Daszczuk, Improving Resilience of Autonomous Moving Platforms by real-time analysis of their Cooperation, Autobusy-TEST. 17 (6), 2016, 1294–1301. arXiv: 1705.04263. I. Andréasson, Reallocation of Empty PRT vehicles en route, in: TRB Annu. Meet. Washingt. DC, 12-16 January 2003, Transportation Research Board, 2003: pp. 1–13. [15] [16] W.B. Daszczuk, W. Choromański, J. Mieścicki, W. Grabski, Empty vehicles management as a method for reducing passenger waiting time in Personal Rapid Transit networks, 6 IET Intell. Transp. Syst. 9(3), 2015 231–239. doi:10.1049/iet- its.2013.0084. [17] J. Lees-Miller, J. Hammersley, R. Wilson, Theoretical Maximum Capacity as Benchmark for Empty Vehicle Redistribution in Personal Rapid Transit, Transp. Res. Rec. J. Transp. Res. Board. 2146 (2010) 76–83. doi:10.3141/2146- 10. [19] [18] S. Kukliński, J. Wytrębowicz, K.T. Dinh, E. Tantar, Application of Cognitive Techniques to Network Management and Control, in: EVOLVE - A Bridg. between Probab. Set Oriented Numer. Evol. Comput. V, Beijing, China, 1–4 July 2014, Springer International Publishing, Cham, Switzerland, 2014: pp. 79–93. doi:10.1007/978-3-319-07494-8_7. I. Bluemke, M. Tarka, Learning Algorithms in the Detection of Unused Functionalities in SOA Systems, in: K. Saeed, R. Chaki, A. Cortesi, S. Wierzchoń (Eds.), 12th IFIP TC8 Int. Conf. Emerg. Trends Comput. Informatics, Syst. Sci. Eng. Krakow, Poland, 25-27 Sept. 2013, LNCS 8104, Springer- Verlag, 389–400. doi:10.1007/978-3-642-40925-7_36. Heidelberg, Berlin 2013: pp. [20] W.B. Daszczuk, J. Mieścicki, Distributed management of Personal Rapid Transit (PRT) vehicles under unusual transport conditions, Logistyka. 4 (2015) 2896–2901. arXiv: 1705.04497 [21] P. Zheng, D. Jeffery, M. McDonald, Development and evaluation of traffic management strategies for personal rapid transit, in: Ind. Simul. Conf. 2009, Loughborough, UK, 1-3 June 2009, 2009: pp. 191–195. [22] J.D. Lees-Miller, R.E. Wilson, Proactive empty vehicle redistribution for personal rapid transit and taxis, Transp. Plan. Technol. 17–30. doi:10.1080/03081060.2012.635414. 35(1), 2012, [23] W.-H. Steeb, The Nonlinear Workbook Chaos, Fractals, Cellular Automata, Neural Networks, Genetic Algorithms, Gene Expression Programming, Support Vector Machine, Wavelets, Hidden Markov Models, Fuzzy Logic with C++, Java and SymbolicC++ Programs, World Scientific, 2008. ISBN: 978-981-4583-46-6 [24] S. Russell, P. Norvig, Artificial Intelligence: A Modern Approach, Person Education Limited, Essex, UK, 2014. ISBN: 0-13-629841-9 [25] Machine Learning Repository, University of California,. http://archive.ics.uci.edu/ml/ . [26] M. Egmont-Petersen, D. de Ridder, H. Handels, Image processing with neural networks-a review, Pattern Recognit. 35 (2002) 2279–2301. doi:10.1016/S0031-3203(01)00178-9. [27] D.L. Reilly, L.N. Cooper, C. Elbaum, A neural model for learning, Biol. Cybern. 45(1), 1982, 35–41. category doi:10.1007/BF00387211. [28] M. Riedmiller, H. Braun, A direct adaptive method for faster backpropagation learning: the RPROP algorithm, in: IEEE Int. Conf. Neural Networks, San Farncisco, CA, 28 March-1 April 1993, 1993: 586–591. doi:10.1109/ICNN.1993.298623. IEEE, pp. [29] F. Bajramovic, C. Gruber, B. Sick, A comparison of first- and second-order training algorithms for dynamic neural networks, in: 2004 IEEE Int. Jt. Conf. Neural Networks, Budapest, Hungary, 25-29 July 2004, IEEE, 2004: pp. 837–842. doi:10.1109/IJCNN.2004.1380038. [30] T. Hadzibeganovic, S.A. Cannas, A Tsallis' statistics based neural network model for novel word learning, Phys. A Stat. Mech. 732–746. doi:10.1016/j.physa.2008.10.042. 388(5), 2009, Appl. Its I [31] D.P. Mandic, J.A. Chambers, Recurrent Neural Networks for Prediction: Learning Algorithms, Architectures and Stability, Wiley, 2001. ISBN: 978-0-471-49517-8 [32] J.J. Hopfield, Neural networks and physical systems with emergent collective computational abilities., Proc. Natl. Acad. Sci. 79(8), 1982, 2554–2558. doi:10.1073/pnas.79.8.2554. [33] M.A. Arbib, The Handbook of Brain Theory and Neural Networks, Second Edition, MIT Press, Cambridge, MA, 2002. ISBN: 9780262011976 [34] K. Fukushima, Cognitron: A self-organizing multilayered neural network, Biol. Cybern. 20(3-4), 1975, 121–136. doi:10.1007/BF00342633. Zastosowanie Uczenia Maszynowego (Machine Learning) do poprawy jakości działania zautomatyzowanych sieci transportowych (Automatic Transit Network-PRT) poprzez dostrajanie rozwiązaniom, suboptymalnym W artykule omówiono nowe techniki usprawniania zautomatyzowanych sieci transportowych (ATN, wcześniej nazywanych Personal Rapid Transit - PRT), opartych na narzędziach sztucznej inteligencji. Głównym kierunkiem jest poprawa współpracy autonomicznych modułów, które używają protokołów negocjacyjnych w paradygmacie IoT. Jednym z celów jest zwiększenie przepustowości systemu transportowego współpracy autonomicznych pojazdów. Uczenie maszynowe (ML) jest wykorzystywane do poprawy algorytmów opracowanych przez programistów. Użyliśmy "istniejącego sterowania", odpowiadającego i skonstruowaliśmy modele dostrajania, aby dokładniej odnieść dynamikę systemu do jego wydajności. Mechanizm, który wpływa głównie na wydajność ATN to Zarządzanie Pustymi Pojazdami (Empty Vehicle Management - EVM). Wykorzystano algorytmy opracowane przez programistów: wzywanie pustych pojazdów dla oczekujących pasażerów i równoważenie w oparciu o realokację pustych pojazdów w celu osiągnięcia lepszej regularności ich rozmieszczenia. W tym artykule omówimy, jak można poprawić te algorytmy (i dostroić je do aktualnych warunków), używając ML do dostosowania behawioralnych. Wykorzystanie technik ML było możliwe, ponieważ nasz jest na zbiorze parametrów. Zestaw algorytm oparty współczynników i progów może zostać dostrojony do podejmowania lepszych decyzji o planowaniu ruchu pustych pojazdów na torze. Authors: indywidualnych zasad Bogdan Czejdo, PhD – Department of Mathematics and Computer Science, Fayetteville State University, Fayetteville, NC 28301, USA, [email protected] Wiktor B. Daszczuk, PhD – Warsaw University of Technology, Institute of Computer Science, Nowowiejska str. 15/19, 00-665 Warsaw, Poland, [email protected] Mikołaj Baszun, PhD – Warsaw University of Technology, Institute of Microelectronics and Optoelectronics, Nowowiejska str. 15/19, 00-665 Warsaw, Poland, [email protected] 7
1504.04811	1	1504	2015-04-19T08:37:40	Socializing Autonomous Units with the Reflexive Game Theory and Resonate-and-Fire neurons	[ "cs.MA" ]	In this study the concept of reflexia is applied to modeling behavior of autonomous units. The relationship between reflexia, on the one hand, and mirror neuron system and perception of emotions, on the other hand, is introduced. The main method of using reflexia in a group of autonomous units is Reflexive Game Theory (RGT). To embody RGT in a group of autonomous agents a communication system is employed. This communication system uses frequency domain multiplexing by means of Izhikevich's resonate-and-fire neural models. The result of socialization of autonomous units by means of RGT and communication system is illustrated in several examples.	cs.MA	cs	Socializing Autonomous Units with the Reflexive Game Theory and Resonate-and-Fire neurons Sergey Tarasenko 1 5 1 0 2 r p A 9 1 ] A M . s c [ 1 v 1 1 8 4 0 . 4 0 5 1 : v i X r a Abstract -- In this study the concept of reflexia is applied to modeling behavior of autonomous units. The relationship between reflexia, on the one hand, and mirror neuron system and perception of emotions, on the other hand, is introduced. The main method of using reflexia in a group of autonomous units is Reflexive Game Theory (RGT). To embody RGT in a group of autonomous agents a communication system is employed. This communication system uses frequency domain multiplexing by means of Izhikevich's resonate- and-fire neural models. The result of socialization of autonomous units by means of RGT and communication system is illustrated in several examples. Keywords -- reflexive game theory, multiagent systems, resonate- and-fire neurons I. INTRODUCTION The notion of ref lexia in the psychological context was fisrt introduced by Lefebvre in late 60s [1], [2], [3]. Ref lexia means projection of the external world on one's mental state. More specifically, if a human being stands in the field of barley he/she can imagine onesself standing in the field of barley from the 3rd person's perspective. Thus, preserving the egocentric point of view humans are capable of imaganing their own allocentric representation. 1 Therefore, the gist of reflexia is an ability to imagine self perception in the allocentric reference frame (external2 point of view) being operating in the one's egocentric reference frame. Reflexia is an ability to penetrate into the deeper layers of one's psychological state. An abstract example of penetration into the deeper layers is when subject a can imagine another person (subject b), who is imaganing subject a, the world around and himself imagaging it. The overall results about the justification of reflexia and its application for modeling human behavoir has been summa- rized by Vladimir Lefebvre, the principle investigator in this field, in his book "Algebra of Consciece" [11], [16]. Here we would like to refer two more examples of potential implementation of reflexia: mirror neuron system and percep- tion of emotions. Mirror neuron system: The key concept of the Mirror Neuron System discovered by Rizzolatti and his colleagues [4], [5] is that there are neurons in primate brain that activate in both cases when primate is doing a particular action itselt or observe someone else doing the same action. Therefore mirror neuron system translates external state of another agent into S. Tarasenko is an independent researcher. Email: [email protected] 1Term egocentric means perception from the 1st person perspective, while term allocentric means the perception from the 3rd person perspective. 2Extenal means here outside of ones body or any other feasable entity the internal state of the current agent. Therefore primates can repeat the observed action. This functionallity is very close to the notion of ref lexia. Perception of emotions: Usually, the emotions are character- ized by some physiological patterns of body activity on the one hand, and external expression by face mimic or gestures on the other hand. A reproduction by the onesself of physical part of emotional pattern, i.e., just making an angry face, can elicit the anger as emotional state itself [6], [7]. This suggests that there is a mechanism that helps mapping someones internal state to the self. For instance, Edgar Allan Poe, in his story "The Purloined Letter", describes how one character is attemping to understand the intensity of emotional experience of another character by self-mimicing (or imitating) facial expressions of another one. Most recently the principle of reflexia have been reconsid- ered in the shape of the Reflexive Game Theory (RGT). The expectations are that since reflexia is intrinsic ability of human being and the principles and models proposed by Lefebvre have been proved to be true, the Reflexive Game Theory can deliver the human-like decision-making processes. In contrast to the Game Theory based on purely utilitary and rational principle, the Reflexive Game Theory is based on the human's decision making. The conner stone of the Reflexive Game Theory is the egoism forbiddeness principle, while Game Theory is build upon pure egoism stemming from (Min Loss - Max Utility) principle. Most recently RGT has been applied for theoretical model- ing of human-robot groups [13], in which robots were sucess- fuly refraining people from doing risky actions. However, the issue of how to embody RGT into the system of robotic agents is still an open questions. The goal of this paper is to illustrate how application of RGT algorithms for modeling purely robotic groups can literary humanize the robots by given them human sensitivity instead of cold rationality, which is usually attributed to machines. A certian structure of communication system to enable robots "talk" to each other is proposed. II. BRIEF OVERVIEW OF THE REFLEXIVE GAME THEORY A. Representation of groups: graphs, polynomials, stratifica- tion tree and decision equation The exhaustive desription of the Reflexive Game Theory (RGT) and tutorial of RGT application have been presented by Lefebvre [8], [9], [?]. Here, we present a brief overview of RGT enough to understand its basic concept and formulate the tasks solved in this paper. The RGT deals with groups of abstract subjects (individuals, humans, autonomous agents etc). Each subject is assigned a unique variable (subject variable). Any group of subjects is represented in the shape of fully connected graph, which is called a relationship graph. Each vertex of the graph corresponds to a single subject.The name of each vertex is a unique subject variable. The RGT uses the set theory and the Boolean algebra as the basis for calculus. Therefore the values of subject variables are elements of Booleans algebra. All the subjects in the group can have either alliance or conflict relationship. The relationships are illustrated with graph ribs. The solid-line ribs correspond to alliance, while dashed ones are considered as conflict. For mathematical analysis alliance is considered to be conjunction (multipli- cation) operation (·), and conflict is defined as disjunction (summation) operation (+). The decomposable relationship graphs [8], [9], [10] can be presented in the analytical form of a corresponding polyno- mial. Any relationship graph of three subjects is decompos- able. Consider three subjects a, b and c. Let subject a is in alliance with other subjects, while subjects b and c are in conflict (Fig. 1). The polynomial corresponding to this graph is ab + c. Regarding a certain relationship, the polynomial can be stratified (decomposed) into sub-polynomials [8], [9]. Each sub-polynomial belongs to a particular level of stratification. If the stratification regarding conflict (alliance) was first built, then the stratification regarding alliance (conflict) is imple- mented on the next step The stratification procedure finalizes, when the elementary polynomials, containing a single variable, for each variable are obtained after a certain stratification step. The result of stratification is the Polynomial Stratification Tree (PST). It has been proved that each non-elementary polynomial can be stratified in an unique way, i.e., each non- elementary polynomial has only one corresponding PST (see [10] considering one-to-one correspondence between graphs and polynomials). Each higher level of the tree contains polynomials simpler than the ones on the lower level. For the purpose of stratification the polynomials are written in square brackets. The PST for polynomial ab + c is presented in Fig.2. We omit the branches of the PST and from each non- elementary polynomial write in top right corner its sub- polinomials. The resulting tree-like structure is called a di- agonal form[8], [9], [11], [16]. Consider the diagonal form corresponding to the PST presented in Fig. 2: [a][b] [ab] +[c] [ab + c] We introduce the universal set, which contains the elemen- tary actions. For example, these actions are actions α and β. The Boolean algebra of the universal set includes four elements: 1 = {α, β}, {α}, {β} and the empty set 0 = {} = Ø. The diagonal form is considered to be a function defined on the Boolean algebra. Accroding to definition given by Lefebvre [11], we present 2 Fig. 1. Relationship graph. Fig. 2. elementary polynomials. Polynomial Stratification Tree. Polynomials [a], [b] and [c] are here exponential operation defined by formula P W = P + W , (1) where W stands for negation of W [8], [9], [?]. It can be shown that function in eq. (1) is equivalent to implication function [16]. This operation is used to fold the diagonal form. During the folding, round and square brackets are considered to be interchangeable. Next we implement folding of diagonal form of polynomial ab + c: [a][b] [a][b] +[c] [ab + c] = ab + c B. The Decision Equation: definition and solution Each subject in the group should choose an alternative (element) of the Boolean algebra. We consider the decision equations. Each equation corresponds to a single subject. The solution of such equation defines the choice of each subject in the group. This equation contains subject variable in the left-hand side and the result of diagonal form folding in the right-hand side: a = ab + c b = ab + c c = ab + c (2) (3) (4) To solve the decision equation, one should first transform it into canonical form [8], [13] defined as: x = Ax + Bx , (5) where x is the subject variable, and A and B are some sets. This equation has solution if and only if the set B is contained in set A: A ⊇ B. If this requirement is satisfied, then eq.(5) has at least one solution from the interval A ⊇ x ⊇ B [8], [9], [?]. Otherwise, the decision equation has no solution, and it is considered that subject cannot make a decision. The state of inability to make decision (choice) is called a f rustration. The explicit tranformation of decision equation TABLE I INFLUENCE MATRIX a a {β} {β} b {α} b {β} c {β} {β} c a b c for subjects a, b and c into canonical form is consider in [14]. Therefore, here we only provide the resulting canonical forms: a = (b + c)a + ca b = (a + c)b + cb c = c + abc (6) (7) (8) Next we consider two tasks, which can be formulated regarding the decision equation in the canonical form and briefly discuss methods to solve each task. C. The Forward and Inverse Tasks In this section, we only illustrate the introductory examples of the Forward and Inverse task of the RGT. The comprehen- sive explaination how to solve the forward task can be found here [8], [9], [?]. The issues regarding the Inverse task are discussed in details in [14]. The Forward Task. The variable in the left-hand side of the decision equation in canonical form is the variable of the equations, while other variables are considered as influences on the subject from the other subjects. The forward task is formulated as a task to find the possible choices of a subject of interest, when the influences on him from other subjects are given. The mutual influences in the forward task are presented in the Influence matrix (Table I). The main diagonal of influence matrix contains the subject variables. The rows of the matrix represent influences of the given subject on other subjects, while columns represent the influences of other subjects on the given one. The influence values are used in decision equations. I illustrate solution of the forward task using subjects a, b and c. By using canonical forms of decision equations for each subject (eqs. 6-8) and Influence matrix, we obtain the choice of each subject: subject a: a = ({β} + {β})a + a ⇒ a = {β}a + a. subject b: b = b + ({α}{β} + {α})b ⇒ b = b + {β}b. subject c: c = c + ({β}{β} + {β})c ⇒ c = c + ({β} + {α})c ⇒ c = 1. Equation for subject a does not have any solutions, since set A = {β} is contained in set B = 1: A ⊂ B. Therefore, subject a cannot make any decision. Therefore he is considered to be in frustration state. Equation for subject b has at least one solution, since A = 1 = {α, β} ⊇ B = {β}. The solution belongs to the interval 1 ⊇ b ⊇ {β}. Therefore subject b can choose any alternative from Boolean algebra, which contains alternative {β}. These alternatives are 1 = {α, β} and {β}. 3 Equation for subject c turns into equality c = 1. This is possible only in the case, when A = B = 1. The solution of the Forward task can be algorithmized as 1) formalize of actions in terms of Boolean algebra of follows: alternatives; 2) represent a group as relationship graph; 3) represent relationship graph The Inverse Task. The inverse task is formulated as a task to find all the simultaneous (or joint) influences of all the subjects together on the subject of interest that result in choice of a particular alternative or subset of alternatives. We call the subject of interest to be a controlled subject. Let subject a is the controlled subject and a∗ is a fixed value representing an alternative or subset of alternatives, which subjects b and c want subject a to choose. By substituting subject variable a in decision equation for fixed value a∗, we obtain the influence equation: a∗ = (b + c)a∗ + ca∗ , (9) In contrast to the decision equation, which is equation of a single variable, the influence equation is the equation of multiple variables. We need to find all the joint influences of subjects b and c in form of pairs (b, c). Let a∗ = {α}, then we need to solve the system of equations (cid:40) b + c = {α} c = {α} (10) Consequently, we have to solve equation b + {α} = {α} (11) and to find all the pairs (b, c), results in solution a∗ = {α}. These pairs are solutions of the system (10). Therefore, we run all the possible values of variable b and check if the first equation of the system (11) turns into true equality: b = 1 : 1 + {α} = 1 ⇒ 1 (cid:54)= {α}; b = {α} : {α} + {α} = {α} ⇒ {α} = {α} b = {β} : {β} + {α} = 1 ⇒ 1 (cid:54)= {α}; b = 0 : 0 + {α} = {α} ⇒ {α} = {α}. Therefore, out of four possible values, only two values {α} and 0 are appropriate. Thus, we obtain two pairs (b, c): ({α},{α}) and ({α}, 0). These pairs represent the strategies of reflexive control. This conludes the overview of the Reflexive Game Theory. As a final remark, we show Basic Control Schema of Abstract Individual (BSCAS) (Fig. 3). The detailed decription of the BSCAS can be found in [14]. Summarizing this section, I emphasize the information needed for the RGT to be applied. First of all, we need to define the universal set of actions and the corresponding Boolean algebra. Second, we need to know the structure of a group. Finally, we need to know the mutual influences of the group members. This imposes the following requirements of functionallity of autonomous units. Autonomous units have to be able to 1) code each alternative in the Boolean algebra and relationships; 2) trasmit this information to each other in a manner that each 4 Euler method, we striaghtforwardly obtain difference equation (14) from differential equation (13): z(t + τ ) = z(t) + τ (b + iω)z(t), (14) where τ is a small time step. We set τ to 0.005 in all our simulations. Iterating difference equation (14) with z(0) = z0, we can approximate the analitical solution of eq.(13) and, conse- quently, of system (12). Now to obtain value of voltage variable y(t) at time t, we only need to take imaginary part of z(k). The real nearons produce a spike, once value of y(t) equal to or exceeds some preset threshold. However, this feature is not provided by the linear model. Therefore we slightly modify the original model by adding the 'spiking' condition: if y(t) ≥ threshold y(t) = 1.5 y(t + τ ) = 0.1 (15) (cid:40) A threshold value is to 1 throughout the simulations. There- fore, the ultimate model of resonate-and-fire neuron used in this study is described by the system (16): z(t + τ ) = z(t) + τ (b + iω)z(t) if y(t) ≥ 1 y(t) = 1.5 y(t + τ ) = 0.1 (16) We present the sample dynamics of two resonate-and- fire neurons, described by system (16), with different eigen- frequencies ω1 = 3π/2 and ω2 = 4π/3 in Fig. 4. It is illustrated that neurons with eigen-frequency ω1 spikes for the series of pulses with the same frequency and does not respond to the series of pulses with frequency ω2. The same is true for the second neurons regarding shift in roles of frequencies ω1 and ω2. Thus, we have described the mechanism of frequency se- lectivity. This can be used to enable multiple neurons to talk to each other via the same medium by means of Frequency Domain Multiplexing. However, the linear model has other important properties. The inhibitory pulses can also make resonate-and-fire neurons to spike, if the inhibitory pulses are applied with the eigen- frequency of the neuron (Fig. 5). However, it is not the final feature of this model. It is possible to make the neuron fire with series of pulses of different magnitudes. For example, let the magnitudes of the first, second and third pulses are 0.1, 0.4 and 0.6, respectively (Fig. 6). The same result will occur for the inhibitory pulses (Fig. 7). Since the neurons are selective to a certian frequency, it is possible to transfer signals of several frequencies through the same communication channel. This concludes the description of communication system. Next we consider the framework to manage the groups of autonomous units. Fig. 3. The Basic Control Schema of an Abstract Subject (BSCAS). unit could 'understand' the information address to it and at the same time all the units should be aware about all the information transfered from any unit to any unit. How to code and transfer this information in the groups of autonomous units is the main question discussed in the rest of this study. Therefore, the material presented hereafter is dedicated to the matter of how an autonomous unit can obtain the required information. III. INTRODUCING COMMUNICATIONS BETWEEN AUTONOMOUS UNITS In this study, we use the abstract autonomous units, which are capable of communicating with each other in the frequency domain. Therefore, these units can distinguish between several frequencies. Each frequency can be used as a carring frequency to transmit the information between the autonomous units. To be capable of distinguishing between various frequencies the autonomous units are supplied with frequency selective devices - resonators. There are many possible implimentations of the resonators. In this study, we suggest to used the Resonate-and-Fire linear neural model proposed by Izhikevich [15]. The choice of this model is justified by its simplicity in implementation and low computational complexity. Next, we provide brief overview of the Resonate-and-Fire linear model. A. Resonate-and-fire neurons The original linear model of resonate-and-fire neuron pro- posed by Izhikevich [15] is described by the system of two differential equations:(cid:40) x = bx − ωy y = ωx + by (12) where x is current-like or recovery vairable, and y is voltage- like or action potential variable, in terms of neuroscience. Both variables x and y are functions of time: x = x(t) and y = y(t). Notation x means derivative of x with respect to time t: x = ∂x/∂t. Paramater ω is the eigen-frequency, which is preferred or resonant frequency of the system, and represents frequency of subthreshold oscillations; parameter b is ananalog of damping factor in the linear damped oscillator. The value of paramter b is set to -0.1 throughout our simulations. For the purpose of numerical intergration the system (12) can be transformed into the form: z = (b + iω)z . (13) Then variables x and y are real and imaginary parts of complex variable z, respectively. Using eq.(13) and first-order Decision equation of a robotDecision Module Solutions : DBoolean Algebra ofAlternativesEnvironmentDecision equation of a humanInfluence ModuleSolution: DhRealization of an alternativeReflexive controlInfluencesSystem of Influence eqs.Structure of a GroupSolution: ZhSolutions5 Fig. 4. The Resonate-and-Fire neurons. T op: solid line illustrates resonanse with the input frequency ω1 = 3π/2, dashed line shows only subthreshold oscillations meaning that neuron does not respond to the frequency ω2 = 4π/3. Bottom: solid line illustrates resonanse with the input frequency ω2 = 4π/3, dashed line shows only subthreshold oscillations meaning that neuron does not respond to the frequency ω1 = 3π/2. The green and red vertical lines indicate the equal input pulses of magnitude 0.4. Green and red pulses are provided with frequencies ω1 = 3π/2 and ω2 = 4π/3, respectively. Each series of pulses starts 1 ms after the system onset. Threshold is set to 1. Parameter b is -0.1. Fig. 6. The selective responses of resonate-and-fire neurons to the excitatory series {0.1, 0.4, 0.6}. Fig. 5. The selective responses of resonate-and-fire neurons to the series of equal inhobitory pulses (magnatude -0.4). IV. BUILDING THE GROUPS OF AUTONOMOUS UNITS In this section we introduce a sketch of communication system to socialize autonomous units. A. Information Coding Each autonomous unit has several resonators tuned to par- ticular frequencies. Each resonator corresponds to a certian unit in the group. Therefore, the total number of resonators equals the total number of units in the group. For each unit, we reserve its unique frequency. Once the information is obtained from resonators with a frequency accosiated with this unit, it is that this unit is an addressee. Fig. 7. The selective responses of resonate-and-fire neurons to the inhobitory series {-0.1, -0.4, -0.6}. Using the resonate-and-fire neurons presented in the pre- vious section, we can transfer different types of information throught the network of autonomous units. In fact, we can transmit two types of information coded by 1) the kind of pulses (exictatory vs inhibitory), and 2) selecting different magnitude of pulses in the series. These two types of information are enough to model the groups in the Reflexive Game Theory. We consider a certain frequency to be the unique identifier of the autonomous unit in the group. Next, if the series pulses contains the excita- tory impulses, it is considered that two units are in alliance relationships, or they are in conflict otherwise. Finally, we can define a certian alternative of the Boolean Algebra by a certain seris of pulses. B. Receiving Informaiton in the Group So far, we understand how a certian unit can recieve information. The question remains how autonomous unit can INTERNAL STATE OF EACH UNIT REGARDING RELATIONSHIPS WITH TABLE II 6 OTHERS a - 0.63 0.09 b 0.81 - c 0.92 0.12 0.27 - a b c (conflict or alliance), it wants to install with other units. For example, we consider that the relationships are decided at random, meaning that at the very begining the units do not have any infomation about each other, except for the preferred frequencies. Therefore, this condition can be assumes as guessing. The human guessing based on no prior information has been describe from both theoretical and experimental points of views in [16], [17]. In the case of two options, one option (positive pole) is chosen with probability p ≈ 0.61, while another option (negative pole) is chosen with probability 1 − p ≈ 0.39. The concept of option's polarity has been first introduced by Lefebvre [16]. We consider alliance relationship to be positive pole and conflict relationship to be a negative pole. The alliance and conflict relationships are coded with codes {0.4,0.4,0.4} and {-0.4,-0.4,-0.4}, respectively. We call codes {0.4,0.4,0.4} and {-0.4,-0.4,-0.4} to be alliance and conf lict codes, respectively, if they are transmitted NOT on the pre- ferred frequency of the unit, which sends it. The alliance and conflict codes are chosen with probabilities 0.61 and 0.39, respectively. However, to install the relationship, the decisions of both units are needed. In other words, since units chosen the type of relationships independently from each other, it is possible that, for example, unit a sends conflict code to unit b, but unit b sends alliance code to unit a. Therefore, each unit has decided its own relationship, which is different from the one chosen by counterparty. Thus, the codes are different. We define that the alliance relationship is installed if and only if both units send alliance code to each other, the conflict relationship is installed otherwise. Thus, the relationship between units a and b is conflict. If we consider codes {0.4,0.4,0.4} and {-0.4,-0.4,-0.4} as logic 1 and 0, respectively, the alliance relationship can be defined as logic conjunction (AND) function, and conflict relationships as disjunction (OR) function. Example 1. Let us generate a group of three units with randomly chosen relationships. We use a uniform random number generator with interval (0,1). If the value of random variable exceeds 0.61, unit x generates conflict code {-0.4,- 0.4,-0.4} to some other unit, otherwise it generates alliance code {0.4,0.4,0.4} (Table II). The rows of Table II contain the decisions about relationships that each unit generated itself, but have not yet transmitted to other units. Therefore, Table II illustrates internal state of each unit. This internal state is yet not known by other units in the group. According to Table II, unit a will send conflict code to both Fig. 8. Receiving messages in the network. understand where the signal comes from or which unit sends it? For this purpose, we reserve a series of equal pulses {0.4,0.4,0.4}. Hereafter, we refer to the series of pulses as code or message. In particular, we call the code {0.4,0.4,0.4} to be identification-code (ID-code), if this code is transmitted by a certain unit on its own preferred frequency. Suggest, we have three units a, b and c. Each unit is char- acterized by its preferred frequency: ωa = 3π/2, ωb = 4π/3 and ωc = 5π/3. If autonomous unit a with eigen-frequency ωa decides to send some code to another unit, it first sends ID-code on the frequency ωa. Therefore, the corresponding neuron spikes in each autonomous units, and units b and c 'understand' that unit a wants to send a code. This can be considered as unit a attracts attention of the other units in the group. Then, after a short delay (0.5 sec) after spike on the frequency ωa, unit a sends a certain code on the frequency ωsubject, where subject can be either b or c. As an example, we consider that unit a wants to send its ID-code to unit c. Therefore, unit a first sends ID-code {0.4,0.4,0.4} on the frequency ωa to attract attention of other units: in units b and c the channels with frequency ωa show a spike (Fig. 8, top). Then, 0.5 sec after a spike on the frequency ωa , unit a sends the ID-code on the frequency ωc: in units b and c the channels with frequency ωc show a spike (Fig. 8, center). Since, frequency ωc is the frequency reserved for unit c, unit c receives ID-code from unit a. At the same time channel with frequency ωb shows no spike (Fig. 8, bottom), and unit b 'understands' that ID-code is not addressed to it. This way each unit in the groups can become completely awear about the whole information transmitted between any two units. Therefore, such communication schema provides all necessary information for application of the Reflexive Game Theory. C. How to Install Relationships in the Group Now we consider how to install relationship between units. Each unit decides on its own, which type of relationship TABLE III TRANSMITTED RELATIONSHIP CODES a - 0 1 b 0 - 1 c 0 1 - a b c 7 TABLE IV INFLUENCE MATRIX a a {α} {β} b {α} b {} c {} {β} c a b c units b and c (Fig. 9). Unit b will send conflict code to unit a and alliance code to unit c (Fig. 10). Unit c will send alliance code to both units b and c (Fig. 11). After the codes have been transmitted from each unit to each unit, the information from Table II become available to each unit in the group. Using the correlation between conflict and alliance codes and logic 1 and 0, we can rewrite Table II. Therefore, using informaiton from Table III together with conjunction and disjunction functions, we obtain the relation- ships installed between the units: units b and c are in alliance, while unit a is in conflict with both units b and c. Since, the informaiton about the relationships between units is now known by each unit in the group , we can construct the relationship graph (Fig. 1) of the Reflexive Game Theory and obtain the polynomial corresponding to this graph, which is a + bc. D. How to Transfer Information about the Influences In this section, we illustrate how to transmit influences of unit on each other. We use the same approach described in the previous section. The only difference is that instead of the alliance or conflict codes, unit transmits some code associated with a particular alternative. Example 2. Suggest, we have Boolean algebra of four alternatives: 1 = {α, β}, {α}, {β} and 0 = {}. We arbitrary assign a certain code to each alternative: 1) code {0.2, 0.3, 0.7} corresponds to alternative 1 = {α, β}; 2) code {0.7, 0.3, 0.2} corresponds to alternative 0 = {}; 3) code {0.5, 0.2, 0.5} corresponds to alternative {α}; and 4) code {0.3, 0.6, 0.3} corresponds to alternative {β}. To make reference easier, we refer to each code as 'altern- code {0.7, 0.3, 0.2} is called unit-code; code {0.2, 0.3, 0.7} is zero-code; code {0.5, 0.2, 0.5} is referred as {α}-code; and code {0.3, 0.6, 0.3} is {β}-code. We assume that unit a makes influences {α} and 0 = {} on units b and c, respectively (Fig. 12). Unit b makes influence {α} on both units a and c (Fig. 13). Unit c makes influences {α} and {β} on units a and b, respectively (Fig. 14). tive name'-code: E. RGT Inference Therefore, after all the influences have been transmitted, we obtain the influence matrix (Table IV). Thus, each unit now has complete information to apply the RGT inference schema based on the decision equations [13]. The canonical form of decision equation for unit a is a = a + bca and the corresponding solution interval is 1 ⊇ a ⊇ bc. The canonical form of decision equation for unit b is b = (a+c)b+ab and the corresponding solution interval is (a+c) ⊇ b ⊇ a. The canonical form of decision equation for unit c is c = (a + b)c + ac and the corresponding solution interval is (a + b) ⊇ c ⊇ a. Under the given influences, the choice of unit a is define by the interval 1 ⊇ a ⊇ {}. The solution interval for unit b turns into equality b = {α}. The choice of unit a is define by the interval {β} ⊇ a ⊇ {}. V. COOPERATIVE BEHAVOIR OF THE AUTONOMOUS UNITS CONTROLLED BY THE RGT ALGORITHMS Until now, we have briefly described the gist of the RGT, communication system for the autonomous units and illus- trated how the units can use the information obtained by means of the communication system for the RGT inference. In this section we consider how units can make mutual influences in order to achieve a particular goal in a cooperative behavior task. Example 3. Consider four autonomous units (robots). Let these robots are functioning by using the electric accumulators. There's a chaging pool in the restricted perimeter (castle), which has only one entrance (gate). There are three robots a, b and c in the castle, and one robot d outside the perimeter. Let robots a, b and c are in alliance with each other and in conflict with robot d. Robots a, b and c are locked inside the perimeter, but they can open the gate if each robot agrees so. The power source has a limited capacity is only 25% full. The accumulators of robots a, b and c are 50% full, while accumulator of robot d is only 10% full. The power source should be regenerated. However, the time left until regeneration exceed the life-time of the robot d's accumulator. Consider 'open the gates' is action 1 (unit-code). Then Boolean algebra of alternatives contains two elements 1 - to open the gates, and 0 - 'don't open'. From the point of view of Game Theory, it is clearly out of utility to share the power source with 'exhausted' enemy robot, but let's have look what should happen if RGT inference is applied. The diagonal form transformation is present as follows: [a][b][c] [abc] +[d] [abc + d] = abc + d 8 Fig. 9. Activity of unit a. Fig. 10. Activity of unit b. Fig. 11. Activity of unit c 9 Fig. 12. Transmission of influnces produced by unit a. Fig. 13. Transmission of influnces produced by unit b. Fig. 14. Transmission of influnces produced by unit a. obviously demostrate principles of the RGT on the particular autonomous units. 10 REFERENCES [1] Lefebvre, V.A. (1965) The basic ideas of reflexive game's logic. Problems of research of systems and structures, pp. 73 -- 79 (in Russian). [2] Lefebvre, V.A. (1967) Conflicting structures [in Russian]. [3] Lefebvre, V.A. (2015) Conflicting structures. Leaf & Oaks, Los Angeles. [4] Rizzolatti, R. (2005) The mirror neuron system and its function in humans. Anat Embryol, 210, pp. 419-421. [5] Gallese, V., Fadiga, L., Fogassi, L. and Rizzolatti, G. (1996) Action recognition in the premotor cortex. Brain, 119, pp. 593-609. [6] Ekman, P., and Davidson, R. J. (1993) Voluntary changes regional brain activity. Psychological Science, 4, 5, pp. 342-345. [7] Levenson, R. W., Ekman, P., and Friesen, W. V. (1990) Voluntary facial action generates emotion specific autonomic nervous system activity. Psychophysiology, 27, 4, pp. 363-384. [8] Lefebvre, V.A. (2010) Lecture on Reflexive Game Theory. Leaf & Oaks, Los Angeles. [9] Lefebvre, V.A. (2009) Lecture notes on the Reflexive Game Theory. Cognito-Center, Moscow (in Russian). [10] Batchelder, W.H., Lefebvre, V.A. (1982) A mathematical analysis of a natural class of partitions of a graph. Journal of Mathematical Psychology, 26, 124-148. [11] Lefebvre, V.A. (1982) Algebra of Conscience. D. Reidel, Holland. [12] Lefebvre, V.A. (2001) Algebra of Conscience. 2nd Edition. Holland: Kluwer. [13] Tarasenko, S. (2011) Modeling mixed groups of humans and robots with Reflexive Game Theory. In Lamers, M.H. and Verbeek, F.J. (Eds): HRPR 2010, LNICST 59, pp. 108-117. [14] Tarasenko, S. (2010) The Inverse Task of the Reflexive Game Theory. [15] Izhikevich, E. M. (2001) Resonate-and-fire neurons. Neural Networks, 14, pp. 883-894. Mellen, Ney York. [16] Lefebvre, V.A. (2006) Research on Bipolarity and Reflexivity. Edwin [17] Tarasenko, S., and Inui, T. (2009) Blind Choice. Perceptual and Motor Skills, 109, 3, pp. 791-803. Fig. 15. Relationship graph for Example 3. The resultant decision equations in canoncal forms are: a = (bc + d)a + da b = (ad + c)b + db c = (ab + d)c + dc d = d + abcd (17) (18) (19) (20) All three robot inside the perimeter are willing not to open the gates and they inlfuence by zero-code on each other and on robot d. Then the deicision intervals for each robot are (bc + d) ⊇ a ⊇ d ⇒ (0 + d) ⊇ a ⊇ d ⇒ a = d; (ac + d) ⊇ b ⊇ d ⇒ (0 + d) ⊇ b ⊇ d ⇒ b = d; (ab + d) ⊇ c ⊇ d ⇒ (0 + d) ⊇ b ⊇ d ⇒ c = d; 1 ⊇ d ⊇ abc ⇒ 1 ⊇ d ⊇ 0. Therefore, the decision of all three robot inside the perime- ter is defined by the influence of robot d. Therefore, it robot d makes influence 1, then all three robot will agree to open the gate, and since this is required condition, they will open the gate. On the other hand, robot d has a freedom of choice. Therefore, if robot d asks for help, other three robots should open the gate and allow access to the power supply source. VI. DISCUSSION In this study, we have presented the structure of autonomous units, which allows these units to install communication with each other and create groups. As the basis for communication network, we use resonate-and-fire neurons, which are em- ployed as signal receivers. The main feature of resonate-and- fire neurons is their selectivity to a particular frequency, which is eigen-frequency of the neuron. Therefore, it is possible to send different codes through the same network and be sure that each unit understands the message addressed exclusisvely to it. I do not discuss here physical mechanisms of generating signals. We illustrated how it is possible to arrage a group of three units as a communication network. I also showed how to code different messages such as sender identification and Boolean algebra alternatives. We concluded with examples of how a simple group can be arranged based on the information about relationships between units and showed how to transmit the information about influences in the group. Thus, having received the information about the structure of the group and the mutual influences, each autonomous unit can apply algorithms of RGT inferences. Thus, each unit can make both its own choice and also predict the possible choices of other members of the group. Therefore, the fusion of the proposed sketch of communication network with the RGT inference allows to
1012.1651	1	1012	2010-12-08T00:23:53	The Rule Responder eScience Infrastructure	[ "cs.MA" ]	To a large degree information and services for chemical e-Science have become accessible - anytime, anywhere - but not necessarily useful. The Rule Responder eScience middleware is about providing information consumers with rule-based agents to transform existing information into relevant information of practical consequences, hence providing control to the end-users to express in a declarative rule-based way how to turn existing information into personally relevant information and how to react or make automated decisions on top of it.	cs.MA	cs	Rule Responder: A Rule-Based Semantic eScience Service Infrastructure Adrian Paschke and Zhili Zhao Freie Universitaet Berlin, Germany [email protected] Abstract. To a large degree information and services for chemical e- Science have become accessible -anytime, anywhere -but not necessarily useful. The Rule Responder eScience middleware is about providing in- formation consumers with rule-based agents to transform existing infor- mation into relevant information of practical consequences, hence pro- viding control to the end-users to express in a declarative rule-based way how to turn existing information into personally relevant information and how to react or make automated decisions on top of it. 1 Introduction The Semantic Web builds upon XML as the common machine-readable syntax to structure content and data, upon RDF as a simple language to express prop- erty relationships between arbitrary resources identified by URIs, and ontology languages such as RDFS or OWL as a means to define rich vocabularies (on- tologies) which are then used to precisely describe resources, their relations and their semantics. This prepares an infrastructure to share the relevant meaning of content and leads to a more machine-processable and relevant Web. Many bioinformatics projects, such as UniProt, Tambis, FungalWeb, Yeast- Hub, BioPax have meanwhile adopted the Semantic Web approach (in particular the RDF standard) and large ontologies such as the Gene Ontology are provided as RDFS or OWL ontologies. This has been utilized by several bioinformatics projects, such as W3C HCLS RDF or Bio2RDF, to solve the old problem of distributed heterogeneous data integration in health care and life sciences. The goal of this article is to show how the Rule Responder approach can be used to build a flexible, loosely-coupled and service-oriented eScience infras- tructure which allows wrapping the existing web data sources, services and tools by rule-based agents which access and transform the existing information into relevant information of practical consequences for the end-user. 2 A Rule-Based Pragmatic Agent Web Model for Virtual eScience Infrastructures A virtual eScience infrastructure consists of a community of independent and often distributed (sub-) organizations which are typically represented by an or- ganizational agent and a set of associated individual agents. The organizational 0 1 0 2 c e D 8 ] A M . s c [ 1 v 1 5 6 1 . 2 1 0 1 : v i X r a 2 Rule Responder: A Rule-Based Semantic eScience Service Infrastructure agent might act as a single agent towards other internal and external individual or organizational agents. In particular, a virtual organization's agent can be the single (or main) point of entry for communication with the "outer" world. In the architecture of the eScience Agent Web model(Figure 1), the syntactic level controls the appearance and access of syntactic information resources such as HTML pages. The representation languages such as XML, RDF and OWL on the semantic level make these Web-based resources more readable and process- able not only to humans, but also to computers to infer new knowledge. Finally, the pragmatic and behavioral level defines the rules that how information is used and describes the actions in terms of its pragmatic aspects. These rules e.g. transform existing information into relevant information of practical conse- quences, trigger automated reactions according to occurred complex events, and derive answers from the existing syntactic and semantic information resources. Fig. 1. A Pragmatic Agent Web for Virtual Organizations In this paper we focus on the pragmatic and behavioral layer and build it upon existing technologies and common language formats of the Semantic Web such as HTML/XML Web pages, RDF/RDFS, OWL and etc. We assume that there is already a critical mass of such data sources on the semantic and syntactic layer. Furthermore, we integrate data and functionality from legacy applications. 3 Distributed Rule Responder Agent Services The core parts of the distributed Rule Responder Architecture for the eScience Agent Web are the common platform-independent rule interchange format (RuleML), the communication middleware (ESB) and the execution environments (Prova). Rule Responder: A Rule-Based Semantic eScience Service Infrastructure 3 The Rule Markup Language (RuleML) [1] is a modular, interchangeable rule specification on standard to express both forward and backward rules for deduc- tion, reaction, rewriting, and further inferential-transformational tasks. Reaction RuleML [2] is a sublanguage of RuleML and incorporates various kinds of pro- duction, action, reaction, and KR temporal/event/action logic rules as well as (complex) event/action messages. To seamlessly handle message-based interactions between the responder agents and with other applications, an enterprise service bus (ESB), the Mule open- source ESB [3] is used. The ESB allows deploying the rule-based agents as highly distributable rule inference services installed as Web-based endpoints in the Mule object broker and supports the Reaction RuleML based communication between them. Mule is based on ideas from ESB architectures, but goes beyond the typical definition of an ESB as a transit system for carrying data between applications. Prova [4],which is a highly expressive Semantic Web rule engine to the ref- erence implementation for complex agents with complex reaction workflows, de- cision logic and dynamic access to external Semantic Web data sources.The current version of Prova follows the spirit and design of the recent W3C Seman- tic Web initiative and combines declarative rules, ontologies and inference with dynamic object-oriented Java API calls and access to external data sources such as relational databases or enterprise applications and IT services. 4 Rule Responder Use Case The discovery process for a researcher to find the Alzheimer's drug target candi- ates is very complex and time-consuming.He/she first discovers from Uniprot, the W3C HCLS KB and the SWAN data that Beta amyloidal in various forms, and in particular ADDLs, which are good therapeutic targets. He/she then searches the PubMed database about articles on ADDLs and ranks the results to find the top location, which is Evanston, and the top author, who is William Klein. From this, the researcher makes the hypothesis that William Klein works in Evanston, and simply proves it using Google. Finally, the researcher queries the EMBI-EBI database for the patents addressing ADDLs as therapeutic target for AD and concludes that William Klein who also holds two patents is one of the top experts in ADDLs research.Implicitly, the researcher executes the following rule: IF a Person has most publications in the Field and one or more Patents in the field THEN the Person is an expert for this Field. Figure 6 shows how this rule can be implemented in terms of Rule Responder agents. The HCLS Rule Responder agent service (Figure 2) implements the main logic of the eScience infrastructure and acts as the main communication endpoint for external agents. Its the rule code defines the public interfaces to receive requests (queries, tasks) to the eScience infrastructure and the logic to look up the respective source agents and delegate requests to them in order to answer the queries and fulfill the tasks. Each existing legacy data sources / service is wrapped by a Rule Responder source agent which runs a Prova rule engine. The agents rule base comprises the local rule interface descriptions, i.e. the rule 4 Rule Responder: A Rule-Based Semantic eScience Service Infrastructure functions which can be queried by other agents of the eScience infrastructure, the respective transformation rules to issue queries to the platform-specic services and access the heterogeneous local data sources, and the rule logic to process incoming requests and derive answers / information from the local knowledge. Fig. 2. Rule Responder HCLS eScience Infrastructure 5 Conclusion With Rule Responder HCLS we have evolved a rule-based approach which facilitates easy heterogeneous systems integration and provides computation, database access, communication, web services, etc. This approach preserves local anonymity of local agent nodes including modularity and information hiding and provides much more control to users with respect to the relatively easy declar- ative rule-based programming techniques. The rules allow specifying where to access and process information, how to present information and automatically react to it, and how to transform the general information available from existing data sources on the Web into personally relevant information accessible via the eScience infrastructure. The Rule Responder eScience infrastructure is available online at responder.ruleml.org. References 1. H. Boley. The rule-ml family of web rule languages. In 4th Int. Workshop on Prin- ciples and Practice of Semantic Web Reasoning, Budva, Montenegro, 2006. 2. A. Paschke, et al. Reaction ruleml, http://ibis.in.tum.de/research/reactionruleml/ 3. Mule. Mule enterprise service bus, http://mule.codehaus.org/display/mule/home,2006. 4. A. Kozlenkov, et al. Prova, http://prova.ws, 2006.
0712.0744	1	0712	2007-12-05T15:02:19	Computational Chemotaxis in Ants and Bacteria over Dynamic Environments	[ "cs.MA", "cs.AI", "q-bio.PE", "q-bio.QM" ]	Chemotaxis can be defined as an innate behavioural response by an organism to a directional stimulus, in which bacteria, and other single-cell or multicellular organisms direct their movements according to certain chemicals in their environment. This is important for bacteria to find food (e.g., glucose) by swimming towards the highest concentration of food molecules, or to flee from poisons. Based on self-organized computational approaches and similar stigmergic concepts we derive a novel swarm intelligent algorithm. What strikes from these observations is that both eusocial insects as ant colonies and bacteria have similar natural mechanisms based on stigmergy in order to emerge coherent and sophisticated patterns of global collective behaviour. Keeping in mind the above characteristics we will present a simple model to tackle the collective adaptation of a social swarm based on real ant colony behaviors (SSA algorithm) for tracking extrema in dynamic environments and highly multimodal complex functions described in the well-know De Jong test suite. Later, for the purpose of comparison, a recent model of artificial bacterial foraging (BFOA algorithm) based on similar stigmergic features is described and analyzed. Final results indicate that the SSA collective intelligence is able to cope and 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, while outperforming BFOA in adaptive speed. Results indicate that the present approach deals well in severe Dynamic Optimization problems.	cs.MA	cs	 Computational Chemotaxis in Ants and Bacteria over Dynamic Environments Vitorino Ramos, Carlos Fernandes, Agostinho C. Rosa and Ajith Abraham Abstract— Chemotaxis can be defined as an innate behavioural response by an organism to a directional stimulus, in which bacteria, and other single-cell or multicellular organisms direct their movements according to certain chemicals in their environment. This is important for bacteria to find food (e.g., glucose) by swimming towards the highest concentration of food molecules, or to flee from poisons. Based on self-organized computational approaches and similar stigmergic concepts we derive a novel swarm intelligent algorithm. What strikes from these observations is that both eusocial insects as ant colonies and bacteria have similar natural mechanisms based on stigmergy in order to emerge coherent and sophisticated patterns of global collective behaviour. Keeping in mind the above characteristics we will present a simple model to tackle the collective adaptation of a social swarm based on real ant colony behaviors (SSA algorithm) for tracking extrema in dynamic environments and highly multimodal complex functions described in the well-know DeJong test suite. Then, for the purpose of comparison, a recent model of artificial bacterial foraging (BFOA algorithm) based on similar stigmergic features is described and analyzed. Final results indicate that the SSA collective intelligence is able to cope and 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, while outperforming BFOA in adaptive speed. Results indicate that the present approach deals well in severe Dynamic Optimization problems. Index Terms—Swarm Intelligence and Perception, Social Cognitive Maps, Social Foraging, Self-Organization, Distributed Search and Optimization in Dynamic Environments. I. INTRODUCTION S WARM Intelligence (SI) is 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 Vitorino Ramos, Carlos Fernandes and Agostinho Rosa, are with LaSEEB-ISR Evolutionary Systems and BioMedical Eng. Lab. , IST - Technical University of Lisbon (IST), Av. Rovisco Pais, 1, TN 6.21, 1049- 001, : e-mails author (corresponding PORTUGAL Lisbon, [email protected], [email protected] , [email protected]). Second author work was supported in part by FCT-PRAXIS XXI, Ministério da Ciência, Tecnologia e Ensino Superior , under a PhD fellowship. Ajith Abraham is with the School of Computer Science and Engineering, Chung- Ang University, Seoul, South Korea. (e-mail: [email protected] ). provision of a global model (Stan Franklin, Coordination without Communication, talk at Memphis Univ., USA, 1996). The well-know bio-inspired computational paradigms know as ACO (Ant Colony Optimization algorithm [5]) based on trail formation via pheromone deposition / evaporation, and PSO (Particle Swarm Optimization [14]) are just two among many successful examples. Yet, and in what specifically relates to the biomimicry of these and other computational models, much more can be of useful employ, namely the social foraging behavior theories of many species, which can provide us with consistent hints to algorithmic approaches for the construction of social cognitive maps, self-organization [1,6], coherent swarm perception and intelligent distributed search, with direct applications in a high variety of social sciences and engineering fields [25→30]. In the present work, we will address the collective adaptation of a social community to a cultural (environmental, contextual) or informational dynamical landscape, represented here – for the purpose of different experiments – by several 3D mathematical functions that change over time. Our precise and final goal will be to keep track of extrema on those environments. For instance, typical applications of evolutionary optimization in static environments involve the approximation of the extrema of functions. On the contrary, for dynamic environments, the interest is not to locate the extrema but to follow it as closely as possible [12]. Flocks of migrating birds and schools of fish are familiar examples of spatial self-organized patterns formed by living organisms through social foraging. Such aggregation patterns are observed not only in colonies of organisms as simple as single-cell bacteria, as interesting as social insects like ants and termites as well as in colonies of multi-cellular vertebrates as complex as birds and fish but also in human societies [8]. Wasps, bees, ants and termites all make effective use of their environment and resources by displaying collective “swarm” intelligence. For example, termite colonies build nests with a complexity far beyond the comprehension of the individual termite, while ant colonies dynamically allocate labor to various vital tasks such as foraging or defense without any central decision-making ability [5]. Slime mould is another perfect example. These are very simple cellular organisms with limited motile and sensory capabilities, but in times of food shortage they aggregate to form a mobile slug capable of transporting the assembled individuals to a new feeding area. Should food shortage persist, they then form into a fruiting body that disperses their spores using the wind, thus ensuring the survival of the colony [18]. New research suggests that microbial life can be even richer: highly social, intricately networked, and teeming with interactions. Bassler [2] and other researchers have determined that bacteria communicate using molecules comparable to pheromones, as ant colonies so often do. By tapping into this cell-to-cell network, microbes are able to collectively track changes in their environment, conspire with their own species, build mutually beneficial alliances with other types of bacteria, gain advantages over competitors, and communicate with their hosts - the sort of collective strategizing typically ascribed to bees, ants, and people, not to bacteria. Eshel Ben-Jacob [4] indicate that bacteria have developed intricate communication capabilities (e.g. quorum-sensing, chemotactic signalling and plasmid exchange) to cooperatively self-organize into highly structured colonies with elevated environmental adaptability, proposing that they maintain linguistic communication. Meaning-based communication permits colonial identity, intentional behaviour (e.g. pheromone-based courtship for mating), purposeful alteration of colony structure (e.g. formation of fruiting bodies), decision-making (e.g. to sporulate) and the recognition and identification of other colonies – features we might begin to associate with a bacterial social intelligence. Such a social intelligence, should it exist, would require going beyond communication to encompass unknown additional to generate intracellular processes inheritable colonial memory and commonly shared genomic context. Moreover, Eshel [3] argues that colonies of bacteria are able to communicate and even alter their genetic makeup in response to environmental challenges, asserting that the lowly bacteria colony is capable of computing better than the best computers of our time, and attributes to them properties of creativity, intelligence, and even self-awareness. These self- organizing distributed capabilities were also found in plants. Peak and co-workers [23] point out that plants may regulate their uptake and loss of gases by distributed computation – using information processing that involves communication between many interacting units (their stomata). As described, leaves have openings called stomata that open wide to let CO2 in, but close up to prevent precious water vapour from escaping. Plants attempt to regulate their stomata to take in as much CO2 as possible while losing the least amount of water. But they are limited in how well they can do this: leaves are often divided into patches where the stomata are either open or closed, which reduces the efficiency of CO2 uptake. By studying the distributions of these patches of open and closed stomata in leaves of the cocklebur plant, Peak et al. [23] found specific patterns reminiscent of distributed computing. Patches of open or closed stomata sometimes move around a leaf at constant speed, for example. What’s striking is that it is the same form of mechanism that is widely thought to regulate how ants forage. The signals that each ant sends out to other ants, by laying down chemical trails of pheromone, enable the ant community as a whole to find the most abundant food sources. Wilson [32] showed that ants emit specific pheromones and identified the chemicals, the glands that emitted them and even the fixed action responses to each of the various pheromones. He found that pheromones comprise a medium for communication among the ants, allowing fixed action collaboration, the result of which is a group behaviour that is adaptive where the individual’s behaviours are not. II. SELF-ORGANIZATION AND STIGMERGY Many structures built by social insects are the outcome of a process of self-organization [27,28], 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 [11]. A major mediating factor is stigmergy [31], 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 adaptive learning and perception [25→30]. memory, cooperative 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. The phrasing of his introduction of the term is worth noting (translated to English in [11]): The coordination of tasks and the regulation of constructions do not depend directly on the the workers, but on 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). Keeping in mind the above characteristics (section I and II) we will present a simple model to tackle the collective adaptation of a social swarm based on real ant colony behaviors (Swarm Search Algorithm SSA - section III / results on section IV). Then, and for the purpose of comparison, a recent model of artificial bacterial foraging [22,17] (Bacterial Foraging Optimization Algorithm - BFOA) based on similar stigmergic features is described and analyzed (section V). Final results indicate that the SSA collective intelligence is able to cope and 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, outperforming BFOA. 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. TABLE I HIGH-LEVEL DESCRIPTION OF THE SWARM SEARCH ALGORITHM PROPOSED F0a - 3D view F0a - 2D view F0b - 3D view F0b - 2D view F1 - 3D view F1 - 2D view F2 - 3D view F2 - 2D view F3 - 3D view F3 - 2D view F4 - 3D view F4 - 2D view F6 - 2D view F6 - 2D view Fig. 1. Three-dimensional views (3D) and respective landscapes views (2D) of several test functions used in our analysis [38]. White pixels correspond to high peaks, while darker ones represent deep (F0-F4) or holes (F6). valleys Check table II in section 4. /* Initialization / For all agents do Place agent at randomly selected site End For / Main loop / For t = 1 to tmax do For all agents do / According to Eqs. 1 and 2 (section 3) / Compute W(σ) and Pik Move to a selected neighboring site not occupied by other agent / According to Eq. 3 (section 3) / Increase pheromone at site r: Pr= Pr+[+p(Δ[r]/Δmax)] End For Evaporate pheromone by K, at all grid sites End For Print location of agents Print pheromone distribution at all sites / Values of parameters used in experiments / k = 0.015,  = 0.07, =3.5, γ=0.2, p = 1.9, tmax = 500, 600, 1000 or 1150 steps. / Useful references / Check [25], [27], [7], [21] and [20]. 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 [7], 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 . 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,) to the next (r,*) an instant later. In previous works by Millonas [21,20], 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.1): (1) t = 0 t = 1000 t = 1000 Fig.2. Pheromone maxF0a. distribution (Social Cognitive Maps) for t=0, 50, 100, 500 and 1000 time steps, of 3000 ants exploring function F0a on a 100 x 100 toroidal grid (1st and 3rd column: darker pixels correspond to higher concentrations). Columns 2 and 4 correspond to the geographical place where agents are situated (each black pixel is an ant). At t=100, the highest peak is already surrounded by agents while convergence proceeds. Processing time equals to 54 s (1200 Mhz Intel Processor). t = 50 t = 100 t = 0 t = 50 t = 100 t = 500 t = 500 III. A SWARM MODEL FOR FORAGING IN DYNAMIC ENVIRONMENTS 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 [29]. There is no direct communication between the organisms but a type of indirect communication through the pheromonal field. t = 0 t = 50 t = 0 t = 50 t = 500 t = 500 t = 150 t = 150 11W t = 1000 t = 1010 t = 1080 t = 1100 t = 1000 t = 1010 t = 1080 t = 1100 t = 250 t = 250 t = 300 t = 300 t = 350 t = 400 t = 350 t = 400 t = 1150 t = 1150 Fig. 3. maxF0a => maxF0b. Social evolution from maximizing function F0a to maximizing function F0b. In the first 1000 time steps the ant colony explores function F0a, while suddenly at t=1001, function F0b is used as the new habitat. Pheromone distribution (Social Cognitive Maps) for t = 0, 500, 1000, 1010, 1050, 1080, 1100 and 1150 time steps, of 3000 ants exploring function F0a and F0b on a 100 x 100 toroidal grid are shown. Already at t=1010, the old highest peak on the right suffers a radical erosion, on the presence of ants (they start to explore new regions). t = 500 t = 500 Fig. 4. maxF0a => minF0a. Maximizing function F0a during 250 time steps and then min imizing t  251. Pheromone it for distribution (Social Cognitive Maps) for t = 50, 150, 250, 300, 350, 400, 450 and 500 time steps, of 2000 ants exploring function F0a on a 100 x 100 toroidal grid are shown. Already at t=300, the highest peak on the right suffers a radical erosion, on the presence of ants starting to explore new regions. As time passes the majority of the colony moves to the deep valley, on the left. Parameters are different from those used in Figs. 2-3 (check table III). t = 20 t = 400 t = 400 t =2 0 t = 100 t = 100 t = 500 t = 500 t = 300 t = 300 t = 600 t = 600 t = 320 t = 320 Fig. 5. minF6 => maxF0a. Minimizing function F6 during 300 time steps and then maximizing function F0a t  301. for (Social distribution Pheromone Cognitive Maps) for t = 20, 100, 300, 320, 400, 500, and 600 time steps, of 3000 ants exploring function F6 and F0a on a 100 x 100 toroidal grid are shown. Parameters are different from those used in Figs. 2-3 (check table III). 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 (r). The parameter  is associated with the osmotropotaxic sensitivity, recognised by Wilson [32] as one of two fundamental different types of ant’s sense-data 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 [7]). 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. (2) (3) In addition to the former equation, there is a weighting factor w(), where  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  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 cell k to cell i are given by Pik (Eq. 2, [7]), where the notation j/k indicates the sum over all the surrounding cells j which are in the local neighbourhood of k. i measures the magnitude of the difference in orientation for the previous direction at time t-1. That is, since we use a neighbourhood composed of the cell and its eight neighbours, i 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 =0.07 (pheromone deposition k=0.015 rate), =3.5 evaporation (pheromone rate), (osmotropotaxic sensitivity) and =0.2 (inverse of sensorycapacity), where the emergence of well defined networks of trails were possible. Except when indicated, these values will remain in the following framework. As an additional condition, each individual leaves a constant amount kjjjiiikwWwWP/maxipT  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   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  used in [7], a term not constant is included. This upgrade can significantly change the expected ant colony cognitive map (pheromonal field). The strategy fo llows an idea implemented earlier by Ramos [25,26], 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 (fig. 1) instead of a 2D habitat, and the pheromone update takes in account not only the some local pheromone distribution as well as characteristics 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 3. In this equation, Δmax = zmax – zmin , being zmax the maximum altitude found by 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 landscape, then we should instead use Δ[i] = zi – zmin . Finally, please 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]/max 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 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, memory and distributed learning by the community [29]. TABLE II CLASSICAL TEST FUNCTIONS USED IN OUR ANALYSIS FROM MATLAB [24] Function ID Equation F0a F0b F1 F2 F3 F4 F5 F6 IV. EXPERIMENTAL SETUP AND RESULTS In order to test the dynamical behaviour of this new Swarm Search algorithm presented earlier in section 3 (pseudo-code in table I), we have used classical test functions (table II) drawn from the literature in Genetic Algorithms, Evolutionary strategies and global optimization [24], several of them graphically accessible in fig. 1. Function F0a represents one deep valley and one peak, while F0b his the opposite. Function F1 represents De Jong’s function 1 and his one of the simplest. It is continuous, convex and unimodal; xi is in the interval [- TABLE III PARAMETERS USED FOR DIFFERENT TEST SETS Fig. 2 3 4 5 N ants 3000 3000 2000 3000 tmax 1000 1150 500 600 k 0.015 0.015 1.000 1.000  0.07 0.07 0.10 0.01  3.5 3.5 3.5 3.5 γ 0.2 0.2 0.2 0.2 p 1.93 1.93 1.90 1.90 5.12; 5.12] and the global minimum is at xi=0. Function F2 represents an axis parallel hyper-ellipsoid similar to De Jong’s function 1. It is also know as the weighted sphere model. Again it is continuous, convex and unimodal in the interval xi → [-5.12; 5.12], with global minimum at xi=0. Function F3 represents an extension of the axis parallel hyper-ellipsoid (F2), also know as Schwefel’s function 1.2. With respect to the coordinate axes this function produces rotated hyper- ellipsoids; xi is in the interval [-65.536; 65.536] and the global minimum is at xi=0. Likewise F2, it is continuous, convex and unimodal. Function F4 represents the well-know Rosenbrock’s valley or De Jong’s function 2. Rosenbrock’s valley is a niixaexxf122.010.niixbexxf122.010.niixxf121niiaxixf121.niijjbxxf121111222121.100niiiixxxxfniiixxnxf126..2cos.10.10niiixxxf17sin. classic optimization problem. The global optimum is inside a long, narrow, parabolic shaped flat valley. To find the valley is trivial, however convergence to the global optimum is difficult and hence this problem has been repeatedly used in assess the performance of optimization algorithms; xi is in the interval [- 2.048; 2.048] and the global minimum is at xi=0. Function F5 represents the Rastrigin’s function 6. This function is based on De Jong’s function 1 with the addition of cosine modulation to produce many local minima. Thus, the test function is highly multimodal. However, the location of the minima are regularly distributed. As in F1, xi is in the interval [-5.12; 5.12] and the global minimum is at xi=0. Finally, F6 represents Schwefel’s function 7, being deceptive in that the global minimum is geometrically distant, over the parameter space, from the next best local minima. Therefore, the search algorithms are potentially prone to convergence in the wrong direction; xi is in the interval [-500; 500] and the global minimum is at xi=420,9687 while f(x)=n.418,9829. In our tests, n=2. Within this specific framework we have produced several run tests using different test functions, some of which are presented here trough figures 2 to 5. The parameters used are shown on table 3. The simplest test was the first one (fig.2) where we forced the colony to search for the maximal peak in function F0a, during 1000 time steps. The other tests were harder , that is dynamic, since they include not only different purposes simultaneously (maximizing and minimizing), tracking different extrema, as well as different landscapes that changed dynamically on intermediate swarm search stages (e.g., fig. 3, 4 and 5). V. SWARM SEARCH VERSUS BACTERIAL FORAGING ALGORITHMS In order to further analyze the collective behavior of the present proposal, we performed a comparison between the ant- like Swarm Search Algorithm (SSA) and the Bacterial Foraging Optimization Algorithm (BFOA), on the dominion of function optimization. BFOA was selected since it represents an earlier proposal for function optimization as well based on natural foraging capacities. Presented by Passino at IEEE Control Systems Magazine in 2002 [22] and later that year in the Journal of Optimization Theory and Applications [17], the author for the purpose of a simple but powerful illustrative example, used his algorithm to find the minimum of two complex functions Jcc, described in [22], page 60. Further material, as the MATLAB code of his algorithm and the tri- dimensional functions experimented, can also be found on the web address of a recent book from the same author (Biomimicry for Optimization, Control and Automation, 2005), London, Springer-Verlag, at UK, http://www.ece.osu.edu/ ~passino/ICbook/ ic_index.html. Passino uses S=50 bacteria-based agents, during four genera- tions. In each generation, and has a requirement of his algorithm, each agent enters a chemotaxis loop (see page 61 [22]), performing Nc=100 chemotactic (foraging) steps. Passino F1 3D Passino F1 2D Passino F2 3D Passino F2 2D t = 100 t = 100 t = 200 t = 200 t = 100 t = 200 t = 100 t = 200 t = 300 t = 400 t = 300 t = 400 t = 300 t = 300 t = 400 t = 400 t = 100 t = 400 t = 300 t = 200 Fig. 6. In the first row the test functions used by Passino [22,17]. In the second and third rows, BFOA minimizing results respectively for F1 and F2. The graphics show the bacterial motion trajectories (using 50 bacteria -like agents). In the fourth and fifth rows, SWARM -SEARCH algorithm (SSA) minimizing results respectively for F1 and F2, and for the same foraging time period. The graphics shows the pheromone distribution. In the last row, SSA is requested to deal with two contradictory goals, i.e. to min imize F1 and then to maximize it. In all these tests, SSA has used 50 ant-like agents. Check main text for the parameters used. Hab itat size equals 2 x [0,30]. Thus the algorithm – for the precise application – runs for t=400 time steps, which make us believe that a fair comparison can be make in regard of the parameter values we use. The two functions represent what Passino designates by nutrient concentration landscapes (see fig. 6, first row – the web address also contains his MATLAB code used in the two functions, where Nutrientsfunc.m and Nutrientsfunc1.m are function F2 represented by different weights). His (Nutrientsfunc1.m) has a zero value at [15,15] and decreases to successively more negative values as you move away from that point, reaching a plateau with the same value. Moreover, and for the purpose of discrete function optimization, Passino [22,17] represented both functions by a discrete lattice (as well as us in our past tests) with a size of 30 x 30 cells over the optimization domain (each cell has a correspondent z or Jcc value). For these reasons and in order to keep a coherent comparison, we shall use 50 ant-like agents in our SSA, on a 30 x 30 tri-dimensional habitat, for t=400 time steps, on both functions. We then run 3 tests. The first is requested to minimize Passino’s function F1. The second test is requested to minimize Passino’s function F2. Finally, and in order to prove the highly adaptive features of our model, we requested SSA to deal with two contradictory goals, i.e. to minimize F1 and then to maximize it, over the same period of 400 time steps. As visible, SSA quickly adapts to the different purposes. Over function F1, the pheromone concentration is already intensely allocated at the right point at t=100 (and not in other areas), while BFOA, at this moment, still explores different regions on the optimization domain. Over function F2, the swarm quickly separates in different foraging groups, since there are a large number of points with the same minimal value. Finally over function F1 again, in the final test (last row – fig. 6), SSA is able to process two different demands (minimization followed by maximization) over the same foraging time period that BFOA uses for F1 minimization. The in our experiments follows: Nants=50, parameters used tmax=400, k =1 (pheromone evaporation rate), =0.1 (pheromone deposition rate), =7 (this parameter controls how ants follow the pheromone gradient), γ=0.2, and p=1.9. Exception made for test 1, where =6. VI. CONCLUSIONS Evolution of mass behaviours on time are difficult to predict, since the global behaviour is the result of many part relations operating in their own local neighbourhood. The emergence of network trails in ant colonies, for instance, are the product of several simple and local interactions that can evolve to complex patterns, which in some sense translate a meta- behaviour of that swarm [29]. Moreover, the translation of one kind of low-level (present in a large number) to one meta-level is minimal. Although that behaviour is specified (and somehow constrained), there is minimal specification of the mechanism required to generate that behaviour; global behaviour evolves from the many relations of multiple simple behaviours, without global coordination (i.e. from local interactions to global complexity. There is some evidence that our brain as well as many other complex systems, operates in the same way, and as a consequence collective perception capabilities could be derived from emergent properties, which cannot be neglected in any pattern search algorithm. These systems show in general, interesting and desirable features as flexibility (e.g. the brain is able to cope with incorrect, ambiguous or distorted information, or even to deal with unforeseen or new situations without showing abrupt performance breakdown) or versability, robustness (keep functioning even when some parts are locally damaged), and they operate in a massively parallel fashion. Present results point to that type of interesting features. Although the current model is far from being consistent with real ones, since only some type of real mechanisms were considered, swarm pheromonal fields reflect some convergence towards the identification of a common goal in a purely decentralized form. Moreover, the present model shows important adaptive capabilities, as in the presence of sudden changes in the habitat - our test landscapes (fig. 1). Even if the model is able to quickly adapt to one specific environment, evolving from one empty pheromonal field, habitat transitions point that, the whole system is able to have some memory from past environments (i.e. convergence is more difficult after learning and perceiving one past habitat). On the other hand this feature can have some advantage, for instance in the case where the original or similar environments are back in place . This emerged feature of résistance, is somewhat present in many of the natural phenomena that we find today in our society. In a certain sense, the distribution of pheromone represents the collective solutions found so far (memory, risk avoidance, exploitation behavior), while evaporation enables the system to adapt (tricks a decision, explorative behavior), not only as in normal situations (a complex but static search environment), as well as when the landscape suddenly changes, moving the colony’s new target to a new unexplored region and keep tracking of it. One crucial aspect observed here, as noted in the past by Langton [16] and present in many complex systems, only at the right intermediary regime, in here between contradictory behaviors of exploration and exploitation, the swarm is able to quickly converge. The recognizable results indicate that the collective intelligence is able to cope and 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. All these above mentioned aspects show how vital can be the study of social foraging for the development of new distributed search algorithms, and the construction of social cognitive maps, with interesting properties in collective memory, collective decision-making and swarm-based pattern detection and recognition. But the work could have important consequences in other areas. Perhaps, one of the most valuable relations to explore is that of social foraging and evolution. For two reasons; First, as described by Passino [22], natural selection tends to eliminate animals with poor “foraging strategies” (methods for locating, handling, and ingesting food) and favor the propagation of genes of those animals that have successful foraging strategies since they are more likely to enjoy reproductive success (they obtain enough food to enable them to reproduce). Logically, such evolutionary principles have led scientists in the field of foraging theory to hypothesize that it is appropriate to model the activity of foraging as an optimization process: A foraging animal takes actions to maximize the energy obtained per unit time spent foraging, in the face of constraints presented by its own physiology and by the environment. Second, because there is an increasing recognition that natural selection and self-organization work hand in hand to form evolution, as defended by Kauffmann [13]. For example, anthropologist Jeffrey McKee [19,14] has described the evolution of human brain as a self-organizing process. He uses the term autocatalysis to describe how the design of an organism’s features at one point in time affects or even determines the kinds of designs it can change into later. For example the angle of the skull on the top of the spine left some extra space for the brain to expand. Thus the evo lution of the organism is determined not only by selection pressures but by constraints and opportunities offered by the structures that have evolved so far. Also, and back again in what regards the evolution of collectives, it is known that during the evo lution of life, there have been several transitions in which individuals began to cooperate, forming higher levels of organization and sometimes losing their independent reproductive identity (insect societies are one example). Several factors that confer evolutionary advantages on higher levels of organization have been proposed, such as Division of Labor and Increased Size. But recently, a new third factor was added: Information Sharing [15]. Lachmann et al., illustrate with a simple model how information sharing can result in individuals that both receive more information about their environment and pay less for it. Being social foraging essentially a self-organized phenomenon, the study of computational foraging embedded with GA (Genetic Algorithm) like natural selection can much probably enhance our understanding on the detailed forms of the hypothetical equation: Evolution = Natural Selection + Self-Organization, and in the precise role of each “variable”. As an example, current work in the same area [10], include the research of variable population size swarms, as used similarly in Evolutionary Computation [9], where each individual can have a probability of making a child, as well to die, depending on his accumulated versus spent energetic resources. The system as a whole, then proceeds on the search space as a kind of distributed evolutionary swarm. Finally and in parallel, an effort is being made in order to understand the societal memory and his speed on tracking extrema over dynamic environments using self-regulatory swarms based on the present model [30,10,29]. REFERENCES [1] Bak, P., How Nature Works – The Science of Self-Organized Criticality, Springer-Verlag, 1996. [2] Bassler, B.L, “Small Talk: Cell-to-Cell Communication in Bacteria”, Cell, Vol. 109, pp. 421-424, May 2002. [3] Ben-Jacob, E., Shochet, O., Tenenbaum, A., Cohen, I., Czirók, A., Vicsek, T., “Generic Modelling of Cooperative Growth in Bacterial Colonies”, Nature, 368, pp. 46-49, 1994. [4] Ben-Jacob, E., Becker, I., Shapira, Y., Levine, H., “Bacteria l Linguistic Communication and Social Intelligence”, Trends in Microbiology, Vol. 12/8, pp. 366-372, 2004. [5] Bonabeau, E., Dorigo, M., Theraulaz, G., Swarm Intelligence: From Natural to Artificial Systems, Santa Fe Institute in the Sciences of Complexity, Oxford Univ. Press, New York, Oxford, 1999. [6] Camazine, S., Deneubourg, J.-L., Franks, N.R., Sneyd, J., Theraulaz, G., Bonabeau, E., Self-Organization in Biological Systems, Princeton Studies in Complexity, Princeton University Press, 2001. [7] 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. [8] Chowdhury, D., Nishinari, K., Schadschneider, A., “Self -Organized Patterns and Traffic Flow in Colonies of Organisms: from Bacteria and Social Insects to Vertebrates”, special issue on Pattern Formation, in Phase Transitions, Taylor and Francis, vol. 77, 601, 2004. [9] Fernandes, C., Rosa, A.C., “Study on Non-random Mating and Varying Population Size in Genetic Algorithms using a Royal Road Function”, IEEE CEC´01, Proc. of the 2001 IEEE Congress on Evolutionary Computation, pp. 60-66, 2001. [10] Fernandes, C., Ramos, V, Rosa, A.C., “Varying the Population Size of Artificial Foraging Swarms on Time Varying Landscapes”, to appear in ICANN-05, Int. Conf. on Artificial Neural Networks, Springer-Verlag, LNCS Series, Warsaw, Poland, Sept. 11-15, 2005. [11] Holland, O., Melhuish, C.: Stigmergy, “Self-Organization and Sorting in Collective Robotics”, Artificial Life, Vol. 5, n. 2, MIT Press, 173 , 1999. [12] Huang, C.-F., Rocha, L.M., “Tracking Extrema in Dynamic Environments using a Coevolutionary Agent-based Model of Genotype Edition”, in GECCO-05, Genetic and Evolutionary Computation Conf., Washington, D.C., USA, 25-29 June, 2005. [13] Kauffmann, S.A., The Origins of Order: Self-Organization and Selection in Evolution, New York: Oxford University Press, 1993. [14] Kennedy, J. Eberhart, Russel C. and Shi, Y., Swarm Intelligence, Academic Press, Morgan Kaufmann Publ., San Diego, London, 2001. [15] Lachmann, M., Sella, G., Jablonka, E., “On Information Sharing and the Evolution of Collectives”, Proc. of the Royal Society: Biological Sciences, 267, pp. 1265-1374, 2000. [16] Langton, C.G., “Computation at the Edge of Chaos”, Physica D, 42, pp. 12-37, 1990. [17] Liu, Y., Passino, K.M., “Biomimicry of Social Foraging Bacteria for Distributed Optimization: Models, Principles, and Emergent Behaviors”, Journal of Optimization Theory and Applications, Vol. 115, nº3, pp. 603-628, Dec. 2002. [18] Maree, A.F.M., Hogeweg, P., “How Amoeboids Self-Organize into a Fruiting Body: Multicellullar Coordination in Dictyostelium discoideum”, PNAS, vol. 98, nº 7, pp. 3879 -3883, 2001. [19] McKee, J.K., The Riddled Chain: Change, Coincidence, and Chaos in Human Evolution, Piscataway, NJ: Rutjers University Press, 2000. [20] Millonas, M.M., “A Connectionist-type model of Self-Organized Foraging and Emergent Behavior in Ant Swarms”, J. Theor. Biol., nº 159, 529, 1992. [21] 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. [22] Passino, K.M., “Biomimicry of Bacterial Foraging for Distributed Optimization and Control”, IEEE Control Systems Magazine, pp. 52-67, June 2002. [23] Peak, D.A., West, J.D., Messinger, S.M., Mott, K.A., “Evidence for Complex, Collective Dynamics and Emergent, Distributed Computation in Plants”, PNAS, Proc. of the National Academy of Sciences, USA, 101, pp. 918-922, 2004. [24] Pohlheim, H, “Genetic Algorithm MATLAB Toolbox Test Functions”, MATLAB reference manual, version 1.2, Mathworks, 1997. [25] 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. [26] Ramos, V., “On the Implicit and on the Artificial - Morphogenesis and Emergent Aesthetics in Autonomous Collective Systems”, in ARCHITOPIA Book, Art, Architecture and Science, Institut D’Art Contemporain, J.L. Maubant et al. (Eds.), pp. 25 -57, Chapter 2, ISBN 2905985631 – EAN 9782905985637, France, Feb. 2002. [27] Ramos,V., Merelo, Juan J., “Self-Organized Stigmergic Document Maps: Environment as a Mechanism for Context Learning” , in AEB’2002 – 1st Spanish Conf. on Evolutionary and Bio-Inspired Algorithms, E. Alba, F. Herrera, J.J. Merelo et al. (Eds.), pp. 284 -293, Centro Univ. de Mérida, Mérida, Spain, 6 -8 Feb. 2002. [28] Ramos, V., Abraham, A., “Evolving a Stigmergic Self-Organized Data- Mining”, in ISDA-04, 4th Int. Conf. on Intelligent Systems, Design and Applications, Budapest, Hungary, ISBN 963-7154-30-2, pp. 725-730, August 26-28, 2004. [29] 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, 2005. [30] Ramos, V., Fernandes, C., Rosa, A.C., “Societal Memory and his Speed on Tracking Extrema over Dynamic Environments using Self- Regulatory Swarms”, invited paper at NiSIS-05, 1st European Symp. on Nature-inspired Smart Information Systems, Portugal, 3-5 Oct., 2005. [31] Theraulaz, G., Bonabeau, E., “A Brief History of Stigmergy”, Artificial Life, Vol. 5, n. 2, MIT Press, 97-116, 1999. [32] Wilson, E.O., The Insect Societies, Cambridge, MA., Belknap Press, 1971.
1811.02921	2	1811	2019-04-01T21:53:09	Flexible Representative Democracy: An Introduction with Binary Issues	[ "cs.MA" ]	We introduce Flexible Representative Democracy (FRD), a novel hybrid of Representative Democracy (RD) and direct democracy (DD), in which voters can alter the issue-dependent weights of a set of elected representatives. In line with the literature on Interactive Democracy, our model allows the voters to actively determine the degree to which the system is direct versus representative. However, unlike Liquid Democracy, FRD uses strictly non-transitive delegations, making delegation cycles impossible, and maintains a fixed set of accountable elected representatives. We present FRD and analyze it using a computational approach with issues that are binary and symmetric; we compare the outcomes of various democratic systems using Direct Democracy with majority voting as an ideal baseline. First, we demonstrate the shortcomings of Representative Democracy in our model. We provide NP-Hardness results for electing an ideal set of representatives, discuss pathologies, and demonstrate empirically that common multi-winner election rules for selecting representatives do not perform well in expectation. To analyze the behavior of FRD, we begin by providing theoretical results on how issue-specific delegations determine outcomes. Finally, we provide empirical results comparing the outcomes of RD with fixed sets of proxies across issues versus FRD with issue-specific delegations. Our results show that variants of Proxy Voting yield no discernible benefit over RD and reveal the potential for FRD to improve outcomes as voter participation increases, further motivating the use of issue-specific delegations.	cs.MA	cs	Flexible Representative Democracy: An Introduction with Binary Issues Ben Abramowitz and Nick Mattei April 3, 2019 Abstract We introduce Flexible Representative Democracy (FRD), a novel hybrid of Representative Democracy (RD) and direct democracy (DD), in which voters can alter the issue-dependent weights of a set of elected representatives. In line with the literature on Interactive Democracy, our model allows the voters to actively determine the degree to which the system is direct versus representative. However, unlike Liquid Democracy, FRD uses strictly non-transitive delegations, making delegation cycles impossible, and maintains a fixed set of accountable elected representatives. We present FRD and analyze it using a computational approach with issues that are binary and symmetric; we compare the outcomes of various democratic systems using Direct Democracy with majority voting as an ideal baseline. First, we demonstrate the shortcomings of Representative Democracy in our model. We provide NP-Hardness results for electing an ideal set of representatives, discuss pathologies, and demonstrate empirically that common multi-winner election rules for selecting representatives do not perform well in expectation. To analyze the behavior of FRD, we begin by providing theoretical results on how issue-specific delegations determine outcomes. Finally, we provide empirical results comparing the outcomes of RD with fixed sets of proxies across issues versus FRD with issue- specific delegations. Our results show that variants of Proxy Voting yield no discernible benefit over RD and reveal the potential for FRD to improve outcomes as voter participation increases, further motivating the use of issue-specific delegations. 1 Introduction Since the Athenian Ecclesia in 595 BCE Direct Democracy (DD) as an ideal collective decision making scheme has loomed large in the western imagination [18]. While DD may be desirable it becomes impractical at scale because it places too much burden on individual decisions makers: everyone must be well-informed on every issue and always available to vote [24]. In addition to the attention requirements, voters are also required to know and be able to articulate their preferences at the time of every vote. While preferences and preference learning are large research areas in AI [17, 22], every voter may not have enough knowledge, information, time, or energy to participate, particularly when issues are complex. Given the prohibitive costs of implementing a large-scale Direct Democracy in both human and agent societies, we often resort to forms of representation, relying upon a set of proxies to decide 1 on the voters' behalf. Countries have parliaments, companies have shareholders, and even groups of agents select leaders to represent them [45]. Sets of representatives have been used in many contexts and disciplines to reduce the computation and communication burden of decision makers. Within computer science, many applications face the task of selecting representatives for down- stream decision making. In portfolio selection a particular set of algorithms and hyper-parameters are selected from a large pool of candidates and then used as representatives for later problems [30], in multi-agent systems the role assignment problem uses distributed voting to decide on tasks for agents [46], and in group recommendation settings this can correspond to picking a set of experts to later make decisions. The COMSOC community [11] has produced a large body of research on how to select and weight representatives. Indeed, using multi-winner voting [41], we can view the winners as a set of exemplars that may be used to decide some downstream application -- e.g., we select a set of points in space and then aggregate these points (votes) over the set. Often it is beneficial to elect fixed committees which meet certain axiomatic criteria. For ex- ample, committees should be proportional and have justified representation of the voters [4]. In- tuitively, these difficulties in electing committees carry through to the setting of Representative Democracy (RD) where the committee makes decisions in the interest of the voters/agents who elect them [40]. As the prevalence and security of the internet improve, many scholars and com- panies are turning toward computer systems to address issues with democratic decision making systems; some going so far to suggest that we create an AI-based Direct Democracy or an "Aug- mented Democracy" 1. Since DD is impractical and RD comes with inherent tradeoffs and limitations, hybridizations of the two have arisen under the umbrella of interactive democracy. This idea, coupled with mod- ern communication technologies, has spawned a large number of proposed democratic decision making systems, and interactive democracy has become an important area of research and appli- cation for AI [12]. Perhaps the most popular version of this today is Liquid Democracy. Liquid Democracy has received significant attention in the political science [24], AI [28] and agents com- munities [13], and has been implemented in both corporate [27] and political settings [10, 7]. In contrast to existing interactive democracy proposals, our model of Flexible Representative Democracy maintains a set of expert representatives while allowing voters to guarantee their own representation without raising the minimum required burden on them. In an FRD voters elect a set of representatives to serve a term during which they decide the outcomes over a set of issues. Each voter, by default, allocates a fraction of their voting power to each member of the committee. If this allocation is uniform and we stop here, we are left with a traditional model of RD where each representative has equal power. However, for each issue under consideration in FRD, the voters can deviate from this default by delegating their voting power over any subset the committee. If all voters use their option to delegate on each issue, as long as there is at least one representative who agrees with each voter's view, the outcome can become exactly that of DD. Voters have both the election and the flexible delegation option as tools for achieving representation and holding representatives accountable. In an FRD, voters have great flexibility in determining how they are represented and the man- dated disclosure of representatives' votes guarantees that an attentive voter can be fully informed about how their voting power will be and was used. For example, the day after the election an inattentive voter might choose a few elected representatives they trust, apportion the power of their 1https://www.peopledemocracy.com/ 2 vote to these few for all future issues and pay no attention until the next election. A more attentive voter might alter their allocations on an issue-by-issue basis as issues arise, reacting to represen- tatives' votes. In general, voters determine the granularity with which they privately express their preferences over issues via the representatives. Thus, the degree to which Direct Democracy is em- ulated by a Flexible Representative Democracy depends both on the caliber of the representatives and the fastidiousness of the voters. 1.1 Contributions We introduce Flexible Representative Democracy (FRD), a new model of interactive democracy which smoothly transitions, at the discretion of the voters, between Representative Democracy and Direct Democracy. Our proposal for FRD solves standing issues in the literature on interactive democracy including maintaining a fixed, elected committee to generate legislation and making delegation cycles impossible. We analyze our model in decision making scenarios involving bi- nary, symmetric issues and (1) show that electing an optimal set of representatives is hard for any large-scale Representative Democracy that uses a multi-winner voting rule, (2) investigate the performance of various deterministic multi-winner voting rules to select committees, (3) demon- strate the theoretical ability of issue-specific delegations under FRD to overcome the limitations of Representative Democracy, and (4) provide empirical results demonstrating that FRD outperforms both Representative Democracy and Proxy Voting for representing the will of the voters. 2 Model and Preliminaries We primarily consider three democratic decision systems: Direct Democracy (DD), Representative Democracy (RD), and our model of Flexible Representative Democracy (FRD), which we define as follows. Given a set of voters V with preferences over the alternatives for each issue in a set of issues S, we represent their collective preferences by a preference profile PV,S. In a direct democracy, a decision rule RS is applied directly to the voters' preference profile to obtain a set of issue outcomes, RS(PV,S) → ODD.2 By contrast, in a representative system voters' preferences on the issues may never be directly elicited. Rather, voters report their preferences over a set of candidates seeking election C. We denote the collective preferences of the voters over the candidates by the electoral profile PV,C. An election rule (i.e. multi-winner voting rule) is then used to aggregate these preferences and select a subset of candidates to serve as representatives, RE(PV,C) → D ⊆ C. In a standard representative democracy, a decision rule is then applied to the preferences of the representatives to determine the outcomes on all issues, RS(PD,S) → ORD. Clearly, RD may produce different outcomes than DD, and may leave accessible information about voter preferences unsolicited and unused. Flexible Representative Democracy endeavors to use this information if and when it is available without relying entirely upon it. 2For clarity, we will refer to a social choice function RS which determines the outcome of an inanimate issue as a decision rule, and a social choice function RE which selects from among candidates to serve as voting representatives as an election rule. Our definitions are easily extended to social welfare functions. 3 As with RD, in FRD the voters elect a set of representatives RE(PV,C) → D ⊆ C. However, for every issue, divisible units of voting power (votes) are given to the voters rather than simply giving a vote to each representatives. Automatically after the election, the voters' issue-specific votes are distributed among the representatives according to some default distribution mechanism. Subsequently, every voter has the option to alter how their voting power is assigned to the repre- sentatives on an issue-by-issue basis and we refer to this active process of allocating voting power to representatives as delegation. We let W i jl represent the voting power allocated by voter vj to candidate cl on issue si, yielding a collective matrix of weights W . To determine the outcomes of a set of issues in FRD, a decision rule is applied to the representatives' preferences which takes these weights into account RS(PD,S, W ) → OF RD. 2.1 Model Specification Our objective is to compare the extent to which RD and FRD can emulate the decision which would be made under DD with binary issues. In our specification, each voter in the set of voters V = {v1, . . . , vN} has a preferred outcome vi j ∈ {0, 1} for every issue si in the set of issues S = j} represent the full preferences of voter vj over the issues, {s1, . . . , sr}. We let (cid:126)vj = {v1 yielding the collective approval profile PV,S = {(cid:126)vj : vj ∈ V}. Similarly, in our representative systems we can represent the preference profile of the candidates PC,S = {(cid:126)cl : cl ∈ C} where candidate cl has preferences (cid:126)cl ∈ {0, 1}r. Without loss of generality, we label the outcome preferred (cid:80)r by the weak majority of voters 1 and the other 0, breaking ties randomly (when N is even). Generally, we define the agreement between any two outcome vectors O1,O2 as L(O1,O2) = 2. Thus we will often refer to the agreement between a voter and candidate i=1 Oi 1 − 1 L((cid:126)vj, (cid:126)cl) and the agreement between the outcomes produced by different democratic systems, i.e. L(ODD, ORD). 1 − Oi j , . . . , vr r We consider three possible ways voters might express their preferences over the candidates: approvals, total orderings, and normalized weights. In our simulation and analysis we make a large assumption about these preferences to give RD the greatest chance of maximizing L(ODD, ORD). Namely, we assume each voters' preferences over each candidate is induced by their level of agreement. When voters submit approval ballots, we assume vj approves of cl if and only if L((cid:126)vj, (cid:126)cl) > 1/2. When voters report total orderings ((cid:31)j), we assume they order all candidates so that cl (cid:31)j ch implies L(vj, cl) ≥ L(vj, ch) where ties are broken privately (e.g. randomly). When ch∈C L(vj ,ch).3 The collective voters report their preferences as normalized weights, wl preferences of the voters over the candidate set yield the profile PV,C, which may consist of ap- proval, ordinal, or normalized cardinal preferences. For Direct Democracy, we only consider the simple majority rule as our decision rule RS be- cause our issues are binary and symmetric. However, for our representative systems, we compare several common, anonymous election rules RE with a fixed, odd committee size k so that the set of elected representatives is D ⊆ C where D = k. All rules considered are deterministic other than randomized tie-breaking. In the setting where voters submit approval ballots, we consider Approval Voting and Re-weighted Approval Voting (AV, RAV). When voters submit their pref- erences over candidates as total orderings, we consider Single-Transferrable Vote (STV), Borda, j = L(vj ,cl)/(cid:80) 3These normalized weights reported during the election process should not be confused with the weights (voting power) assigned to representatives later by default and through delegation. 4 k-Median, and Chamberlin-Courant (CC). When voters submit their ballots as normalized weights over the candidates, we consider the rule which selects the k candidates who receive the largest total weight. Lastly, we compare these rules to sortition, selecting k representatives uniformly at random from the candidates. Formal definitions of AV, RAV, STV, and Borda can be found in the book chapter by [47] and definitions of CC and k-Median can be found in [40]. Given a set D of k elected representatives we want to evaluate the capability of this set to represent the will of the voters, i.e., recover the outcome of a Direct Democracy. To this end we introduce the notions of coverage, full coverage, and majority agreement. Let ki 0 represent 0 = k. the number of representatives who prefer 1 and 0 on issue si respectively, such that ki Majority Agreement. There is majority agreement on an issue if the majority of representatives 1 and ki 1 + ki agree with the majority of voters (ki 1 > k 2 > ki 0). Coverage. An issue is covered if at least one representative agrees with the voter majority (0 < 1). ki Full Coverage. An issue is fully covered if the representatives are not unanimous (0 < ki 1 < 1). Bear in mind that in practice the number of representatives on either side of the issue is known, but it is not known which side corresponds to the voter majority. Therefore the majority agreement of any set of representatives for an issue may be unknown, but it is always known on which issues they achieve full coverage. Full coverage implies coverage, but otherwise the status of coverage is unknown. For FRD, we allocated each voter one divisible vote for each independent issue, maintaining the principle of "one person, one vote". Once the representatives have been elected, each voter's unit of voting power is distributed among the representatives on each issue. In this paper we distribute this power uniformly by default, so initially W i l = N/k for each of the k candidates and r issues. Various distributions from the literature on voting power [38, 6] and Proxy Voting [1] are worth consideration. We do not consider abstentions by representatives nor voter abstentions, whereby a voter assigns less than a full vote to the representatives as a whole. The total voting power held by the representatives remains N collectively for all issues. 0, αi di 5 the delegations rates may differ for the majority and minority (αi weight W i from default and delegation(cid:80) We refer to the fraction of voters who use their delegation option as the delegation rate αi, and 1). For our purposes, the total l assigned to representative dl on issue si is the sum of the voting power they receive jl. Consequently, the total weight assigned to representatives l . In this paper our decision rule for FRD is 2 , and who agree with the voter majority is X i weighted majority with random tie breaking. That is, Oi = 1 if X i Oi = 1 with probability 1/2 if X i 2 , Oi = 0 if X i 1 =(cid:80) l=1 W i 1 > N 1 < N j W i 2 . 1 = N In our theoretical analysis we assume that voters who delegate do so optimally after the repre- sentatives have voted. A voter delegates optimally if they only delegate voting power to represen- tatives who agree with their preferred outcome on an issue. This is equivalent to restricting voters to only delegate to a single representative, or to vote directly on the issue (if there is full cover- age). We relax the assumption of optimal delegations in our simulations and consider voters who delegate only to their most preferred candidate(s) or divide their delegation evenly across their ap- proved set. If an issue is not fully covered because the representatives are unanimous, the outcome is already determined since opposing voters have no one to whom they can delegate. We discuss this further in Section 7. Example 1. Consider an FRD instance with issues s1 and s2, three voters, and three represen- tatives. Below, the solid arrows from voter to representative indicate delegations, and any voter without an arrow defaults on that issue. The voter and representative preferences are given in the tables above and below the agents. Notice that both delegations are optimal. Issue s1 Issue s2 vi j vj dl di l d1 1 d2 1 d3 0 W i l 2/3 2/3 5/3 1 v1 1 v2 0 v3 1 v1 1 v2 0 v3 d1 1 d2 0 d3 0 5/3 2/3 2/3 X 1 1 = 4/3 < N/2 O1 = 0 X 2 1 = 5/3 > N/2 O2 = 1 On issue s1, the representative majority agrees with the voter majority, so RD would yield O1 RD = 1 3 = 0) delegates, the weighted majority as desired. However, since only the voter in the minority (v1 1 < 1/2). This can oc- of representatives now decides the outcome in favor of the voter minority (X 1 cur if the number of voters in the minority is large enough, the number of representatives who agree with the voter minority is large enough, and the voters in the minority delegate at a substantially higher rate than the voters in the majority. On issue s2 the representative majority disagrees with the voter majority so the RD outcome (without delegations) would be, regrettably, O1 RD = 0. Looking again at the figure we see the delegations flip the result to what would be achieved by Direct Democracy (X 2 1 > 1/2). Hence, FRD can improve the outcomes over RD as measured against DD. Fortunately, for both s1 or s2, if any two or all of the voters delegate optimally, the outcome will always agree with the voter majority. 3 Related Work [33], inspired by [43] and shareholder proxy voting, suggested an interactive democratic system for legislation that could take place at scale using computers. Miller lamented the lack of flexibility in traditional Representative Democracy and sought to remedy this using a dynamic system of proxies, although admitted this was not conducive to creating legislation. Soon after, [39] warned that electronic systems may accelerate the legislative process in undesirable ways and suggested holding every referendum twice to guarantee time for sufficient public deliberation. Our use of a fixed, elected set of representatives answers Miller's question of how to produce legislation, and rather than holding redundant referenda we give the voters sufficient time to continue deliberation and alter their delegations after the representatives vote. Just before the dawn of the internet, [44] revisited the ideas in a proposal that motivates the 6 default distribution and delegation mechanism in FRD [43]. The notion of the default distribution is also similar to that proposed by [1], which suggests that the weights of representatives be based on the preferences of voters expressed in the election, but these weights are fixed during their term. By contrast, in FRD the weight of each representative on each issue is not strictly determined by the election. [16] took an analytical approach to studying a Proxy Voting model very close to that of [1] for decision making with no election, infinite voters, spatial preferences, and assuming agents lie in a metric space. The hallmark of an interactive democracy is that rather than adjudicating whether a direct or representative system is better for expressing the will of the voters and asserting it by fiat, the extent to which the system is direct or representative is itself a function of the will of the voters. Currently, the most well-known and well-studied form of Interactive Democracy is Liquid Democracy, which has been studied from an algorithmic perspective as a decision-making process in the AI and COMSOC literature [13, 28, 9, 14] and elsewhere [24, 21, 10, 12, 27]. Unlike Liquid Democracy, FRD does not allow transitive delegations nor delegations to another voter, thereby violating the second axiom proposed by [24]. However, as we discuss in Section 7, the notion of voluntary representatives can be maintained if desired for a particular application. Fractional delegations in FRD serve a similar function to that of the virtual committees proposed by [24], although in theory FRD could incorporate virtual committees as well as many other mechanisms for delegating voting power. The design of FRD is also largely based on work in probabilistic voting, binary aggregation, statistical decision theory, and computational social choice. In particular, work on the optimal weighting of experts [5, 35, 25, 34, 8], the Condorcet Jury Theorem [26], variable electorates [20, 42, 36], and optimal committee sizes [3, 29, 32]. In FRD, one can view the voter delegations as a pseudo-tie breaking mechanism for the representatives or, conversely, see the default distribution as a way to dampen the variance in the outcome which occurs in Direct Democracy when the sample of participating voters is small or biased. Another view is that electing representatives is analogous to a compression algorithm [37], which is the algorithmic version of John Adams's alleged intuition that the representatives should be a microcosm of the population (taken from [1]). In this view, the delegations in FRD are a decompression mechanism where a higher delegation rate reduces the "loss" of representation. Our evaluations are similar to those of [40], however, in their approval model the quality of the committee is measured as the sum of the voter proportion being represented for each issue, while we focus only on the total number of issues correct according to DD. 4 Difficulties of Representative Democracy Electing good committees is hard. Electing a set of representatives which maximizes majority agreement on binary issues is NP-Hard even if we know the view of every voter on every issue. But suppose we wanted to solve the easier problem of maximizing coverage. Even maximizing coverage is NP-Hard, as is maximizing full coverage. If the majority view of the voters were known, maximum coverage could be approximated deterministically in polynomial time within a factor of 1 − 1/e by a greedy algorithm, and this bound is tight [19]. Therefore, none of our deterministic polynomial-time election rules can provide a better guarantee than this. Worse yet, even for small instances where the problem is computationally tractable, there are pathological 7 examples for which truthful voters whose derived preferences over the candidates are perfectly consistent with their preferences over the issues will elect horrible representatives. We refer to the problems of selecting k representatives to maximize coverage, full coverage, and majority agreement as Max k-Coverage, Max k-Full Coverage, and Max k-Majority Agreement, respectively. Below we provide complexity results and pathologies, followed by simulated results to show how well our various polynomial time multi-winner voting rules perform in terms of majority agreement when voter and candidate preferences are generated uniformly at random. Theorem 2. If the outcome preferred by the majority of voters is known for every issue, Max k-Coverage is NP-hard. The theorem can be stated more explicitly as follows. Consider a set of binary issues S = l ∈ {s1, . . . , sr} and a set of candidates C = {c1, . . . , cm} where each candidate has preference ci {0, 1} on each issue si, the problem of selecting the subset of k candidates D ⊆ C that maximizes the number of issues si on which(cid:80) dl∈D di l > 0 is NP-Hard. Proof. Our proof of the hardness of Max k-Coverage is a reduction from the NP-Hard problem of MAX K-COVER [23, 19]. The input to MAX K-COVER is a set U = {x1, . . . , xr} of r points, a collection S = {s1, . . . , sm} of subsets of U, and an integer k. The objective of MAX K-COVER is to select k subsets from S such that their union has maximum cardinality. Given an instance (U, S, k) of MAX K-COVER we create an instance of Max k-Coverage as follows. For every point xi ∈ U create an issue si and for every subset sl ∈ S create a candidate cl. For all points xi and subsets sl, if xi ∈ sl then let ci l = 0. And let k be the number of representatives we will elect. There is a one-to-one correspondence between the number of issues covered by our k representatives and the cardinality of the corresponding subsets in the original max k-cover instance. Therefore, any set of k candidates that maximizes coverage corresponds exactly to a collection of k subsets in our MAX K-COVER instance whose union has maximum cardinality. l = 1, otherwise let ci Note that if the majority view of the voters were known, coverage could be approximated deterministically in polynomial time within a factor of 1 − 1/e by a greedy algorithm, and this bound is tight [19]. Therefore, none of our deterministic polynomial-time election rules can provide a better guarantee than this, although they may provide decent approximations in expectation. The proofs for the two theorems below follow directly, although non-trivially, from our proof for Max k-Coverage. Theorem 3. If the outcome preferred by the majority of voters is known for every issue, Max k-Full Coverage is NP-hard. Proof. We now prove the hardness of Max k-Full Coverage by polynomial-time reduction from Max k-Coverage. To do this we construct an instance of Max k-Full Coverage by adding an addi- tional candidate c, adding r + 1 additional issues to the original r issues, and desire a set of k + 1 candidates. We show that in this new instance of Max k-Full Coverage the additional candidate must be selected in any optimal solution because they are uniquely required to cover the r + 1 added issues, and the remaining k candidates in the solution set correspond exactly to the optimal k candidates in the solution to our original Max k-Coverage instance. 8 l = ci Given an instance (S = {s1, . . . , sr},C = {c1, . . . , cm}, k) of Max k-Coverage we construct an instance of Max k-Full Coverage as follows. Create a set of binary issues S = S = {s1, . . . , sr} and augment it with r + 1 additional binary issues so that S = {s1, . . . , sr, . . . , s2r+1}. Create a l for all cl ∈ C, for all 1 ≤ i ≤ r and ci = 0 for all set of candidates C = C ∪ c where ci l = 0 for all cl ∈ C\{c} for issues r + 1 ≤ i ≤ 2r + 1 and let ci = 1 issues 1 ≤ i ≤ r. Let ci for all issues r + 1 ≤ i ≤ 2r + 1. Our objective is to select a set D ⊆ C of k + 1 candidates from C which maximizes full coverage. We will now prove that for all solutions D to our new Max k-Full Coverage problem, D = {c} ∪ D where D is a set of k candidates whose corresponding counterparts maximize coverage over issues {s1, . . . , sr} in our original Max k-Coverage instance. Lemma 4. D must contain c Proof. Clearly, the set {c} ∪ {cl} achieves full coverage for issues {sr+1, . . . , s2r+1} for any cl ∈ C\{c}, and any set which does not contain c cannot fully cover (or cover) {sr+1, . . . , s2r+1}. Since {sr+1, . . . , s2r+1} comprises more than half the issues, any set of k + 1 candidates for k ≥ 1 which maximizes the number of issues fully covered, must contain c. Given that ci = 0 for all issues {s1, . . . , sr}, the set of candidates D = D\{c}, which max- imizes full coverage for issues {s1, . . . , sr} is the set of k candidates which maximizes coverage over issues in {s1, . . . , sr}. Therefore, the k candidates corresponding to D are the solution to our original instance of Max k-Coverage and given the solution D to Max k-Coverage we simply add c to find D. Theorem 5. If the outcome preferred by the majority of voters is known for every issue, Max k-Majority Agreement is NP-hard. Proof. We now prove the hardness of Max k-Majority Agreement by polynomial-time reduction from our problem of Max k-Coverage. Similar to our proof for Max k-Full Coverage, we replicate the instance of Max k-Coverage and add r + 1 issues to the original r issues such that S = 2r + 1. However, we now augment the candidate set with k + 1 additional candidates who must be included in any committee which maximizes majority agreement. The objective is to select the 2k + 1 candidates which maximize majority agreement. The k + 1 additional candidates must be in the solution set for Max k-Majority Agreement, and the remaining k candidates selected will correspond exactly to the k candidates in the solution to our original instance of Max k-Coverage. Given an instance of Max k-Coverage with input (S = {s1, . . . , sr},C = {c1, . . . , cm}, k) we construct an instance of Max k-Majority Agreement with input ( S = {s1, . . . , sr}, C = {c1, . . . , c m}, k) as follows. Create a set of binary issues S = {s1, . . . , s2r+1} and a set of candi- dates C = {c1, . . . , cm+k+1}. We can think of C as being made up of three sets of candidates based on how we will construct their preferences over S, that is, C = {c1, . . . , cm}∪{cm+1, . . . , cm+k}∪ {cm+k+1}. The first set {c1, . . . , cm} has identical preferences to the m candidates in the original problem over the first r issues, and prefers 0 on the rest. The second set {cm+1, . . . , cm+k} unan- imously prefer the outcome of 1 on all issues. The last candidate cm+k+1 prefers 0 on the first r issues, and 1 on the remaining r + 1 issues. Formally, for l ≤ m, ci l = 0 for r < i ≤ 2r + 1. For m < l < m + k + 1, ci l = 0 for i ≤ r and ci Lemma 6. Any set D ⊆ C of 2k + 1 representatives which maximizes majority agreement must contain {cm+1, . . . , cm+k+1}. l = 1 for all i. And for l = m + k + 1, ci l = 1 for r < i ≤ 2r + 1. l = ci l for i ≤ r and ci 9 Proof. Clearly, D agrees with the voter majority on issues {sr+1, . . . , s2r+1} if and only if D con- tains {cm+1, . . . , cm+k+1}, because this is the only way at least k+1 out of the 2k+1 representatives can agree with the voter majority on any of these issues. This directly implies there is agreement on more than half the issues if and only if {cm+1, . . . , cm+k+1} ⊆ D. Selecting candidates {cm+1, . . . , cm+k+1} provides exactly k representatives who agree with the voter majority on issues {s1, . . . , sr}. Since we are selecting 2k + 1 representatives in total, on any of these first r issues we need only 1 more representative who agrees with the voter major- ity on each issue to achieve majority agreement. Therefore, selecting k additional representatives from {c1, . . . , cm} which maximize coverage over issues {s1, . . . , sr}, maximizes the majority agreement of the 2k + 1 representatives over S. Clearly, these k representatives are a one-to-one correspondence to the k representatives in the solution to our original Max k-Coverage problem. Likewise, given the solution to the original Max k-Coverage problem, taking the corresponding candidates and adding {cm+1, . . . , cm+k+1} maximizes majority agreement. (a) Varying number of issues with k = 21, C = 60. (b) Varying number of candidates with k = 21, S = 150. (c) Varying the committee size with C = 100, S = 150. Figure 1: Agreement of the elected committee with the outcomes of a Direct Democracy as a function of various properties of an democratic system. Across all treatments the weighted voting, approval voting, and repeated approval voting (RAV) select the best committees. Theorem 7. No Condorcet-consistent election rule using approvals or total orderings can approx- imate Max k-Majority Agreement. Our proof by example for Theorem 5 in the Appendix is derived from an example found in [2] with 11 voters and 11 issues. This example is particularly pathological, because the worst conceivable candidate gets elected over the best conceivable candidate. 10 020406080100120140160Number of Issues0.40.50.60.70.80.91.0Agreement MeanAV AgreementBorda AgreementSTV AgreementRAV AgreementRandom AgreementWeights Agreement20406080100Number of Candidates0.50.60.70.80.91.0Agreement MeanAV AgreementBorda AgreementSTV AgreementRAV AgreementRandom AgreementWeights Agreement020406080100Committee Size (k)0.50.60.70.80.91.0Agreement MeanAV AgreementBorda AgreementSTV AgreementRAV AgreementRandom AgreementWeights AgreementFigure 2: Comparison of NP-hard rules with our polynomial rules for C = 17,S = 80,V = 51. We cannot scale this graph in the same was Figure 1 due to the high computational cost of computing the winning sets for k−Median and CC. However, from this small sample we see that Weights, STV, and AV all strictly dominate both CC and k−Median in terms of agreement. 4.1 Simulation Results We investigate the properties of coverage and majority agreement as functions of the numbers of candidates, issues, and committee size. In all our simulations, for all issues s1 we let vi j = 1 and 2 for all voters and candidates. In all of our runs, coverage was 1.0 for all l = 1 with probability 1 ci combinations, hence we omit it from the graphs in Figures 1 and 5. For all simulations we perform 50 iterations at each datapoint and plot the mean of these runs. Variance for all points is ≤ 0.002 so our results are robust to noise [15]. For a first set of simulations we included rules that have NP-hard winner determination prob- lems: Chamberlin-Courant and k−Median. We implemented these rules in Gurobi 8.1 using an ILP formulation based off one given by [31] for OWA based assignments. Even with a relatively optimized implementation it still took almost 24 hours to generate Figure 5 on a server with 16 cores and 32 GB of memory. Looking at Figure 5 we see that both CC and k−Median are strictly dominated by Weights, STV, and AV at all committee sizes. Hence, to run experiments on larger samples of voters and issues we drop the NP-hard rules as they do not seem to offer a particular advantage over the easy to compute rules. For our larger simulations we hold V = 501 fixed as we did not observe a strong dependence on the number of voters as long as it was sufficiently larger than the number of candidates. Turning first to Figure 1a we hold C = 60, k = 21 and vary S ∈ {15, . . . , 150} in steps of 15. We see that for a small number of issues the AV, RAV, and the weighted voting rule can be expected to select a committee in agreement with the majority nearly 80% of the time. However, as we add issues to the docket, the voting rules seem to converge around 60%. In Figure 1b we fix k = 21,S = 150 and vary the number of candidates between C ∈ {21, . . . , 100} in steps of 5. We observe again that AV, RAV, and weighted voting are the best followed closely by STV. As we increase the number of candidates it is possible for the system to more frequently recover the will of the majority but this number does not climb above 65% across all treatments. Finally, in Figure 1c we hold C = 100,S = 150 and vary k. These simulations reinforce the idea that electing an ideal committee, i.e., one that represents 11 4681012Committee Size (k)0.400.450.500.550.600.650.70Agreement MeanAV AgreementBorda AgreementSTV AgreementRAV AgreementRandom AgreementWeights AgreementChamberlin-Courant Agreementk-Median Agreementthe will of the majority of the voters on every issue, is a hard problem. In the next section we will explore how FRD can out preform RD and its dependence on the constituent delegation rates (α1, α0). 5 Benefits of Flexibility The flexibility of issue-specific delegations is the motivating feature of FRD. We look first at basic features of FRD in a deterministic setting before considering probabilistic delegations. 5.1 Deterministic Delegation A voter's delegation is optimal if the voter only delegates to representatives who agree with them on that issue. Observe that if the representatives are unanimous only one outcome is possible, but as long as there is some dissent in the committee opinions FRD can take advantage and return decision making power to the voters. 1+N i 0+ki We denote by N i 0 = N the numbers of voters and by ki 2 ≥ N i 1 = k the numbers of candidates who agree and disagree with the voter majority on issue si, respectively. We have labeled the majority view of the voters as 1, so ∀si ∈ S : N i 0. The overall outcome of any resolute democratic process over this set of issues is a single outcome vector O = {O1, . . . ,Or} ∈ {0, 1}r, and our ideal outcome is {1}r. Treating all issues equally and independently, we seek to maximize 1 ∈ Z be the number of voters who delegate in favor of each outcome, assuming they delegate optimally. We drop the i superscript below because we will be talking about a single issue. (cid:80) si∈S Oi. Let λi 0 ∈ Z and λi 0 = αi 0N i 1 = αi 1N i 1 ≥ N . k +λ1. If λ0 = N0, then X1 = (N−N0−λ1) k1 Theorem 8. If all delegations are optimal and the issue is fully covered (0 < k1 < k), the outcome is guaranteed to agree with the voter majority if the number of voters in the majority who delegate (λ1) is greater than N k−2N1k1 2(k−k1) Proof. For the majority to guarantee the outcome in their favor by delegating optimally, this means that even if all voters in the minority delegate optimally, the outcome must still be 1. Recall that X1 is the total weight assigned by default and delegation to representatives who prefer the outcome 1. Let X1 = (N−λ0−λ1) k1 k +λ1 = (N1−λ1) k1 k +λ1. For the outcome to be guaranteed in favor of the voter majority, it must be that X1 = (N1−λ1) k1 k +λ1 > 2 . Solving for λ1 we find that λ1 > N k−2N1k1 . Note that this lower bound may be negative. In this N 2(k−k1) case k1 is so large and N0 is so small that the outcome is guaranteed regardless of delegations of the minority, so λ1 ≥ 0. Theorem 9. If all delegations are optimal and the issue is fully covered (0 < k1 < k), the outcome will favor the minority if λ0 > kλ1+(N−λ1)(2k1−k) Proof. With only optimal delegations, the outcome favors the minority if (N − λ0 − λ1) k0 (N − λ0 − λ1) k1 k1 = k − k0. k + λ0 > k + λ1. The lower bound can be found directly by solving for λ0 and substituting 2k1 . 12 5.2 Probabilistic Delegation Instead of assuming that some fractions (α1, α0) of voters delegate we investigate what happens if each voter chooses to delegate with some fixed individual probability. These results gives us an idea of how motivated or attentive voters must be to improve the outcome of FRD over RD. We assume here that all voter and candidate preferences are independent for all issues. As all issues are independent, we will consider a single issue. Suppose each voter vj ∈ V chooses to delegate (deviate from the default) with independent probability pj and that all del- egations are optimal. Let xj ∈ [0, 1] be the amount of power voter vj assigns to candidates who agree with the voter majority (cl = 1), either by delegation or default. If vj defaults then j = 1) then xj = 1, and xj = k1 xj be the if vj delegates optimally and is in the voter minority then xj = 0. Let X1 = (cid:80) (cid:80) vj =1(pj + (1 − pj)k1/k) +(cid:80) total power assigned to these candidates via both delegation and default. Let µ = E[X1] = vj =0(1 − pj)k1/k be the expected value of the total power assigned k , if vj delegates optimally and is in the voter majority (vi vj∈V to representatives who agree with the voter majority. agree with the voter majority on an issue and X1 =(cid:80) Theorem 10. Consider an FRD with an odd number of voters N, odd committee size k, and only optimal delegations. Suppose each voter vj ∈ V delegates with probability pj on each issue such that µ > N/2. Then the probability that the outcome agrees with the voter majority is bounded by P (y = 1) ≥ 1 − e−(N−2µ)2/4N. Proof Sketch: The probability that the outcome agrees with the voter majority is P (y = 1) = P (X1 > N/2) + P (y = 1X1 = N/2)· P (X1 = N/2) where P (y = 1X1 = N/2) is due to some tie-breaking mechanism. First we show that with odd voters, odd representatives and only optimal delegations there can be no ties. Namely, X1 (cid:54)= X0 = N − X1. This proof is due to parity and holds regardless of the delegation rate. Without ties, we simply need to determine P (X1 > N/2). We use a Chernoff inequality to provide a lower bound on this value based on the probability of delegation pj of all voters. See Section 7 for discussion about tie-breaking 4. Proof. Recall that xj ∈ [0, 1] is the amount of voting power voter vj assigns to candidates who vj∈V xj. Given some tie breaking rule, we have that P (y = 1) = P (X1 > N/2) + P (y = 1X1 = N/2) · P (X1 = N/2). First we show that P (X1 = N/2) = 0, then we give a lower bound for P (X1 > N/2). Lemma 11. If N is odd, k is odd, and all delegations are optimal, then no ties are possible. j = k · xj where xj ∈ [0, 1] is the amount of weight (voting power) voter vj assigns Proof. Let x(cid:48) to candidates who agree with the voter majority on an issue via default or delegation. If vj defaults then x(cid:48) j = k, and if j ∈ {0, k1, k}. vj delegates optimally and is in the voter minority then x(cid:48) Let X(cid:48) 1, X(cid:48) 0 = kN. Since kN is odd, it must be that X(cid:48) 0 have opposite parity and so they cannot be 4It is an interesting open question how the distribution of these delegation probabilities effects the outcome when voter and representative preferences are correlated or when this distribution changes over time based on the outcomes of previous issues. j = k1, if vj delegates optimally and is in the voter majority (vi j = 1) then x(cid:48) j = 0. Therefore, ∀j : x(cid:48) 0 are non-negative integers and X(cid:48) j). Then X(cid:48) 1 and X(cid:48) 1 = (cid:80) vj∈V 0 = (cid:80) vj∈V j and X(cid:48) x(cid:48) (k − x(cid:48) 1 + X(cid:48) 13 X1 = (cid:80) equal. Therefore X1 = X(cid:48) representatives on either side of the issue cannot be equal, so no ties may occur. 0/k, meaning the total amounts of weight delegated to the 1/k (cid:54)= X0 = X(cid:48) vj∈V Given that no ties are possible, we have that P (y = 1) = P (X1 > N/2). Remember that xj where xj is the total weight that vj delegates to representatives who agree with the j = 0 then E[xj] = (1 − pj) k1 voter majority. If vi k . Let µ = E[X1] be the expected total weight assigned to representatives who agrees with the voter majority. j = 1 then E[xj] = pj + (1 − pj) k1 k , else if vi (cid:88) vi j =1 µ = (pj + (1 − pj) k1 k ) + (1 − pj) k1 k (cid:88) vi j =0 We now use the fact that P (X1 > N/2) = 1 − P (X1 ≤ N/2). Let δ = (2µ − N )/2µ. If µ > N/2, then δ > 0. This allows us to apply a Chernoff bound to derive our lower bound P (X1 > N/2) = 1− P (X1 ≤ N/2) = 1− P (X1 ≤ (1− δ)µ) ≥ 1− e−δ2µ2/N = 1− e−(2µ−N )2/4N This bound depends on the mild condition that µ > N/2. This requires that the delegation rate of the majority α1 cannot be too small relative to the minority α0. Observe that this condition is satisfied when ∀j : pj = p and k1 > k/2. Furthermore, as an increasing number of voters delegate optimally, we expect µ → N1 > N/2 regardless of k1. Naturally, as the delegation rate increases (α → 1), we observe our lower bound approach the ideal 1− e−(2µ−N )2/4N → 1. To find µ ≤ N/2, the voter majority cannot be too large compared to the voter minority, k1 must be smaller than or somewhat close to k0, and/or the voters in the majority must be significantly more apathetic towards delegation than voters in the minority. Tighter bounds may be achieved when the delegation probabilities are assumed to come from a particular distribution. It is an interesting open question to see how the expected outcome is effected when voters have various motivations to delegate which give rise to different delegation probability distributions. 5.3 Simulated Delegations Given our theoretical results on how FRD can improve the outcomes of a decision making process, we investigate the effect of the overall delegation rate α on recovering the ideal outcomes according to Direct Democracy. We use the same model to generate candidates and voter preferences as used in Section 4. For our simulated delegations we create instances with V = 301, C = 60, S = 150, and k = 21. We vary α ∈ {0, 1.0} in increments of 0.01 and for each setting of α we run 50 iterations. We plot the means in Figure 3 and note again that the variance is ≤ 0.002 at every point. In Figure 3 we can see the agreement of the outcomes of FRD and RD for the weighted voting committee selection rule for several delegation types and delegation rates. For each of the instances we measure against a baseline of the proportion of issues where the outcome is that of DD. A value of 1.0 means that the outcomes of all issues are the same under the democracy as they would be in a Direct Democracy. 14 Figure 3: Weighted Voting Committee Selection We compare RD with four different delegation schemes: (1) Approve where voters delegate evenly to the representatives in the committee of whom they approve and do not update; Best Rep where voters delegate to their single most preferred representative and do not update per issue; Best-3 Rep where voters delegate equally to their three most preferred members of the committee and do not update per issue; and finally Optimal where voters delegate to a single representative with whom they agree per issue. Most surprising is how little delegations that are not active and optimal help emulate Direct Democracy. The Approve delegation system is perhaps closest to the proposal of Proxy Voting espoused by [33] but does not improve RD in a meaningful way. Similarly interesting are the 1-Best and 3-Best delegations, which also do not move the outcome towards the ideal of Direct Democracy. Hence, we can see that the issue-specific flexibility FRD allows can be effectively used to improve outcomes of decision making systems. Another striking result in Figure 3 is how drastically FRD can improve agreement over RD when voters are highly attentive. With as little as 60% of the population delegating we can improve agreement by 10%, and when the delegation rate reaches 80% we see an almost 20% increase, eventually reaching 100% when everyone delegates if the issue is fully covered. 6 Acknowledgements We would like to thank Niccolo Dalmasso for his insight and many constructive conversations. 7 Conclusion We have introduced a novel system of Interactive Democracy called Flexible Representative Democracy which transitions smoothly, at the discretion of the voters, between direct and repre- sentative. We have shown theoretically and empirically that FRD has the potential to overcome the shortcomings of other systems such as Representative Democracy and Proxy Voting. An important point to remember is that in FRD delegations are optional, and not an additional burden imposed on the system or voters. In contrast to Liquid Democracy, voters in FRD have greater certainly about how their vote will be cast ahead of time and delegation cycles are not possible. Furthermore, FRD maintains a fixed, elected set of accountable representatives to produce legislation and hold public debates. This committee of representatives does not need to expand to guarantee proportional or 15 020406080100Alpha Percentage0.50.60.70.80.91.0Agreement with Direct DemocracyFRD: OptimalFRD: Best RepFRD: Best-3 RepsFRD: ApproveRDjustified representation, because as long as there is full coverage of an issue voters have the power to collectively guarantee these properties for themselves. In our analysis, we seek to create a best-case scenario for traditional RD to achieve high agree- ment: we assume the full list of issues is known, all voters participate, candidates are truthful and do not change their preferences after the election, and voter preferences over candidates are con- sistent with their preferences over the issues. Intuitively, relaxing any of these assumptions only strengthens the argument for enabling flexible, issue-specific delegations. Flexible delegation also has the effect of minimizing the role that the choice of election rule plays in the outcome. 8 Extensions In addition to the model laid out here, there are a number of interesting and important extensions to FRD that one could consider. Nothing in our system prevents moving to issues where there are non-binary domains and/or different domains for every issue. One can also consider asymmetric issues with decision rules other than simple majority rule, such as quota rules. Additionally, we have modeled the weight given to representatives as a simple sum of the defaults and delegations, but this could be any function and may treat delegations and defaults differently; though such a function should be monotone with respect to the delegated weights. Note that we can easily relax the assumptions in our analysis that there are an odd number of voters, optimal delegations, and no abstentions. Relaxing any of these requires us only to account for potential ties in the outcomes of our analysis. If ties are broken randomly, then we only need to compute the probability of a tie occurring. This probability is computable, but in general we should not expect the xj values to come from a nice, symmetric, well-behaved distribution. It is also worth noting that ties can be broken in other ways including based on the observed delegation rate. For example, when the delegation rate is low one might break ties in favor of the representative majority, whereas when the delegation rate is high one might break ties in favor of the outcome with more weight delegated to it. Lastly, one of the attractive features of Liquid Democracy is the notion of voluntary represen- tatives who need not be formally elected. These voluntary representatives are not beholden to an election cycle and can be local leaders, with personal relationships to voters who support them. This can be incorporated into a Flexible Representative Democracy by allowing any voter or agent to become a voluntary representative on a single issue by casting their vote publicly by the same deadline as the elected representatives. However, voluntary representatives do not receive any vot- ing power by default, may not receive delegated voting power on any issue before their vote is declared, and can only receive delegations on a per-issue basis. Since no voluntary representatives receive any default voting power, all of our results still hold. In fact, the addition of representatives who receive no power by default constitutes a Pareto improvement because they effectively serve to guarantee full coverage (and coverage) and if the outcome changes as a result it can only change in the direction of the voter majority (assuming optimal delegations). Alternatively, an FRD can automatically add contrarian single-issue dummy representatives whenever the representatives do not achieve full coverage on an issue. Keep in mind that while full coverage can be artificially guaranteed in this way, any attempt to guarantee majority agreement would constitute rigging the election. 16 References [1] Dan Alger. Voting by proxy. Public Choice, 126(1/2):1 -- 26, 2006. [2] Gertrude EM Anscombe. On frustration of the majority by fulfilment of the majority's will. Analysis, 36(4):161 -- 168, 1976. [3] Emmanuelle Auriol and Robert J Gary-Bobo. On the optimal number of representatives. Public Choice, 153(3-4):419 -- 445, 2012. [4] Haris Aziz, Markus Brill, Vincent Conitzer, Edith Elkind, Rupert Freeman, and Toby Walsh. Justified representation in approval-based committee voting. Social Choice and Welfare, 48(2):461 -- 485, 2017. [5] Eyal Baharad, Jacob Goldberger, Moshe Koppel, and Shmuel Nitzan. Beyond condorcet: optimal aggregation rules using voting records. Theory and decision, 72(1):113 -- 130, 2012. [6] John F Banzhaf III. Weighted voting doesn't work: A mathematical analysis. Rutgers L. Rev., 19:317, 1964. [7] Jan Behrens, Axel Kistner, Andreas Nitsche, and Bjorn Swierczek. The Principles of Liquid- Feedback. Interacktive Demokratie, 2014. [8] Ruth C Ben-Yashar and Shmuel I Nitzan. The optimal decision rule for fixed-size committees in dichotomous choice situations: the general result. International Economic Review, pages 175 -- 186, 1997. [9] Daan Bloembergen, Davide Grossi, and Martin Lackner. On rational delegations in liquid democracy. CoRR, abs/1802.08020, 2018. [10] Christian Blum and Christina Isabel Zuber. Liquid democracy: Potentials, problems, and perspectives. Journal of Political Philosophy, 24(2):162 -- 182, 2016. [11] F. Brandt, V. Conitzer, U. Endriss, J. Lang, and A. D. Procaccia, editors. Handbook of Com- putational Social Choice. Cambridge University Press, 2016. [12] M. Brill. Interactive democracy. In Proc. 17th AAMAS, pages 1183 -- 1187, 2018. [13] Markus Brill and Nimrod Talmon. Pairwise liquid democracy. In Proc. 27th IJCAI, pages 137 -- 143, 2018. [14] Zo´e Christoff and Davide Grossi. Binary voting with delegable proxy: An analysis of liquid democracy. In Proc. 16th TARK, pages 134 -- 150, 2017. [15] P. R. Cohen. Empirical Methods for Artificial Intelligence. MIT Press, 1995. [16] Gal Cohensius, Shie Mannor, Reshef Meir, Eli A. Meirom, and Ariel Orda. Proxy voting for better outcomes. In Proc. 16th AAMAS, pages 858 -- 866, 2017. 17 [17] C. Domshlak, E. Hullermeier, S. Kaci, and H. Prade. Preferences in AI: An overview. AI, 175(7):1037 -- 1052, 2011. [18] John Dunn. Democracy: The unfinished journey. Oxford University Press, 1995. [19] Uriel Feige. A threshold of ln n for approximating set cover. Journal of the ACM (JACM), 45(4):634 -- 652, 1998. [20] Scott L Feld and Bernard Grofman. The accuracy of group majority decisions in groups with added members. Public Choice, 42(3):273 -- 285, 1984. [21] Bryan Ford. Delegative democracy. Unpublished Manuscript. Available at http://ww.brynosaurus.com/deleg/deleg.pdf, 2002. [22] J. Furnkranz and E. Hullermeier. Preference Learning. Springer, 2010. [23] M. R. Garey and D. S. Johnson. Computers and Intractability, A Guide to the Theory of NP-Completeness. W.H. Freeman and Company, 1979. [24] James Green-Armytage. Direct voting and proxy voting. Constitutional Political Economy, 26(2):190 -- 220, 2015. [25] Bernard Grofman and Scott L Feld. Determining optimal weights for expert judgment. In Information Pooling and Group Decision Making, pages 167 -- 72. JAI Press Greenwich, CT, 1983. [26] Bernard Grofman, Guillermo Owen, and Scott L Feld. Thirteen theorems in search of the truth. Theory and decision, 15(3):261 -- 278, 1983. [27] Steve Hardt and Lia CR Lopes. Google votes: A liquid democracy experiment on a corporate social network. Technical Disclosure Commons, 2015. [28] Anson Kahng, Simon Mackenzie, and Ariel D Procaccia. Liquid democracy: An algorithmic perspectiven. In Proc. 32nd AAAI, 2018. [29] Drora Karotkin and Jacob Paroush. Optimum committee size: Quality-versus-quantity dilemma. Social Choice and Welfare, 20(3):429 -- 441, 2003. [30] Ashiqur R KhudaBukhsh, Lin Xu, Holger H Hoos, and Kevin Leyton-Brown. SATenstein: Automatically building local search sat solvers from components. AI, 232:20 -- 42, 2016. [31] J. W. Lian, N. Mattei, R. Noble, and T. Walsh. The conference paper assignment problem: Using order weighted averages to assign indivisible goods. In Proc. 33rd AAAI, 2018. [32] Malik Magdon-Ismail and Lirong Xia. A mathematical model for optimal decisions in a representative democracy. arXiv preprint arXiv:1807.06157, 2018. [33] James C Miller. A program for direct and proxy voting in the legislative process. Public choice, 7(1):107 -- 113, 1969. 18 [34] Shmuel Nitzan and Jacob Paroush. Optimal decision rules in uncertain dichotomous choice situations. International Economic Review, pages 289 -- 297, 1982. [35] Shmuel Nitzan and Jacob Paroush. Collective decision making and jury theorems. Oxford handbook of law and economics. Oxford: Oxford University Press. Google Scholar, 2017. [36] Jacob Paroush and Drora Karotkin. Robustness of optimal majority rules over teams with changing size. Social choice and welfare, 6(2):127 -- 138, 1989. [37] Marko Antonio Rodriguez and Daniel Joshua Steinbock. Societal-scale decision making using social networks. 2004. [38] Lloyd S Shapley and Martin Shubik. A method for evaluating the distribution of power in a committee system. American political science review, 48(3):787 -- 792, 1954. [39] Martin Shubik. On homo politicus and the instant referendum. Public Choice, 9, 1970. [40] P. Skowron. What do we elect committees for? A voting committee model for multi-winner rules. In Proc. 24th IJCAI, 2015. [41] P. Skowron, P. Faliszewski, and J. Lang. Finding a collective set of items: From proportional multi-representation to group recommendation. AI, 241:191 -- 216, 2016. [42] John H Smith. Aggregation of preferences with variable electorate. Econometrica: Journal of the Econometric Society, pages 1027 -- 1041, 1973. [43] Gordon Tullock. Toward a mathematics of politics. University of Michigan Press, 1967. [44] Gordon Tullock. Computerizing politics. Mathematical and Computer Modelling, 16(8- 9):59 -- 65, 1992. [45] Chih-Han Yu, Justin Werfel, and Radhika Nagpal. Collective decision-making in multi-agent systems by implicit leadership. In Proc. 9th AAMAS, pages 1189 -- 1196, 2010. [46] Haibin Zhu, MengChu Zhou, and Rob Alkins. Group role assignment via a Kuhn -- Munkres IEEE Transactions on Systems, Man, and Cybernetics-Part A: algorithm-based solution. Systems and Humans, 42(3):739 -- 750, 2012. [47] William S Zwicker. In F. Brandt, V. Conitzer, U. Endriss, J. Lang, and A. D. Procaccia, editors, Handbook of Computational Social Choice, chapter 2. Cambridge University Press, 2015. Introduction to voting theory. 19 9 Appendix 9.1 Preference Representation For clarification, there are two ways voters could report their strict orderings over the candidates. The first is a list where the position in the list denotes the rank and the value in the list denotes the candidate. This is what is typically thought of as an ordering. We use a second, equivalent, representation where the position in the list denotes the candidate and the value denotes their rank. Consider a paper ballot in which you want voters to rank their candidates in order of preference. The first representation corresponds to writing the numbers 1, . . . , m on the paper and having the voters fill in the names of the candidates next to them. Our second representation is like listing the candidates on the page, and having the voters fill in the numbers next to the names to denote their rank. The choice of representation in practice does not impact our analysis, as they represent the exact same information. Similarly with approvals. Agents can just report the subset of candidates of which they approve, or they could write whether they approve next to each candidate name on a ballot. Again, we use the second representation and the two are equivalent. (a) Weighted Voting Committee Selection (b) Approval Voting Committee Selection (c) Borda Committee Selection Figure 4: Performance of FRD with four types of delegation versus RD for various committee selection rules and values of α. k + λ0, or equivalently (N − λ0 − λ1) k1 k + λ1 > N/2. k + λ1 > 9.2 Deterministic Theorems The following theorems are all derived from the fact that y = 1 if (N − λ0 − λ1) k1 (N − λ0 − λ1) k0 Theorem 12. If all delegations are optimal, and the representatives are not unanimous (e.g. the issue is fully covered where 0 < k1 < k), the outcome is guaranteed to agree with the voter majority if the number of voters in the majority who delegate (λ1) is greater than N k−2N1k1 2(k−k1) Proof. For the majority to guarantee the outcome in their favor by delegating optimally, this means that even if all voters in the minority delegate optimally, the outcome must still be 1. Recall that X1 is the total weight assigned by default and delegation to representatives who prefer the outcome 1. k +λ1 = (N1−λ1) k1 Let X1 = (N−λ0−λ1) k1 k +λ1. For the outcome to be guaranteed in favor of the voter majority, it must be that X1 = (N1−λ1) k1 k +λ1 > 2 . Solving for λ1 we find that λ1 > N k−2N1k1 . Note that this lower bound may be negative. In this 2(k−k1) k +λ1. If λ0 = N0, then X1 = (N−N0−λ1) k1 . N 20 020406080100Alpha Percentage0.50.60.70.80.91.0Agreement with Direct DemocracyFRD: OptimalFRD: Best RepFRD: Best-3 RepsFRD: ApproveRD020406080100Alpha Percentage0.50.60.70.80.91.0Agreement with Direct DemocracyFRD: OptimalFRD: Best RepFRD: Best-3 RepsFRD: ApproveRD020406080100Alpha Percentage0.50.60.70.80.91.0Agreement with Direct DemocracyFRD: OptimalFRD: Best RepFRD: Best-3 RepsFRD: ApproveRDcase k1 is so large and N0 is so small that the outcome is guaranteed regardless of delegations of the minority, so λ1 ≥ 0. Theorem 13. If all delegations are optimal, and the representatives are not unanimous (0 < k1 < k), the outcome will favor the minority if λ0 > kλ1+(N−λ1)(2k1−k) Proof. With only optimal delegations, the outcome favors the minority if (N − λ0 − λ1) k0 (N − λ0 − λ1) k1 k1 = k − k0. Theorem 5. No Representative Democracy using a Condorcet-consistent election rule in which voters report approvals or total orderings over the candidates can provide a bounded approxima- tion of agreement with majority voting in a Direct Democracy. k + λ0 > k + λ1. The lower bound can be found directly by solving for λ0 and substituting 2k1 . Proof. The example that proves this statement comes from [2]. We have two candidates; the ideal candidate c1 and the worst conceivable candidate c2. In this case, the majority of the voters prefer the worse candidate because they agree on a greater number of issues. This also means that the majority of voters approve of the worse candidate but not the ideal candidate. This pathology arises because the majority of voters are in the minority on the majority of issues. Thus, if we only have duplicates of these two candidates for any k, there are cases in which the voters will elect representatives who achieve a majority agreement of 0 when a set of candidates exists who would achieve and agreement of 1. That is, L(ODD,ORD) = 0. s1 0 1 1 1 1 0 0 1 1 0 0 1 0 s2 0 1 0 1 1 0 0 1 0 1 1 1 0 s3 0 1 0 0 1 0 0 1 1 1 1 1 0 s4 0 1 0 0 0 1 1 0 1 1 1 1 0 v1 v2 v3 v4 v5 v6 v7 v8 v9 v10 v11 c1 c2 s6 0 0 0 0 0 1 1 1 1 1 1 1 0 s5 0 0 0 0 1 1 1 0 1 1 1 1 0 21 s7 0 0 0 0 0 1 1 1 1 1 1 1 0 s8 1 0 0 0 0 0 1 1 1 1 1 1 0 s9 1 0 1 0 0 0 0 1 1 1 1 1 0 s10 1 0 1 1 0 0 0 1 1 0 1 1 0 s11 1 0 1 1 0 0 0 1 0 1 1 1 0 Figure 5: Comparison of NP-hard rules with our polynomial rules for C = 17,S = 80,V = 51. We cannot scale this graph in the same was Figure 1 due to the high computational cost of computing the winning sets for k−Median and CC. However, from this small sample we see that Weights, STV, and AV all strictly dominate both CC and k−Median in terms of agreement. 9.3 Additional Simulated Results For a first set of simulations we included rules that have NP-hard winner determination problems but are optimal according to a different utility metric : Chamberlin-Courant and k−Median [40]. Even with a relatively optimized implementation it still took almost 24 hours to generate Figure 5 on a server with 16 cores and 32 GB of memory. Looking at Figure 5 we see that both CC and k−Median are dominated by Weights, STV, and AV at all committee sizes. Hence, to run experiments on larger samples of voters and issues we drop the NP-hard rules as they preform similarly to the easier to compute rules. The results in Figure 5 are interesting in light of the theoretical results obtained by [40]. [40] show that it is optimal if we elect a committee using k-Median and use majority voting by the representatives. However, their optimality criteria explicitly incorporates the proportion of voters who are (mis)represented in the objective function. For agreement we do not trade off the weighting of how many voters are (mis)represented, we only look at the total agreement. It is interesting that CC and k-median are not optimal and an interesting direction for future work is investigating this phenomona. In Figure 6 we can see the agreement of the outcomes of FRD and RD for the weighted voting, AV, and Borda election rules for several delegation types and delegation rates. For each of the instances we measure against a baseline of the proportion of issues where the outcome is that of DD. A value of 1.0 means that all issues are the same under the democracy as they would be in a Direct Democracy. Overall, this plots show that RD under any of these rules cannot achieve above ≈ 62% agreement with Direct Democracy in expectation. Similarly interesting are the 1-Best and 3-Best delegations, which also do not move the outcome towards the ideal of Direct Democracy. Hence, we can see that the issue-specific flexibility FRD allows can be effectively used to improve outcomes of decision making systems. Another striking result in Figure 6 is how drastically FRD can improve agreement over RD when voters are highly attentive. With as little as 60% of the population delegating we can improve agreement by 10%, and when the delegation rate reaches 80% we see an almost 20% increase, eventually reaching 22 4681012Committee Size (k)0.400.450.500.550.600.650.70Agreement MeanAV AgreementBorda AgreementSTV AgreementRAV AgreementRandom AgreementWeights AgreementChamberlin-Courant Agreementk-Median Agreement(a) Weighted Voting Committee Selection (b) Approval Voting Committee Selection (c) Borda Committee Selection Figure 6: Performance of FRD with four types of delegation versus RD for various committee selection rules and values of α. 100% when everyone delegates if the representatives are not unanimous. It is an interesting open question to see how this behavior changes when voters have specific motivations to delegate such as the makeup of the representatives or outcomes of past issues. 23 020406080100Alpha Percentage0.50.60.70.80.91.0Agreement with Direct DemocracyFRD: OptimalFRD: Best RepFRD: Best-3 RepsFRD: ApproveRD020406080100Alpha Percentage0.50.60.70.80.91.0Agreement with Direct DemocracyFRD: OptimalFRD: Best RepFRD: Best-3 RepsFRD: ApproveRD020406080100Alpha Percentage0.50.60.70.80.91.0Agreement with Direct DemocracyFRD: OptimalFRD: Best RepFRD: Best-3 RepsFRD: ApproveRD
1709.02556	1	1709	2017-09-08T06:35:10	Game Theory Models for the Verification of the Collective Behaviour of Autonomous Cars	[ "cs.MA", "cs.GT" ]	The collective of autonomous cars is expected to generate almost optimal traffic. In this position paper we discuss the multi-agent models and the verification results of the collective behaviour of autonomous cars. We argue that non-cooperative autonomous adaptation cannot guarantee optimal behaviour. The conjecture is that intention aware adaptation with a constraint on simultaneous decision making has the potential to avoid unwanted behaviour. The online routing game model is expected to be the basis to formally prove this conjecture.	cs.MA	cs	Game Theory Models for the Verification of the Collective Behaviour of Autonomous Cars L´aszl´o Z. Varga Faculty of Informatics ELTE Eotvos Lor´and University Budapest, Hungary http://people.inf.elte.hu/lzvarga The collective of autonomous cars is expected to generate almost optimal traffic. In this position paper we discuss the multi-agent models and the verification results of the collective behaviour of autonomous cars. We argue that non-cooperative autonomous adaptation cannot guarantee optimal behaviour. The conjecture is that intention aware adaptation with a constraint on simultaneous de- cision making has the potential to avoid unwanted behaviour. The online routing game model is expected to be the basis to formally prove this conjecture. 1 Introduction Autonomous cars open new possibilities and offer several benefits. These benefits include: increased mobility of the elderly and disabled people; better utilisation of travel time; finding urban places faster; more efficient traffic flow; less congestion; increased fuel efficiency. The last three benefits imply that the collective behaviour of autonomous cars will be close to a kind of optimum on the collective level. We are going to discuss the formal verification of this promise. Because we the verification is aimed at proving the properties of the collective behaviour of decentralised autonomous entities, the formal proofs are somewhat different from classical formal verification methods. Autonomous cars detect their environment using different sensors like radar, LIDAR, GPS, computer vision, digital map, real-time traffic information and shared information. The planning unit of the au- tonomous car merges and interprets this sensory information to determine the necessary control actions to navigate the car to its destination and to avoid obstacles. As long as we focus on a single autonomous car, we can say that the planning unit executes centralised adaptation, because there is only one actor that senses the environment and takes actions to adapt to the changing environment. However if the road network is populated by several autonomous cars, then the overall traffic will emerge as the result of the collective behaviour of several autonomous cars. If there are several actors that sense the environment and take autonomous actions, then it is decentralized adaptation. If autonomous cars are designed with only centralized adaptation in focus, then they may be able to avoid obstacles and navigate to their desti- nation, but if they meet other autonomous cars, then their joint actions may generate unwanted behaviour in some situations. The human driven vehicle population may also bring about unwanted behaviour sometimes, but the issue will be more critical in the case of autonomous cars, because of two major differences between human driven and autonomous cars. One difference is that human drivers may be psychologically influ- enced, while autonomous cars always make rational decisions. Human drivers may follow their habits, although these habits may not be optimal and these habits may be unwanted for the overall traffic. The psychologically influenced decision of human drivers may sometimes result in preferable altruistic be- haviour, but sometimes it may result in unwanted panic-like behaviour. On the contrary, autonomous cars L. Bulwahn, M. Kamali, S. Linker (Eds.): First Workshop on Formal Verification of Autonomous Vehicles (FVAV 2017). EPTCS 257, 2017, pp. 27–34, doi:10.4204/EPTCS.257.4 c(cid:13) L.Z. Varga This work is licensed under the Creative Commons Attribution-Noncommercial License. 28 GameTheory Models for the Verification of the Collective Behaviour of Autonomous Cars always follow their designed rational preferences. The other difference is that human drivers may not always be aware of the relevant information, while autonomous cars always make informed decisions. Although humans are better in many cognitive tasks than machines, the machines have wider sensory capabilities than humans. Machines can use telecommunication technologies to "see" beyond objects (e.g. the approaching car behind the corner) and to "see" much farther away (e.g. congestion along the planned route on the other side of the city). As more and more information services are deployed to provide real-time traffic information to traffic participants, autonomous cars will have real-time and more precise information than humans. On the other hand, informed decision raises the issue of security and dependence on technology. Information might be provided by malicious sources, or the data may not be reliable or accurate. Up-to-date information is a critical issue according to a recent study [10], but we presume that autonomous cars will receive exact enough real-time data. The verification of centralised adaptation is necessary to ensure that autonomous cars can safely move on the street. Verification of decentralised adaptation is complementary to the verification of centralised adaptation, and it makes sure that the collective of autonomous cars do not produce unwanted behaviour, like for example inefficient usage of the road infrastructure. In Section 2 we overview the main non-cooperative game theory models of decentralised adaptation and we highlight the main conclusions of the verification results from these models. In Section 3 we discuss methods that improve the properties of decentralised adaptive systems, we present their models, and we highlight the main conclusions of the verification results from these models. In Section 4 we conclude the paper with the conjecture that unwanted behaviour in decentralised adaptation could be avoided with the help of intention aware predictions, however this needs further research, which is among the goals of the investigations within the EFOP-3.6.3-VEKOP-16 project in connection with the RECAR [14] project. 2 Non-cooperative Decentralised Autonomous Adaptation The basic characteristic of decentralised autonomous adaptation is that there are several autonomous actors, called agents, which make decisions on which action to perform. The execution of the action of an agent uses limited resources, which are shared by several agents. The more agents use a resource, the less the agents prefer to use that resource. This is usually modelled with a cost function of the resource. The result of the action of an agent depends not only on its own action, but also on the action of all the other agents in its environment, because other agents may decide autonomously to use the same resource. An agent may not know which actions the other agents intend to do, therefore an agent may be uncertain which is the best action to perform. The decentralised decision making of the autonomous agents is the subject of multi-agent research [27]. The current belief [15] is that the best model of multi-agent decision making is founded in game theory [18]. In accordance with game theory [13], the agents prefer some states of the environment to other states of the environment, which is modelled with a utility function. If the agents do not cooperate, then the rational agent selects that action which has the highest value among the worst possible utility of the outcomes of its own action and the actions of other agents. The actions of the agents are in equilibrium if no agent can benefit by changing its action while the other agents keep their actions unchanged. The efficiency of the multi-agent system can be measured as a combination of the utilities of all of the agents. A simple efficiency measure is the sum of the utilities of all the agents. The multi-agent system is optimal if the sum of the utilities is maximal. Non-cooperative decentralized autonomous decision making may not lead to optimal result. This inefficiency is measured with the price of anarchy which is the ratio L.Z. Varga 29 Figure 1: A simple minority game like situation between efficiency measure of the equilibrium and the optimum. Game theory verified that the price of anarchy has an upper limit in some routing problems. The routing problem is a network with source routing, where end users simultaneously choose a full route to their destination and the traffic is routed in a congestion sensitive manner. If the cost functions are linear functions of the traffic flow, then the price of anarchy is at most 4 ÷ 3 [16], i.e. this is how bad the overall traffic is when decentralised autonomous decision making is applied by the traffic flow. There are games with many equilibria. In this case, if there is no coordination, then agents do not know which equilibrium is the goal of the collective and they may not select the right action. There are games where the equilibrium is not symmetric, i.e. some of the agents are not happy with the equilibrium, like in minority games [4]. In a minority game, the agents choose one of two choices independently, and the agents who end up on the minority side win. If every agent deterministically chooses the same action, then every agent is guaranteed to fail. The solution to this problem in game theory is to permit each agent to use a mixed strategy, where a choice is made with a particular probability. This may be good to model the macroscopic properties of diverse human behaviour, but random action must not be allowed to control autonomous cars in situations like for example on Figure 1. Cars B and C are parked. The road is narrow and only one of the autonomous cars A and D can pass at a time. If both decide to go first (programmed randomly or deterministically), then they block each other. If both decide to wait for the other, then the collective of cars A and D end up in a deadlock. Classic game theory is concerned with the equilibrium, like traffic engineers who assume that the traffic is always assigned in accordance with the equilibrium [25] [1]. They assume that all agents have complete knowledge about the game, and the agents come to the equilibrium with full rationality. This is not realistic, especially if agents do not cooperate and they have to make different actions depending on the actions of the others, because there are several equilibria. Therefore game theory investigated the evolutionary dynamics where the agents receive feedback by observing their own and other agents' actions and utility, and change their own actions based on these observations. It is verified that with this feedback assumption, the above mentioned routing game converges to the equilibrium [17] [9]. Another type of feedback is used in regret minimisation, where agents compare their actually experienced utility with the best possible utility in retrospect. It is verified that if the agents of the above routing game select actions to minimize their regret, then their behaviour will converge to the equilibrium [2]. However the investigations of these game dynamics have the following assumptions: the decision making is on the flow level; the game is repeated; and the decision is based on experiences from the previous games. This is not realistic in autonomous cars, where the decision making is done at the individual car level, and the decision is based on the real-time situation instead of previous experiences. Nowadays more and more services provide real-time information about the overall traffic situation. If we want to model that each subsequent agent of the traffic flow may select different route, depend- ing on the current traffic situation, then the above game theory models cannot be used. Autonomous cars continuously enter the road network, and each agent of the traffic flow decides its optimal route when it 30 GameTheory Models for the Verification of the Collective Behaviour of Autonomous Cars enters the road network, and the decision is based on the real-time information about the status of the road network. The outcome travel time for a given agent depends not only on the current characteristic of the network and the route choice of all the agents simultaneously entering the road network, but also on the trip schedule of other agents that have entered the network previously, enter the network simultaneously, or will enter the network later. The online routing game model (developed in [22], and later refined in [20]) models this case when each autonomous car may select different route, based on real-time traffic information. In order to measure the efficiency of real-time data usage, the benefit of online real-time data concept was defined in [22]. The benefit of online real-time data is the ratio between the travel time with real-time data based planning and the travel time without real-time data based planning. The agents are happy with real-time data, if the benefit value is below 1. Three types (worst/average/best) of benefit of online real-time data are needed in case an equilibrium traffic distribution cannot be achieved. It is proved in [22], that if the agents try to maximise their utility, then the following properties are true: equilibrium is not guaranteed; "single flow intensification" is possible; and the worst case benefit value of online real-time data is not guaranteed to be below 1. Equilibrium may not be reached, because the traffic may fluctuate. "Single flow intensification" happens when vehicles entering the road network later may select alternative faster routes, and they may catch up with the vehicles already on the road, and this way they cause congestion. All-in-all, sometimes the traffic may produce strange behaviour [21] and the collective of agents may be worse off by exploiting real-time information than without exploiting real-time information. The above results from the verification with the online routing game model indicates, that equilibrium cannot be verified if the agents autonomously follow their preferences to adapt to their environment. This is in line with experiments as well. For example the media supplement of [12] demonstrates that a small disturbance may bring about the fluctuation of the traffic and serious traffic jams are formed, if the agents apply non-cooperative decentralised adaptation. A specific algorithm can eliminate this effect if an autonomous car can proactively force speed on others [19]. However in this experiment the autonomous car plays the role of a kind of central controller. We can conclude from the above models, proofs and verification results, that non-cooperative decen- tralised autonomous adaptation to real-time data cannot guarantee to avoid unwanted behaviours of the collective of autonomous cars. The optimal traffic flow is not guaranteed, and an equilibrium traffic flow, which is worse than the optimal, is not guaranteed either. 3 Improved Decentralised Autonomous Adaptation Some form of cooperation should be built into autonomous cars to improve their collective behaviour. Cooperation is not only information exchange, but also coordinated actions within an agent commu- nity, which means that an agent's behaviour is influenced by the intentions and results of other agents. Cooperation proved to be useful in other application areas as well [7]. There are several ways to coordinate the actions of agents. One way is to centralise the multi- agent system and assign a control authority above the agents. For autonomous cars this would mean for example that each geographical territory would be under the control of a control authority. When an autonomous car reaches such territory, then it checks in at the control authority, and after the check-in, the control authority would tell the autonomous car the exact route to follow to its destination. This is somewhat similar to the operation of airports and how airplanes move in the area of the airport. The con- trol authority is centralised, therefore it can be verified with verification methods of centralised systems. In this control authority approach the autonomous car becomes something similar to a remote controlled L.Z. Varga 31 car. The users of autonomous cars may not accept this concept, especially if the commands of the central authority are not in line with the personal preferences and the individual has to suffer major drawback to facilitate the collective benefit. The central control authority is a reliability risk and a performance bottleneck as well. If the idea of the central control authority is given up, then another approach with some autonomy is when the coordination and communication is fostered by some kind of central service, but the control remains at the agents. The agents do not communicate directly with each other, but they communicate their intentions with the central service. The central service aggregates data about the agent collective and sends feedback to the agents [5]. The intention aware [26] and intention propagation [6] approaches are based on this scheme. When an agent has made a decision on its planned route, then it sends its selected intention to the central service. The central service is able to make a forecast of the future traffic situation based on the current traffic state and the communicated intentions of the agents. The central service provides the traffic forecast back to those agents who are still planning their trips, and these agents use this information to plan their trip, and when they have made a decision, then they also communicate their intentions to the central service. In theory, the online navigation software like Google Maps and Waze (Figure 2) know the intentions of the agents and could use this information to make predictions. The online routing game model was used to formally verify if the prediction power of the central service can solve the problem of avoiding unwanted behaviours of the collective of autonomous cars. The verification results [20] show that there is no guaran- tee on the value of the worst case benefit of online data and there is no guarantee on the equilibrium, i.e. the traffic may fluctu- ate. This means that if the agents selfishly exploit intention aware prediction, then in some networks and in some cases the traffic may be worse off by exploiting real-time information and predic- tion than without. This is due to the fact that the central service cannot take into account the decisions of those agents that make decisions simultaneously, and the decisions may depend on each other. However, it is verified [23] that in a small but complex enough network, where there is only one traffic flow and there- fore the agents that follow each other do not make decisions at the same time, then there is a guarantee on the value of the worst case benefit of online real-time data with prediction. In this case the agents might just slightly be worse off with real-time data and prediction in some cases. This verification result shows, that in the network of [23], the intention aware prediction establishes enough coordination among the agents. Currently the conjecture is that if simultaneous decision making among the agents is prevented, then intention aware prediction can limit the fluctuation in the multi-agent sys- tem [24] and the traffic converges to the equilibrium in bigger networks as well. This conjecture is an important challenge for the verification of the collective behaviour of autonomous cars. Figure 2: A service from the Google Play Store with prediction potential A critical issue of intention aware prediction is the trust in the intention submissions. If agents can profit from misleading other agents with revealing false intention, then they might be tempted to exploit this. In case of autonomous cars, the intention submission is done by the software built into the car. Car manufacturers will have to certify that their software submits its intention truthfully and correctly, and the software cannot be modified. If the idea of a central coordination service is given up, then the agents have to coordinate their activ- 32 GameTheory Models for the Verification of the Collective Behaviour of Autonomous Cars ities on a peer-to-peer basis. Two possible approaches are coalition formation and gossiping. Coalition formation may improve the behaviour of the agent collective, however if an agent can benefit from break- ing the coalition agreement, then an authority is needed to make sure that the coalition agreement is kept by the agents. The modelling and verification of coalition formation of autonomous cars can be founded in cooperative game theory [3]. Gossiping may be a means for spreading information [8] or aggregating information [11], however it is up to the agents how they use this information. The most likely usage is to predict future traffic situation as in the case of intention awareness, but gossiping does not need a central service. As we have seen, the guaranteed benefits of the prediction of future traffic situation has not yet been verified. We can conclude from the above discussion, that cooperation techniques based on intention aware- ness have been proposed, but the verification of the preferred collective behaviour of the agents is still a challenge. 4 Conclusion One of the main promises of autonomous cars is that they produce better and closer to optimum collec- tive behaviour than human drivers do. Each member agent of the collective of autonomous cars tries to adapt to the changing environment, therefore they execute decentralised autonomous adaptation. The verification of the preferred decentralised autonomous adaptation is important, because the collective of agents may produce unwanted behaviour, especially if they exploit real-time information about the whole collective. We have discussed models of collective agent behaviour, and discussed the verification results from these models. In this position paper, we argued that guaranteeing the avoidance of unwanted collective behaviour of non-cooperative agent collectives cannot be verified. Improved decentralised autonomous adaptation techniques try to establish some kind of cooperation among the agents, mainly through intention awareness. We have discussed models and verification results of intention aware col- lective agent behaviour. The verification process started recently, and the verification results from these models show that intention awareness improves the collective behaviour, but simultaneous decision mak- ing may still cause problems. The aggregation of intentions to predict future traffic state needs further research, too. We argued that these are critical issues, because if the wanted behaviour cannot be verified, than the only viable approach to ensure the close to optimum behaviour of the collective of autonomous agents is centralised control. We aim to discuss this at the workshop and investigate this issue in a future project. References [1] Martin J. Beckmann, C. B. McGuire & Christopher B. Winsten (1956): Studies in the economics of trans- portation. Yale University Press. [2] Avrim Blum, Eyal Even-Dar & Katrina Ligett (2006): Routing Without Regret: On Convergence to Nash Equilibria of Regret-minimizing Algorithms in Routing Games. In: Proceedings of the Twenty-fifth Annual ACM Symposium on Principles of Distributed Computing, PODC '06, ACM, New York, NY, USA, pp. 45–52, doi:10.1145/1146381.1146392. [3] Georgios Chalkiadakis, Edith Elkind & Michael Wooldridge (2011): Computational Aspects of Cooperative Game Theory (Synthesis Lectures on Artificial Inetlligence and Machine Learning), 1st edition. Morgan & Claypool Publishers, doi:10.2200/S00355ED1V01Y201107AIM016. L.Z. Varga 33 [4] Damien Challet, Matteo Marsili & Yi-Cheng Zhang (2005): Minority Games: Interacting Agents in Financial Markets (Oxford Finance Series). Oxford University Press, Inc., New York, NY, USA. [5] R. Claes & T. Holvoet (2014): Traffic Coordination Using Aggregation-Based Traffic Predictions. IEEE IntelligentSystems 29(4), pp. 96–100, doi:10.1109/MIS.2014.73. [6] R. Claes, T. Holvoet & D. Weyns (2011): A decentralized approach for anticipatory vehicle routing using delegate multi-agent systems. IEEE Transactions on Intelligent TransportationSystems 12(2), pp. 364–373, doi:10.1109/TITS.2011.2105867. [7] David Cockburn, Laszlo Z. Varga & Nick R. Jennings (1992): Cooperating Intelligent Systems for Electricity Distribution. In M A Bramer & R W Milne, editors: BCSExpertSystems92Conference(ApplicationTrack): Churchill College, Cambridge, UK, 15-17 December 1992., pp. 1–12. Available at https://eprints. soton.ac.uk/252130/. [8] Alan Demers, Dan Greene, Carl Hauser, Wes Irish, John Larson, Scott Shenker, Howard Sturgis, Dan Swine- hart & Doug Terry (1987): Epidemic Algorithms for Replicated Database Maintenance. In: Proceedings of the Sixth Annual ACM Symposium on Principles of Distributed Computing, PODC '87, ACM, New York, NY, USA, pp. 1–12, doi:10.1145/41840.41841. [9] Simon Fischer & Berthold Vcking (2004): On the Evolution of Selfish Routing. In: In Proc. of the 12th European Symposium on Algorithms (ESA 04, Springer-Verlag, pp. 323–334, doi:10.1007/ 978-3-540-30140-0_30. [10] Jim Gorzelany (2017): Available Says. bad-gps-routes-cost-drivers-29-hours-a-year-study-says/. Bad GPS Routes Cost Drivers at Study https://www.forbes.com/sites/jimgorzelany/2017/05/30/ 29 Hours A Year, [11] David Kempe, Alin Dobra & Johannes Gehrke (2003): Gossip-Based Computation of Aggregate Information. In: Proceedings of the 44th Annual IEEE Symposium on Foundations of Computer Science, FOCS '03, IEEE Computer Society, Washington, DC, USA, pp. 482–, doi:10.1109/SFCS.2003.1238221. Available at http://dl.acm.org/citation.cfm?id=946243.946317. [12] Akihiro Nakayama, Minoru Fukui, Macoto Kikuchi, Katsuya Hasebe, Katsuhiro Nishinari, Yuki Sugiyama, Shin ichi Tadaki & Satoshi Yukawa (2009): Metastability in the formation of an experimental traffic jam. New Journal of Physics 11(8), p. 083025, doi:10.1088/1367-2630/11/8/083025. Available at http:// iopscience.iop.org/article/10.1088/1367-2630/10/3/033001/data. [13] Noam Nisan, Tim Roughgarden, Eva Tardos & Vijay V. Vazirani (2007): Algorithmic Game Theory. Cam- bridge University Press, New York, NY, USA, doi:10.1017/CBO9780511800481. [14] RECAR (2017): Research Center for Autonomous Road Vehicles. Available at http://recar.bme.hu/ eng/. [15] Jeffrey S. Rosenschein (2013): Multiagent Systems, and the Search for Appropriate Foundations. In: Pro- ceedings of the 12th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2013), International Foundation for Autonomous Agents and Multiagent Systems (www.ifaamas.org), pp. 5–6. Available at http://www.ifaamas.org/Proceedings/aamas2013/docs/p5.pdf. [16] Tim Roughgarden: Routing games, pp. 461–486. In [13]. [17] William H. Sandholm (2001): Potential Games with Continuous Player Sets. Journal of Economic Theory 97(1), pp. 81 – 108, doi:10.1006/jeth.2000.2696. Available at http://www.sciencedirect.com/ science/article/pii/S0022053100926966. [18] Yoav Shoham & Kevin Leyton-Brown (2009): Multiagent Systems: Algorithmic, Game-Theoretic, and Log- ical Foundations. Cambridge University Press, New York, NY, USA. [19] Raphael E. Stern, Shumo Cui, Maria Laura Delle Monache, Rahul Bhadani, Matt Bunting, Miles Churchill, Nathaniel Hamilton, R'mani Haulcy, Hannah Pohlmann, Fangyu Wu, Benedetto Piccoli, Benjamin Seibold, Jonathan Sprinkle & Daniel B. Work (2017): Dissipation of stop-and-go waves via control of autonomous vehicles: Field experiments. CoRR abs/1705.01693. Available at http://arxiv.org/abs/1705.01693. 34 GameTheory Models for the Verification of the Collective Behaviour of Autonomous Cars [20] L´aszl´o Varga (2015): On Intention-Propagation-Based Prediction in Autonomously Self-adapting Naviga- tion. Scalable Computing: Practice and Experience 16(3), pp. 221–232. Available at http://www.scpe. org/index.php/scpe/article/view/1098. [21] L´aszl´o Z. Varga (2015): Paradox Phenomena in Autonomously Self-Adapting Navigation. Cybernetics and InformationTechnologies 15(5), pp. 78–87, doi:10.1515/cait-2015-0018. [22] L´aszl´o Zsolt Varga (2014): Online Routing Games and the Benefit of Online Data. In Franziska Klugl, Giuseppe Vizzari & Jir´ı Vokr´ınek, editors: ATT 2014 8th International Workshop on Agents in Traffic and Transportation, Held at the 13th Int. Conf. on Autonomous Agents and Multiagent Systems (AAMAS 2014),May5-6,2014,Paris, France, pp. 88–95. Available at http://www.ia.urjc.es/ATT/documents/ ATT2014proceedings.pdf. [23] L´aszl´o Zsolt Varga (2016): Benefit of Online Real-Time Data in the Braess Paradox with Anticipatory Rout- ing. In Samuel Kounev, Holger Giese & Jie Liu, editors: 2016IEEEInternationalConferenceonAutonomic Computing, ICAC 2016, Wuerzburg, Germany, July 17-22, 2016, IEEE Computer Society, pp. 245–250, doi:10.1109/ICAC.2016.68. [24] L´aszl´o Zsolt Varga (2016): How Good Is Predictive Routing in the Online Version of the Braess Paradox? In Gal A. Kaminka, Maria Fox, Paolo Bouquet, Eyke Hullermeier, Virginia Dignum, Frank Dignum & Frank van Harmelen, editors: ECAI 2016 - 22nd European Conference on Artificial Intelligence, 29 August-2 September2016,TheHague,TheNetherlands, FrontiersinArtificialIntelligenceandApplications 285, IOS Press, pp. 1696–1697, doi:10.3233/978-1-61499-672-9-1696. [25] John Glen Wardrop (1952): Some theoretical aspects of road traffic research. Proceedingsof the Institution of Civil Engineers,PartII 1(36), pp. 352–378. [26] M. M. de Weerdt, S. Stein, E. H. Gerding, V. Robu & N. R. Jennings (2016): Intention-Aware Routing of Electric Vehicles. IEEE Transactions on Intelligent Transportation Systems 17(5), pp. 1472–1482, doi:10. 1109/TITS.2015.2506900. [27] Michael Wooldridge (2009): An Introduction to MultiAgent Systems, 2nd Edition. John Wiley & Sons.
1911.05907	1	1911	2019-11-14T02:45:50	A Dynamic Preference Logic for reasoning about Agent Programming	[ "cs.MA", "cs.LO" ]	In this work, we investigate the use of Dynamic Preference Logic to encode BDI mental attitudes. Further, exploring this codification and the representation of preferences over possible worlds by preferences over propositional formulas, here called priority graphs, we comment on how to interpret BDI agent programs in this logic. Also, using the connection between dynamic operations defined over preference models and their encoding as transformations on priority graphs, we show how our logic can be used not only to reason about agent programs, but as a tool to specify reasoning mechanisms to guarantee certain properties in the theory of rationality for the programming language.	cs.MA	cs	A Dynamic Preference Logic for reasoning about Agent Programming Marlo Souza ´Alvaro Moreira Federal University of Bahia - UFBA Federal University of Rio Grande do Sul - UFRGS Salvador, Brazil Email: [email protected] Renata Vieira Pontifical Catholic University of Rio Grande do Sul - PUCRS Porto Alegre, Brazil Email: [email protected] Porto Alegre, Brazil Email: [email protected] John-Jules Ch. Meyer Utrecht University Uthrecht, the Netherlands Email: [email protected] Abstract -- In this work, we investigate the use of Dynamic Preference Logic to encode BDI mental attitudes. Further, ex- ploring this codification and the representation of preferences over possible worlds by preferences over propositional formulas, here called priority graphs, we comment on how to interpret BDI agent programs in this logic. Also, using the connection between dynamic operations defined over preference models and their encoding as transformations on priority graphs, we show how our logic can be used not only to reason about agent programs, but as a tool to specify reasoning mechanisms to guarantee certain properties in the theory of rationality for the programming language. Index Terms -- Dynamic Epistemic Logic; Agent Programming; Formal Semantics; BDI Logics; I. INTRODUCTION The formalisation of mental attitudes have been the object of much discussion in Logic and Philosophy and many such formalisations have been proposed. One of the most influential semantic frameworks in agent specification is the so-called BDI framework, which focuses on the Belief, Desire and Intention attitudes inspiring the development of many agent- oriented programming languages. While the engineering of such languages has been much discussed, the connections between the theoretical work on Philosophy and Artificial Intelligence and its implementations in programming languages are not so clearly understood yet. This distance between theory and practice has been acknowl- edged in the literature for agent programming languages and is commonly known as the "semantic gap". Interpretations for the mental attitudes of the BDI frame- work have been constructed for some BDI agent-oriented languages such as AgentSpeak(L) [1], GOAL [2], etc. These attempts follow the approach of constructing a logic of mental attitudes based on the formal semantics of the programming language. While this approach has the clear advantage of highlighting the meaning of mental attitudes diffused in the programming language, it is not clear how one can use such logics to construct programs - or propose changes in the language semantics - that guarantee certain desired properties. More yet, one crucial limitation in these attempts to connect agent programming languages and BDI logics, in our opinion, is that the connection is mainly established at the static level, i.e. they show how a given program state can be interpreted as a BDI mental state. Since mental change is not expressible in many of these logics, it is not clear how the execution of a program may be understood as changes in the mental state of the agent. The reason for this, in our opinion, is that the formalisms employed to construct BDI logics are usually static, i.e. cannot represent actions and change, or can only represent ontic change, not mental change. In this work, we will explore the use of Dynamic Preference Logic [3] to encode mental attitudes. This logic was shown to have a strong connection with syntactic representations, known as priority graphs [4], which can be used as a computational representation of a possible worlds model. We exploit this connection to show how the programming language semantics can be specified by means of mental attitude changes in the corresponding logics and vice-versa. The structure of this work is as follows: Section II presents Dynamic POreference Logic that will be the foundation of our logic for agents; in Section III, we show how Dynamic preference Logic can be used to create a logic for reasoning about an agent's mental state and show how the BDI mental at- titudes can be encoded in this logic; in Section IV, we present some syntactic representations for the models discussed in Sections II and III, and how these representations can be connected to agent programs, as commonly defined in various agent programming languages in the literature. In Section V, we discuss the related literature and, finally, in Section VI, we present some final considerations about our work. II. THE DYNAMIC LOGIC OF PREFERENCES Preference Logic is a modal the class of transitive and reflexive frames. It has been applied to model a plethora of phenomena in Deontic Logic [5], Logics of Preference [6][7], Logics of Belief [8] etc. logic about Dynamic Preference Logic (DPL) [3][9] is the result of "dy- namifying" Preference Logic, i.e. extending it with dynamic modalities - usually represented by programs in Propositional Dynamic Logic (PDL). In this section, we introduce Dynamic Preference Logic and some fundamental properties of this logic. This language will be the base for the construction of a logic for BDI reasoning in Section III. Let's first introduce the language of (static) Preference Logic. Definition 2.1: Let P be a finite set of propositional letters. We define the language L≤(P ) by the following grammar (where p ∈ P ): ϕ ::= p ¬ϕ ϕ ∧ ϕ Aϕ [≤]ϕ [<]ϕ We will often refer to the language L≤(P ) simply as L≤, by supposing the set P is fixed. Also, we will denote the language of propositional formulas by L0(P ) or simply L0 Definition 2.2: A preference model is a tuple M = hW, ≤, vi where W is a set of possible worlds, ≤ is a reflexive, transitive relation over W , and v : P → 2W a valuation function. In such a model, the accessibility relation ≤ represents an ordering of the possible worlds according to the preferences of a certain agent. As such, given two possible worlds w, w′ ∈ W , we say that w is at least as preferred as w′ if, and only if, w ≤ w′. The interpretation of the formulas over these models is defined as usual. M, w (cid:15) Aϕ M, w (cid:15) [≤]ϕ M, w (cid:15) [<]ϕ iff iff iff ∀w′ ∈ W : M, w′ (cid:15) ϕ ∀w′ ∈ W : w′ ≤ w ⇒ M, w′ (cid:15) ϕ ∀w′ ∈ W : w′ < w ⇒ M, w′ (cid:15) ϕ In the definition above, w < w′ if, and only if, w ≤ w′ and w′ 6≤ w. As usual, we will refer as h<iϕ to the formula ¬[<]¬ϕ. Also, given a model M and a formula ϕ, we use the notation JϕKM to denote the set of all the worlds in M satisfying ϕ. When it is clear to which model we are referring to, we will denote the same set by JϕK. Given a set of worlds JϕK and a (pre-)order ≤, we will denote the minimal elements of JϕK, according to ≤, by the notation M in≤JϕK. This corresponds to the notion of 'most preferred worlds satisfying ϕ' in the model.This notion can be defined in this logic by the following formula: µϕ ≡ ϕ ∧ ¬h<iϕ A. Dynamics of preferences In this section, we "dynamify" Preference Logic by in- troducing dynamic modalities representing standard mental change operations such as revisions and contractions. In this work, we present the operations of public announcement [12], radical upgrade [13] and natural contraction [14]. The choice for these three operations was motivated by the fact that they are each a representative of a large class of important mental changing operations studied in the literature, namely expansion, revision and contraction, and for the fact that these operations have been well studied in the framework of Dynamic Preference Logic before [15][4][11]. The first operation we present is the well-known public announcement of Plaza [12]. A public announcement of ϕ is a truthful and knowledge increasing announcement of ϕ. Definition 2.3: [3] Let M = hW, ≤, vi be a preference model and ϕ a formula of L0. We say the model M!ϕ = hW!ϕ, ≤!ϕ, v!ϕi is the result of public announcement of ϕ in M , where: = {w ∈ W M, w (cid:15) ϕ} = ≤ ∩ (W 2 W!ϕ ≤!ϕ v!ϕ(p) = v(p) ∩ W!ϕ The radical upgrade of a model by an information ϕ results in a model such that all worlds satisfying ϕ are deemed preferable than those not satisfying it. !ϕ) Definition 2.4: Let M = hW, ≤, vi be a preference model and ϕ a formula of L0. We say the model M⇑ϕ = hW, ≤⇑ϕ, vi is the result of the radical upgrade of M by ϕ, where ≤⇑ϕ= (≤ \{hw, w′i ∈ W 2 M, w 6(cid:15) ϕ and M, w′ (cid:15) ϕ})∪ {hw, w′i ∈ W 2 M, w (cid:15) ϕ and M, w′ 6(cid:15) ϕ} Natural contraction is a conservative contraction operation, in the sense that it aims to achieve some form of "minimal change" in the belief state. In other words, the preference relation is changed only in regards to the minimal worlds not satisfying the property to be contracted. We define this oper- ation by means of the following transformation on preference models. Definition 2.5: Let M = hW, ≤, vi be a preference model and ϕ a formula of L0. We say the model M↓ϕ = hW, ≤↓ϕ, vi is the natural contraction of M by ϕ, where: w ≤↓ϕ w′ iff (w ∈ M in≤W or w ∈ M in≤J¬ϕKM or w ≤ w′ and w′ 6∈ M in≤J¬ϕKM The existence of minimal worlds satisfying ϕ is not always guaranteed, since infinite descending chains may exist in the model. If the relation < in a model M is well-founded, how- ever, we can always guarantee that JµϕKM = M in≤JϕKM 6= ∅. A complete axiomatization for the logic restricted to well- founded models has been provided by Souza [10]. For each operation ⋆ defined above, we introduce in our language a new modality [⋆ϕ]ψ in our language, meaning "after the operation of ⋆ by ϕ, ψ holds". which can be interpreted as M, w (cid:15) [⋆ϕ]ψ iff M⋆ϕ, w (cid:15) ψ Moreover, Souza et al. [11] showed that if preference models are well-founded some important operations over preference relations, such as some forms of iterated belief revision and contraction, are well-defined in Preference Logic, expanding the results of Liu [4]. An important result about Dynamic Preference Logic with it has the same expressibility as these operations is that Preference Logic studied before. In fact, the formulas [!ϕ]ψ, [⇑ ϕ]ψ and [↓ ϕ]ψ are definable in the language of Preference Logic. III. A DYNAMIC LOGIC OF FOR BDI AGENT PROGRAMMING ϕ, it is most plausible that, ψ holds.' We propose the following codification of conditional belief: In this section, we use Dynamic Preference Logic to model the mental attitudes of the BDI framework. Preference Logic has been used to encode several different mental attitudes in the literature before, among them knowledge, beliefs [8] and goals or desires [6][7]. In this section, we propose a logic encoding both notions. For that, we introduce two (box) modalities in the language one for encoding the notion of plau- sibility, written [≤P ], and one for preference or desirability, written [≤D]. As such, we construct the language of this logic below. Definition 3.1: We define the language L≤P ,≤D (P ) by the following grammar (where p ∈ P a set of propositional letters): ϕ ::= p ¬ϕ ϕ ∧ ϕ Aϕ [≤P ]ϕ [<P ]ϕ [≤D]ϕ [<D]ϕ As before, we will define Eϕ ≡ ¬A¬ϕ and h≤(cid:3)iϕ ≡ ¬[≤(cid:3)]¬ϕ with (cid:3) ∈ {P, D}. The formula [≤D]ϕ ([≤P ]ϕ) means that in all words equally or more desirable (plausible) than the current one, ϕ holds and [<D]ϕ ([<P ]ϕ) that in all words strictly more desirable (plausible) than the current one, ϕ holds. To interpret these formulas, we will introduce a new kind of Kripke model containing two accessibility relations - one for plausibility and one for desirability. We will call this new model an agent model. Definition 3.2: An agent model is a tuple M = hW, ≤P , ≤D, vi where W is a set of possible worlds, and both ≤D and ≤P are pre-orders over W with well-founded strict parts <P and <D and v is a valuation function. Notice that an agent model is an amalgamation of two differ- ent preference models encoding the orderings for plausibility and desirability. The interpretation of the formulas is defined as usual, with each modality corresponding to an accessibility relation. We will call µP ϕ (µDϕ) the formula with the same structure as µϕ when using the modality <P (resp. <D), i.e. µP ϕ ≡ ϕ ∧ ¬h<P iϕ Similar to what was done in Preference logic, we can dynamify our agent logic by including dynamic modalities such as [⇑P ϕ]ψ to mean that "after the radical upgrade of the plausibility relation by ϕ, ψ holds". Once we introduced the language we will use in our work, let's encode the notions of mental attitudes. In Philosophical Logic, particularly Deontic Logic, it has been argued that men- tal attitudes are conditional in nature [16]. These conditional attitudes have been traditionally expressed by means of dyadic modalities of the form C(ψϕ) to represent 'in the context of ϕ, Cψ.' Conditionals are common in planning and practical reasoning, being used, for example, to express dependency relations among the agent's desires. In this work, we will encode mental attitudes by conditional modalities. Let's start with encoding beliefs. We want to define a conditional modality B(ψϕ) meaning that 'in the context of B(ψϕ) ≡ A(µP ϕ → ψ) Clearly, the semantics of B(ψϕ) implies that the most i.e. M in≤P JϕK ⊆ JψK. plausible ϕ-worlds are ψ-worlds, Finally, we define the unconditional belief B(ψ), meaning 'it is most plausible that ψ holds', as B(ψ) ≡ B(ψ⊤). Encodings of the notion of desire are numerous in the literature with various meanings according to the intended application. For the sake of our modeling, we will require that agent's desires are consistent with each other - a common requirement in logical modelling of desires. Hindriks et al. [2] argues that, since desires are future-directed in nature, such restriction is not necessary, for an agent needs not to desire to achieve ϕ and ¬ϕ at the same time. We agree with their criticism and point out that the kinds of desires they aim to represent can be expressed in our language as well, but for the aim of encoding intentions this simple kind of desires will suffice. Similar to belief, we propose a codification of desires as everything that is satisfied in all most desirable worlds. In other way, we want to encode a formula G(ψϕ) meaning that "in the most desirable ϕ-worlds, ψ holds". As such, we can encode goals as: G(ψϕ) ≡ A(µDϕ → ψ) Our encoding of desires is similar to [6]'s ideals in Quali- tative Decision Theory. It is our belief that Boutilier's ideals model quite faithfully the notion of overwhelming desire, i.e. a desire that is always preferred to its alternatives. As such, the formula G(ψ) ≡ G(ψ⊤) models the fact that the agent 'necessarily wants that ψ', i.e. in the most desirable worlds ψ holds. There is no consensus on which properties a theory of intentions should satisfy to properly describe the notions of intentional action, intentionality, etc. In the Artificial Intel- ligence research, Cohen and Levesque's [17] desiderata for intentions based in Bratman's[18] work has become the official benchmark for any theory aiming to formalise such notions. in our understanding, are two distinctive roles of intention in practical reasoning: the role of intention as a constraint in the possible actions/desires entertained by the agent and intention as a product of practicality, i.e. intentions as intrinsically connected to plans. to Cohen and Levesque's requirements, Central Since our logic does not possess the expressibility to refer to ontic actions, i.e. actions that change the current state of the world, we propose an initial codification of 'admissible intention', AdmInt(ψϕ), i.e. a property that satisfies the consistency requirements of Bratman, and may be eventually adopted as a prospective intention. This notion later will be refined, when we enrich the language to include ontic actions. Bratman's [18], simplified by Cohen and Levesque's desiderata, expresses the relationship between the attitudes of intention, desire and belief. Particularly, according to this requirements, an intention is a desire that the agent believes to be possible and that has not yet been achieved. We can model this relation in the following way, where AdmInt(ψϕ) means that 'in the context of ϕ it is admissible to intend to achieve ψ ': AdmInt(ψϕ) ≡ G(ψϕ) ∧ E(ψ ∧ ϕ) ∧ ¬B(ψϕ) In the following we will extend our logic to include ontic actions. With that extention, we can propoerly express the relationship between intentions and practical reason, i.e. how intention and actions are connected. A. Intentions and practicality The relationship between intention and practicality is quite a different aspect than what we have been treating before. In our framework we do not have the machinery to represent ontic actions - i.e. actions that change the environment. To allow the representation of practicality, we must extend the language of L≤P ,≤D to incorporate ontic actions, or simply plans. Definition 3.3: We call P = hΠ, pre, posi an action library, or plan library, iff Π is a set finite set of plans symbols, pre : Π → L0 is a function that maps each plan to a propositional formula representing its preconditions and pos : Π → L0 the function that maps each plan to a propositional formula representing its post-conditions. We further require that the post-conditions of any plan is a consistent conjunction of propositional literals. We say α ∈ P for any plan symbol α ∈ Π. To model the effect of performing an ontic action α ∈ P given an agent model M , we will define the notion of model update, as commonly used in the area of Dynamic Epistemic Logic. Definition 3.4: Let P = hΠ, pre, posi be a plan library, α ∈ P an action (or plan) and M = hW, ≤P , ≤D, vi an agent model. The product update of model M by action α is defined as the model M ⊗ [P, α] = hW ′, ≤′ D, v′i where P , ≤′ W ′ ≤′ P ≤′ D = {w ∈ W M, w (cid:15) pre(α)} = ≤P ∩ W ′ × W ′ = ≤D ∩ W ′ × W ′ W ′ ∅ v(p) ∩ W ′ × W ′ if pos(α) (cid:15) p if pos(α) (cid:15) ¬p otherwise v′(p) =   Bratman [18] defends the thesis that intentions are intrin- sically connected to plans, in the sense that intentions are the plans that the agent adopts to achieve a certain desired state of the world. These (procedural) intentions, however, are constrained by a series of consistency requirements, most of which we encoded by means of the formula AdmInt(ψϕ). We define, thus, when a set of plans are considered admissible as the (procedural) intentions of an agent in a given state of mind. plans is P-consistent with M if for all α ∈ I, M (cid:15) B(pre(α)) and M (cid:15) AdmInt(pos(α)). If I is P-consistent with M , we say M ′ = hW, ≤P , ≤D, I, vi is a practical agent model. With that, we can expand our language to include actions and procedural intentions, i.e. formulas of the sort [α]ϕ and Iα, meaning that 'after the execution of α, ϕ holds' and 'it is intended to α,' respectively. Definition 3.6: Let P be a plan library and M = hW, ≤P , ≤D, I, vi be a practical agent model. For any α ∈ P, we introduce the formulas [α]ϕ and Iα and define M, w (cid:15) [α]ϕ iff M, w (cid:15) pre(α) ⇒ M ⊗ [P, α], w (cid:15) ϕ M, w (cid:15) Iα iff α ∈ I With the addition of actions, we can represent the notion of ability. An agent can achieve ϕ if there is an executable action α, i.e. pre(α) holds, that causes ϕ to hold. More yet, Bratman requires that Intentions are intrinsically connected to plans, meaning that if an agent intends to achieve a state of affairs, she must have a plan to achieve it. With these requirements we propose the following codification for intentions: Int(ψϕ) ≡ AdmInt(ψϕ)∧ _α∈P (Iα∧B (pre(α) ∧ [α]ψ ϕ)) Meaning that "in the context of ϕ the agent intends to achieve ψ". As before we define Int(ϕ) ≡ Int(ϕ⊤). It is easy to see by our construction that procedural in- tentions, i.e. intentions to do, and prospective intentions, i.e. intentions to be, are well-connected. Proposition 3.7: Let P be a plan library and M = hW, ≤P , ≤D, I, vi be a practical agent model, it holds that M, w (cid:15) Iα ⇒ M, w (cid:15) B(pre(α)) ∧ Int(pos(α)) IV. AGENT LOGIC AND AGENT PROGRAMMING Now, we will focus our attention to the connection between our logic and agent programs. To understand this connection, we will explore some representation results relating preference models and a syntactic structure to encode preference relations, known as priority graphs. Definition 4.1: [4] Let L0(P ) be the propositional language constructed over the set of propositional letters P , as usual. A P-graph is a tuple G = hΦ, ≺i where Φ ⊂ L0(P ), is a set of propositional sentences and ≺ is a strict partial order on Φ. A priority graph is a partial order over a set of propositional formulas. The connection between these preferences over formulas and preferences over possible worldshas been studied in the literature [7]. In our work, following [4], we use the lexicographic ordering to provide this connection. Definition 4.2: [4] Let G = hΦ, ≺i be a P-graph, W be a non-empty set of states or possible worlds, and v : P → 2W be a valuation function. The order relation ≤G ⊆ W 2 is defined as follows: Definition 3.5: Let P be a plan library and M = hW, ≤P , ≤D, vi be an agent model. We say a set I ⊂ Π of w ≤G w′iff∀ϕ ∈ Φ : (w′ (cid:15) ϕ ⇒ w (cid:15) ϕ)∨ (∃ψ ≺ ϕ : (w (cid:15) ψ and w′ 6(cid:15) ψ)) From Definition 4.2, we can say a model M = hW, ≤G , vi is induced by a given priority graph G when its preference relation is constructed as above. Definition 4.3: Let G = hΦ, ≺i a P-graph and M = hW, ≤, vi a preference model. We say M is induced by G iff ≤ = ≤G, where ≤G is the relation defined in Definition 4.2 over the set W considering the valuation v. Liu [4] shows that any model with a reflexive, transitive relation is induced by some priority graph. Theorem 4.4: [4] Let M = hW, R, vi a modal model. The following two statements are equivalent: 1) M = hW, Ri is a preference frame; 2) There is a priority graph G = (Φ, ≺) and a valuation v on M s.t. ∀w, w′ ∈ W : wRw′ iff w ≤G w′. More yet, if W is finite, then so is Φ. Since in this work agent models are nothing more that the union of two preference models, we know that there must be a similar syntactic representation for agent models as well. We will define, thus, the notion of an agent structure, which will serve as this syntactic counterpart for agent models. Definition 4.5: Let L0(P ) be the propositional language constructed over the set of propositional letters P , as usual. An agent structure is a pair G = hGP , GDi, where both GP = hΦP , ≺P i and GD = hΦD, ≺Di are P-graphs. From agent structures we define the notion of induced agent model, similar to what was done to preference models in Definition 4.3. We just need to take the P-graphs that induce the plausibility and desirability relations (≤P and ≤D, respectively) which are guaranteed to exist by Theorem 4.4. Definition 4.6: Let G = hGP , GDi be an agent structure and M = hW, ≤P , ≤D, vi an agent model. We say M is induced by G iff ≤P = ≤GP and ≤D = ≤GD . From Definition 4.6, it is clear that every agent model is induced by some agent structure. Corollary 4.7: Let M = hW, ≤P , ≤D, vi be an agent model. There is an agent structure G = hGP , GDi s.t. M is induced by G. In most BDI agent programming languages, an agent pro- gram is defined by means of a tuple ag = hK, B, D, Ii, where K, B and G are sets of (ranked) propositional for- mulas representing the agent's knowledge, beliefs and desires, respectively, and I is a set of plans adopted by the agent. Since a set of (ranked) formulas is nothing more than an order over formulas, we can construct an agent structure G which induces an agent model MG representing the mental state of the agent program ag. Definition 4.8: Let ag = hK, B, D, Ii be a tuple, where K is a consistent set of propositional formulas, B and D are priority graphs and I is a set of actions in an action library P. We define the model induced by ag as M = hJKK, ≤B, ≤D, vi where JKK ⊂ 2P are all the propositional valuations that satisfy the set K, ≤B⊂ JKK × JKK and ≤D⊆ JKK × JKK are the preference relations induced by the graphs B and D, and w ∈ v(p) iff p ∈ w. If I is P-consistent with M , then Mag = hJKK, ≤B , ≤D, I, vi is the practical agent model induced by ag. In Section II, we introduced three dynamic operations in the logic of Preference Logic. In our agent logic these operations gain an interpretation of mental change, based on the results in the agent's mental attitudes. As such, public announcements can be understood as knowledge acquisition, while radical upgrade and natural contraction can be understood as either belief revision/contraction or preference revision/contraction. Theorem 4.9 ([10]): All the dynamic operations presented in Subsection II-A are definable by means of operations in P-graphs, if we consider induced models defined in Defini- tion 4.8. Since the semantics of agent programming languages (and deliberation mechanisms) can be specified by means of the transformation on the agent's mental state, if we can specify a desirable property one wishes the programming language semantics (or deliberation mechanism) to satisfy by means of these actions, we can automatically generate the corresponding semantic rule by means of transformation of agent programs, using the established correspondence between operations on preference models and operations in agent structures. For example, if one wishes to implement a belief revision such that every time an agent comes to believe ϕ she will drop her intentions to ¬ϕ. We can define such operation, let's call it M↑ϕ as a composition of the operations of preference contraction and belief revision (M↓D ¬ϕ)⇑P ϕ, which can be translated as an operation in priority graphs. V. RELATED WORK From the Agent Programming perspective, the two most important works on modelling BDI mental attitudes are, in our opinion, the seminal work of Cohen and Levesque [17] and the work of Rao and Georgeff [19] describing the logic BDI- CTL. While their contribution to the area is undeniable, much criticism has been drawn to both approaches. Particularly, both approaches have proven to be difficult to connect with agent programming languages, by the use of a possible-world model semantics - vastly different from the syntactical representations used in agent programming. Other work have also been proposed for studying the declarative interpretation of mental attitudes in concrete agent programming languages. Works as that of Wobcke [20] and of Hindriks and Van der Hoek [21] propose ways to connect the semantics of a given programming language to some appropriate logic to reason about agent's mental attitudes. While they are important in allowing us to analyse the mental attitudes diffused in the semantics of the language, since these logics cannot represent mental actions, the transformations in the agent program, which are defined in the programming language semantics, cannot be understood within the logic used to analyse these mental attitudes and thus the dynamic properties of these attitudes cannot be reasoned about in the logic. Also, in this approach, it is not clear how to establish the contrary connection, i.e. how to create or change programs to guarantee a certain property in the theory of intentions. In our work, since we can translate both ways, from the logic to agent programs and back, this is not an issue. On the other way, works as that of Bordini and Moreira [1] present a declarative interpretation of BDI attitudes based on the actual implementation of these concepts in a concrete agent programming language. The aim of their work is to analyse Rao and Georgeff's [19] asymmetry properties in the formal semantics of the language AgentSpeak(L). The result is that, due to implementation considerations of the programming lan- guage, the logic suffers from a great expressibility limitation, not being able to represent several important properties about mental states. Perhaps the work most related to ours in spirit is that of [22]. They propose a dynamic logic for agents and show that this logic can be understood as a verification logic, i.e. it has an equivalent state-based semantics based on the an operational semantics. The main difference of their approach to ours is that the authors choose to work in a framework closely related to situation calculus. The mental actions involved in decision making and in mental change are, thus, only implicitly defined, while the inclusion of such actions in the language is exactly the main advantage advocated by us. In some sense, our work can be seen as a generalisation of their work, since by employing Dynamic Preference Logic the equivalence they seek between operational semantics and declarative semantics can be automatically achieved by the results of Liu [4]. Recently, Herzig et al. [23] pointed out some deficiencies in the formal frameworks for specifying BDI agents which are available in the literature. The authors point out the advantages of a formal theory with a close relationship with the work in belief dynamics and with agent programming. VI. FINAL CONSIDERATIONS Our work has proposed a logic for reasoning about BDI agents and a connection between this logic and agent pro- grams, as usually described in agent programming languages. Our logic is flexible enough to specify different mechanisms for agent deliberation as well as different properties for beliefs, desires and intention from the codifications proposed in this work. As such, we believe this logic to be applicable to reason about programs for many agent programming languages. Regarding the requirements proposed by Herzig et al. [23] for a formal theory of agent programming, we believe our work tackles most of the problems identified by those authors. It remains, however, to provide a greater connection of our logics with the work areas as planning and game theory. We point out, however, that we have powerful evidences that such connections can be done. For example, the work of Andersen et al. [24] explore how to integrate planning in the dynamic logics as the one we propose. For the connection with decision theory and game theory, we point out that utilitarian interpre- tations of agent models have been provided by Boutilier [6]. Also, the work of Roy [25] provides codification of intentions in epistemic game theory using possible worlds models related to ours in which each possible world is a strategy. We believe we can provide a connection between our agent models and Roy's semantics without many difficulties. REFERENCES [1] R. Bordini and A. Moreira, "Proving BDI properties of agent-oriented programming languages: The asymmetry thesis principles in AgentS- peak (L)," Annals of Mathematics and Artificial Intelligence, vol. 42, no. 1, pp. 197 -- 226, 2004. [2] K. V. Hindriks, F. S. De Boer, W. Van der Hoek, and J.-J. C. Meyer, "Agent programming with declarative goals," in Intelligent Agents VII Agent Theories Architectures and Languages. New York, US: Springer, 2001, pp. 228 -- 243. [3] P. Girard, "Modal logic for belief and preference change," Ph.D. dissertation, Stanford University, 2008. [4] F. Liu, Reasoning about preference dynamics. New York, US: Springer, 2011, vol. 354. [5] J. Van Benthem, D. Grossi, and F. Liu, "Priority structures in deontic logic," Theoria, vol. 80, no. 2, pp. 116 -- 152, 2014. [6] C. Boutilier, "Toward a logic for qualitative decision theory," in Proceed- ings of the 4th International Conference on Principles of Knowledge Representation and Reasoning. New York, US: Morgan Kaufmann, 1994, pp. 75 -- 86. [7] J. Lang, L. Van der Torre, and E. Weydert, "Hidden uncertainty in the logical representation of desires," in Proceedings of the 18th international joint conference on Artificial intelligence. New York, US: Morgan Kaufmann, 2003, pp. 685 -- 690. [8] A. Baltag and S. Smets, "A qualitative theory of dynamic interactive belief revision," Texts in logic and games, vol. 3, pp. 9 -- 58, 2008. [9] J. Van Benthem, P. Girard, and O. Roy, "Everything else being equal: A modal logic for ceteris paribus preferences," Journal of philosophical logic, vol. 38, no. 1, pp. 83 -- 125, 2009. [10] M. Souza, "Choices that make you change your mind: a dynamic epistemic logic approach to the semantics of bdi agent programming languages," Ph.D. dissertation, Universidade Federal do Rio Grande do Sul, 2016. [11] M. Souza, A. Moreira, R. Vieira, and J.-J. C. Meyer, "Preference and priorities: A study based on contrction," in Proceedings of the Fifteenth International Conference on Principles of Knowledge Representation and Reasoning. AAAI Press, 2016, pp. 155 -- 164. [12] J. Plaza, "Logics of public communications," Synthese, vol. 158, no. 2, pp. 165 -- 179, 2007. [13] J. Van Benthem, "Dynamic logic for belief revision," Journal of Applied Non-Classical Logics, vol. 17, no. 2, pp. 129 -- 155, 2007. [14] R. Ramachandran, A. C. Nayak, and M. A. Orgun, "Three approaches to iterated belief contraction," Journal of philosophical logic, vol. 41, no. 1, pp. 115 -- 142, 2012. [15] J. Van Benthem and F. Liu, "Dynamic logic of preference upgrade," Journal of Applied Non-Classical Logics, vol. 17, no. 2, pp. 157 -- 182, 2007. [16] B. Hansson, "An analysis of some deontic logics," Nous, pp. 373 -- 398, 1969. [17] P. R. Cohen and H. J. Levesque, "Intention is choice with commitment," Artificial Intelligence, vol. 42, no. 2-3, pp. 213 -- 261, 1990. [18] M. E. Bratman, Intention, plans, and practical reason. Cambridge, US: Harvard University Press, 1999. [19] A. S. Rao and M. P. Georgeff, "Decision procedures for BDI logics," Journal of Logic and Computation, vol. 8, no. 3, pp. 293 -- 343, 1998. [20] W. Wobcke, "Model theory for PRS-like agents: Modelling belief update and action attempts," in Proceedings of the 8th Pacific Rim International Conference on Artificial Intelligence. Berlin, DE: Springer-Verlag, 2004, pp. 595 -- 604. [21] K. Hindriks and W. Van der Hoek, "Goal agents instantiate intention logic," in Logics in Artificial Intelligence. New York, US: Springer, 2008, pp. 232 -- 244. [22] K. V. Hindriks and J.-J. C. Meyer, "Toward a programming theory for rational agents," Autonomous Agents and Multi-Agent Systems, vol. 19, no. 1, pp. 4 -- 29, 2009. [23] A. Herzig, E. Lorini, L. Perrussel, and Z. Xiao, "BDI logics for BDI ar- chitectures: old problems, new perspectives," KI-Kunstliche Intelligenz, pp. 1 -- 11, 2016. [24] M. B. Andersen, T. Bolander, and M. H. Jensen, "Don't plan for the unexpected: Planning based on plausibility models," Logique et Analyse, vol. 1, no. 1, 2014. [25] O. Roy, "A dynamic-epistemic hybrid logic for intentions and informa- tion changes in strategic games," Synthese, vol. 171, no. 2, pp. 291 -- 320, 2009.
1709.00951	3	1709	2017-11-23T09:16:50	Consensus of second order multi-agents with actuator saturation and asynchronous time-delays	[ "cs.MA", "eess.SY", "eess.SY" ]	This article presents the consensus of a saturated second order multi-agent system with non-switching dynamics that can be represented by a directed graph. The system is affected by data processing (input delay) and communication time-delays that are assumed to be asynchronous. The agents have saturation nonlinearities, each of them is approximated into separate linear and nonlinear elements. Nonlinear elements are represented by describing functions. Describing functions and stability of linear elements are used to estimate the existence of limit cycles in the system with multiple control laws. Stability analysis of the linear element is performed using Lyapunov-Krasovskii functions and frequency domain analysis. A comparison of pros and cons of both the analyses with respect to time-delay ranges, applicability and computation complexity is presented. Simulation and corresponding hardware implementation results are demonstrated to support theoretical results.	cs.MA	cs	IET Research Journals Brief Paper: This paper is a postprint of a paper submitted to and accepted for publication in IET Control Theory & Applications and is subject to Institution of Engineering and Technology Copyright. The copy of record is available at the IET Digital Library Consensus of second order multi-agents with actuator saturation and asynchronous time-delays Venkata Karteek Yanumula1 Indrani Kar1 Somanath Majhi1 1 Dept. of EEE, IIT Guwahati-781039, India. * E-mail: [email protected] ISSN 1751-8644 doi: 0000000000 www.ietdl.org Abstract: This article presents the consensus of a saturated second order multi-agent system with non-switching dynamics that can be represented by a directed graph. The system is affected by data processing (input delay) and communication time-delays that are assumed to be asynchronous. The agents have saturation nonlinearities, each of them is approximated into separate linear and nonlinear elements. Nonlinear elements are represented by describing functions. Describing functions and stability of linear elements are used to estimate the existence of limit cycles in the system with multiple control laws. Stability analysis of the linear element is performed using Lyapunov-Krasovskii functions and frequency domain analysis. A comparison of pros and cons of both the analyses with respect to time-delay ranges, applicability and computation complexity is presented. Simulation and corresponding hardware implementation results are demonstrated to support theoretical results. 1 Introduction In the recent past, multi-agent systems have attracted a lot of atten- tion due to their wide range of application in robotics, unmanned air and underwater vehicles, automated traffic signal control, wireless sensor networks, etc.. One of the most important problems in coor- dinated control is consensus of a multi-agent system, which deals with algorithms required for the convergence of agents [1, 2]. After the initial study by Vicsek et al.[3] on self-ordered motions in bio- logically motivated particles, Jadbabaie et al. [1] gave the theoretical explanation. Olfati-Saber et al. [2] provided mathematical analysis of consensus behaviour in linear first order agents with time-delay using graph theory concepts. Multi-agent consensus problems with higher order agents, switching topologies, time-delays, nonlineari- ties etc., started receiving more attention [4, 5] after the initial results given by authors in [1, 2]. However, the majority of control laws are designed to solve con- sensus problems in linear multi-agent systems [1, 2, 6–12]. For linear systems, it has been shown that eigenvalues of graph laplacian play an important role in estimating whether the network of agents converge. Since nonlinearities are unavoidable in most of the prac- tical applications, nonlinear agents and the corresponding control laws are being considered recently [13–15]. Mobile agents gener- ally have limited capability due to factors like actuator saturation, moment of inertia, maximum limit on velocity, etc.. Actuator satura- tion is frequently encountered due to limitations in hardware. Some of the researchers focused on consensus in multi-agent systems with saturation in first order [16] and second order agents [17–22]. Apart from eigenvalues of graph laplacian, time-delays play ma- jor role in stability of multi-agent network. In practical applications, time-delays are inevitable and are classified into two categories; communication and input time-delays. The amount of time taken by agents to communicate is defined as communication time-delay and the amount of time taken by agents to process the information received from other agents is called input time-delay. Olfati-Saber et al.[2] started the analysis of time-delay effects on multi-agent systems and gave an upper bound for first order agents considering constant uniform communication and input time-delays. Later on it was extended to systems with first order agents and uniform time- varying delay[6] multiple delays [9, 23], second order agents with constant time-delays [8, 24] and system with second order agents with non-uniform delay [11]. The majority of research is confined to linear agents with time- delays and recently nonlinear agents with time-delays are receiving attention [14, 22]. Furthermore, nonlinearities are common in mobile agents and actuator saturation is the most frequent hard nonlinearity affecting them. For example, the acceleration of an agent is constant over certain range and cannot be maintained after the agent attains its maximum velocity. Recently, saturation nonlinearity is receiving considerable attention. Li et al. [16] considered a first order system with saturation and without time-delays. For second order agents with saturation and without time-delay, a differential gain feedback control is used by authors in [17, 18]. Adaptive control laws with an observer are used by Chu et al. [19] and nonlinear agents are consid- ered by Cui et al. [21]. The effects of synchronous time-delays are taken into consideration by You et al. [22] for a network of second order saturated agents. It is evident from the literature that, there is very little focus on consensus of second order saturated multi-agent system with asyn- chronous communication and input time-delays. In this contribution, a multi-agent system is considered with asynchronous time-delays and hard saturation nonlinearities. The objective of the article is to extend the results of Liu et al. [14] for velocity saturated nonlin- ear multi-agent system with time-delays using a different approach. Describing function analysis [25] is used to break agents into ap- proximate linear and nonlinear elements, with the nonlinear element represented by an appropriate describing function. The difference among position states and velocity states of agents is defined as error dynamics. The system achieves consensus when the error dynamics are asymptotically stable. Here, the existence of limit cycles in the multi-agent system is estimated with the help of describing functions and stability of linear element. Lyapunov-Krasovskii functions and frequency domain analysis are used to prove the stability of linear element and further estimate the stability of limit cycles. Consensus is achieved when there are no limit cycles. Some necessary and suf- ficient conditions for consensus in terms of linear matrix inequalities and explicit expressions are derived. The major contributions of the paper can be summarised as, 1. Deriving various conditions for four consensus control laws with asynchronous time-delays; 2. Describ- ing function analysis is used to estimate the limit cycle behaviour of the system; 3. Stability analysis of the linear element using Lya- punov-Krasovskii and frequency domain approaches is performed; 4. A comparison of pros and cons of both the stability analyses is presented; 5. Simulations and further validation of results on a IET Research Journals, pp. 1–10 c(cid:13) The Institution of Engineering and Technology 2015 1 7 1 0 2 v o N 3 2 ] A M . s c [ 3 v 1 5 9 0 0 . 9 0 7 1 : v i X r a four-agent and a five-agent networks are demonstrated to support theoretical analysis. The rest of the paper is organised as follows, Section 2 explains graph theory preliminaries. Section 3 elaborates the system model with four control laws given in Eqns. (5) to (8). Stability analysis is performed using Lyapunov-Krasovskii functions for control laws in Eqns. (5) to (8), using the Nyquist stability criterion for control laws in Eqns. (5) and (6). Furthermore, simulation and implementa- tion of the control laws on two networks are explained. Depiction of results and comparison of the two stability procedures are performed in Section 4. 2 Preliminaries 2.1 Graph theory Graph theory is widely used to study multi-agent systems. A net- work of agents and the underlying communication topology can be represented by a graph G = (V, E,A). If the communication among agents could be unidirectional, a directed graph is used to describe the multi-agent network. The vertex set V = {v1, v2, ...., vn} where vertices are analogous to agents and an edge set E = {(i, j) : i, j ∈ V} where edges are analogous to the branches of directed network with (i, j) representing information flowing from jth vertex to ith. Edge set has distinct ordered pairs of vertices which depict existence and direction of information flow among the vertices. An adjacency matrix A = (aij )n×n also represents communication topology with aij = 1 if (i, j) ∈ E and aij = 0 otherwise. A weighted adjacency matrix will have entries other than zero and unity weights depend- ing on the assumptions of cost of communication. If there exists at least one vertex which has a directed path to all the other vertices, the graph is said to form a spanning tree and if all the vertices have directed paths to all the other agents, it is called strongly connected. A spanning tree condition is a necessary condition for consensus but not sufficient when time-delays and higher order systems are in- volved [4, 5]. The sum of weights of inward branches at a vertex is called in-degree din(vi) and the weight sum of outward branches is called out-degree of the vertex dout(vi). 2.2 Notations The following notations are used throughout the paper, Rn repre- sents an n-dimensional Euclidean space. Rm×n represent a space of m × n matrices. Position and velocity of n agents are rep- resented by x = [x1 x2 ... xn]T and x = v respectively. X = [X1 X2 ... Xn Xn+1 ... X2n]T represent the states of a multi-agent system with [X1 X2 ... Xn]T = x and [Xn+1 ... X2n]T = x. In and I2n represent identity matrices of sizes n × n and 2n × 2n re- spectively. 1n is a vector ones of size 1 × n. For {A, B} ∈ Rn×n, if A (cid:60) B, then A − B is positive semidefinite; if A (cid:31) B, then A − B is positive definite. D represents a matrix with diagonal elements as row-sum of adjacency matrix A and rest of the elements as zero. A matrix A is defined with elements aij = and [λ1, λ2, ..., λn] are the eigenvalues of matrix A. aij(cid:80)n j=1 aij 3 System model and analysis Consider a multi-agent network of homogeneous second order agents with ith agent dynamics given in Eqn. (1), xi (t) = sat (vi (t)) vi (t) = ui (t) ui2 = (1) For mobile agents, the position of an agent is represented by xi (t) and the velocity by xi (t). Various control protocols used in the anal- ysis are given in Eqns. (5) to (8). It is assumed that ∀t ∈ (−∞, 0], xi (t) = x (0) and vi (t) = 0. Saturation nonlinearity used in the system is defined in Eqn. (2) with ±∆ as bounds. 2 −∆, α, ∆, sat (α) = if α ≤ −∆ if − ∆ < α < ∆ if α ≥ ∆ (2) The ith agent dynamics are depicted using a block diagram given (cid:90) (cid:90) ui vi −∆ ∆ vi xi Communication & process (aij )(xj , vj ) (aji)(xi, vi) Fig. 1: Block diagram of ith agent. in Fig. 1. Using the concepts of describing function to estimate limit cycles [25], the system can be approximately transformed as shown in Fig. 2. Since a single-valued nonlinearity is considered, its ap- proximate describing function for the saturation is given in Eqn. (3) [25], (cid:114) (cid:35) N (A) = 2 π arcsin + ∆ A 1 − ∆2 A2 (3) (cid:18) ∆ (cid:19) A (cid:34) where, the limit cycles' amplitude is represented by A. The describing function is real valued and −1/N (A) ∈ [−1,∞), it can be estimated that the limit cycles are stable when the transfer function of linear element in Fig. 2 encircles (−1, 0) in a complex plane. In other words, limit cycles are exhibited when the linear ele- ment is unstable in the multi-agent system. Stability analysis of the linear element is performed using Lyapunov-Krasovskii approach in Section 3.1 and Nyquist stability approach given in Section 3.2. Nonlinear element Communication & process ui Linear element (xi, vi) Fig. 2: Rearranged block diagram of ith agent. The approximate linear element given in Fig. 2 is represented by Eqn. (4), xi (t) = vi (t) vi (t) = ui (t) (4) Various control laws considered from the literature for analysis are given in Eqns. (5) to (8), ui1 (t) = −vi (t − τ1) j=1 aij + 1(cid:80)n 1(cid:80)n j=1 aij j=1 j=1 n(cid:88) (cid:0)xj (t − τ2) − xi (t − τ1)(cid:1)(cid:3) (cid:2)aij n(cid:88) (cid:0)vj (t − τ2) − vi (t − τ1)(cid:1) (cid:2)aij (cid:0)xj (t − τ2) − xi (t − τ1)(cid:1)(cid:3) (cid:0)xj (t − τ2) − xi (t − τ1)(cid:1)(cid:3) (cid:2)aij n(cid:88) + + aij ui3 (t) = − vi (t − τ1) (5) (6) (7) j=1 IET Research Journals, pp. 1–10 c(cid:13) The Institution of Engineering and Technology 2015 (cid:2)aij n(cid:88) j=1 ui4 = (cid:0)vj (t − τ2) − vi (t − τ1)(cid:1) (cid:0)xj (t − τ2) − xi (t − τ1)(cid:1)(cid:3) + aij (cid:20)1n×n 0n×n Lemma 1. Consider Φ01 = 1 n Φ01, then the following statements hold true: 0n×n 1n×n (cid:21) and E = I2n − (8) where τ1 and τ2 represent input and communication time-delays re- spectively. With any of the control laws in Eqns. (5) to (8), consensus is said to be reached if (xi(t) − xj (t)) → 0 and ( xi(t) − xj (t)) → 0 ∀{i, j} ∈ [1, n]. Control laws in Eqns. (5) and (6) generate lesser magnitude of control input ui which result in slightly larger conver- gence time compared to the ones in Eqns. (7) and (8). The averaging in control laws given by Eqns. (5) and (6) have better time-delay tol- erance due to smaller Fiedler eigenvalue compared to control laws in Eqns. (7) and (8) at the expense of convergence time. With control laws in Eqns. (5) and (7), the state xi(t) → 0 when the consensus is achieved since they do not consider difference in velocity. State xi(t) → 0 is not guaranteed with control laws in Eqns. (6) and (8). 3.1 Lyapunov-Krasovskii approach Consider the linear element represented in Eqns. (4) to (8), which can be represented as given in Eqn. (9). X (t) = A0X (t) + A1X (t − τ1) + A2X (t − τ2) (9) Where A0, A1 and A2 are as given in Eqns. (10) to (13). For ui1 given in Eqn. (5), (cid:20)0n×n 0n×n −In −In (cid:21) ; A1 = ; (10) For ui2 given in Eqn. (6), 0n×n In 0n×n 0n×n In 0n×n 0n×n (cid:20)0n×n (cid:20)0n×n 0n×n (cid:101)A (cid:20)0n×n (cid:20)0n×n 0n×n (cid:101)A (cid:101)A (cid:20)0n×n (cid:20)0n×n 0n×n (cid:20)0n×n (cid:20)0n×n 0n×n In 0n×n 0n×n In 0n×n 0n×n 0n×n A (cid:21) (cid:21) (cid:21) (cid:21) (cid:21) (cid:21) (cid:21) (cid:21) A A A0 = A2 = A0 = A2 = A0 = A2 = A0 = A2 = For ui3 given in Eqn. (7), For ui4 given in Eqn. (8), (cid:20)0n×n 0n×n −In −In (cid:21) ; ; A1 = (cid:20)0n×n 0n×n −D −In (cid:21) ; ; A1 = (cid:20)0n×n 0n×n −D −D (cid:21) ; ; A1 = (11) (12) (13) Some definitions and lemmas analogous to the ones in [26] are given below, Definition 1. Balanced graph: A graph is said to be balanced if in- degree equals to out-degree for all vertices in the graph, din(vi) = dout(vi), ∀i ∈ [1, n]. Definition 2. k-regular graph: It is a balanced graph with all the vertices having in-degree and out-degree equal to k, din(vi) = dout(vi) = k, ∀i ∈ [1, n] Control laws given in Eqns. (5) and (6) make the multi-agent system behave like a system connected by 1-regular graph. IET Research Journals, pp. 1–10 c(cid:13) The Institution of Engineering and Technology 2015 (cid:20)1n×n 0n×n (cid:21) 1.A multi-agent system with k-regular graph communication topol- ogy and with inputs in Eqns. (5) to (8) produces balanced matrices E (A0 + A1 + A2), EA0, EA1 , EA2 and E (A1 + A2) with maximum rank 2n − 2 and eigenvalues 0 of multiplicity atleast two. 2.A multi-agent system with a spanning tree in communication topol- ogy and with inputs in Eqns. (5) and (6) produces balanced matrices E (A0 + A1 + A2), EA0, EA1 , EA2 and E (A1 + A2) with maximum rank 2n − 2 and eigenvalues 0 of multiplicity atleast two. Definition 3. Balanced matrix: A square matrix M ∈ Rn×n is said to be balanced iff M1T n = 0 and 1nM = 0. 0n×n 1n×n and E = I2n − Lemma 2. Consider Φ01 = 1 n Φ01, then the following statements hold true for k-regular graph with inputs in Eqns. (5) to (8) and for spanning tree graph with inputs in Eqns. (5) and (6): 1.E (A0 + A1 + A2) is a balanced matrix with rank 2n − 2 and eigenvalues 0 of multiplicity 2. 2.Matrices EA0, EA1 , EA2 and E (A1 + A2) are all balanced with eigenvalues 0 of multiplicity atleast 2. 3.There is a matrix U, an orthogonal matrix of eigenvectors of E which satisfies, U T EU = 4.Let EA0 = F0, EA1 = F1 and EA2 = F2. E, F0, F1 and F2 have maximum rank 2n − 2 and with zero row sums, then, U T FiU = , i ∈ [0, 2]. 5.Also, for cases of τ1 = τ2 > 0 and τ1 = τ2 = 0, U T E (A1 + A2) U = U T E (A0 + A1 + A2) U = (cid:20)((cid:101)F1 + (cid:101)F2)(2n−2)×(2n−2) (cid:20)((cid:101)F0 + (cid:101)F1 + (cid:101)F2) (cid:20)(cid:101)E(2n−2)×2 (cid:20)(cid:101)Fi(2n−2)×(2n−2) 02×2 0(2n−2)×2 02×(2n−2) 0(2n−2)×2 0(2n−2)×2 02×(2n−2) 02×(2n−2) 0(2n−2)×2 02×2 02×2 (cid:21) (cid:21) (cid:21) (cid:21) 02×(2n−2) 02×2 Let the difference in position and velocity among the agents be assumed as error Ψ, each element of Ψ is given by Eqn. (14)  1 n 1 n (cid:1) (cid:0)Xi − Xj n(cid:80) (cid:0)Xi − Xj 2n(cid:80) j=1 j=n+1 Ψi = ∀i ∈ [1, n] (cid:1) ∀i ∈ [n + 1, 2n] From the assumption in Lemma 1, Ψ = EX Lemma 3. When error Ψ → 0, then xi → xj and vi → vj. Con- versely when xi → xj and vi → vj, then Ψ → 0. Proof. Consider a matrices,  γn×n = 0 1 −1 0 ... ... 0 −1 ... 0 1 −1 ... 0 ... ... ... ... ... ... 1 −1 ... 1 0 0 ... ... 0 ... (cid:20)γn×n 0n×n (cid:21) 0n×n γn×n Γ2n×2n =  (16) (17) 3 (14) (15) Multiplying with Γ on both sides of Eqn. (15), Ψ1 − Ψ2 Ψ2 − Ψ3 X1 − X2 X2 − X3   . . . . Ψn − Ψ1 Ψn+1 − Ψn+2 Ψn+2 − Ψn+3 = Xn − X1 Xn+1 − Xn+2 Xn+2 − Xn+3 Ψ2n − Ψn+1 X2n − Xn+1 . . . .   (18) When Ψ → 0, left side of Eqn. (18) becomes 02n×1. Which implies, Xi → Xj , ∀{i, j} ∈ [1, n] and Xi → Xj , ∀{i, j} ∈ [n + 1, 2n]. From Eqn. (14), when xi → xj and vi → vj, then Ψ → 0. (cid:3) A control input is said to have solved the consensus problem in a globally asymptotic manner when xi → xj and vi → vj, in other words, Ψ → 0. Stability of linear element with the control inputs estimates the existence of limit cycles in the system. 12 34 12 13 13 14 23 24 23 12 13 (20) (19) of size (2n − 2) × (2n − 2) such that, Theorem 1. Consider the linear element in Eqn. (4) with time- delays (τ1, τ2) ≥ 0, τ1 ≤ τ2. The control inputs for a k-regular graph given in Eqns. (5) to (8) and the control inputs for a span- ning tree graph given in Eqns. (5) and (6) globally asymptotically solve consensus problem, if there exist matrices (cid:101)P > 0, (cid:101)Q1 > 0, (cid:101)Q2 > 0, (cid:101)Zi > 0, ∀i ∈ [1, 3], (cid:101)Fi from Lemma 2 ∀i ∈ [1, 3] and arbi- ij , (cid:101)Jij , (cid:101)J T ij ,(cid:101)Iij ,(cid:101)I T trary matrices {(cid:101)Hij , (cid:101)H T ij} ∀i ∈ [1, 3] ∀j ∈ [1, 4] (cid:101)G11  ≺ 0 (cid:101)G12 (cid:101)G13 (cid:101)G23 (cid:101)G22 (cid:101)GT (cid:101)GT (cid:101)GT (cid:101)G33   (cid:60) 0 (cid:101)H11 (cid:101)H12 (cid:101)H13 (cid:101)H14 (cid:101)H24 (cid:101)H23 (cid:101)H22 (cid:101)H T (cid:101)H T (cid:101)H T (cid:101)H33 (cid:101)H34 (cid:101)H T (cid:101)H T (cid:101)H T (cid:101)Z1   (cid:60) 0 (cid:101)I11 (cid:101)I12 (cid:101)I13 (cid:101)I14 (cid:101)I24 (cid:101)I23 (cid:101)I22 (cid:101)I T (cid:101)I T (cid:101)I T (cid:101)I33 (cid:101)I34 (cid:101)I T (cid:101)I T (cid:101)I T (cid:101)Z2   (cid:60) 0 (cid:101)J14 (cid:101)J13 (cid:101)J12 (cid:101)J11 (cid:101)J T (cid:101)J22 (cid:101)J23 (cid:101)J24 (cid:101)J T (cid:101)J T (cid:101)J33 (cid:101)J34 (cid:101)J T (cid:101)J T (cid:101)J T (cid:101)Z3 0 (cid:101)P(cid:101)E +(cid:101)E T (cid:101)P (cid:101)F0 +(cid:101)E T (cid:101)Q1(cid:101)E +(cid:101)E T (cid:101)Q2(cid:101)E + (cid:101)F T + τ1(cid:101)E T (cid:101)H11(cid:101)E +(cid:101)E T (cid:101)H14(cid:101)E +(cid:101)E T (cid:101)H T (cid:101)E T(cid:101)I14(cid:101)E +(cid:101)E T(cid:101)I T (cid:101)G12 =(cid:101)E T P (cid:101)F1 + (cid:101)F T (cid:101)E T (cid:101)H T 24(cid:101)E + τ2(cid:101)E T(cid:101)I12(cid:101)E +(cid:101)E T(cid:101)I T 24(cid:101)E+ (τ2 − τ1)(cid:101)E T (cid:101)J12(cid:101)E +(cid:101)E T (cid:101)J14(cid:101)E (cid:101)E T(cid:101)I T 34(cid:101)E + τ2(cid:101)E T(cid:101)I13(cid:101)E −(cid:101)E T(cid:101)I14(cid:101)E+ (τ2 − τ1)(cid:101)E T (cid:101)J13(cid:101)E −(cid:101)E T (cid:101)J14(cid:101)E 14(cid:101)E + (τ2 − τ1)(cid:101)E T (cid:101)J11(cid:101)E 0 (cid:101)Ξ(cid:101)F1 + τ1(cid:101)E T (cid:101)H12(cid:101)E −(cid:101)E T (cid:101)H14(cid:101)E+ 0 (cid:101)Ξ(cid:101)F2 + τ1(cid:101)E T (cid:101)H13(cid:101)E +(cid:101)E T (cid:101)H T 34(cid:101)E+ 0 (cid:101)Ξ(cid:101)F0 14(cid:101)E + τ2(cid:101)E T(cid:101)I11(cid:101)E+ (cid:101)G13 =(cid:101)E T (cid:101)P (cid:101)F2 + (cid:101)F T where,(cid:101)G11 =(cid:101)F T 1 (cid:101)Ξ(cid:101)F1 + τ1(cid:101)E T (cid:101)H22(cid:101)E −(cid:101)E T (cid:101)H24(cid:101)E− (cid:101)G22 = −(cid:101)E T (cid:101)Q1(cid:101)E + (cid:101)F T (23) (24) (25) (21) (22) 12 13 23 24 23 24 34 34 14 14 34(cid:101)E + τ2(cid:101)E T(cid:101)I23(cid:101)E− 24(cid:101)E + τ2(cid:101)E T(cid:101)I22(cid:101)E + (τ2 − τ1)(cid:101)E T (cid:101)J22(cid:101)E (cid:101)E T (cid:101)H T +(cid:101)E T (cid:101)J24(cid:101)E +(cid:101)E T (cid:101)J T 24(cid:101)E (cid:101)G23 =(cid:101)F T 1 (cid:101)Ξ(cid:101)F2 + τ1(cid:101)E T (cid:101)H23(cid:101)E −(cid:101)E T (cid:101)H T (cid:101)E T(cid:101)I24(cid:101)E + (τ2 − τ1)(cid:101)E T (cid:101)J23(cid:101)E− (cid:101)E T (cid:101)J24(cid:101)E +(cid:101)E T (cid:101)J T 34(cid:101)E 2 (cid:101)Ξ(cid:101)F2 + τ1(cid:101)E T (cid:101)H33(cid:101)E −(cid:101)E T(cid:101)I34(cid:101)E (cid:101)G33 = −(cid:101)E T (cid:101)Q2(cid:101)E + (cid:101)F T −(cid:101)E T(cid:101)I T 34(cid:101)E + τ2(cid:101)E T(cid:101)I33(cid:101)E + (τ2 − τ1)(cid:101)E T (cid:101)J33(cid:101)E −(cid:101)E T (cid:101)J34(cid:101)E −(cid:101)E T (cid:101)J T 34(cid:101)E (cid:101)Ξ =τ1(cid:101)Z1 + τ2(cid:101)Z2 + (τ2 − τ1)(cid:101)Z3 (26) (27) (28) (29) Proof. Let P , Q1, Q2, Zi i ∈ [1, 3] be balanced positive semi- definite matrices of rank 2n − 2 and Ψ = EX using E from Lemma 1. The Lyapunov-Krasovskii functional is assumed as, (cid:90) V (Ψ (t)) = ΨT (t) P Ψ (t) + t t−τ1 ΨT (s) Q1Ψ (s) ds ΨT (s) Q2Ψ (s) ds T Ψ T Ψ (s) Z1 Ψ (s) dsdθ (s) Z2 Ψ (s) dsdθ T Ψ (s) Z3 Ψ (s) dsdθ (30) (cid:90) (cid:90) (cid:90) (cid:90)−τ1 t t−τ2 0 −τ1 0 −τ2 (cid:90) (cid:90) (cid:90) t t t+θ t+θ t −τ2 t+θ + + + + T V (Ψ (t)) = Ψ (t) Z1 Ψ (t) T (t) Z3 Ψ (t) T T (t) P Ψ (t) + ΨT (t) P Ψ (t) + ΨT (t) Q1Ψ (t) + ΨT (t) Q2Ψ (t)− ΨT (t − τ1) Q1Ψ (t − τ1)− ΨT (t − τ2) Q2Ψ (t − τ2) + τ1 Ψ (t) Z2 Ψ (t) + (τ2 − τ1) Ψ + τ2 Ψ − (cid:90) (cid:90) (cid:90)−τ1 X (t)T X (t − τ1)T X (t − τ2)T(cid:105) (cid:104) (t + θ) Z3 Ψ (t + θ) dθ (t + θ) Z1 Ψ (t + θ) dθ (t + θ) Z2 Ψ (t + θ) dθ 0 −τ1 0 −τ2 T Ψ T Ψ T Ψ −τ2 − − (cid:99)X = Let, then, Ψ = E X = E(cid:2)A0 A1 A2 (t) P Ψ (t) = (cid:99)X(cid:2)A0 A1 A2 (cid:3)(cid:99)X (cid:3)T E T PEX (t) (cid:3)(cid:99)X ΨT (t) P Ψ (t) = X (t)T E T PE(cid:2)A0 A1 A2 T Ψ T T ΨT (t) Q1Ψ (t) = X (t)T E T Q1EX (t) (31) (32) (33) (34) (35) (36) 4 IET Research Journals, pp. 1–10 c(cid:13) The Institution of Engineering and Technology 2015 ΨT (t) Q2Ψ (t) = X (t)T E T Q2EX (t) ΨT (t − τ1) Q1Ψ (t − τ1) = X (t − τ1)T E T Q1EX (t − τ1) ΨT (t − τ2) Q2Ψ (t − τ2) = For i = {1, 2, 3}, X (t − τ2)T E T Q2EX (t − τ2) T Ψ (t) Zi Ψ (t) = (cid:99)X(cid:2)A0 A1 A2 Consider a set of matrices, (cid:3)(cid:99)X T H T H T H T 12 H22 H23 H24 13 H T 23 H33 H34 14 H T 24 H T 34 Z1 (cid:3)T E T ZiE(cid:2)A0 A1 A2 H11 H12 H13 H14  (cid:60) 0 I11  (cid:60) 0  (cid:60) 0 J11 J12 J13 J14 I14 I13 I24 I23 I33 I34 I T 34 Z2 J T 12 J22 J23 J24 J T 13 J T 23 J33 J34 J T 14 J T 24 J T 34 Z3 I12 I22 I T 23 I T 24 I T 12 I T 13 I T 14 (37) (38) (39) (40) (41) (42) (43) AT 0 E T ΞEA0 + τ1E T H11E + E T H14E + E T H T τ2E T I11E + E T I14E + E T I T 14E+ 14E + (τ2 − τ1)E T J11E 0 E T ΞEA1 + τ1E T H12E − E T H14E G12 =E T PEA1 + AT 24E + τ2E T I12E + E T I T + E T H T (τ2 − τ1)E T J12E + E T J14E 24E+ G13 =E T PEA2 + AT 0 E T ΞEA2 + τ1E T H13E+ E T H T 34E + τ2E T I13E + E T I T (τ2 − τ1)E T J13E − E T J14E 34E − E T I14E+ (50) (51) G22 = − E T Q1E + AT 1 E T ΞEA1 + τ1E T H22E − E T H24E G23 =AT 24E + τ2E T I22E + (τ2 − τ1)E T J22E − E T H T + E T J24E + E T J T 24E 1 E T ΞEA2 + τ1E T H23E − E T H T 34E + τ2E T I23E − E T I24E + (τ2 − τ1)E T J23E − E T J24E + E T J T 34E (52) (53) G33 = − E T Q2E + AT 2 E T ΞEA2 + τ1E T H33E − E T I34E 34E + τ2E T I33E + (τ2 − τ1)E T J33E − E T I T − E T J34E − E T J T 34E Ξ =τ1Z1 + τ2Z2 + (τ2 − τ1) Z3 (54) (55) (cid:104) Let,(cid:98)Ψ θ = (cid:90) then, Where Hij , Iij , Jij ∀i ∈ [1, 3] ∀j ∈ [1, 4] are some arbitrary matri- ces to be found by an LMI solver with size 2n × 2n. Ψ (t + θ)T(cid:105) Ψ (t)T Ψ (t − τ1)T Ψ (t − τ2)T (cid:98)Ψ θ H T H T H T 12 H22 H23 H24 13 H T 23 H33 H34 14 H T 24 H T 34 Z1 (cid:98)Ψ H11 H12 H13 H14 (cid:98)Ψ I11 (cid:98)Ψ θ (cid:98)Ψ J11 J12 J13 J14 (cid:98)Ψ θ I14 I13 I24 I23 I33 I34 I T 34 Z2 J T 12 J22 J23 J24 J T 13 J T 23 J33 J34 J T 14 J T 24 J T 34 Z3 I12 I22 I T 23 I T 24 I T 12 I T 13 I T 14 T 0 −τ1 0 −τ2 (cid:90) (cid:90)−τ1 −τ2 T θ dθ ≥ 0 θ dθ ≥ 0 T θ dθ ≥ 0 (44) (45) (46) (47) The matrices in Ineqs. (41) to (43) are chosen to satisfy expression in Ineqs. (45) to (47), which further simplify V in Eqn. (31). Parts of Eqn. (31) consisting integrals with multiplication two variable in terms of θ are eliminated when added with Ineqs. (45) to (47), since V + {positive semidef inite} < 0 implies V < 0. Substi- tuting Eqns. (32) to (40) in Eqn. (31), adding Ineqs. (45) to (47) and further solving leftover integrals, Ineq. (48) is obtained. V ≤ (cid:99)X G11 G12 G13 12 G22 G23 13 GT 23 G33 GT GT (cid:99)X T where, G11 =AT 0 E T PE + E T PEA0 + E T Q1E + E T Q2E+ (49) IET Research Journals, pp. 1–10 c(cid:13) The Institution of Engineering and Technology 2015 The matrices P , Q1, Q2, Zi, Hij, H T (cid:101)P , (cid:101)Q1, (cid:101)Q2, (cid:101)Zi, (cid:101)Hij, (cid:101)H T ij and Fi ij and (cid:101)Fi ∀i ∈ [1, 3], ∀j ∈ from Lemma 2 ∀i ∈ [1, 3], ∀j ∈ [1, 4] will generate corresponding [1, 4] of size (2n − 2) × (2n − 2) when multiplied with eigenvec- tor matrices U T , U at appropriate positions. The corresponding Gij is as given in Eqn. (56). ij, (cid:101)Jij, (cid:101)J T ij, Jij, J T ij, Iij, I T ij,(cid:101)Iij,(cid:101)I T (cid:20)(cid:101)Gij(2n−2)×(2n−2) 02×(2n−2) (cid:21) {i, j} ∈ [1, 3]; (cid:101)Gji = (cid:101)GT 0(2n−2)×2 02×2 Gij = U (56) U T ji ∀i (cid:54)= j the LMIs given in (cid:3) With above set of reduced order matrices, Eqns. (19) to (22) can be obtained. Feasibility of LMIS in Eqns. (19) to (22) determine the consensus reachability of the multi-agent system affected by time-delays. They can be solved by using solvers like SeDuMi [27], Matlab LMI Lab solver etc.. 3.2 Nyquist stability approach Stability analysis of the linear element for different control inputs is performed using frequency domain analysis and Nyquist stability criterion which is discussed in Sections 3.2.1 and 3.2.2. First control law: Consider a control input ui1 with input 3.2.1 delay τ1 and communication delay τ2 as given in Eqn. (5) assuming τ1 ≤ τ2. (48) Theorem 2. The system represented by Eqns. (4) and (5) is stable if and only if, (cid:113) λk ω4 − 2ω3 sin (ωτ1) + ω2 (1 − 2 cos (ωτ1)) + 1 < 1 (57) 5 where ω satisfies −π = − ωτ2 + arg (cid:17) (cid:16)−λk (cid:18) − arctan ω cos (ωτ1) − sin (ωτ1) −ω2 + cos (ωτ1) + ω sin (ωτ1) (cid:19) (58) Proof. The system in Eqns. (4) and (5) can also be represented as, (cid:0)xj (t − τ2) − xi (t − τ1)(cid:1) (59) xi (t) = − xi (t − τ1) + n(cid:88) j=1 aij j=1 aij 1(cid:80)n (cid:12)(cid:12)(cid:12)s2I + (s + 1) Ie (cid:16) n(cid:89) −sτ1 − Ae −sτ2 s2I + (s + 1) Ie −sτ1 − λke (cid:12)(cid:12)(cid:12) = −sτ2(cid:17) Converting it into s domain will give the characteristic expression as, k=1 Consider ∀λk (cid:54)= 0, −λke−sτ2 Gk (s) = s2 + (s + 1) e−sτ1 Magnitude expression of Eqn. (61) is given by, (cid:12)(cid:12)Gk (jω)(cid:12)(cid:12) = (cid:112)ω4 − 2ω3 sin (ωτ1) + ω2 (1 − 2 cos (ωτ1)) + 1 λk Phase expression of Eqn. (61) is given by, (cid:18) Gk (jω) = − ωτ2 + arg arctan ω cos (ωτ1) − sin (ωτ1) −ω2 + cos (ωτ1) + ω sin (ωτ1) (cid:16)−λk (cid:17)− (60) (61) (62) (cid:19) (63) Let us assume at ω = ω, the Nyquist plot intersects with negative real axis. The phase at ω = ω is −π, −π = − ωτ2 + arg (cid:18) arctan ω cos (ωτ1) − sin (ωτ1) −ω2 + cos (ωτ1) + ω sin (ωτ1) (cid:19) (64) (cid:16)−λk (cid:17)− By applying Nyquist stability criterion, magnitude given in Eqn. (62) should satisfy the condition as given in Ineq. (65). λk ω4 − 2ω3 sin (ωτ1) + ω2 (1 − 2 cos (ωτ1)) + 1 < 1 (65) (cid:3) 3.2.2 Eqn. (6), where, Second control law: Consider control input ui2 given in Theorem 3. The system represented in Eqns. (4) and (6) is stable if and only if, (cid:113) (cid:113) 1 + ω2 λk(cid:112) (cid:16)−λk (cid:17) (cid:18) ω4 − 2ω3 sin (ωτ1) + ω2 (1 − 2 cos (ωτ1)) + 1 where ω satisfies, −π = −ωτ2 + arg + arctan (ω)− ω cos (ωτ1) − sin (ωτ1) −ω2 + cos (ωτ1) + ω sin (ωτ1) < 1 (66) (cid:19) (67) arctan 6 Proof. The proof follows a similar procedure as given in Theorem 2 (cid:3) 4 Simulation and Implementation Results The communication topologies considered for simulation and im- plementation are depicted in Figs. 3a and 3b. The graph in Fig. 3a is undirected, strongly connected and 2-regular balanced with each node receiving states' information from two neighbours and send- ing states' information to the same neighbours. The graph in Fig. 3b is directed, has a spanning tree and unbalanced. Using the results obtained in Theorems 1 to 3, limits on communication time-delay for given input-delays are calculated for both the topologies. The feasibility of LMIs given in Theorem 1 is solved using SeDuMi [27] solver for Matlab/Octave. The expressions in Theorems 2 and 3 have three unknowns (ω, τ1, τ2), a unique solution can be obtained if it is assumed that τ1 = τ2 or else, stable range of τ2 for a given τ1 have to be found. Dominant pole for both the topologies in Figs. 3a and 3b is −1, Nyquist plot used in one of the cases with λ = −1 and assumption τ1 = 0.4 is shown in Fig. 4. It can be observed that the system is stable if τ2 < 0.48 (Fig. 4). The time-delay tolerances are tested with simulations and on a hardware setup (Fig. 5). The results obtained using Lyapunov-Krasovskii and Nyquist approaches from Theorems 1 to 3 are tabulated in Tables 1 and 2, corresponding plots of τ1 vs τ2 depicting stable regions are given in Figs. 6a and 6b. (a) Four agents. (b) Five agents. Fig. 3: Graphs of communication topologies. Fig. 4: Nyquist plot with ui1 in Eqn. (5) for different time delays. From the results in Tables 1 and 2 and Figs. 6a and 6b, it can be deduced that the Lyapunov-Krasovskii approach is conservative compared to Nyquist approach with respect to time-delay. Conser- vativeness of Lyapunov-Krasovskii approach is more evident for topology in Fig. 3b with control law in Eqn. (6). Multi-agent systems connected by topologies in Figs. 3a and 3b reach consensus with full range of time-delay given by Nyquist approach for control laws in Eqns. (5) and (6). Nyquist approach in Section 3.2 is not applicable IET Research Journals, pp. 1–10 c(cid:13) The Institution of Engineering and Technology 2015 123412345−1.25−1.00−0.75−0.50−0.250.000.250.50Real−1.50−1.25−1.00−0.75−0.50−0.250.000.250.50ImaginaryNyquistplotforλ=-1withτ1=0.4τ2=0.478τ2=0.479τ2=0.48−1.01−1.00−0.990.010.00−0.01Fig. 5: Hardware setup connected by LAN. to control laws given in Eqns. (7) and (8). Solving LMIs is com- putationally more intensive compared to solving of equations from Theorems 2 and 3. Moreover, the increase in computational time of solving LMIs is exponentially as the number of nodes are increased and the increase with Nyquist approach is linear. Table 1 Maximum value of τ2 for a given τ1 (τ1 ≤ τ2) with topology in Fig. 3a and control inputs given in Eqns. (5) to (8). τ1 (sec) τ2 (sec) Lyapunov Approach ui1 ui2 ui3 ui4 Nyquist Approach ui1 ui2 0 0.1 0.2 0.3 0.4 0.5 0.6 τ1 = τ2 1.414 1.063 0.791 0.572 0.413 - - 0.405 1.154 0.994 0.851 0.720 0.598 - - 0.492 0.577 0.395 0.252 - - - - 0.223 0.894 0.656 0.462 0.30 - - - 0.3 1.570 1.308 1.035 0.755 0.479 - - 0.421 1.351 1.178 0.999 0.821 0.664 0.541 - 0.520 Table 2 Maximum value of τ2 for a given τ1 (τ1 ≤ τ2) with topology in Fig. 3b and control inputs given in Eqns. (5) and (6). τ1 (sec) τ2 (sec) Lyapunov Approach Nyquist Approach 0 0.1 0.2 0.3 0.4 0.5 0.6 τ1 = τ2 ui1 1.414 1.063 0.791 0.572 0.413 - - 0.405 ui2 1.154 0.572 0.567 0.605 0.598 - - 0.492 ui1 1.570 1.308 1.035 0.755 0.479 - - 0.421 ui2 1.351 1.178 0.999 0.821 0.664 0.541 - 0.520 (a) Plot of τ1 vs τ2 for topology in Fig. 3a. described earlier, pulse width of both the agents can be observed in the display of two channel DSO. (b) Plot of τ1 vs τ2 for topology in Fig. 3b. Fig. 6: Plots representing stable regions with τ1 ≤ τ2. Simulations are performed using scripts written in C to have uniformity with hardware implementation. Implementations of cor- responding simulations are performed on a network of four agents. Four Arduino-Uno boards for topology in Fig. 3a and five Arduino- Uno boards for topology in Fig. 3b are considered as nodes of sensor networks, all of them are connected to host computers us- ing serial interface. Host computers are connected by LAN switch locally and communication topology is based on the graphs shown in Figs. 3a and 3b. All the implementations are performed after time synchronization at the start of each run with a local server through ntp protocol. UDP packet switching is used for communication and appropriate precautions like time-stamping of packets are taken to ensure that packets are received in order. Owing to limitations in ca- pability of hardware, the step-size is chosen as 10ms. Fig. 5 depicts hardware setup consisting of two agents with connection settings as Fig. 7: Plots with ui3 in Eqn. (8) for asynchronous time delays. IET Research Journals, pp. 1–10 c(cid:13) The Institution of Engineering and Technology 2015 7 0.000.050.100.150.200.250.300.350.400.450.500.55Inputdelayτ1(sec)0.00.10.20.30.40.50.60.70.80.91.01.11.21.31.41.51.6Communicationdelayτ2(sec)LMIui1LMIui2Nyquistui1Nyquistui2LMIui3LMIui4τ1=τ2τ1<τ20.000.050.100.150.200.250.300.350.400.450.500.55Inputdelayτ1(sec)0.00.10.20.30.40.50.60.70.80.91.01.11.21.31.41.51.6Communicationdelayτ2(sec)LMIui1LMIui2Nyquistui1Nyquistui2τ1=τ2τ1<τ2020406080100120140160180Time(s)050100150200Position(units)Hardwareandsimulationplotofxi(t)withτ1=0.1,τ2=0.39hwagent1hwagent2hwagent3hwagent4simagent1simagent2simagent3simagent4141144114117020406080100120140160180Time(s)−2−1012Velocity(units/s)Hardwareandsimulationplotofxi(t)withτ1=0.1,τ2=0.39hwagent1hwagent2hwagent3hwagent4simagent1simagent2simagent3simagent4Input time-delay τ1 and communication time-delay τ2 are user- defined and implemented from code. A minor deviation is visible between simulation and implementation results due to additional delay (≈ 4ms) in implementation due to actual processing and com- munication. The pulse width of PWM wave generated from an Arduino-Uno is considered as state xi and the rate of change is considered as vi of each agent. Initial conditions for both simulation and implementation for topology in Fig. 3a are assumed to be, {xi0} = [0, 230, 110, 40] and {vi0} = [0, 0, 0, 0]. Plots with overlapping simulation and hardware results are given in Figs. 7 to 10 to show the effectiveness of theoret- ical results given in Tables 1 to 3 and Figs. 6a and 6b. For control law in Eqn. (7) with input delay τ1 = 0.1s and communication delay of τ2 = 0.39s, Fig. 7 depict states xi(t) and xi(t) vs time in seconds respectively. Similarly, Fig. 8 show the plots of states for control law in Eqn. (8) with τ1 = 0.2s and of τ2 = 0.46s. It can be observed that the difference in states, Ψ is asymptotically converging to zero and reinforcing the effectiveness of the theoretical results given in Table 1 and Fig. 6a. Fig. 9: Plot with ui1 in Eqn. (5) for uneven time delays. as discussed earlier, corresponding results are depicted in Fig. 10. Limit cycles are exhibited with τ1 = 0.2 and τ2 ≥ 1.04 as shown in Fig. 10. The corresponding numerical values for Nyquist approach for λ = −1 are provided in Table 3. Using results in Theorems 1 to 3, stable regions with respect to time-delays for both the topologies in Figs. 3a and 3b are calcu- lated. Few of them are validated with the help of simulations and corresponding implementations as given above. Figs. 6a and 6b de- pict the calculated stable regions for topologies in Figs. 3a and 3b respectively. Table 3 Values of ω &(cid:12)(cid:12)Gi (ω)(cid:12)(cid:12) for different values of τ1 and τ2 with ui1 in (cid:12)(cid:12)Gi (ω)(cid:12)(cid:12) Eqn. (5) and communication topology in Fig. 3b. τ1 (sec) τ2 (sec) ω Fig. 8: Plots with ui4 in Eqn. (8) for asynchronous time delays. 0.2 0.2 0.2 0.79 1.03 1.04 1.375 1.249 1.244 0.841 0.997 1.002 Similarly, some simulations and corresponding hardware valida- tions are performed on a five-agent system with communication topology given in Fig. 3b, the corresponding adjacency matrix A and A are as given below,  A = 0 1 0 0 0 1 0 0 1 1 0 0 0 0 0 1 0 0 0 1 1 0 1 1 0  , A =  0 1 2 0 0 0  . 0 0 1 3 1 0 1 2 0 1 3 0 0 0 1 2 0 0 0 1 2 0 1 3 0 1 The initial values for the five-agent system are considered as x1(0) = 0, x2(0) = 230, x3(0) = 110, x4(0) = 40, x5(0) = 170 and xi(0) = 0, i ∈ [1, 5]. From Fig. 9, it can be observed that the system converges at a faster rate with τ1 = 0.2 and τ2 = 0.79. With τ1 = 0.2 and τ2 = 1.03 given by Nyquist approach, the sys- tem converges at a very slow rate in simulation and further slower in implementation due to added delay from communication links 4.1 Remarks 1.Lyapunov-Krasovskii is applicable to all the four control laws pro- vided the communication topology satisfy the conditions mentioned in Theorem-1, but Nyquist approach is not applicable. Lyapunov- Krasovskii approach is more conservative with respect to time- delay tolerance, whereas Nyquist approach gives the full range of time-delay. 2.Describing function analysis allows us to use Nyquist approach on approximated nonlinear multi-agent system, which is better at providing time-delay tolerance ranges compared to Lyapunov approach. 3.Compared to work in [14], we have considered saturation and time- delays in the system. An approximate analysis with the help of de- scribing function is performed. Some conditions in Theorems 1 to 3 8 IET Research Journals, pp. 1–10 c(cid:13) The Institution of Engineering and Technology 2015 020406080100120140160180Time(s)050100150200Position(units)Hardwareandsimulationplotofxi(t)withτ1=0.2,τ2=0.46hwagent1hwagent2hwagent3hwagent4simagent1simagent2simagent3simagent4141144114117020406080100120140160180Time(s)−2−1012Velocity(units/s)Hardwareandsimulationplotofxi(t)withτ1=0.2,τ2=0.46hwagent1hwagent2hwagent3hwagent4simagent1simagent2simagent3simagent4020406080100120140160180Time(sec)050100150200250Position(units)Hardwareandsimulationplotofxi(t)withτ1=0.2,τ2=0.79hwagent1hwagent2hwagent3hwagent4hwagent5simagent1simagent2simagent3simagent4simagent5127130103.5107.5020406080100120140160180Time(sec)−2−10123Velocity(units/sec)Hardwareandsimulationplotofxi(t)withτ1=0.2,τ2=0.79hwagent1hwagent2hwagent3hwagent4hwagent5simagent1simagent2simagent3simagent4simagent55 Conclusion The consensus problem for second order saturated multi-agent sys- tem with asynchronous communication and input time-delays is presented in the paper. An approximate system with separate linear and nonlinear elements is derived using describing function analy- sis to study the limit cycle behaviour. The instability of limit cycles or consensus reachability is estimated using describing functions, stability of linear elements with the help of Lyapunov-Krasovskii function and Nyquist stability criterion. Stable ranges of input and communication time-delays are calculated for different control laws using both the approaches and comparative results are presented. Justification to the theoretical results is done with the help of simula- tions and corresponding implementations on hardware. With current control laws, the system is not immune to external disturbances in the state information. Noise in the state information and its mitigation strategies will be considered in the future research. References 6 1 Jadbabaie, A., Lin, J., Morse, A.S.: 'Coordination of groups of mobile autonomous agents using nearest neighbor rules', IEEE Transactions on Automatic Control, 2003, 48, (6), pp. 988–1001 2 Olfati.Saber, R., Murray, R.M.: 'Consensus problems in networks of agents with switching topology and time-delays', IEEE Transactions on Automatic Control, 2004, 49, (9), pp. 1520–1533 3 Vicsek, T., Czirók, A., Ben.Jacob, E., Cohen, I., Shochet, O.: 'Novel type of phase transition in a system of self-driven particles', Phys Rev Lett, 1995, 75, pp. 1226– 1229 4 Cao, Y., Yu, W., Ren, W., Chen, G.: 'An overview of recent progress in the study of distributed multi-agent coordination', IEEE Transactions on Industrial Informatics, 2013, 9, (1), pp. 427–438 5 Wang, X., Zeng, Z., Cong, Y.: 'Multi-agent distributed coordination control: Devel- opments and directions via graph viewpoint', Neurocomputing, 2016, 199, pp. 204 – 218 6 Xiao, F., Wang, L.: 'Asynchronous consensus in continuous-time multi-agent sys- tems with switching topology and time-varying delays', IEEE Transactions on Automatic Control, 2008, 53, (8), pp. 1804–1816 7 Ren, W.: 'On consensus algorithms for double-integrator dynamics', IEEE Trans- actions on Automatic Control, 2008, 53, (6), pp. 1503–1509 8 Hu, J., Lin, Y.S.: 'Consensus control for multi-agent systems with double- integrator dynamics and time delays', IET Control Theory Applications, 2010, 4, (1), pp. 109–118 9 Münz, U., Papachristodoulou, A., Allgöwer, F.: 'Delay robustness in consensus problems', Automatica, 2010, 46, (8), pp. 1252 – 1265 10 Meng, Z., Ren, W., Cao, Y., You, Z.: 'Leaderless and leader-following consensus with communication and input delays under a directed network topology', IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 2011, 41, (1), pp. 75–88 11 Zhang, W., Liu, J., Zeng, D., Yang, T.: 'Consensus analysis of continuous-time second-order multi-agent systems with nonuniform time-delays and switching topologies', Asian Journal of Control, 2013, 15, (5), pp. 1516–1523 12 Meng, X., Meng, Z., Chen, T., Dimarogonas, D.V., Johansson, K.H.: 'Pulse width modulation for multi-agent systems', Automatica, 2016, 70, pp. 173 – 178 13 Yu, W., Chen, G., Cao, M., Kurths, J.: 'Second-order consensus for multiagent systems with directed topologies and nonlinear dynamics', IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 2010, 40, (3), pp. 881–891 14 Liu, K., Xie, G., Ren, W., Wang, L.: 'Consensus for multi-agent systems with in- herent nonlinear dynamics under directed topologies', Systems & Control Letters, 2013, 62, (2), pp. 152 – 162 15 Li, J., Guan, Z.H., Chen, G.: 'Multi-consensus of nonlinearly networked multi- agent systems', Asian Journal of Control, 2015, 17, (1), pp. 157–164 20 16 Li, Y., Xiang, J., Wei, W.: 'Consensus problems for linear time-invariant multi- agent systems with saturation constraints', IET Control Theory Applications, 2011, 5, (6), pp. 823–829 17 Meng, Z., Zhao, Z., Lin, Z.: 'On global leader-following consensus of identical linear dynamic systems subject to actuator saturation', Systems & Control Letters, 2013, 62, (2), pp. 132 – 142 18 Wei, A., Hu, X., Wang, Y.: 'Tracking control of leader-follower multi-agent sys- tems subject to actuator saturation', IEEE/CAA Journal of Automatica Sinica, 2014, 1, (1), pp. 84–91 19 Chu, H., Yuan, J., Zhang, W.: 'Observer-based adaptive consensus tracking for lin- ear multi-agent systems with input saturation', IET Control Theory Applications, 2015, 9, (14), pp. 2124–2131 Su, H., Chen, M.Z.Q.: 'Multi-agent containment control with input saturation on switching topologies', IET Control Theory Applications, 2015, 9, (3), pp. 399–409 21 Cui, G., Xu, S., Lewis, F.L., Zhang, B., Ma, Q.: 'Distributed consensus track- ing for non-linear multi-agent systems with input saturation: a command fil- tered backstepping approach', IET Control Theory Applications, 2016, 10, (5), pp. 509–516 22 You, X., Hua, C., Peng, D., Guan, X.: 'Leader following consensus for multi-agent systems subject to actuator saturation with switching topologies and time-varying delays', IET Control Theory Applications, 2016, 10, (2), pp. 144–150 Sun, Y.G., Wang, L., Xie, G.: 'Average consensus in networks of dynamic agents with switching topologies and multiple time-varying delays', Systems & Control Letters, 2008, 57, (2), pp. 175 – 183 24 Lin, P., Jia, Y.: 'Consensus of a class of second-order multi-agent systems with time-delay and jointly-connected topologies', IEEE Transactions on Automatic 23 Fig. 10: Plot with ui1 in Eqn. (5) for uneven time delays. are derived for reaching consensus and estimation of non-existence of limit cycles. 4.Compared to the research presented by authors in [16–19, 21], we have considered time-delays along with saturation in second order multi-agent systems. Compared to the work presented by You et al. [22], asynchronous time-delays are considered rather than single delay. IET Research Journals, pp. 1–10 c(cid:13) The Institution of Engineering and Technology 2015 9 050100150200250Time(sec)050100150200250Position(units)Hardwareandsimulationplotofxi(t)withτ1=0.2s,τ2=1.03shwagent1hwagent2hwagent3hwagent4hwagent5simagent1simagent2simagent3simagent4simagent5127130103.5107.5050100150200250Time(sec)−2−10123Velocity(units/sec)Hardwareandsimulationplotofxi(t)withτ1=0.2s,τ2=1.03shwagent1hwagent2hwagent3hwagent4hwagent5simagent1simagent2simagent3simagent4simagent5050100150200250Time(sec)050100150200250Position(units)Hardwareandsimulationplotofxi(t)withτ1=0.2s,τ2=1.04shwagent1hwagent2hwagent3hwagent4hwagent5simagent1simagent2simagent3simagent4simagent5127130103.5107.5050100150200250Time(sec)−2−10123Velocity(units/sec)Hardwareandsimulationplotofxi(t)withτ1=0.2s,τ2=1.04shwagent1hwagent2hwagent3hwagent4hwagent5simagent1simagent2simagent3simagent4simagent525 Control, 2010, 55, (3), pp. 778–784 Slotine, J.J.E., Li, W. 5, Describing Function Analysis. control'. (Englewood Cliffs (N.J.): Prentice Hall, 1991. pp. 157–190 26 Lin, P., Jia, Y.: 'Average consensus in networks of multi-agents with both switch- ing topology and coupling time-delay', Physica A: Statistical Mechanics and its In: 'Applied nonlinear 27 Applications, 2008, 387, (1), pp. 303 – 313 Sturm, J.F.: 'Using SeDuMi 1.02, a MATLAB toolbox for optimization over sym- metric cones', Optimization Methods and Software, 1999, 11–12, pp. 625–653. version 1.05 available from http://fewcal.kub.nl/sturm 10 IET Research Journals, pp. 1–10 c(cid:13) The Institution of Engineering and Technology 2015
1208.3691	1	1208	2012-08-17T21:10:56	On the genericity properties in networked estimation: Topology design and sensor placement	[ "cs.MA", "cs.IT", "cs.IT" ]	In this paper, we consider networked estimation of linear, discrete-time dynamical systems monitored by a network of agents. In order to minimize the power requirement at the (possibly, battery-operated) agents, we require that the agents can exchange information with their neighbors only \emph{once per dynamical system time-step}; in contrast to consensus-based estimation where the agents exchange information until they reach a consensus. It can be verified that with this restriction on information exchange, measurement fusion alone results in an unbounded estimation error at every such agent that does not have an observable set of measurements in its neighborhood. To over come this challenge, state-estimate fusion has been proposed to recover the system observability. However, we show that adding state-estimate fusion may not recover observability when the system matrix is structured-rank ($S$-rank) deficient. In this context, we characterize the state-estimate fusion and measurement fusion under both full $S$-rank and $S$-rank deficient system matrices.	cs.MA	cs	On the genericity properties in networked estimation: Topology design and sensor placement Mohammadreza Doostmohammadian and Usman A. Khan 1 2 1 0 2 g u A 7 1 ] A M . s c [ 1 v 1 9 6 3 . 8 0 2 1 : v i X r a Abstract In this paper, we consider networked estimation of linear, discrete-time dynamical systems monitored by a network of agents. In order to minimize the power requirement at the (possibly, battery-operated) agents, we require that the agents can exchange information with their neighbors only once per dynamical system time-step; in contrast to consensus-based estimation where the agents exchange information until they reach a consensus. It can be verified that with this restriction on information exchange, measurement fusion alone results in an unbounded estimation error at every such agent that does not have an observable set of measurements in its neighborhood. To over come this challenge, state-estimate fusion has been proposed to recover the system observability. However, we show that adding state-estimate fusion may not recover observability when the system matrix is structured-rank (S-rank) deficient. In this context, we characterize the state-estimate fusion and measurement fusion under both full S-rank and S-rank deficient system matrices. The main results of this paper are the following. Firstly, we show that when the system matrix has full S-rank, state-estimate fusion alone (with no measurement fusion) can recover the observability. Subsequently, we characterize the minimal topology for inter-agent communication required for a stable networked estimator. Secondly, we provide methodologies to recover (networked) estimator observability when the system matrix is S-rank deficient. In particular, we introduce a novel agent classification based on their local measurements and identify the agents that are crucial for stable estimation error. We then provide topology modifications and sensor placement techniques to recover observability in the S-rank deficient scenario. Finally, we provide an iterative method to compute the local estimator gain at each agent that results into a stable estimation error once the observability is ensured using the aforementioned techniques. Keywords: Networked estimation, Observability, Structured system theory, Generic rank I. INTRODUCTION Estimation of dynamical systems with observations distributed among a network of agents is an important field of research, where the idea is to assign a group of agents to monitor a certain system Department of Electrical and Computer Engineering, Tufts University, {mrd,khan}@ece.tufts.edu. August 21, 2018 DRAFT 2 Fig. 1. A group of mobile ground-robots are tracking a large flock of quad-copters. Every ground-robot is monitoring part of the flock, and then shares its partial data with other ground-robots through a communication network. or phenomenon of interest. Agents are distributed in the sense that each agent can only measure some of the states of a dynamical system, referred to as local measurements. For example, a group of sensors spread geographically over a large region to monitor daily temperature evolution. The measurement data and dynamical models are further corrupted by noise and disturbances. The objective is to enable each agent to make an unbiased decision on the global state relying only on its own measurement and the measurements from its immediate neighbors. Such a scheme is often referred to as networked estimation where the term network implies that the information is restricted on a sparse network. Networked estimation is preferable to a wide range of applications as it is scalable and further requires less communication load at each individual agent, in contrast to the centralized case where each agent may require repeated long-distance communication to a central location. Applications of networked estimation include social networks [1] to learn global beliefs based on partial understanding of the state of the society, market, politics, etc., monitoring physical processes and environmental spatio-temporal fields [2], [3], state estimation in power systems [4] -- [6], and multi-agent systems such as collaborative target tracking and flocking of mobile robots [7], see for example Fig. 1. A variety of solutions exists for networked estimation starting from the earlier work in [8], [9] and references therein on parallel Kalman filter architectures for all-to-all connected networks, to more recent diffusion-based schemes via least mean square implementation, such as in Kalman filtering and smoothing [10] and distributed binary detection [11]. Meanwhile, incremental adaptive distributed strategies can be found in [12], [13] along with distributed moving horizon estimation [14] to minimize estimation error variance for constrained problems. State estimators based on low-cost single-bit data transmission is proposed in [15] with binary sign of innovations (sign of difference of measurement and estimated value). Information theoretic approach based on consensus over the Kullback-Leibler average of Gaussian PDFs is exploited in [16]. The literature can also be classified into static and dynamic estimation. In static August 21, 2018 DRAFT 3 Fig. 2. (a) The traditional two-time scale consensus-based approach. (b) single time-scale approach. estimation [2], [10] -- [12], [17], [18], the target state to be estimated does not change over time, while dynamic estimation [4], [8], [14] -- [16], [19] -- [23] takes into account the time-evolution of the system1. Consensus-based strategies have recently found a lot of interest in the context of sparsely-connected networks, where the main focus is to reduce the uncertainty of individual estimates by averaging on collaborative data. Early work in [19] -- [23] considers a two time-scale method where consensus is implemented at a time-scale different than the system dynamics. These results require that a consensus is reached within every two time-steps of the system dynamics, and is thus, challenged with a large number (infinite, in general) of consensus iterations between every two steps of the dynamics. To elaborate this, consider Fig. 2(a), where a large number (→ ∞) of data fusion iterations are implemented between every two successive time-steps, k and k + 1, of the dynamics. This approach requires communication over a much faster rate than the sampling of the dynamics, and thus, in general, becomes practically infeasible when the underlying system is operating under power constraints and has restricted communication and computation budgets. In contrast to the two time-scale approach to distributed estimation, recently References [1], [16] -- [18], [23] -- [25] studied the behavior of networked estimators when the communication time-scale is the same 1As stated in [17], diffusion algorithms can be extended for non-stationary (dynamic) tracking when the target is not moving too fast, i.e. its state is relatively stationary over a period such that the algorithm can converge. August 21, 2018 DRAFT k k+1 Dynamics/Estimator time-scale Fusion time-scale (consensus) k k+1 Dynamics/Estimator time-scale Fusion time-scale (a) (b) 4 as the time-scale of the dynamics, as shown in Fig. 2(b). This method is practically feasible for real-time applications and computationally efficient as compared to the two time-scale approach. A preliminary study on this single time-scale estimator is carried out in [24], where it is shown that a particular linear networked estimator has a bounded estimation error if the two-norm of the system matrix is less than the network tracking capacity -- a function of the communication network and observation models. Notice that in the two time-scale method, the communication network becomes irrelevant due to more information exchanges among the individuals (the information in a sparsely connected graph is equivalent to the information in a fully connected graph when a large number of information exchanges are carried out). Therefore, the performance and properties of the underlying estimator depends only on the data fusion principles among the agents. However, in the single time-scale scenario of Fig. 2 (b), the underlying agent network remains sparse and an arbitrary communication network may not suffice to make the networked estimation error stable (e.g., see [24], [26], [27]). In this context, the key problem is to design the structure of the inter-agent communication according to the underling fusion rules in order to recover the observability of the networked estimator. In this paper, we use a variant of the Networked Kalman-type Estimator (NKE) protocol, initially introduced in [24]. The main contribution is to determine the communication network among the agents to recover the observability of the underlying estimation protocol, given that each agent may not be locally (in its neighborhood) observable. We study the observability with a structural point of view [28] -- [34] in the sense that we explore the generic properties of the system. The generic properties are applicable to any choice of system parameters as long as the sparsity structure (zeros and non-zeros) is not violated. The generic approach is helpful when the underlying system parameters may change depending on the system operating point (linearization of non-linear dynamics) and is further significant in communication network design as the approach is independent of the exact value of the weights chosen for data fusion. Moreover, this implies that for smooth non-linear systems with fixed structure Jacobian matrix, similar analysis of the networked observability can be applied. Comparing with other work in the literature, we consider single time-scale estimation, as opposed to the multi time-scale estimation in [19] -- [22] and the vanishing time-step algorithms proposed in [35], [36]. Unlike [37], [38], we do not impose an agent hierarchy (i.e., we assume the processing/communication duties at all agents are the same). Avoiding agent hierarchy increases the reliability of node/link failure. We further do not require the communication network to be (strongly) connected [10], [11], [16], [17], [20], [26] or for it to include a cyclic path [12], [18]. Our goal is to design the network with minimal communication. Specifically, we use methodologies that are independent of exact system values and rely August 21, 2018 DRAFT 5 only on the structure of the underlying system. This leads to a robust estimator design where the analysis is not algebraic, as in the conventional Grammian or PBH observability tests, but graph-theoretic [30], [31]. We now describe the rest of the paper. Section II provides preliminary material on basic dynamical system estimation and structured systems theory, whereas Section III presents our problem formulation. Section IV enlists our assumptions and describes a novel agent classification method. Section V covers the main results of this paper on state and output fusion, whereas local gain design is explored in Section VI. We provide an illustrative example and simulations in Section VII, and finally, Section VIII concludes the paper. II. BACKGROUND AND PRELIMINARIES We consider the system model to be a discrete-time linear dynamical system: (1) where xk ∈ Rn is the state vector, A = {aij} ∈ Rn×n is the system matrix, and vk ∼ N (0, V ) is the system noise. We note that the main emphasis of this paper is not on modeling but on structural xk+1 = Axk + vk, characteristics of the underlying system and the results we present hold for any phenomenon following Eq. (1). We assume that the dynamical system is monitored by a network of N agents such that each sensor i has the following observation model: where yi k ∈ Rpi is the output vector at agent i, ri k = Cixk + ri yi k, (2) k ∼ N (0, Ri) is the output noise, and Ci is the output matrix at agent i. With this notation, we can write the global observation model as where yk =  , rk =  r1 ... rN k k  , (3) (4) yk = Cxk + rk,  y1 ... yN k k  , C =  C1 ... CN rk ∼ N (0, R) is the global observation noise with R = blockdiag[R1, . . . , RN ], and C = {cij} is the global output matrix. August 21, 2018 DRAFT kk be the centralized Kalman estimator [39] at time k given all the observations, yk, up to time k. A. Centralized estimator Let(cid:98)xc It can be shown that the error,(cid:98)ec kk = xk −(cid:98)xc (cid:98)ec kk = (A − KcCA)(cid:98)ec kk, in this estimator is given by k−1k−1 + ηk, 6 (5) independent of(cid:98)ec where Kc is the centralized Kalman gain and the vector ηk collects the remaining terms that are kk is stable if and only if all the unstable modes (eigenvalues) of the system are observable. For the ease of explanation, we k−1k−1. It is well known that the centralized Kalman error,(cid:98)ec assume that there are no stable unobservable nodes. In other words, detectability and observability are equivalent throughout this paper. In the traditional sense of n-step (A, C)-observability, the observability Gramian is given by O =(cid:2)CT AT CT ... (An−1)T CT(cid:3) . (6) Algebraic tests for observability check the Gramian, O, for being full-rank or the matrix OTO for being invertible. An alternative method is the PBH (Popov-Belevitch-Hautus) observability test [40], which requires the matrix, [AT − sI CT ], to be full-rank for all s. The matrix [AT − sI] is full rank for all values of s other than the eigenvalues of A and, therefore, the PBH test is needed to be checked only for these values. Note that, both these algebraic methods rely on the knowledge of exact values of each element in the matrices A and C. However, in many dynamical systems, only the sparsity (zero and non-zero pattern) of these matrices may remain fixed while the non-zero elements are subject to change. For example, when the elements of the concerned matrices depend on certain parameters or operating points. Hence, these conventional methodologies fail to check for observability in such cases and graph-theoretic techniques are to be employed. We introduce such graph-based methods below. B. Graph notations Let X = {x1, . . . , xn} denote the state set, and let Y = {y1, . . . , yp} denote the output set. We define the system digraph as GA = (V, E), where V = X ∪ Y is the vertex set, and E is the edge set containing directed edges, (v1, v2) ∈ E, of the form v1 → v2 with v1, v2 ∈ V . The edge set E is defined as EA ∪ EC, where EA = {(xj, xi) aij (cid:54)= 0} and EC = {(xj, yi) cij (cid:54)= 0}. A path of length (cid:96) from v1 ∈ V to v(cid:96) ∈ V is such that there exists a sequence of vertices, v1, v2, . . . , v(cid:96) with each subsequent edge, (v1, v2), (v2, v3), . . . , (v(cid:96)−1, v(cid:96)) ∈ E. Here v1 is the begin-vertex of the path and v(cid:96) is August 21, 2018 DRAFT 7 Fig. 3. This digraph is an example of (A, C) observable system based on the Theorem 1. its end-vertex. Here, we assume that each vertex contained in a path occurs only once (simple path). A path is said to be Y -topped if it ends at a vertex in Y . A digraph is called strongly connected if there exist a directed path from each vertex to every other vertex in the digraph. In a not strongly connected digraph, define Strongly Connected Components (SCC) as its maximal strongly connected partitions or sub-graphs. A cycle is a simple path where the begin and end vertices are the same. Since the nodes in Y have no outgoing link, nodes included in a cycle all belong to X. As an example, consider Fig. 3 which shows the system digraph of a dynamical system with n = 7 states (encircled) and N = 3 measurements (or agents) denoted by squares. We now introduce some new concepts on SCCs over state vertices. These will be helpful in describing our results. Definition 1 (Parent SCC): A state SCC, is called a parent SCC, if it has no outgoing link to any state vertex not belonging to itself. Definition 2 (Child SCC): Any SCC that is not a parent SCC is a child SCC. Notice that the set of disjoint state SCCs in system matrix A can be explicitly characterized as either a parent or a child. As an example, the SCC containing vertices {4, 5, 6} in Fig. 3 is a parent SCC, since there is no outgoing edges from its states to other states {1, 2, 3, 7} not included in it. Furthermore, {1, 2} and {7} are child SCCs. More details on parent/child SCCs and efficient algorithms for computing SCCs in a digraph can be found in [27] and [41], respectively. We now use the concepts from this section to formally introduce structured systems theory and generic properties. August 21, 2018 DRAFT 8 C. Structured systems theory Structural analysis deals with system properties that do not depend on the numerical values of the parameters but only on the underlying structure (zeros and non-zeros) of the system [28] -- [34]. It turns out that if a structural property is true for one admissible choice of non-zero elements as free parameters it is true for almost all choices of non-zero elements and, therefore, is called a generic property of the system [42]. Furthermore, it can be shown that those particular (non-admissible) choices for which the generic property does not hold lie on some algebraic variety with zero Lebesgue measure, for more details see [42], [43]. Definition 3 (S-rank): The structural rank (also called generic rank) is the maximum rank for all numerical values of the non-zero entries of the matrix A. It is, in fact, an upper-bound on the numerical rank of A. The S-rank as a generic property holds for almost all choices of nonzero parameters of the matrix, A. It is equal to the cardinality of the maximum matching associated to the bipartite graph associated to the matrix, A. In the algebraic sense, this is the maximum number of non-zero elements in distinct rows and columns of the matrix, A [44]. Details on the generic rank implication in graph theoretic sense and algorithms on maximum matching can be found in [30], [41]. Among other generic properties, controllability/observability are of interest in the context of this paper, see [29], [30], [33], [34], [45], [46] for details. We extend the following theorem from the generic controllability results in [31]. Theorem 1: A dynamical system is generically observable if and only if in the system digraph: (i) Every state is the begin-node of a path that ends in an output (termed as a Y -topped path); (ii) There exist a disjoint union of Y -topped paths and cycles that cover all the state vertices. The following lemma is from [28]. Lemma 1: The condition (ii) in Theorem 1 on the generic observability of (An×n, Cm×n) is equivalent  A C  = n. to the following: S-rank (7) The proof of Theorem 1 for generic controllability and Lemma 1 is given in [28], [31], where other equivalent graph-theoretic conditions to generic controllability (observability) are also defined that we omit here. As an example, consider the system shown in Fig. 3. It can be verified that each state is a begin-vertex of a Y -topped path, and {7}, {4, 5, 6}, {1, 2, 3, a} constitute a disjoint union of cycles and Y -topped paths that cover all the state vertices in X. Thus, satisfying both conditions in Theorem 1 and the system in Fig. 3 is observable for almost all choices of non-zero elements. August 21, 2018 DRAFT 9 D. Corresponding graphs In this paper, we deal with two different graph representations: system digraph, GA, representing states of the dynamic system (1) and (2), and digraph GW defining the agentcommunication network. Let GW = (VW , EW ), where VW = {1, . . . , N} is the vertex set consisting of N agents, EW = {(i, j) i ← j} is the edge set, and Di = {i} ∪ {j (i, j) ∈ E} denote the extended neighborhood of agent i. Notice that, unlike many works in the literature we do not constrain GW to be undirected. In fact, no assumption on the topology is considered here, as designing GW is a contribution of this paper. Example: To shed more light on this, we give an example here. Consider the flocking motion example given in the Fig. 1. The position, velocity or acceleration of every quad-copters can be considered as a state of the system. having ground robots as agents, the coordination law among them, typically following dynamics (1) [47], defines the dynamical system, A, and system digraph, GA. The system outputs/measurements are the states tracked by the ground robots, and the communication network (to be designed) among these ground-robots is GW . III. PROBLEM FORMULATION We employ a variant of the Networked Kalman-type Estimator (NKE) proposed in [26], [27]. Let(cid:98)xi km be the state estimate of agent i at time k given the outputs up to time m, (m ≤ k), from its neighboring agents, j ∈ Di. Each agent implements the following: (i) Predictor and state fusion: (8) (9) kk−1 = k−1k−1 (cid:98)xi kk = (cid:98)xi (cid:88) wijA(cid:98)xj (cid:88) j∈Di j∈Di (ii) Estimator and output fusion:(cid:98)xi where W = {wij} is the state fusion weight matrix such that wij ≥ 0 with (cid:80) k − Cj(cid:98)xi kk−1 + Ki k j (yj CT kk−1) j∈Di wij = 1 (W is stochastic), and Ki k is the local estimator gain at agent i. Remark 1: Following are some useful remarks: (i) The diagonal entries of W are all nonzero, since every agent is in its own extended neighborhood and uses its own information. (ii) The first equation (8) is a local prediction fusion where each agent i fuses the neighboring estimates from time k − 1 and then implements a local predictor. (iii) In the second equation (9), each agent i updates its local prediction with an innovation term. We define this innovation as the difference between the state prediction of agent i and the state measurements obtained via agents, j ∈ Di. Adding this term, agent, i, makes its August 21, 2018 DRAFT 10 final estimate, (cid:98)xi kk, for the current time step. (iv) The protocol given in Eqs. (8) -- (9) takes place at the same time-scale as the system dynamics, see Fig. 2. Notice that both Eqs. (8) -- (9) can be combined into one equation; we give separate equations for the ease of explanation. Let the estimation error at agent i and time k be defined as k = xkk −(cid:98)xi ei kk, (10) and let ek = [(e1 k)T , . . . , (eN k )T ]T be the networked estimation error derived in the following. Proposition 1: Let qi k be some function of the system and measurement noise, vk and ri k, independent of ek−1 and let (cid:80)  DC = Kk = blockdiag[K1 k , . . . , KN k ], j∈D1 CT j Cj 0 ... (cid:80) 0 j∈DN CT j Cj qk = [(q1 k)T , . . . , (qN k )T ]T . Then we get the following networked error dynamics, ek = (W ⊗ A − KkDC(W ⊗ A))ek−1 + qk.  , (11) (12) The derivation requires some straightforward manipulations and is omitted here. Comparing this to Eq. (5), we note that the networked estimation error, ek, can be stabilized if and only if, the following pair, (W ⊗ A, DC), (13) is observable. In other words, a gain matrix, Kk, exists such that ρ(W ⊗ A − KkDC(W ⊗ A)) < 1 (i.e., it is a Schur matrix), if and only if (W ⊗ A, DC) is observable, where ρ(·) denotes the spectral radius of a matrix. As it can be seen from Eq. (13), the communication network, W , plays a major role in distributed estimation as opposed to the multiple time-scale approach where W is irrelevant. The role of W in observability is because of the single time-scale nature of the estimator, see Fig. 2. Remark 2: • The variables DC and Kk are block-diagonal matrices. • Every block diagonal,(cid:80) j∈Di CT j Cj, in the matrix DC, can be thought of as a representation of all of the measurements in the extended neighborhood of agent i. August 21, 2018 DRAFT 11 (Left) Matrix structure of the distributed system with no data fusion. Every block diagonal Wii ⊗ A is a subsystem Fig. 4. associated with the output/agent i. (Right) Adding data fusion, the intra-connections among these subsystems depends on the non-diagonals Wij ⊗ A, i (cid:54)= j. We refer to (W⊗A, DC) as the distributed system and GW⊗A as the graph associated with the matrix W⊗ A. For better understanding of the structural relevance of the estimator in (8) -- (9), we first consider W = I and DC defined as follows,  CT 1 C1 DC =  , ... CT N CN (14) implying no information exchange among the agents. This distributed system, (I⊗A, DC), can be thought of as N subsystems each of them associated to an n × n block diagonal in the matrix W ⊗ A, see Fig.4 (Left). Now consider W to have some non-zero non-diagonal entries. As it is shown in Fig.4 (Right), these entries define the inter-connections among these subsystems. To shed more light on this, consider the example given in Fig. 3 where we show a n = 7-state dynamical system with N = 3 agents, {a, b, c} such that agent a measures x3, agent b measures x5 and agent c measures x7. Each agent is required to estimate the entire n = 7 dimensional state-vector. Without any information fusion each agent only has a partial observation of the system as it is shown in Fig. 5. Therefore, each agent has to acquire the missing information via communicating with agents in its immediate neighborhood. However, in this illustration, no agent finds any observation in its neighborhood in addition to what it already possesses. Information sharing among the agents by applying state and output fusion provides more links among the subsystems in the distributed system digraph. This extra August 21, 2018 DRAFT (cid:1849)(cid:2869)(cid:2869)(cid:4670)(cid:1827)(cid:4671)(cid:4670)(cid:1829)(cid:2869)(cid:3021)(cid:1829)(cid:2869)(cid:4671)(cid:4670)(cid:1829)(cid:3015)(cid:3021)(cid:1829)(cid:3015)(cid:4671)(cid:1849)(cid:3015)(cid:3015)(cid:4670)(cid:1827)(cid:4671)SubsystemsSubsystem of Agent NSubsystems Intra‐connection(cid:1849)(cid:2869)(cid:2869)(cid:4670)(cid:1827)(cid:4671)(cid:3533)(cid:4670)(cid:1829)(cid:3037)(cid:3021)(cid:1829)(cid:3037)(cid:4671) (cid:3037)∈(cid:3005)(cid:3117)(cid:1849)(cid:3015)(cid:3015)(cid:4670)(cid:1827)(cid:4671)(cid:1849)(cid:3015)(cid:2869)(cid:4670)(cid:1827)(cid:4671)(cid:1849)(cid:2869)(cid:3015)(cid:4670)(cid:1827)(cid:4671)Subsystems Intra‐connectionSubsystem of Agent N(cid:3533)(cid:4670)(cid:1829)(cid:3037)(cid:3021)(cid:1829)(cid:3037)(cid:4671) (cid:3037)∈(cid:3005)(cid:3263)12 Fig. 5. The graph associated with distributed system, (I ⊗ A, DC ), with no data fusion represented as a sub-system at each agent. According to the Theorem 1, each sub-system (agent) is not observable with no data fusion. linking among subsystems and outputs, captured by the non-zeros in W and the summation in DC, has the potential to improve the generic observability of the system. In this regard, the main objective is to define the structure of the communication matrix W (graph GW ) such that the distributed system (W ⊗ A, DC) is generically observable. We seek a general method to make each subsystem observable. First, we describe how adding a link between two agents changes the graph structure of the distributed system. We explain this by considering the same example as in Fig. 3 and Fig. 5. In the case of output-fusion, a link between two agents, for example, from agent b to agent a (a ← b), implies that agent a has access to agent b's measurement, that is measurement of state x5. However, the state-fusion case is more involved. For example, adding a link from agent b to agent a implies a nonzero entry in W , i.e., wab(cid:54)=0, which reflects in the networked system matrix, W ⊗ A, as adding edges to some states in the subsystem associated to agent a from some states in the subsystem associated to agent b. This will be discussed in more detail in the next sections. We now enlist our assumptions and provide a novel agent classification that will help in establishing IV. OUR APPROACH our results. A. Assumptions In the rest of the paper, we make the following assumptions: (ii) For every agent, i, the pairs, (A, Ci) or (A,(cid:80) (i) The communication between the agents is stable, i.e., the the network is static; j∈Di CT j Cj), are not necessarily observable; August 21, 2018 DRAFT 13 (iii) The system is globally (A, C)-observable, i.e., if we collect all the sensor measurements at a center then the dynamical system is observable. Assumption (ii), in practice, makes the networked estimation problem more challenging and is where this work becomes significantly different from current approaches, see, for example, [7] and references therein. Assumption (iii) is a typical assumption in distributed estimation implying the observability of centralized estimator; without this, no estimation scheme will work. B. State and Agent classification To describe our approach, we provide a novel agent classification. Since the system is (A, C)-observable, (iii) in Section IV-A, we can enlist a disjoint set of cycles and Y -topped paths that covers all the state vertices (existence is ensured from condition (i) -- Theorem 1). We are interested in a listing that involves the maximal cycles and we denote this set as L. For example, from Fig. 3, the disjoint set of cycles and Y -topped paths that covers all the state vertices includes {(4, 6, 4), (5, b), (1, 2, 3, a), (7, c)}, and {(4, 5, 6, 4), (1, 2, 1), (7, 7), (3, a)}, among others. However, the latter includes the maximal cycles and thus L = {(4, 5, 6, 4), (1, 2, 1), (7, 7), (3, a)}. The following classification is with respect to L. (i) Type-α agent is an agent that appears in the Y -topped paths in L. For example, agent a in Fig. 3. (ii) Type-β agent is an agent that measures a state in the parent cycles cycles in L; a parent cycle is a cycle that does not have an outgoing link to any other state not belonging to itself. For example, agent b in Fig. 3. (iii) Type-γ agent is an agent that measure a state in the child cycles in L; a child cycle is a cycle that is not a parent cycle. For example, agent c in Fig. 3. The above agent classification leads to the following definition and lemma. Definition 4 (Crucial observation): A crucial observation is a measurement such that removing it renders the dynamical system unobservable. Lemma 2: The agents of Type-α and Type-β make "crucial" observations while the measurements at Type-γ agents are not crucial. Proof: Since the Type-β agents monitor the parent cycles and there is no outgoing link from a parent cycle to any other state outside this cycle, the states in the parent cycles can only be the begin vertices of a Y -topped path (in order to satisfy condition (ii) in Theorem 1) when any one of these states is connected to an output. Hence all the Type-β agents make crucial observations. On the other hand, removing a Type-α agent violates the condition (i) in Theorem 1 as the attached state vertex is not included in L anymore. Hence, Type-α agents are also crucial. Finally, the only location for the Type-γ August 21, 2018 DRAFT 14 THIS TABLE SHOWS DISTRIBUTED SYSTEM FOR DIFFERENT FUSION LEVELS. TABLE I Fusion level No data fusion Only state fusion Only output fusion Both measurement and state fusion equivalent distributed system (I ⊗ A, DC ) (W ⊗ A, DC ) (I ⊗ A, DC ) (W ⊗ A, DC ) agent is monitoring a child SCC, which either has a directed path to a Type-α agent or to a Type-β agent and hence is redundant. While the contribution of this state remains in L due to the cycle present there. For example, in Fig. 3, if either agent a or agent b is removed, then the system becomes unobservable. It can also be verified that agent c is non-crucial. Having defined types of agents, we note that the observability of the distributed system can be recovered via either W ⊗ A matrix (state-fusion) or DC (output-fusion). Here, we provide the minimal sufficient number of communication among the agents. Unlike our previous works [26], [27], we do not impose any constraint on the system matrix, A. Furthermore, the generic approach is further robust to uncertain systems and to linearized approximation of nonlinear models where the structure is fixed while the values are a function of the operating point [29]. V. RECOVERING OBSERVABILITY In this section, we first present some helpful results for the development of the paper and then find a general solution for (W ⊗ A, DC) observability. We first discuss the role of state fusion, related to the structure of matrix W , and then the role of output fusion, related to the structure of matrix DC (see Table I). The reason is to get more intuitive and separate solutions for state and output fusion; obviously, in real applications if two agents are linked together they nay share all of their information, including both their measurement and state estimates, to maximally improve their current state estimates. The results and proofs in this section are mainly graph theoretic that is a direct consequence of using the generic approach. A. Results on rank genericity The result below follows from Theorem 1 and Lemma 1 as provided in Section II. August 21, 2018 DRAFT 15 Corollary 1 (Full S-rank): A system matrix, A, is full S-rank if and only if its associated digraph has a disjoint union of cycles covering all the state vertices. Notice that non-zero diagonals of a matrix can be represented as a disjoint union of self-cycles in its associated digraph. From Corollary 1 and by Remark 1 (Section III) we obtain the following result. Corollary 2: The communication matrix, W , has a disjoint union of self-cycles and, (15) This is always true because W has non-zero diagonals, i.e., wii (cid:54)= 0, ∀i. Consequently, we state the following lemma for the networked system (W ⊗ A). S-rank(W ) = N. Lemma 3: For the communication matrix, WN×N , and system matrix, An×n, the networked sys- tem W ⊗ A is structurally full-rank if and only if A is structurally full-rank. Mathematically, S-rank(W ⊗ A) = N × n ⇐⇒ S-rank(A) = n Proof: Recall that for two matrices, W and A, rank(W ⊗ A) = rank(W ) × rank(A) (16) (17) From Corollary 2, we have rank(W ) = N for almost all numerical values, and for any full rank matrix An×n, we have max(rank(W ⊗ A)) = N × n, (18) Based on the definition of the S-rank, we can conclude that (W ⊗ A) is generically full rank for almost all choices of numerical values. This proves the necessity. On the other hand, if rank(A) < n for any choice of W , then we have max(rank(W ⊗ A)) < N × n, S-rank(W ⊗ A) < N × n. (19) (20) which implies that, This proves the sufficiency. B. State fusion We now explore Eq. (8) in NKE protocol and assume that there is no output fusion. In particular, we analyze the structure of the matrix W for (W ⊗ A, DC) observability according to Table I. First, we provide some special cases where the system matrix, A, is structurally full rank. This is the case, for August 21, 2018 DRAFT 16 example, in linearization and discretization of non-linear systems where the system matrix almost always has non-zero diagonal entries. Lemma 4: For full S-rank system matrix, (A, C) is centrally observable if and only if every parent SCC is output connected, i.e., monitored by (at least) one agent. The proof is straightforward and omitted here. Interested readers may see our previous work in [27]. The following theorem establishes conditions on the communication network, GW , over full S-rank systems. Theorem 2: With a full S-rank system, A, the pair (W ⊗ A, DC) is generically observable when for every parent-SCC in A, say K, if agent i does not have an observation of a state in K, then in the communication network, GW , there must be a directed path from agent i to any agent j, which has a state observation in2 K. Proof: The system matrix A being full S-rank ensures the condition (i) in Theorem 1. This is because from Corollary 1, there exists a disjoint union of cycles alone that cover all of the state vertices and the Y -topped paths are not needed to verify condition (i). To satisfy condition (ii), all state vertices in a subsystem associated to every agent, say i, must be a begin vertex of a Y -topped path. This condition, according to Lemma 4, is satisfied by having every parent-SCC in W ⊗ A be output-connected. Since in communication matrix W there is a path from agent i to j, in GW⊗A graph, subsystem of agent i is connected to subsystem of agent j. Therefore, every state vertex in parent-SCC K in subsystem i is connected to parent-SCC K in subsystem j (see Fig. 6). Since every state of parent-SCC K in subsystem j is Y -topped path to output j, every state of parent-SCC K in subsystem i is also connected to output j. With this for every parent-SCC K in every subsystem i of GW⊗A, all parent-SCCs are output connected and the theorem follows. For example, consider again the three-output system in Fig. 5. Having vertices {4, 5, 6} as parent-SCC, agent b is the Type-α agent. According to the above theorem any other agent without any observation in {4, 5, 6}, like agent a, must be connected to agent b. This provides a connection from parent- SCC {4, 5, 6} in subsystem a to its counterpart SCC in subsystem b in distributed system graph GW⊗A, and in turn its output connectivity. A very important point to mention here is that for full S-rank systems, there only exist Type-β and Type-γ agents. We prove this in the following lemma. 2If there is more than one agent observing SCC K, say agents j, k, a directed path from agent i to only one of them is sufficient. August 21, 2018 DRAFT 17 Fig. 6. This figure illustrates the proof of the Theorem 2 showing that: A directed path from agent i to agent j in W matrix (left) implies a directed path from states in subsystem i to subsystem j and consequently agent j in (W ⊗ A) matrix. Lemma 5: If a system matrix has full S-rank then we only have Type-β or Type-γ agents. Proof: The proof is straightforward and relies on Corollary 1. Since A is full S-rank, there exists a set of disjoint cycles that covers all of the state vertices. Hence, the set L introduced for agent classification in Section IV-B does not include any Y -topped paths and thus we cannot have any Type-α agent. However, when the system matrix is not full S-rank, we also encounter Type-α agents that possess crucial observations from Lemma 2. It turns out that if the system matrix is not full S-rank, then even using a fully-connected communication network (complete GW graph) does not recover observability. We now provide our main result on state fusion. Theorem 3: Assume that (A, Ci) is not observable at any agent i. If system, A, is not full S-rank, then the NKE (8) -- (9) is not observable with state fusion alone, i.e., (W ⊗ A, DC) is not observable for any choice of the communication matrix W . Proof: Let i be an agent for which condition (i) in Theorem 1 does not hold, i.e.,  A  < n. S-rank CT i Ci (21) Such an agent always exists because: (i) based on the assumption (ii) in Section III, the entire system is not observable at any agent; and (ii) the matrix A is not full S-rank. Now consider (W ⊗ A, DC) for the best-case scenario where GW is an all-to-all network. Let Wi be the ith column of the adjacency matrix W . Obviously, Wi ⊗ A is the ith block column of (W ⊗ A), and contains block matrices WjiA August 21, 2018 DRAFT 18 (22) (23) (24) for all j = 1, ..., N as wji (cid:54)= 0 and scalar multiplication does not change the structure and the S-rank (maximum possible rank over all values). Since A is not full S-rank, Wi ⊗ A has rank less than n as stacking matrices with the same structure on top of each other (see Fig.7-Left) does not improve the S-rank, which immediately results in Consequently, according to Fig.7-right, the structure of the matrix W ⊗ A is given as the side-by-side concatenation of the matrices Wi ⊗ A. Thus we have, CT i Ci  Wi ⊗ A  W ⊗ A  < n.  < N n. DC S-rank S-rank Fig. 7. The figures illustrates the structure of Wi ⊗ A (left) and matrix W ⊗ A (right) in the proof of Theorem 3. where Wji (cid:54)= 0 is the element in jth row and ith column of the full matrix W . It follows that  WjiA CT i Ci  < n, S-rank This holds for almost all choices of non-zero elements in the W matrices. Clearly, for any lower S- rank W the rank of W ⊗ A cannot be recovered as well. Therefore, according to Lemma 1, the condition (i) in Theorem 1 is violated. The above theorem shows that when A is S-rank deficient, then using state fusion cannot always guarantee the observability of the system and thus, the agents need access to more measurement data to August 21, 2018 DRAFT 19 recover observability, which is discussed next. In contrast, Theorem 2 shows that when A is full S-rank, then state fusion alone can recover observability and further provides a method for the required agent communication. Clearly, these two results are to be viewed as a direct consequence of Assumptions (ii) and (iii) in Section IV-A. To the best of our knowledge, these conditions have not been developed before. C. Output fusion The other solution to make the NKE observable is output fusion, that is the second update level given in the equation (9). According to the formulation, each agent shares its measurement with its direct neighbors and implements this as an innovation to update its prediction. According to Table I, for output fusion the structure of the matrix DC has to be determined such that I ⊗ A, DC is observable. Based on the definition of DC, the ith n × n diagonal block of DC contains all of the measurements in the extended neighborhood of agent i. In the distributed system graph G(I⊗A,DC ), say for the agent i, this is equivalent to adding all measurements in the neighborhood Ni to the subsystem i. We now provide our main result on output fusion. Theorem 4: The system (I ⊗ A, DC) is observable if and only if: (i) The sub-graph of all Type-α and Type-β agents is a complete graph, i.e. all these crucial agents are needed to be directly linked together; (ii) Every Type-γ agent is directly connected to all Type-α and Type-β agents. Proof: Sufficiency: With the given conditions (i) and (ii), each agent has access to all crucial measurements. This makes every agent generically observable. Necessity: If an agent is not connected to one of the crucial agents, then it is missing a crucial measurement and the statement follows. It can be verified that if a system is not (A, C) observable then even using a fully-connected communi- cation network does not recover observability. Clearly, the only way to get a stable estimation error is by increasing the number of state observations and recovering the (A, C) observability. An interesting result on how to recover (A, C) observability can be found in [32]. D. Main result Finally, we consolidate our results in previous subsections on state and output fusion. Theorem 2 sets the condition for state fusion for full S-rank systems, i.e., conditions for (W ⊗ A, DC) generic observability, while Theorem 3 states that for general S-rank deficient systems networked observability August 21, 2018 DRAFT 20 cannot be achieved via the state fusion alone. Output fusion, i.e., generic observability of (I ⊗ A, DC), is discussed in Theorem 4. Combining these results, we now provide the main theorem on generic observability of the single time-scale NKE protocol in Eqs. (8) -- (9). Theorem 5: For (W ⊗ A, DC) observability with minimal number of communications, each agent needs: 1) A direct link from all the Type-α agents (output-fusion); 2) A directed path to (at least) one Type-β agent for every parent SCC of A. This means, if there is two or more agents observing the same parent SCC, a directed path to any one of them is sufficient (state-fusion). Proof: The proof is a direct consequence of Theorems 2, 3, and 4. The following is a complementary remark to the Theorem 5. Remark 3: In the case of Type-β agents, every agent requires either a directed path to each Type-β agent (as stated in the Theorem 5) or a direct link from each Type-β agent (as stated in the Theorem 4); either one of these two conditions is sufficient for observability. However, the first strategy requires less number of links compared to the latter one, and therefore, it is preferred in terms of the minimal number of links in the communication network. Notice that compared to the typical assumptions on the agents' network in the literature, like strong connectivity or having a cyclic path, here we provide milder condition on the non-crucial agents; as there is no need for connectivity to these agents but from these agents. Furthermore, an agent may have no measurement of the system and still be able to estimate the state of the system via the proposed strategies. Such agents, for example, may play a role to provide and maintain the connectivity of the agent communication network [48], or even, maintaining directed paths to Type-β agents as stated in the second condition of the Theorem 5. VI. DESIGN OF LOCAL ESTIMATOR GAIN In this subsection, we consider the design of the estimator gain matrix, Kk. Notice that having (W ⊗ A, DC) observable guarantees a full gain matrix, Kk, to stabilize the NKE error. However, according to protocol (9), we need a local gain matrix, Kk, which is block-diagonal with N blocks of n× n matrices. For this section, we assume a constant estimator gain matrix is applied, i.e., the matrix Kk is independent of time, k, and denote it by K. A partial list of references devoted to find constrained estimator gain for control and estimation is [26], [49] -- [52]. Here, we use the Linear Matrix Inequality (LMI) approach in [50], [51]. However, in general, August 21, 2018 DRAFT 21 the corresponding LMIs do not have a solution, because of the structural constraints (block-diagonal) on the gain matrix, K. This is the main difficulty in distributed estimation and control as convex/semidefinite approaches are not directly applicable. To this end, we implement an iterative procedure to solve LMIs under structural constraints. In this regard, the following lemma presents the optimization approach to solve the estimator gain design problem. Interested readers may find more details in [26], [50]. Lemma 6: If the NKE protocol (8) -- (9) is observable, then estimator gain matrix, K, is the solution of the following optimization problem. min trace(XY ) subject to X, Y > 0,  X (cid:98)AT (cid:98)A Y  > 0,  X I I Y  > 0, K is block-diagonal. (cid:98)A = W ⊗ A − KDC(W ⊗ A) (25) (26) where, In fact, we need a block-diagonal K such that (cid:98)A is Schur (i.e. ρ((cid:98)A) < 1). Notice that, the solution to the second LMI is equivalent to X = Y −1, which gives the minimum trace and the optimal value as nN. The nonlinear product of X and Y can be replaced with a linear approximation [50], [51], [53], φlin(X, Y ) = trace(Y0X + X0S) and an iterative algorithm [51] can be used to minimize trace(XY ) under the given constraints. Algorithm 1 Iterative calculation of local gain estimator, K. 0: Find feasible points X0, Y 0, K. If no such points exist, Terminate. 0: At iteration t > 0 minimize trace(YtX + XtY ) under the constraints given in (25) and find X, Y, K. 0: If ρ((cid:98)A) < 1 terminate, otherwise set Yt+1 = Y, Xt+1 = X and run the step 2 for next iteration t = t + 1. It is shown in [51] that trace(YtX +XtY ) is a non-increasing sequence that converges to 2nN. In this regard, a stopping criterion in step 3 of the above algorithm can also be established in terms of reaching within 2nN + of the trace objective. The iterative procedure given above is centralized, however, the center has to implement this process only once, off-line; then it transmits the estimator gains to each August 21, 2018 DRAFT 22 agent and plays no further role in the implementation of local estimators; each agent, subsequently, observes and performs in-network operations to implement the estimator. A single time-scale algorithm can also be employed, where the above iterative procedure is implemented at the same time-scale k as of the dynamical system. With this approach, the estimator gain iterations, Kt, at each t is applied to the estimator at time-step, k, and may be transmitted to each agent at each step k. This is helpful when the implementation is assumed in real-time. VII. ILLUSTRATIVE EXAMPLE AND SIMULATION Consider the system, (A, C), given in Fig. 3. The structure of these matrices is given by 0 0 0 0 0 0 0 0  0 × 0 × 0 0 0 × 0 0   = 0 0 0 A =  Ca Cb Cc 0 0 0 0 0 × 0 × × 0 0 × 0 0 0 0 × × 0 × 0 × 0 0 0 0 0 0  0 0 × 0 0 0 0 0 0 0 0 0 0 × 0 0 0 0 0 × , 0 (27) (28)  . Now, recall that based on Theorem 1, the system is globally observable by collection of the three measurements. The state partitioning with maximal cycles so that this partitioning covers all the states (in light of condition (i) -- Theorem 1) is L = {(4, 5, 6, 4), (1, 2, 1), (7, 7), (3, a)}. By definition, agent a is Type-α, agent b is Type-β and agent c is Type-γ. It can further be verified that agents a and b possess crucial observations. To better illustrate the networking effect, we first note that the networked system with no information sharing, i.e., the graph associated to (I⊗A, Dc), is not observable at any of the agents individually (Fig. 5). To make the networked system observable at each agent, we propose the following communication matrices W1 and W2, and their associated graphs GW in Fig. 8 as two minimal networks making the system observable. August 21, 2018 DRAFT 23 (Left) communication network for the three agents using output fusion (Theorem 4), (Right) both measurement and Fig. 8. state fusion (Theorem 5).  × × 0 × × 0 × × ×  , W1 =  × 0 × × × 0 × 0 ×  . W2 = (29) The graph associated to W1 is proposed based on the Theorem 4. In this network crucial agents, {a, b}, are directly linked among each other and both have a directed link to agent, c, with no crucial observation. The second communication network W2 is based on Theorem 5; there is a direct link from agent a (Type- α) to all other agents, and there is a path from every other agent to agent b (Type-β). It can be verified that for both topologies (W ⊗ A, DC) is generically observable. Note that the solution for the network design problem is not unique, and there maybe other examples of communication network satisfying the conditions in the last section. In addition, any network including one of these two topologies as a sub-graph is also a solution to the NKE problem. For example, under the full structured rank assumption of system A, any strongly connected network among agents suffices for individual observability [26]. For simulation, we consider a random valued matrix, A, with the structure in Eq. (27), and an output matrix, Eq. (28), with all non-zero entries equal to 1. The system eigenvalues are as follows. eig(A) = [−1.0838, 1.0838, 0.6511,−0.5571, . . . −0.0940, 0.0000, 1.3072] Clearly, the dynamical system is unstable, since ρ(A) = 1.0838. We choose the agents' network according to Fig. 8 (Right) with random link weights such that it remains stochastic. We use Algorithm 1 to find the block-diagonal gain matrix, K. Using this gain matrix, the eigen-values of the error dynamics, i.e., August 21, 2018 DRAFT 24 Performance of the networked estimator at each agent. The error is squared and then summed over n = 7 states at Fig. 9. each iteration. of the matrix, (cid:98)A, are as follows, eig((cid:98)A) = [0.8190, 0.6511,−0.5571, 0.0073 ± 0.3159i, . . . −0.2700, 0.1643,−0.1406,−0.0940, −0.0788, 0.0214,−0.0237, 0, 0, 0, 0, 0, 0, 0, 0, 0], which are all stable. The system and output noise are, respectively, vk ∼ N (0, 0.05) and ri k ∼ N (0, 0.2). As system initial state we choose a random initial value between 0 and 3. The system error evolution over 100 iterations for three agents are given in Fig. 9. For the visual clarity, we have squared the errors at each iteration and then summed them over the n = 7 states of the dynamical system. As it can be seen, the estimation error at all agents is bounded despite the fact that system is not stable. VIII. CONCLUSION In this paper, we study the role of the agent communication network towards error stability of the NKE protocol (8) -- (9) in the context of single time-scale distributed estimation. As opposed to multiple time-scale strategies where the communication network is irrelevant and diffusion strategies where the estimator error is irrespective of system dynamics, here, we take into account both system dynamics and communication network. We show that the NKE is able to track even potentially unstable dynamical systems, i.e., the networked estimator is observable for all stable and unstable eigenvalues. We show that under sufficient communication among the agents, the system state is generically observable at August 21, 2018 DRAFT 0204060801000510152025Iterations,kSquarederrossummedoverallstates AgentaAgentbAgentc25 every agent. Here, we provide minimal sufficient network connectivity applicable for multi-agent systems subject to constrained communication, e.g., out-of-range geographical distances or costly communication. We define three types of agents/measurement where two types are crucial for observability. We provide two main results on recovering networked observability: (i) with state-fusion, and (ii) with output-fusion. Furthermore, we determine dynamical systems (S-rank deficient) for which no state-fusion results in an observable networked estimator and one has to rely on output-fusion as well. Our results are on the existence of a network structure for bounded estimation error and further finding such network with minimal links. Because of the genericity, the link weights are free parameters and results are independent of any particular fusion rule (e.g., Metropolis-Hastings [54]) chosen in (8) -- (9). Nevertheless, the structure of the underlying agent communication remains relevant and leads to network/infrastructure design. Furthermore, link weights can be optimized to reduce the error, which is a direction for the future work. It is worth noting that, in general, S-rank and other generic properties are easily verified. For example, there are efficient graph theoretic, [28], flow theoretic, [55], and linear programming, [56], methods that can be employed to check for generic properties. REFERENCES [1] D. Acemoglu, A. Nedic, and A. Ozdaglar, "Convergence of rule-of-thumb learning rules in social networks," in 47th IEEE Conference on Decision and Control, Dec. 2008, pp. 1714 -- 1720. [2] J. Cortes, "Distributed kriged kalman filter for spatial estimation," Automatic Control, IEEE Transactions on, vol. 54, no. 12, pp. 2816 -- 2827, dec. 2009. [3] H. Sayyaadi and M. Moarref, "A distributed algorithm for proportional task allocation in networks of mobile agents," Automatic Control, IEEE Transactions on, vol. 56, no. 2, pp. 405 -- 410, feb. 2011. [4] U. A. Khan and M. Doostmohammadian, "A sensor placement and network design paradigm for future smart grids," in 4th International Workshop on Computational Advances in Multi-Sensor Adaptive Processing, San Juan, Puerto Rico, Dec. 2011, pp. 137 -- 140. [5] Le Xie, Dae-Hyun Choi, and S. Kar, "Cooperative distributed state estimation: Local observability relaxed," in Power and Energy Society General Meeting, 2011 IEEE, july 2011, pp. 1 -- 11. [6] Jinghe Zhang, G. Welch, G. Bishop, and Zhenyu Huang, "Reduced measurement-space dynamic state estimation (remedyse) for power systems," in PowerTech, 2011 IEEE Trondheim, june 2011, pp. 1 -- 7. [7] R. Olfati-Saber and P. Jalalkamali, "Collaborative target tracking using distributed kalman filtering on mobile sensor networks," in 30th IEEE American Control Conference, San Francisco, CA, Jun. 2011. [8] H. Hashemipour, S. Roy, and A. Laub, "Decentralized structures for parallel Kalman filtering," IEEE Trans. on Automatic Control, vol. 33, no. 1, pp. 88 -- 94, Jan. 1988. [9] A. Mutambara, Decentralized estimation and control for multisensor systems, CRC Press, Boca Raton, FL, 1998. [10] F.S. Cattivelli and A.H. Sayed, "Diffusion strategies for distributed kalman filtering and smoothing," Automatic Control, IEEE Transactions on, vol. 55, no. 9, pp. 2069 -- 2084, sept. 2010. August 21, 2018 DRAFT 26 [11] F.S. Cattivelli and A.H. Sayed, "Distributed detection over adaptive networks using diffusion adaptation," Signal Processing, IEEE Transactions on, vol. 59, no. 5, pp. 1917 -- 1932, may 2011. [12] C.G. Lopes and A.H. Sayed, "Incremental adaptive strategies over distributed networks," Signal Processing, IEEE Transactions on, vol. 55, no. 8, pp. 4064 -- 4077, aug. 2007. [13] D. P Bertsekas, "Incremental gradient, subgradient, and proximal methods for convex optimization: A survey," Tech. Rep., MIT, 2010, LIDS Technical Report. [14] M. Farina, G. Ferrari-Trecate, and R. Scattolini, "Distributed moving horizon estimation for linear constrained systems," Automatic Control, IEEE Transactions on, vol. 55, no. 11, pp. 2462 -- 2475, nov. 2010. [15] A. Ribeiro, G.B. Giannakis, and S.I. Roumeliotis, "Soi-kf: Distributed kalman filtering with low-cost communications using the sign of innovations," Signal Processing, IEEE Transactions on, vol. 54, no. 12, pp. 4782 -- 4795, dec. 2006. [16] G. Battistelli, L. Chisci, S. Morrocchi, and F. Papi, "An information-theoretic approach to distributed state estimation," in 18th IFAC World Congress, 2011, number 1, pp. 12477 -- 12482. [17] Sheng-Yuan Tu and Ali H. Sayed, "Diffusion strategies outperform consensus strategies for distributed estimation over adaptive networks," May. 2012, available at http://arxiv.org/abs/0809.0009. [18] J. Chen and A. H. Sayed, "Diffusion adaptation strategies for distributed optimization and learning over networks," submitted for publication, available at http://arxiv.org/abs/0809.0009, Oct. 2011. [19] R. Olfati-Saber, "Distributed Kalman filters with embedded consensus filters," in 44th IEEE Conference on Decision and Control, Seville, Spain, Dec. 2005, pp. 8179 -- 8184. [20] R. Carli, A. Chiuso, L. Schenato, and S. Zampieri, "Distributed Kalman filtering using consensus strategies," in Proceedings of the 46th IEEE Conference on Decision and Control, 2007, pp. 5486 -- 5491. [21] Usman A. Khan and Jos´e M. F. Moura, "Distributing the Kalman filter for large-scale systems," IEEE Transactions on Signal Processing, vol. 56(1), no. 10, pp. 4919 -- 4935, Oct. 2008. [22] E. J. Msechu, S. D. Roumeliotis, A. Ribeiro, and G. B. Giannakis, "Decentralized quantized Kalman filtering with scalable communication cost," IEEE Transactions on Signal Processing, vol. 56, no. 8, pp. 3727 -- 3741. [23] U.A. Khan and A. Jadbabaie, "On the stability and optimality of distributed kalman filters with finite-time data fusion," in American Control Conference (ACC), 2011, 29 june- 1 july 2011, pp. 3405 -- 3410. [24] U. A. Khan, S. Kar, A. Jadbabaie, and J. M. F. Moura, "On connectivity, observability, and stability in distributed estimation," in 49th IEEE Conference on Decision and Control, 2010, pp. 6639 -- 6644. [25] S. Kar, J. M. F. Moura, and K. Ramanan, "Distributed parameter estimation in sensor networks: Nonlinear observation models and imperfect communication," submitted for publication, available at http://arxiv.org/abs/0809.0009, Aug. 2008. [26] U. A. Khan and A. Jadbabaie, "Coordinated networked estimation strategies using structured systems theory," in 49th IEEE Conference on Decision and Control, Orlando, FL, Dec. 2011, pp. 2112 -- 2117. [27] M. Doostmohammadian and U. A. Khan, "Communication strategies to ensure generic networked observability in multi- agent systems," in 45th Annual Asilomar Conference on Signals, Systems, and Computers, Pacific Grove, CA, Nov. 2011, pp. 1865 -- 1868. [28] K. J. Reinschke, Mutivariable control, a graph theoretic approach, Berlin: Springer, 1988. [29] Y.Y. Liu, J.J. Slotine, and A.L. Barabsi, "Controllability of complex networks," Nature, vol. 473, no. 7346, pp. 167 -- 173, May 2011. [30] T. Boukhboza, F. Hamelin, S. Martinez-Martinez, and D. Sauter, "Structural analysis of the partial state and input August 21, 2018 DRAFT 27 observability for structured linear systems: Application to distributed systems," European Journal of Control, vol. 15, no. 5, pp. 503 -- 516, Oct. 2009. [31] J.-M. Dion, C. Commault, and J. van der Woude, "Generic properties and control of linear structured systems: a survey," Automatica, vol. 39, pp. 1125 -- 1144, Mar. 2003. [32] T. Boukhobza and F. Hamelin, "State and input observability recovering by additional sensor implementation: A graph- theoretic approach," Automatica, vol. 45, no. 7, pp. 1737 -- 1742, 2009. [33] M. Ji and M. Egerstedt, "Observability and estimation in distributed sensor networks," in Decision and Control, 2007 46th IEEE Conference on, dec. 2007, pp. 4221 -- 4226. [34] M. Egerstedt, "Complex networks: Degrees of control," Nature, vol. 473, pp. 158 -- 159, May 2011. [35] S. Sundhar Ram, A. Nedi, and V.V. Veeravalli, "Distributed stochastic subgradient projection algorithms for convex optimization," Journal of Optimization Theory and Applications, vol. 147, pp. 516 -- 545, 2010. [36] P. Bianchi, G. Fort, W. Hachem, and J. Jakubowicz, "Convergence of a distributed parameter estimator for sensor networks with local averaging of the estimates," in Acoustics, Speech and Signal Processing (ICASSP), 2011 IEEE International Conference on, may 2011, pp. 3764 -- 3767. [37] I.D. Schizas, G. Mateos, and G.B. Giannakis, "Distributed lms for consensus-based in-network adaptive processing," Signal Processing, IEEE Transactions on, vol. 57, no. 6, pp. 2365 -- 2382, june 2009. [38] W. Abbas and M. Egerstedt, "Hierarchical assembly of leader-asymmetric, single-leader networks," in American Control Conference (ACC) 2011, San Francisco, CA, June 29-July 01 2011. [39] R. Kalman and R.Bucy, "New results in linear filtering and prediction theory," ASME Journal of Basic Engineering, vol. 83, pp. 95 -- 108, 1961. [40] M.L.J. Hautus, "Controllability and observability conditions of linear autonomous systems," Ned. Akad. Wetenschappen, vol. Ser. A, no. 72, pp. 443 -- 448, 1969. [41] T.H. Cormen, C.E. Leiserson, R.L. Rivest, and C. Stein, Introduction to Algorithms, MIT Press, 209. [42] E. J. Davison and S. H. Wang, "Properties of linear time-invariant multivariable systems subject to arbitrary output and state feedback," IEEE Transactions on Automatic Control, vol. 18, no. 1, pp. 24 -- 32, Feb. 1973. [43] E. J. Davison and S. H. Wang, "Properties and calculation of transmission zeros of linear multivariable systems," Automatica, vol. 10, pp. 643 -- 658, 1974. [44] F. Harary, "The determinant of the adjacency matrix of a graph," SIAM Review, vol. 4, no. 3, pp. 202 -- 210, Jul. 1962. [45] C. Lin, "Structural controllability," Automatic Control, IEEE Transactions on, vol. 19, no. 3, pp. 201 -- 208, jun 1974. [46] S. Sundaram and C.N. Hadjicostis, "Distributed function calculation and consensus using linear iterative strategies," Selected Areas in Communications, IEEE Journal on, vol. 26, no. 4, pp. 650 -- 660, may 2008. [47] V.D. Blondel, J.M. Hendrickx, A. Olshevsky, and J.N. Tsitsiklis, "Convergence in multiagent coordination, consensus, and flocking," in Decision and Control, 2005 and 2005 European Control Conference. CDC-ECC '05. 44th IEEE Conference on, dec. 2005, pp. 2996 -- 3000. [48] A. N. Kopeikin, S. S. Ponda, L. B. Johnson, O. Toupet, and J. P. How, "Real-time dynamic planning to maintain network connectivity in a team of heterogeneous unmanned systems," in Wi-UAV 2011, 2nd International Workshop on Wireless Networking for Unmanned Autonomous Vehicles at the IEEE GlobeComm Conference, Dec 2011. [49] Khan U. A. and A. Jadbabaie, "Networked estimation under information constraints," submitted, available at http://arxiv.org/abs/1111.4580. August 21, 2018 DRAFT 28 [50] M. Pajic, S. Sundaram, J. Le Ny, G.J. Pappas, and R. Mangharam, "The wireless control network: Synthesis and robustness," in Decision and Control (CDC), 2010 49th IEEE Conference on, 2010, pp. 7576 -- 7581. [51] L. El Ghaoui, F. Oustry, and M. Ait Rami, "A cone complementarity linearization algorithm for static output-feedback and related problems," IEEE Trans. on Automatic Control, vol. 42, no. 8, pp. 1171 -- 1176, Aug. 1997. [52] A.I. Zecevic and D.D. Siljak, "Control design with arbitrary information structure constraints," Automatica, vol. 44, no. 10, pp. 2642 -- 2647, 2008. [53] O. L. Mangasarian and Jong-Shi Pang, "The extended linear complementarity problem," SIAM Journal on Matrix Analysis and Applications, vol. 2, pp. 359 -- 368, Jan. 1995. [54] Lin Xiao, Stephen Boyd, and Seung jean Kim, "Distributed average consensus with least-mean-square deviation," Journal of Parallel and Distributed Computing, vol. 67, pp. 33 -- 46, 2005. [55] V. Hovelaque, C. Commault, and J.-M. Dion, "Analysis of linear structured systems using a primal-dual algorithm," Syst. Control Lett., vol. 27, pp. 73 -- 85, February 1996. [56] S. Poljak, "Maximum rank of powers of a matrix of a given pattern," Proceedings of the American Mathematical Society, vol. 106, no. 4, pp. pp. 1137 -- 1144, 1989. August 21, 2018 DRAFT
1503.06124	1	1503	2015-03-20T15:56:17	A Multi-Agent System of Project Bidding Management Simulation	[ "cs.MA" ]	This paper presents a simulation model based on the general framework of Multi-Agent System (MAS) that can be used to investigate construction project bidding process. Specifically, it can be used to investigate different strategies in project bidding management from the general contractors' perspective. The effectiveness of the studied management strategies is evaluated by the quality, time and cost of bidding activities. As an implementation of MAS theory, this work is expected to test the suitability of MAS in studying construction management related problems.	cs.MA	cs	A Multi-Agent System of Project Bidding Management Simulation R. Liu ABSTRACT This paper presents a simulation model based on the general framework of Multi-Agent System (MAS) that can be used to investigate construction project bidding process. Specifically, it can be used to investigate different strategies in project bidding management from the general contractors’ perspective. The effectiveness of the studied management strategies is evaluated by the quality, time and cost of bidding activities. As an implementation of MAS theory, this work is expected to test the suitability of MAS in studying construction management related problems. INTRODUCTION Project bidding is a multi-disciplinary and multi-organizational process that requires the efforts of different project functional units (Kerzner 2009). Unlike intra-team activities, project bidding happens in a cross-functional environment where a formal boundary between responsibilities is set and leads to diverse institutional arrangements (Thomsen et al. 2005). For example, in an EPC (Engineering, Procurement and Construction) project, ideally the estimating team and the engineering team work closely together on developing a proposal; but in reality, the two teams have distinct responsibilities, focuses and procedures. This often results in additional work such as coordination, and without sufficient management, rework is almost inevitable. Another difficulty is the bid/no bid decision. Some scholars have applied machine learning approaches (Du and El-Gafy 2011), statistical modeling, building information modeling (BIM)(Du et al. 2014; Liu et al. 2014; Liu et al. 2014) or Monte Carlo simulation (Du et al. 2014) to support a better decision, but the bid/no bod decision remain a challenge for the construction managers. A root cause of the inefficiency in bidding management is the lack of understanding about the proper management strategies (Du and El-Gafy 2014). One example is the goal incongruence (Du and El-Gafy 2014): the estimating team may make the economy of the proposed design its first priority, while for the engineering team, robustness is more important. Such difference in perception may lead to completely different practices. In order to improve the efficiency and effectiveness of bidding process, it is critical for general contractors to understand the consequences of different management strategies and approaches. It involves the optimization of number of target projects, the job assignment strategies and meetings. Many existing efforts concentrate on only one aspect of human behaviors pertaining to bidding management, assuming that a deeper investigation on a single aspect will lead to better discovery. The rationale of focusing on one important point is well recognized by this study, especially given the difficulties of conceptualizing human behaviors and validating assumptions. Nonetheless, the importance of addressing as many relevant behaviors as possible in the same investigation should not be intentionally overlooked, when the interactions among diverse behaviors play a critical role in understanding how goals are formed and affected and how goal incongruence influences the efficacy and quality of proposal development (Perrow 1986). This paper introduces a simulation model based on Multi-Agent System (MAS) framework, to investigate the implications of management strategies in the bidding management of a small construction project. Worker behaviors pertaining to bidding were captured and investigated to quantify the impacts of different management strategies on the performance of bidding management of a general contractor. LITERATURE REVIEW As a computational modeling approach, Agent Based Modeling (ABM) is a suitable tool for use in social research to study human and organizational issues in a diversity of areas (Du 2012; Du and El-Gafy 2010; Du and Wang 2011). It is a computational method that builds a common environment for heterogeneous and autonomous agents to share, and allows the agents to simultaneously interact with each other for self-interest (Du and El-Gafy 2014; Ligmann-Zielinska and Jankowski 2007). Unlike top-down modeling approaches (e.g., System Dynamics, Discrete Event Simulation etc.), in ABM the collective behavior of the simulated system is not predefined, but emerges from individual agents who act based on what they perceive to be their own interests. Thus, ABM is capable of reproducing the emergent properties of the studied systems (Macal and North 2007). As for the application of ABM in construction engineering and management, recently Du and colleagues have performed a series of representative works (Du 2014; Du and Bormann 2014; Du and El-Gafy 2010; Du and El-Gafy 2012; Du and El-Gafy 2014; Du and El-Gafy 2014; Du et al. 2012; Du et al. 2014; Du and Wang 2011). In one of their works, they developed a comprehensive ABM model called “Virtual Organizational Imitation for Construction Enterprises” or “VOICE”. In the VOICE model, they creatively captured 13 common behaviors in construction management settings, and simulated them under the MAS framework. Unlike other similar works, in the VOICE model, a comprehensive list of work related behaviors are modeled as separate behavioral modules. It suggests a better capture of the sociotechnical process of construction management. Given the features of VOICE, this study mainly builds its simulation experiments on VOICE model. THE MODEL ARCHITECTURE In order to utilize the VOICE framework to investigate problems discussed above, the following basic assumptions were made: First, there are three major agents including president, who is responsible for the overall management and decision-making of bidding; managers who are responsible for information gathering and expectation handling; and helpers who are responsible for processing routine tasks. Figure 1 illustrates the MAS used in the modeling and simulation. Fig.1 MAS architecture of the proposed model Second, the decisions made by the agents may trigger a variety of individual behavioral responses modeled with the behaviors in the VOICE framework. Typical behaviors include routine activities (e.g., processing tasks), communication, and coordination (e.g., assigning tasks). However, when overloaded, reciprocal activities may also be triggered, such as complaining about the overload. These nonproductive activities create inefficiency and affect the capacity of the estimating team. Third, although under the VOICE framework, task characteristics and organizational context can also affect the cooperative behaviors of team members, they will not be considered in this case study because they are less dynamic in Company D, compared to the four issues addressed by the principals. Therefore, the simulation experiments only focused on the controllable variables for a realistic recommendation. Based on the basic assumptions, the proposed modeling framework is shown in Figure 2. Fig. 2 Modeling bidding management with proposed theoretical framework CASE STUDY A case study was performed to investigate how behaviors and the institutional arrangement between members of a single project team affect management actions and team performance in typical Design Bid Build (DBB) projects. The studied case is a small general contractor focused on small commercial projects. Most jobs are Design Bid Build (DBB). The cost estimation is conducted by a single team: three managers work on separate sections/crafts of the project and all report to the principals for the final estimating and bidding decision. In the simulation experiments, it is of particular interest to test:  The influence of task dependency: Among all the task-related factors, dependence among tasks is considered to be most correlated to the level of cooperation (Deutsch 1949; Pinto et al. 1993; Thompson 2003). Task dependence refers to the extent to which team members are dependent on each other to perform individual tasks (Van de Ven et al. 1976). The original work about task dependence can be dated back to Thompson (1967), who grouped task dependence into three types -- pooled, sequential and reciprocal -- with reciprocal dependence at their highest intensity of interaction. Regarding construction as a complex system (Bertelsen 2003), reciprocal task dependence is probably the most common dependence in construction project teams (Thompson 2003). Because reciprocal task dependence means the highest level of interaction intensity (Thompson 1967), intense coordination work is required to adjust the efforts of different actors (Levitt 2007). Building upon Thompson, it was induced that the hierarchy of increasing levels of task dependence between unit personnel can be determined by observing whether the work flow is (1) independent, (2) sequential, (3) reciprocal, or (4) in a team arrangement (Van de Ven et al. 1976). Yilmaz and Hunt (2001) proposed measuring task dependence by the information need of tasks, i.e., whether additional information is needed to perform a particular task. Following the previous work, this research describes task dependence in construction project teams as the workflow relationship between team members, which can be demonstrated by network techniques, such as activity on node (AON).  The influence of goal congruence. In the bidding management process goal congruence plays a vital role in this process, which is demonstrated in the difference of the perceptions of behavioral standards and ranking of management criteria (Thomsen et al. 2005). Goal congruence can affect the quality and amount of the appropriate information contributed by the designers because a higher magnitude of goal congruence is anticipated to enhance the understanding among team members (Witt 1998). Thomsen et al. (2005) model goal congruence as a percentage, with 100% being the most congruent condition and 0% being the least. This case study uses the same definition and assumes a linear relationship between goal congruence and information quality/amount exchanged between an engineer and a project proposal team member. Simulation results 3,300 simulations were conducted to examine the influence of task dependence and goal congruence on the performance of the bidding team. The following figure demonstrates the results. Efficiency Effectiveness Quality C G 1 . 0 2 . 0 3 . 0 5 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0 5 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0 5 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 14% 12% 10% 8% 6% 4% 2% 0% 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 14% 12% 10% 8% 6% 4% 2% 0% 14% 12% 10% 8% 6% 4% 2% 0% 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 5 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0 5 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0 5 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0 5 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0 4 . 0 5 . 0 6 . 0 7 . 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 14% 12% 10% 8% 6% 4% 2% 0% 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 14% 12% 10% 8% 6% 4% 2% 0% 14% 12% 10% 8% 6% 4% 2% 0% 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 5 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0 5 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0 5 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0 5 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0 8 . 0 9 . 0 0 . 1 l l A 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 14% 12% 10% 8% 6% 4% 2% 0% 14% 12% 10% 8% 6% 4% 2% 0% 14% 12% 10% 8% 6% 4% 2% 0% 14% 12% 10% 8% 6% 4% 2% 0% 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Fig. 3 Influences of task dependence and goal congruence on performance of proposal development DISCUSSION AND CONCLUSIONS The case study made two major findings. First, the simulation finds goal congruence to be an influential factor for team productivity, but negligible to the work quality and work pressure of the project team. First, a higher level of goal congruence between the proposal team and engineering team significantly improves the efficiency and effectiveness of proposal development. A likely interpretation is that enhanced goal congruence improves the mutual understanding of objectives, definitions and needs between two teams, and encourages proactive participation of the engineers in proposal development. This in turn reduces the need for additional coordination, and increases the quality of each information exchange between engineers and proposal team members. Second, task dependence can significantly affect the productivity and work quality of the project team; it is able to alter the effects of goal congruence and micro-management. Task dependence is a crucial factor for understanding inter-team cooperation in proposal development. On the one hand, the simulation results find task dependence to be a significant predictor of efficiency, effectiveness, and quality. If tasks are more dependent, the team is less productive and commits more mistakes. This is understandable from an empirical perspective, since dependence often means additional efforts for communication and coordination, a bigger chance of mistakes and conflicts. On the other hand, task dependence may affect the effects of goal congruence and micro-management. Simulation results found that the efficiency difference between levels of goal congruence becomes bigger when tasks are more dependent. In contrast, the effects of micro- management are more significant when tasks are more independent. This finding highlights task dependence to be a vital point of decision making in project team management, especially when managerial and/or behavioral changes are planned. REFERENCES Bertelsen, S. "Construction as a complex system." 11-23. Deutsch, M. (1949). "A theory of co-operation and competition." Human Relations, 2(2), 129. Du, J. (2012). "Investigation of Interpersonal Cooperation in Construction Project Teams: An Agent- Based Modeling Approach." PhD Disseration, Michigan State University, East Lansing, MI. Du, J. (2014). "The “weight” of models and complexity." Complexity. doi: 10.1002/cplx.21612 Du, J., and Bormann, J. (2014). "Improved Similarity Measure in Case Based Reasoning with Global Sensitivity Analysis: An Example of Construction Quantity Estimating." Journal of Computing in Civil Engineering, 28(6), 04014020. Du, J., and El-Gafy, M. "Virtual Organizational Imitation for Construction Enterprises (VOICE): Managing business complexity using Agent Based Modeling." Proc., Construction Research Congress 2010. Du, J., and El-Gafy, M. (2011). "Feasibility Analytical Mapping (FAM) for the Bidding Decision: A Graphic Bidding Decision Making Model Based on Multidimensional Scaling and Discriminant Analysis." International Journal of Construction Education and Research, 7(3), 198-209. Du, J., and El-Gafy, M. (2012). "Virtual Organizational Imitation for Construction Enterprises: Agent- Based Simulation Framework for Exploring Human and Organizational Implications in Construction Management." Journal of Computing in Civil Engineering, 26(3), 282-297. Du, J., and El-Gafy, M. "Modeling organizational behaviors of construction enterpises: an agent based modeling appraoch." Proc., Proceedings of the 2014 Winter Simulation Conference, IEEE Press, 3341- 3362. Du, J., and El-Gafy, M. (2014). "Using Agent-Based Modeling to Investigate Goal Incongruence Issues in Proposal Development: Case Study of an EPC Project." ASCE Journal of Management in Engineering., 10.1061/(ASCE)ME.1943-5479.0000343 , 05014025. Du, J., El-Gafy, M., and Ghanem, A. (2012). "Adding Value to Change Order Management Process Using Simulation Approach: A Case Study." Proc., 48th Associated Schools of Construction (ASC) Annual International Conference, Associated Schools of Construction (ASC), Birmingham, UK. Du, J., Liu, R., and Issa, R. R. (2014). "BIM Cloud Score: Benchmarking BIM Performance." Journal of Construction Engineering and Management, 140(11). Du, J., Liu, R., and Karasulu, Y. "Cloud Based Interactive Probabilistic Simulation for AEC Industry." Proc., ICCCBE 2014, ASCE. Du, J., and Wang, Q. (2011). "Exploring Reciprocal Influence between Individual Shopping Travel and Urban Form: Agent-Based Modeling Approach." Journal of Urban Planning and Development, 137(4), 390-401. Kerzner, H. (2009). Project management: a systems approach to planning, scheduling, and controlling, Wiley. Levitt, R. E. (2007). "The Virtual Design Team (VDT): A Computational Model of Project Teams." <http://cee.stanford.edu/programs/construction/what/pdfs/VDT_Overview_0711.pdf>. (2013). Ligmann-Zielinska, A., and Jankowski, P. (2007). "Agent-based models as laboratories for spatially explicit planning policies." Environment and Planning B: Planning and Design, 34(2), 316-335. Liu, R., Du, J., and Issa, R. R. (2014). "Cloud-based deep immersive game for human egress data collection: a framework." Journal of Information Technology in Construction, 336-349. Liu, R., Du, J., and Issa, R. R. "Human Library for Emergency Evacuation in BIM-based Serious Game Environment." Proc., Proceedings ICCBE/ASCE/CIBW078 2014 International Conference on Computing in Civil and Building Engineering. Macal, C., and North, M. "Agent-based modeling and simulation: desktop ABMS." IEEE Press Piscataway, NJ, USA, 95-106. Perrow, C. (1986). Complex organizations, McGraw-Hill New York. Pinto, M., Pinto, J., and Prescott, J. (1993). "Antecedents and consequences of project team cross- functional cooperation." Management Science, 1281-1297. Thompson, J. (2003). Organizations in action: Social science bases of administrative theory, Transaction Pub. Thompson, J. D. (1967). Organizations in Action, McGraw-Hill, New York. Thomsen, J., Levitt, R. E., and Nass, C. I. (2005). "The Virtual Team Alliance (VTA): Extending Galbraith’s Information-Processing Model to Account for Goal Incongruency." Computational & Mathematical Organization Theory, 10(4), 349-372. Van de Ven, A. H., Delbecq, A. L., and Koenig Jr, R. (1976). "Determinants of coordination modes within organizations." American sociological review, 41(2), 322-338. Witt, L. (1998). "Enhancing organizational goal congruence: A solution to organizational politics." Journal of Applied Psychology, 83(4), 666. Yilmaz, C., and Hunt, S. (2001). "Salesperson cooperation: The influence of relational, task, organizational, and personal factors." Journal of the Academy of Marketing Science, 29(4), 335.
1102.3341	1	1102	2011-02-16T14:12:00	Reasoning about Social Choice Functions	[ "cs.MA" ]	We introduce a logic specifically designed to support reasoning about social choice functions. The logic includes operators to capture strategic ability, and operators to capture agent preferences. We establish a correspondence between formulae in the logic and properties of social choice functions, and show that the logic is expressively complete with respect to social choice functions, i.e., that every social choice function can be characterised as a formula of the logic. We prove that the logic is decidable, and give a complete axiomatization. To demonstrate the value of the logic, we show in particular how it can be applied to the problem of determining whether a social choice function is strategy-proof.	cs.MA	cs	Reasoning about Social Choice Functions Nicolas Troquard Wiebe van der Hoek Michael Wooldridge Computer Science Department, University of Liverpool, UK Abstract We introduce a logic specifically designed to support reasoning about social choice functions. The logic includes operators to capture strategic ability, and operators to capture agent preferences. We establish a correspondence between formulae in the logic and properties of social choice functions, and show that the logic is expressively complete with respect to social choice functions, i.e., that ev- ery social choice function can be characterised as a formula of the logic. We prove that the logic is decidable, and give a complete axiomatization. To demonstrate the value of the logic, we show in particular how it can be applied to the problem of determining whether a social choice function is strategy-proof. 1 Introduction Social choice theory is concerned with collective decision making in situations where the preferences of the decision makers may differ [3]. Social choice theorists have developed a range of procedures, such as voting protocols, to support such collective decision making, and have developed a range of criteria with which to characterise the properties of such procedures. Such criteria are usually expressed axiomatically, and a major concern of social choice theory is to study the extent to which decision making procedures do or do not satisfy these axioms [8, 2, 7, 12]. In short, the aim of the present paper is to develop a logic that is explicitly intended for reasoning about social choice procedures. We focus on social choice functions, a class of social choice procedures that select a single social outcome as a function of individual preferences. Voting procedures of the type used in political elections throughout the democratic world are perhaps the best-known examples of social choice functions. A voting procedure determines the winner of an election as a function of the votes cast; votes can be understood as an expression of voter preferences. One interesting issue that arises in voting procedures is the extent to which voters are incentivised to truthfully report their preferences when voting. For example, sup- pose we have two voters, 1 and 2, who vote among three candidates, x, y, and z for a role that is currently filled by x. The voting procedure used in this example says that, if there is a unanimously preferred candidate, then that will be chosen, otherwise the candidate x remains. Suppose the true preferences of 1 are given by z <1 x <1 y and those of 2 are x <2 y <2 z. If the social choice function was presented with these true preferences, candidate x would be chosen (since there is no consensus). However, if 1 2 z <(cid:48) voter 2 would instead claim his preferences were x <(cid:48) 2 y while 1 revealed its true preferences, then 2 would be better off, since y would be chosen, rather than x, and agent 2 prefers y over x. This issue suggests the following problem: Can we design a voting procedure that is "immune" to such misrepresentation, i.e., in which a voter can never do any better than by truthfully reporting its preferences? The term strategy proof is used to refer to such voting procedures. In fact, fundamental results in social choice theory tell us that there are severe limits to the development of strategy-proof voting procedures [7, 12], and for this reason, developing and analysing social choice procedures is a lively and highly active research area. The long-term aim of our work is to develop formal tools to assist in the analysis and design of social choice procedures. In particular, we hope to develop techniques that will permit the automated analysis of social choice procedures. To this end, we aim to develop logics that allow us to formally express the properties of social choice procedures, such that these languages may be automatically processed. Our view is that logic can provide a powerful tool for the analysis of social choice procedures [11, 16]. Such logics can be used as query languages for social choice procedures: given some property P of a social choice procedure, we aim to be able to encode the property P as an expression ρP of our language, which we then pose as a query to an automated analysis system. Working towards the long-term goal, the present paper presents a logic for reasoning about social choice procedures, and in particular, for analysing strategy proofness. The remainder of the paper is structured as follows. In Section 2 we recall the main concepts from game theory and social choice theory that we use throughout the paper. We then introduce our logic in Section 3. The logic is basically a modal logic [5], which derives inspiration from the Coalition Logic of Propositional Control (CL-PC) [15]. The latter logic includes operators to capture strategic ability. We extend this with operators for capturing agent preferences. The basic idea is to model an agent's preferences via atomic propositions: a proposition pi x>y will be used to represent the fact that agent i has reported that he prefers outcome x at least as much as outcome y. The strategic abilities of agents are captured using a CL-PC-like operator: an agent can choose any assignment of values for its preference variables that corresponds to a preference ordering. After presenting the syntax and semantics of the logic, we show how the logic can be used to characterise social choice functions, and show that the logic is expressively complete with respect to social choice functions, i.e., that every social choice function can be characterised as a formula of the logic. We give a com- plete axiomatization for the logic. To demonstrate the value of the logic, in Section 4 we formalise some properties of social choice functions and in particular, we show how it can be applied to the problem of determining whether a social choice function is strategy-proof. We conclude in Section 5. 2 Background In this section, we present the basic definitions of game theory and social choice upon which we construct our framework [6, 10]. We begin with some notation. We assume that game forms and social choice func- 2 tions (to be defined hereafter) share the same domains of agents and outcomes. We denote by N = {1, . . . , n} the finite set of agents (or players) and by K the finite set of social outcomes (outcomes hereafter). We use the letters a, b, c, . . . as constants of K. We use variables i, j, . . . to denote agents, and outcomes will be denoted by the variables x, y, z, . . .. Typically, one can consider that the agents are the voters and the outcomes are the candidates in some election. We denote by L(K) the set of linear orders over K. (A linear order here is a relation that is reflexive, transitive, antisymmetric and total.) By using a linear order, we are as- suming the players cannot be indifferent between two distinct outcomes. A preference relation is a linear order of outcomes. Given K and N, a preference profile < is a tuple (<i)i∈N of preferences, where <i ∈ L(K) for every i. The set of preference profiles is denoted by L(K)N. Note that we use the symbol <i for a preference relation for agents, which in this case happens to be reflexive (and we do not write ≤i for it). Also, we will use the symbol >i with the obvious meaning, i.e., y >i x iff x <i y. Definition 1 (Social choice function) Given K and N, a social choice function (SCF) is a single-valued mapping from the set L(K)N of preference profiles into the set K of outcomes. For every preference profile, a social choice function describes the desirable outcome (from the point of view of the designer). Definition 2 (Strategic game form) Given the sets N and K, a strategic game form is a tuple (cid:104)N, (Ai), K, o(cid:105) where: Ai is a finite nonempty set of actions (or strategies) for each player i ∈ N; o : ×i∈NAi → K assigns an outcome for every combination of actions. A strategic game form is sometimes called a mechanism. It specifies the agents taking part in the game, their available actions, and what outcome results from each combina- tion of actions. We refer to a collection (ai)i∈N, consisting of one action for every agent in N, as an action profile. Given an action profile a, we denote by ai the action of the player i. Remark 1 There is a direct link between strategic game forms and social choice func- tions. Any social choice function can be viewed as a game form in which the set of actions of every agent is L(K) (think of this as the preference profiles the agent can claim to be his), and the function o represents the social choice function (see [9]). For any SCF F, we denote its associated game form by gF. A strategic game is basically the composition of a strategic game form with a col- lection of preference relations (one for every agent) over the set of outcomes. Definition 3 (Strategic game) A strategic game is a tuple (cid:104)N, (Ai), K, o, (<i)(cid:105) where (cid:104)N, (Ai), K, o(cid:105) is a strategic game form, and for each player i ∈ N, <i is a preference relation over K. 3 In our context, when the actions Ai in a game (cid:104)N, (Ai), K, o, (<i)(cid:105) are preference relations themselves, one should think of those as preferences that i can choose to report, whereas <i, encodes i's real preferences. A solution concept defines for every game a set of action profiles -- intuitively, those that may be played through rational action. Exactly which solution concept is used depends upon the application at hand: we will soon introduce a well-celebrated solution concept of Nash Equilibrium (see Example 1). Definition 4 (Solution concept) A solution concept SC is a function that maps a strate- gic game form (cid:104)N, (Ai), K, o(cid:105) and a preference profile over K to a subset of the action profiles. We now introduce a simple but fundamentally important solution concept: Nash equi- librium. Definition 5 (Nash equilibrium) Given a strategic game form g = (cid:104)N, (Ai), K, o(cid:105) and a preference profile < over K the set of Nash equilibria NE(g, <) is given as the set of action profiles in g such that no player would benefit from deviating unilaterally from his current action. More formally, (a1, . . . an) ∈ NE(g, <) iff for every player k and every a(cid:48) k ∈ Ak, we have o(a1, . . . a(cid:48) k . . . an) <k o(a1, . . . ak . . . an). We can now introduce the notions of implementation and truthful implementation. The problem of implementation arises because a planner does not know the true prefer- ence profile of the players. Given a social choice function F involving a set of players N and a set of outcomes K, the planner only knows that every player i ∈ N has some preference <i, an element of L(K). We first define the case of (standard) implementation. Assuming a pattern of be- haviour -- a solution concept SC -- the role of the planner is then to design a mechanism (or game form) g such that for every possible preference profile < ∈ L(K)N, the strate- gic game (cid:104)g, <(cid:105) admits at least one SC-equilibrium, and every SC-equilibrium leads to the outcome in K which is prescribed by the social choice function for the preference profile at hand, that is, the value of F(<). Definition 6 (Implementation) Given a solution concept SC, we say that the game form g = (cid:104)N, (Ai), K, o(cid:105) SC-implements the social choice function F if for every prefer- ence profile < ∈ L(K)N we have that SC(g, <) (cid:44) ∅ and a∗ ∈ SC(g, <) implies that o(a∗) = F(<) In words: the game form g SC-implements F if for any game form (cid:104)g, <(cid:105) based on g, any outcome associated to a strategy profile in the solution concept SC is the same as what the social choice function would yield for the preference <. Or, more loosely: the game form g implements F if, for every preference profile < that we can associate with it, the outcomes in the game (cid:104)g, <(cid:105) and the result of F(<) agree at least on those claimed preferences that are in the solution concept of the game. The problem of implementation is illustrated in Figure 1. We say that the social choice function is SC-implementable if there is a game form that SC-implements it. 4 <(cid:48) 2 <2 y x <1 <(cid:48) 1 F a2 x z a1 a(cid:48) 1 a(cid:48) 2 y SC x G = (cid:104)g, (<1, <2)(cid:105) a2 x z a1 a(cid:48) 1 a(cid:48) 2 y x SC G(cid:48) = (cid:104)g, (<(cid:48) 2)(cid:105) 1, <(cid:48) Figure 1: Implementation. The preference profiles < and <(cid:48) are two arbitrary members of L(K)N. The left part represents the SCF F. F(<1, <2) = y and F(<(cid:48) 2) = x. The right part represents the strategic game form g instantiated, in the upper part with the preference profile (<1, <2) (game G = (cid:104)g, (<1, <2)(cid:105)) and in the lower part with the 2)(cid:105)). All the SC-equilibria of G (and preference profile (<(cid:48) possibly also some others than (a(cid:48) 1, a(cid:48) 2)) lead to F(<1, <2). In a like manner, all the SC- equilibria of G(cid:48) lead to F(<(cid:48) 1, <(cid:48) 2). This has to be verified for every preference profile in L(K)N and not only < and <(cid:48): if it holds, g is said to SC-implement F. 2) (game G(cid:48) = (cid:104)g, (<(cid:48) 1, <(cid:48) 1, <(cid:48) 1, <(cid:48) In some situations however, an SCF can be implemented by a strategic game form of which the space of action profiles corresponds to the space of preference profiles, and telling the truth is an equilibrium. We call a strategic game form in which the set of strategies of a player i is the set of preferences over K a direct mechanism. Hence, each player is asked to report a preference, but not necessarily the true one. An appealing class of direct mechanisms is that in which reporting the true preference profile is an equilibrium of the game consisting of the direct mechanism composed with the true preference profile. That is, for every <∈ L(K)N, the action profile where every player i reports its true preference <i is an equilibrium of the game (cid:104)g, <(cid:105). We can define this notion for every solution concept SC. Definition 7 (Truthful implementation) The direct mechanism g = (cid:104)N, (Ai), K, o(cid:105) truthfully SC-implements the SCF F if for every true preference profile < and reported 5 <(cid:48) 2 <2 SC y x <1 <(cid:48) 1 <(cid:48) 2 <2 y x SC <1 <(cid:48) 1 G = (cid:104)gF, (<1, <2)(cid:105) G(cid:48) = (cid:104)gF, (<(cid:48) 2)(cid:105) 1, <(cid:48) Figure 2: Truthful implementation. The preference profiles < and <(cid:48) are two arbitrary members of L(K)N. The left part represents the game form gF associated to the SCF F when the preferences of the two players are <1 and <2. The game G = (cid:104)gF, (<1, <2)(cid:105) admits an equilibrium at the action profile (<1, <2). The right part represents gF when 2)(cid:105) admits the preferences of the two players are <(cid:48) 1 and <(cid:48) an equilibrium at the action profile (<(cid:48) 1, <(cid:48) 2). This has to be verified for every preference profile in L(K)N and not only < and <(cid:48): if it holds, gF is said to truthfully SC-implement F. 2. The game G(cid:48) = (cid:104)gF, (<(cid:48) 1, <(cid:48) profile a∗ with a∗ i =<i for every i: a∗ ∈ SC(g, <), and o(a∗) = F(<) In words: g is a truthful SC-implementation of F if, for every profile <, whenever the agents declare that to be their real preferences, this a solution concept SC, and the outcome in the game and the function F are the same. The problem of truthful imple- mentation is illustrated on Figure 2. We say that the social choice function is truthfully SC-implementable if there is a game form that truthfully SC-implements it. Note that truthful implementations only require that the report of the true preference profile is an equilibrium, but it is not required that this equilibrium is unique. In general, other equilibria could be present that would not lead to the outcome prescribed by the SCF. However, this notion of implementation can be motivated. Indeed, it is assumed that playing a direct mechanism, if casting the real preference is an equilibrium strategy, an agent would be sincere. We illustrate the differences between the problems of implementation with some simple examples (a 'minimal' social choice scenario with only two voters and two alternatives), which demonstrates that the two notions are contingent and independent: a game form g can be both a truthful SC-implementation and an SC-implementation of a social function F, it can be both, and it can be either of them without being the other. 6 [a, b] [b, a] [a, b] [b, a] a (cid:8) a a b (cid:104)gH, ([a, b], [a, b])(cid:105) [a, b] [b, a] [a, b] [b, a] [a, b] [b, a] a a a (cid:8) b (cid:104)gH, ([a, b], [b, a])(cid:105) [a, b] [b, a] [a, b] a [b, a] a b [a, b] [b, a] a (cid:100) a a (cid:8) (cid:104)gH, ([b, a], [a, b])(cid:105) a b (cid:8) (cid:104)gH, ([b, a], [b, a])(cid:105) Figure 3: gH does not NE-implement H. But gH truthfully NE-implements H. Example 1 In this example we define some simple social choice functions, for all of them we set N = {1, 2} and K = {a, b}. Also, for the sake of comparison between stan- dard and truthful implementations, we only consider direct mechanisms, since truthful implementations are not defined otherwise. First, consider the function H for which we claim that its associated game form gH truthfully NE-implements H but gH does not NE-implement it. H is the social choice function prescribing the outcome b if and only both agents prefer b over a. We write [a, b] for the individual order of preferences of the outcome a over the outcome b and [b, a] for the individual preference of b over a. Hence, we have: H([a, b], [a, b]) = H([a, b], [b, a]) = H([b, a], [a, b]) = a; H([b, a], [b, a]) = b. Figure 3 represents the four possible games (cid:104)gH, <(cid:105) where <∈ L({a, b}){1,2}. In each of them, the circles indicate the action profiles that are Nash equilibria. The outcomes in bold are the outcomes o(a∗) for which a∗ = <: in those outcomes, players have revealed their true preferences. So for instance, the outcome a in the upper left corner of the game (cid:104)gH, ([a, b], [a, b])(cid:105) reads: 'the outcome in the game here is a and the voters reveal their true preferences'. For every preference profile <, the ticks (cid:8) indicate that the action profile < leads to the outcome prescribed by the social choice function H and is a Nash equilibrium in the game (cid:104)gH, <(cid:105); Hence the game form gH truthfully NE-implements H: all the bold outcomes are ticked. The cross (cid:100) designates a problem with the (standard) implementation of H by gH: in the game (cid:104)gH, ([b, a], [b, a])(cid:105) the action profile ([a, b], [a, b]) is a Nash equilibrium and leads to the outcome a, however H([a, b], [a, b]) = b. Hence, gH does not NE-implement H. 7 [a, b] [b, a] [a, b] [b, a] a (cid:8) b a (cid:8) b (cid:104)gJ, ([a, b], [a, b])(cid:105) [a, b] [b, a] [a, b] [b, a] [a, b] [b, a] a (cid:8) b a (cid:8) b (cid:104)gJ, ([a, b], [b, a])(cid:105) [a, b] [b, a] [a, b] a [b, a] a b [a, b] [b, a] a b b (cid:8) (cid:8) (cid:104)gJ, ([b, a], [a, b])(cid:105) a b (cid:8) (cid:8) (cid:104)gJ, ([b, a], [b, a])(cid:105) Figure 4: gJ both NE-implements and truthfully NE-implements J. Let us next consider the social choice function J which is dictatorial for player 1, i.e., J is defined by J([a, b], [a, b]) = J([a, b], [b, a]) = a; J([b, a], [a, b]) = J([b, a], [b, a]) = b. The four possible games (cid:104)gJ, <(cid:105) for J are depicted in Figure 4. It is easy to see that the circled outcomes in those games are Nash equilibria: they give the preferred outcome for 1 (so 1 cannot improve by deviating) and they are the same in a fixed row (so 2 cannot change the outcome). Moreover, it is also a straightforward check that for all those Nash equilibria, the outcome in the game (cid:104)gJ, <(cid:105) is the same as J(<) (for in- stance, in the top left game, both equilibria yield a which coincides with J([a, b], [a, b]), and in the bottom left game, both equilibria yield b = J([b, a], [a, b]), etc): this justifies the ticks (cid:8). So g NE-implements J. To show that g also truthfully NE-implements J, we need to check that all the bold outcomes in Figure 4 are circled and ticked(cid:8). Next, to give an example of an NE-implementation that is not a truthful one, con- It is mathematically equivalent to the game form gJ: the sider the game form gJ−. outcomes a and b are only inverted. Playing gJ−, the player 1 would simply play the contrary to his true preference. This always yields a Nash equilibrium and the out- comes are always as prescribed by J. Hence, like gJ, the game form gJ− is an NE- implementation of the social choice function J. However, since the player 1 needs to trick the game in order achieve a Nash equilibrium, it is easy to see that gJ− does not truthfully NE-implement J. The crosses (cid:100) on Figure 5 mark the action profiles that correspond to the true preferences of the players, and we can see that their respective outcome always fails to be as prescribed by J. 8 [a, b] [b, a] [a, b] [b, a] [a, b] [b, a] b (cid:100) a b a [a, b] [b, a] b a b (cid:100) a (cid:8) (cid:8) (cid:104)gJ−, ([a, b], [a, b])(cid:105) (cid:8) (cid:8) (cid:104)gJ−, ([a, b], [b, a])(cid:105) [a, b] [b, a] [a, b] [b, a] [a, b] [b, a] b (cid:8) a (cid:100) b (cid:8) a (cid:104)gJ−, ([b, a], [a, b])(cid:105) [a, b] [b, a] b (cid:8) a b (cid:8) a (cid:100) (cid:104)gJ−, ([b, a], [b, a])(cid:105) Figure 5: gJ− NE-implements J but does not truthfully NE-implement it. Finally, we argue that it is possible for a game form to be neither a NE-implementation nor a truthful implementation of a given function: take P such that P(<) = a for all profiles <. Moreover, for all <, let all outcomes in the matrix for (cid:104)gP, <(cid:105) be b. For every <, every outcome in (cid:104)gM, <(cid:105) is a Nash equilibrium (no agent can change the outcome, let alone improve it). At the same time, for all a∗ we have b = o(a∗) (cid:44) P(<) = a, which shows that g does not NE-implement P. It does also not truthfully NE-implement it: take, for any <, a profile a∗ in the game (cid:104)gM, <(cid:105) such that a∗ = <. We have already seen that o(a∗) (cid:44) P(<), which proves our claim. 3 A Logic of social choice functions Following the tradition in implementation theory (cf. Remark 1), we model social choice functions as a particular kind of strategic game form. In [13] we proposed a logic for modelling strategic games on the basis of CL-PC. Every player controls a set of propositional variables and a strategy for a player amounts to choosing a truth value for the variables he controls. We adapt the ideas of [13] to game forms where the strategies of the players correspond to the reports of preferences. 3.1 Semantics Let X be an arbitrary set of propositions. We can see a valuation of X as a subset V ⊆ X where tt (i.e., true) is assigned to the propositions in V and ff (false) is assigned to the propositions in X \ V. We denote the set of possible valuations over X by ΘX. 9 In the presence of a set of players N and a set of outcomes K, the set of propositions controlled by a player i ∈ N is defined as At[i, K] = {pi x>y is a proposition controlled by the agent i which means that i reports that it values the outcome x at least as good as y. We also define At[N, K] = ∪i∈NAt[i, K], which is then the set of all controlled propositions. x>y x, y ∈ K}. Every pi We can 'encode' a particular preference (or linear order) of player i as a valuation of the propositions in At[i, K]. However, conversely, not all valuations correspond to a linear order preference. A strategy of a player i consists of reporting a valuation of At[i, K] encoding a linear order over K. For every player i, we define strategies[i, K] as a set of valuations V ∈ ΘAt[i,K] such that: (i) pi x>y ∈ V iff x>y ∈ V and pi y>z ∈ V then pi y>x (cid:60) V, and (iii) if pi pi Remark 2 Every pi x>y could be seen as a predicative expression p(i, x, y) that would read that agent i reported to prefer the outcome x over y. However, since N and K are finite, we look at these expressions as a finite collection of propositions. The constraints of control in Figure 6 will be their propositional theory corresponding to the three preceding constraints on the valuations. x>x ∈ V, (ii) if x (cid:44) y then pi x>z ∈ V. For every coalition C ⊆ N, let strategies[C, K] be the set of tuples vC = (vi)i∈C where vi ∈ strategies[i, K]. It is the set of strategies of the coalition C. To put it another way, it corresponds to a valuation of the propositions controlled by the players in C, encoding one preference over K for every player in C. A state (or reported preference profile) is an element of strategies[N, K], that is, a strategy of the coalition containing all the players. We now define the models of social choice functions. Definition 8 (Model of social choice functions) A model of social choice functions over N and K is a tuple M = (cid:104)N, K, out, (<i)(cid:105), such that: out : strategies[N, K] −→ K maps every state to an outcome; For every i ∈ N, <i ∈ L(K) is the true order of preferences of i. Hence, every player i has two levels of preferences: (i) a true one, given by (<i) and (2) a reported one, given by a valuation in strategies[i, K]. Taking out the true preference profile from a model of SCF, we obtain a mere instantiation of a pre-Boolean game [4]. It is required to assign every variable to one (actual control) and only one (exclusive control) player, but there are some constraints on the possible valuations ('non-full' control). In [4], actual and exclusive control are grasped by an assignment function (mapping every propositional variable to exactly one player), and the partial control is modelled by a set of constraints given as a set of satisfiable propositional formulae. The language Lscf [N, K] is inductively defined by the following grammar: (cid:95)iϕ ϕ (cid:70) (cid:62) (cid:94)Cϕ p where p is atom of At[N, K], x is an atom of K, i ∈ N, and C is a coalition. Given a model M and a state (i.e., a reported profile v), formula (cid:94)Cϕ reads that provided that the players outside C hold on to their current strategy vC, the coalition C has a strategy, 10 ¬ϕ ϕ ∨ ϕ x i.e., a way to announce their profiles, such that ϕ holds. Formula (cid:95)iϕ reads that i locally (at the current reported profile) considers a reported profile where ϕ is true at least as preferable. Definition 9 (Truth values of Lscf [N, K]) Given a model M = (cid:104)N, K, out, (<i)(cid:105), we are going to interpret formulae of Lscf [N, K] in a state of the model. A state v = (v1, . . . , vn) in M is a tuple of valuations vi ∈ strategies[i, K], one for each agent. The truth definition is inductively given by: p ∈ vi for some i ∈ N iff iff out(v) = x iff M, v (cid:54)= ϕ M, v = p M, v = x M, v = ¬ϕ M, v = ϕ ∨ ψ iff M, v = ϕ or M, v = ψ M, v = (cid:94)Cϕ M, v = (cid:95)iϕ iff iff there is a state u such that vi = ui for every i (cid:60) C and M, u = ϕ there is a state u such that out(v) <i out(u) and M, u = ϕ We assume that player i only makes claims or announcements about its own pref- erences, and i controls nothing else, so the atomic clause could equivalently have read M, v = pi x>y iff pi x>y ∈ vi The truth of ϕ in all models over a set of players N and a set of outcomes K is denoted by =Λscf [N,K] ϕ. The classical operators ∧, →, ↔ can be defined as usual. We also define (cid:3)Cϕ (cid:44) ¬(cid:94)C¬ϕ and (cid:4)iϕ (cid:44) ¬(cid:95)i¬ϕ. Theorem 1 (Decidability) The problem of deciding whether a formula ϕ ∈ Lscf [N, K] is satisfiable is decidable. Proof. It suffices to remark that N and K are finite. Hence, we can enumerate every model of SCF over N and K and check whether ϕ is satisfiable in one state of one model. (cid:4) 3.2 Ballots We think of a particular preference of L(K) encoded in the language of the propositions as a ballot. Definition 10 (Ballot) For every player i ∈ N, we can see every <i ∈ L(K) as a per- mutation [x1, x2 . . .] of the elements of K, where the more to the left the outcome is, the more it is preferred by the player i. We can reify in the language the reported preferences, that is, the ballot casted by the player i: balloti(<) (cid:44) pi x1>x2 ∧ pi x2>x3 ∧ . . . pi xK−1>xK . 11 Then, the formula (cid:94) i∈N balloti(<) ballot(<) (cid:44) is a reification of the reported preference profile < = (<1, . . . , <n), consisting of one ballot for every player i ∈ N. Remark 3 Note that for every < ∈ L(K), the formula ballot(<) is true at one and only one state. The reader familiar with Hybrid Logic [1] may think of the formula ballot(<) as a nominal, viz. a state label available in the object language. Example 2 Suppose that N = {1, 2} and K = {a, b, c}. Let a preference profile 2 ) ∈ L(K)N given by the data of the two permutations [a, c, b] and [c, a, b] repre- (<ex senting respectively the preferences of player 1 and 2. This reported preference profile can be represented in the language Lscf [{1, 2},{a, b, c}] by the formula 1 , <ex ballot(<ex) (cid:44) p1 a>c ∧ p1 c>b ∧ p2 c>a ∧ p2 a>b. c>c ∧ p1 c>c ∧ p2 a>c ∧ p1 c>a ∧ p2 c>b ∧ p1 a>b ∧ p2 c>a ∧¬p1 a>c ∧ ¬p2 b>c ∧¬p1 b>a ∧ b>a ∧ ¬p2 a>b ∧¬p1 c>b ∧ ¬p2 It is easy to verify that the constraints on the elements of strategies[1, K] and strategies[2, K] are sufficient for inferring a complete characterisation of the prefer- ence profile. The following is valid in the models of social choice functions over {1, 2} and {a, b, c}: a>a ∧ p1 ballot(<ex) ↔ p1 a>a ∧ p2 p2 b>b ∧ p1 b>b ∧ p2 3.3 Characterising an SCF Recall that a model of social choice functions is a tuple M = (cid:104)N, K, out, (<i)(cid:105), where <i are the real preferences of the agents and the outcome function o assigns to every valuation an element of K. There is a one-one correspondence between valuations and preference profiles: the preference profile P(v) associated with valuation v is the x>y ∈ v. Likewise, the valuation V(<) associated with relation < for which x >i y iff pi x>y x >i y}, which collect all the atoms form ballot(<). This makes it < is the set {pi possible to relate a model M with a social choice function F as follows. We say that a model M = (cid:104)N, K, out, (<i)(cid:105) and social choice function F : L(K)N → K correspond, if for every strategy profile < and its associated valuation v (i.e., for which V(<) = v and P(v) =<), we have o(v) = F(<). b>c This correspondence can be syntactically defined in a formula ρF: (cid:94) ρF = <∈L(K)N (cid:94)N(ballot(<) ∧ F(<)) Note that (cid:94)N plays the role of the universal/global existential modality often noted E in the literature in modal logic: it allows us to quantify over all the possible valuations in ΘAt[N,K], or ballots. Given the outcomes K, the agents N and the social choice function F, formula ρF says that every profile < together with F(<) as an outcome appears in the model. Since 12 the states of a model are all possible profiles in L(K)N, and every profile occurs exactly once, we might as well have defined ρF as (cid:94) (ballot(<) → F(<)) ρF = <∈L(K)N It is easy to see that the logic is expressively complete wrt. social choice functions. That is, for every SCF F over a set of players N and a set of outcomes K, there exists a formula ρF ∈ Lscf [N, K] characterising it. Even though it may not be optimal in terms of succinctness, it suffices to consider the conjuncts of formulae (cid:94)N(ballot(<) ∧ x), for < ∈ L(K) and F(<) = x. The next example shows, using a simple scenario, that we can sometimes obtain less naıve and more compact characterisations. Example 3 Consider the following model of SCF (or game form) where player 1 chooses rows, player 2 chooses columns and player 3 chooses matrices. There are two outcomes a and b. Hence, every player i controls the set of atoms {pi b>a}. a>a∧pi Every player i has two strategies: pi b>a, that we denote respectively by [a, b] and [b, a]. (In the logic Λscf [{1, 2, 3},{a, b}], they are in fact equivalent to the formulae pi a>b∧¬pi b>a, respectively.) b>b∧pi a>b and pi a>a, pi a>a∧pi b>b, pi b>b∧¬pi a>b, pi a>b∧pi b>a and pi [a, b] [b, a] [a, b] [b, a] [a, b] [b, a] a a a b [a, b] [b, a] a b b b [a, b] We can represent it in the logic Λscf [{1, 2, 3},{a, b}] of social choice functions by [b, a] the formula: ρF (cid:44) a ↔ (p1 a>b ∧ p2 a>b) ∨ (p1 a>b ∧ p3 a>b) ∨ (p2 a>b ∧ p3 a>b). Note that since out is functional, in the models of social choice functions with K = {a, b} the outcome b will hold whenever a does not. Going back to the social choice functions of Example 1, we invite the reader to check that b>a ∧ p2 b>a) ρH = b ↔ (p1 ρJ = a ↔ p1 ρP = a a>b 3.4 True preferences In Section 3.2 we saw how to use the atoms in At[i, K] to encode the reported preference or ballot of a player i. These atoms do not necessarily represent the true preferences of the agents. We handle the true preferences of player i via the (cid:95)i modality. 13 From our basic language Lscf [N, K], we can also define an operator of interest concerning preferences. We can define the global binary operator of preferences ψ (cid:74)i ϕ, corresponding to a preference between propositions. It reads "all ϕ are better than all ψ". ψ (cid:74)i ϕ (cid:44) (cid:3)N (ballot(<) ∧ (ϕ → (cid:3)N(ψ → (cid:95)iballot(<)). (cid:95) < ∈L(K)N Agent i judges the proposition ϕ at least as good as ψ iff when the reported preference profile is < and ϕ holds at the state labeled by ballot(<), then, whenever ψ holds in a state, i would prefer the state labeled by ballot(<) (cf. Remark 3). As in Definition 10 for reported preferences, we can now reify the true preferences. Provided that x and y are two possible outcomes, the formula y (cid:74)i x captures the fact the player i prefers (globally) the outcome y over the outcome x. Hence, from a preference profile < ∈ L(K)N, we reify the preference [x1, x2 . . .] of the player i as follows: truei(<) (cid:44) (xK (cid:74)i xK−1) ∧ . . . ∧ (x3 (cid:74)i x2) ∧ (x2 (cid:74)i x1). Then, the formula true(<) (cid:44) truei(<) (cid:94) i∈N is a reification of the true preference profile <= (<1, . . . , <n). Remark 4 Whenever in a model of social choice function M the true preference of a player i is such that x <i y, then the formula x (cid:74)i y is true at every state of M. (cid:87) However, the other way around does not hold. Indeed, when either x or y is not a possible outcome of a model, the formula x (cid:74)i y is always true for every i. From the < ∈L(K)N (ballot(<) ∧ (y → (cid:3)N(x → (cid:95)iballot(<)). Hence, if y definition, x (cid:74)i y (cid:44) (cid:3)N is not a possible outcome, the main implication y → (cid:3)N(x → (cid:95)iballot(<)) is always true for y being always false. Likewise, if x is not a possible outcome, the implication x → (cid:95)iballot(<) is always true for x being always false. In turn, it makes the main < ∈L(K)N ballot(<) will always be satisfied since a state implication always true. Also,(cid:87) of evaluation represents a ballot by definition. The object language does not allow to talk about true preferences on impossible outcomes. This observation will have a consequence in the way we prove the com- pleteness of the logic. 3.5 Axiomatics The axiomatization of the models of social choice functions is presented in Figure 6. Constraints of control (refl), (antisym-total) and (trans) say that every player casts an appropriate valuation of its controlled atoms: a valuation must encode a linear or- der. (comp∪) defines the local ability of coalitions in terms of local abilities of sub- coalitions. The transitivity of the operator (cid:3)C is the consequence of (comp∪). Hence, together with (T(i)) and (B(i)), it makes of (cid:3)C an S5 modality. (empty) means that the empty coalition has no power. (comp∪) and (confl) together make sure that the agents' choices are independent. (exclu) means that if an atom is controlled by a player i, the other players cannot change its value. (ballot) makes sure that an agent is always 14 y>x y>z → pi x>z , where x (cid:44) y Constraints of control pi (refl) x>y ↔ ¬pi x>x pi (antisym-total) x>y ∧ pi pi (trans) Propositional control (Prop) (K(i)) (T(i)) (B(i)) (comp∪) (confl) (empty) (exclu) (ballot) (comp-At) Outcomes and preferences (func1) (func2) (incl) (K(<i)) (4(<i)) (antisym(cid:48)) (total(cid:48)) (unifPref ) Rules (MP) (Nec((cid:3)i)) , where j (cid:44) i , where ϕ is a propositional tautology (cid:3)C2 ϕ ↔ (cid:3)C1∪C2 ϕ x∈K(x ∧(cid:86) ϕ (cid:3)i(ϕ → ψ) → ((cid:3)iϕ → (cid:3)iψ) (cid:3)iϕ → ϕ ϕ → (cid:3)i(cid:94)iϕ (cid:3)C1 (cid:94)i(cid:3)jϕ → (cid:3)j(cid:94)iϕ (cid:3)∅ϕ ↔ ϕ ((cid:94)ip ∧ (cid:94)i¬p) → ((cid:3)jp ∨ (cid:3)j¬p) (cid:94)iballoti(<) (cid:94)C1 δ1 ∧ (cid:94)C2 δ2 → (cid:94)C1∪C2(δ1 ∧ δ2) (cid:87) (ballot(<) ∧ ϕ) → (cid:3)N(ballot(<) → ϕ) (cid:3)Nϕ → (cid:4)iϕ (cid:4)i(ϕ → ψ) → ((cid:4)iϕ → (cid:4)iψ) (cid:95)i(cid:95)iϕ → (cid:95)iϕ (ballot(<) ∧ (cid:95)iballot(<(cid:48)) → (cid:3)N(ballot(<(cid:48)) → (cid:4)i¬ballot(<) (ballot(<) ∧ (cid:95)iballot(<(cid:48)) ∨ (cid:3)N(ballot(<(cid:48)) → (cid:95)iballot(<) (x ∧ (cid:95)iy) → (x (cid:74)i y) from (cid:96) ϕ → ψ and (cid:96) ϕ infer (cid:96) ψ from (cid:96) ϕ infer (cid:96) (cid:3)iϕ y∈K\{x} ¬y) Figure 6: Logic of social choice functions Λscf [N, K]. i ranges over N, C1 and C2 over 2N, x and y are over K, and < is over L(K)N. δ1 and δ2 are two formulae from Lscf [N, K] that do not contain a common atom from At[N, K]. ϕ represents an arbitrary formula of Lscf [N, K], and p an arbitrary atom in At[N, K]. locally able to cast any preference. From (comp-At), provided that δ1 and δ2 do not contain a commonly controlled atom, if a coalition C1 can locally enforce δ1 and C2 can locally enforce δ2 then they can enforce δ1 ∧ δ2 together. Axiom (func1) forces the fact that for every action profile there is one and only one outcome. (func2) ensures that the outcomes are only determined by the valuations. (incl) ensures that if something is settled, a player cannot prefer its negation. (4((cid:22)i)) characterises transitivity. (antisym(cid:48)) and (total(cid:48)) force that the relation of preference over states is antisymmetric and total (and hence, in particular, this relation is reflex- ive). Finally, (unifPref ) specifies a fundamental interaction between preferences and the outcomes. If the casted preference profile at hand leads to x and agent i prefers an action profile leading to y, then at every action profile leading to x, agent i will prefer every action profile leading to y, that is, all y are better than all x. The logic has a clear flavour of normal modal logic [5]. The presence of (K(i)) with the necessitation rule (Nec((cid:3)i)) gives to the operator (cid:3)i the property of normality. 15 The necessitation rule for the operator (cid:4)i holds because of (Nec((cid:3)i)) and the axioms (comp∪) and (incl). The normality of the modality (cid:4)i then follows from (K(<i)). The axiomatics is largely inspired by the axiomatics of the logic of games and propositional control (henceforth LGPC) presented in [13]. The logic LGPC is de- signed to model strategic games in general. The agents have arbitrary strategies, and preferences allowing for indifference between two different outcomes. On the other hand, in this paper we focus on SCFs and hence on particular strategic games that 'represent' an SCF (cf. Remark 1). While in LGPC we had an axiom saying that every atom was actually controlled by at least one agent, here we are more specific as we know a priori which atoms are controlled by a given agent. This is the role of the axiom (ballot). Constraints of controls are also specific to the present study. The truth values of the controlled atoms cannot be independent of each other as we use them to encode preferences. In LGPC, all valuations of the controlled atoms were permitted. Theorem 2 (Soundness and completeness) Λscf [N, K] is sound and complete with respect to the class of models of social choice functions. Proof. The proof of completeness first gives an equivalent but more standard semantics to the logic: the Kripke models of SCF. Then we build the canonical model. For every consistent formula ϕ, we show how to isolate a sub-model Mϕ that we prove is a Kripke model of SCF that satisfies ϕ. Further details are given in the Appendix. (cid:4) 4 Applications We have already demonstrated that the language allows to completely characterise an SCF. In this section we show how we can express properties of social choice functions in the language and apply the logic to reason about them. The language can be used to characterise requirements on social choice functions. We first illustrate that with some simple properties, namely citizen sovereignty and non-dictatorship. Next, we will characterise a dominant strategy equilibrium. Finally, we provide a formalisation of monotonicity and strategy-proofness, and use standard results of SCT to show how we can use the logic to check whether an SCF is imple- mentable in a dominant strategy. 4.1 Citizen sovereignty and non dictatorship We say that an SCF satisfies citizen sovereignty iff every outcome in K is feasible. That is, no outcome is rejected independently of the individual opinions. It is defined as follows. Definition 11 (Citizen sovereignty) An SCF F satisfies citizen sovereignty iff for ev- ery x ∈ K there is a < ∈ L(K)N such that F(<) = x. 16 The next formula is a straightforward translation of the definition of citizen sovereignty in the language of social choice functions. (cid:94) x∈K CITSOV (cid:44) (cid:94)Nx. We say that an SCF satisfies non dictatorship iff no player can always impose its favourite outcome. Definition 12 (Non-dictatorship) An SCF F is non dictatorial iff for every player i ∈ N there is a ballot < ∈ L(K)N such that F(<) <i y for some y ∈ K \ {F(<)}. This says that for every player, there is a ballot < whose outcome is F(<), and i prefers an outcome that is not F(<). We can rewrite the definition of non dictatorship into the language of social choice functions as follows. (cid:94) i∈N (cid:95) x∈K x ∧ (cid:95) y∈K\{x}   . pi y>x NODICT (cid:44) (cid:94)N The following proposition is immediate. Proposition 1 Consider a social choice function F and ρF a formula characterising it. 1. F has the property of citizen sovereignty iff =Λscf [N,K] ρF → CITSOV. 2. F is non dictatorial iff =Λscf [N,K] ρF → NODICT. 4.2 Dominant strategy equilibrium Citizen sovereignty and non dictatorship are possible properties of a social choice func- tion: their formulations in logic are globally true (or false) in a model of SCF. However, the logic is also able to formalise solution concepts, which are properties of states. In [13], we characterised several solution concepts (dominant strategy equilibrium, Nash equilibrium, core membership. . . ) that are directly applicable in the logic of the present work. In order to formalise strategy-proofness later, we need to characterise a dominant strategy equilibrium. A dominant strategy equilibrium captures a particularly important pattern of behaviour. It arises when every player plays a dominant strategy, that is, a strategy that would represent the best choice whatever the other agents play. We define it directly in our models of SCF. Definition 13 (Dominant strategy equilibrium) Let v∗ be a state in amodel of social choice functions (cid:104)N, K, out, (<i)(cid:105). v∗ is a dominant strategy equilibrium iff for every player i ∈ N and every strategy uN\{i} ∈ strategies[N \ {i}, K], we have out(u0 . . . u(cid:48) i . . . un) <i out(u0 . . . v∗ i ∈ strategies[i, K]. i . . . un) for every u(cid:48) 17 A dominant strategy equilibrium is a strong solution concept: such an equilibrium does not depend on the knowledge of an agent i about the strategies or preferences of other players. It is convenient to introduce the notion of best response by an agent i. (cid:95) (x ∧ (cid:3)i(cid:95)ix). x∈K BRi (cid:44) A player i plays a best response in a state if, x being the outcome, for every deviation of i, i prefers x. We can now define strategy dominance in terms of best response: (cid:94) i∈N DOM (cid:44) (cid:3)N\{i}BRi. We have a strategy dominant state if the current choice of every player ensures them a best response whatever other agents do. Proposition 2 Assume a model of social choice functions M and a state v∗. We have that v∗ is a dominant strategy equilibrium iff M, v∗ = DOM. 4.3 Monotonicity and strategy-proofness One important property of SCF is monotonicity, as this property can affect the imple- mentability of social choice functions. Definition 14 (Monotonicity) An SCF F is monotonic iff for all {<, <(cid:48)} ⊆ L(K)N and x ∈ K, if F(<) = x and if for all i ∈ N, for all y ∈ K we have that that y <i x implies that y <(cid:48) We propose to characterise monotonic social choice functions. We define i x, then, F(<(cid:48)) = x. (cid:86) < ∈L(K)N <(cid:48) ∈L(K)N i∈N y∈K (cid:94)N(ballot(<(cid:48)) ∧ pi (cid:86) (cid:86) MON (cid:44)  (cid:86) (cid:104)(cid:94)N(ballot(<) ∧ x)∧ (cid:86) (cid:16)(cid:94)N(ballot(<) ∧ pi (cid:105) (cid:17) → (cid:94)N(ballot(<(cid:48)) ∧ x) x>y) → x∈K x>y)  . Although it may appear rather complex, the predicate MON is essentially nothing more than the expression of Definition 14 in our language Lscf [N, K]. The following proposition is immediate. Proposition 3 Consider a social choice function F and ρF a formula characterising it. F is monotonic iff =Λscf [N,K] ρF → MON. Monotonicity does not depend on the true preference profile of the players. Ac- cordingly, our definition does not involve the modalities of preference (cid:95)iϕ and ϕ (cid:74)i ψ. Capitalising on standard results from social choice theory, we will show that using the full expressivity of our language (that is, using true preference modalities) we can obtain a much simpler formulation. We say that an SCF is strategy-proof if for every preference profile, telling the truth (reporting the true preference) is a dominant strategy for every player. 18 Definition 15 (Strategy-proofness) An SCF F is strategy-proof iff F is truthfully DOM- implementable. Hence, a choice function is strategy-proof when it is truthfully implementable in domi- nant strategy: for every preference profile, reporting their true preference is a dominant strategy for every player. The revelation principle [7] is a central result in implementation theory. It states that if an SCF is DOM-implementable, then it is truthfully DOM-implementable. It is true in general even if L(K) is based on weaker orders. The revelation principle tells us that if an SCF F is implementable in dominant strategies then there exists a direct mechanism such that for every preference profile <, truth telling (every player i reports <i) is a dominant strategy and the outcome is F(<). Truthful implementations are rather weak; it is easier in general to implement a choice function truthfully than with 'standard' implementations. Indeed, in truthful implementations there might be an equilibrium that leads to an outcome different of the one prescribed by the SCF. But because in this paper we consider linear preferences, and we assume that players cannot be indifferent between two distinct outcomes, such a situation cannot happen. Thus, we can be more specific than the revelation principle. Theorem 3 ([6, Corollary 4.1.4]) A direct mechanism g truthfully implements an SCF F in dominant strategies iff g DOM-implements F. Hence, when working in dominant strategies with linear preferences, the concepts of implementation and truthful implementation coincide. We propose to characterise strategy-proof social choice functions as follows: STRPROOF (cid:44) [true(<) → (ballot(<) → DOM)] (cid:94) < ∈L(K)N The formula STRPROOF is an immediate reformulation of the definition of strategy- proofness in our language of social choice functions. Proposition 4 Consider a social choice function F and ρF a formula characterising it. F is strategy-proof iff =Λscf [N,K] ρF → STRPROOF. This Proposition provides us with a general procedure to check whether a social choice function is strategy-proof. Moreover (because of Theorem 3), because we re- strict our attention to linear preferences, it allows us to check whether an SCF is DOM- implementable. Example 4 We can verify for instance that the social choice function characterised in Example 3 is strategy-proof. =Λscf [{1,2,3},{a,b}] (a ↔ (p1 a>b ∧ p2 → STRPROOF. a>b) ∨ (p1 a>b ∧ p3 a>b) ∨ (p2 a>b ∧ p3 a>b)) Monotonicity sometimes implies implementability and this is actually the case in our setting. Since we are working with rich domains of preferences1 and linear order- ings the following result holds. 1The notion of a rich domain is some tangential to the purposes of this paper. Briefly, our domain of preferences is rich because we allow every linear order of K. See [6, Sec. 3.1] 19 Theorem 4 ([6, Cor. 3.2.3, Th. 4.3.1]) An SCF is truthfully implementable in domi- nant strategies iff it is monotonic. This standard result of implementation theory shows that in our setting, the notions of monotonicity and of strategy-proofness match. Trivially we are actually able to substantially simplify the formula MON, our characterisation of monotonicity in the formal language. Indeed, as a consequence of Theorem 4, we have the following. Proposition 5 =Λscf [N,K] MON ↔ STRPROOF. 5 Discussion and perspectives We have presented the problem of direct implementation in social choice theory and proposed a logical formalisation of it. We were able to give a sound and complete ax- iomatization to the logic. We showed how we can characterise social choice functions and properties of social choice functions. And finally, we have demonstrated the value of the logic by proposing a general logical procedure for checking whether a social choice function is strategy-proof. Our logical language is a formal counterpart of the language of "natural mathemat- ics" that is typically used in social choice theory. There are however two features that make it particularly useful: (i) it is supported by a non ambiguous semantics; and (ii) the resulting logic is decidable. Section 4 suggests a logical methodology for reasoning about problems of social choice theory with the logic of social choice functions. Let a collection of properties of social choice theory Pi, i ∈ {1, . . . n} be characterised in the logic Λscf [N, K] by ρPi, respectively. 1. We can use the logic in order to check whether an SCF satisfies a certain prop- erty. An SCF F characterised by ρF has the property P1 iff ρF → ρP1 is derivable in Λscf [N, K]. 2. We can use the logic in order to evaluate the strength of constraints in SCT. P1 is a property weaker than P2 iff the formula ρP2 → ρP1 is derivable in Λscf [N, K]. For instance, instead of using a result of SCT to prove Proposition 5, we could actually use the logic to automatically verify that monotonicity and strategy- proofness coincide in the current setting. More interestingly, we could use it to prove new theorems. 3. We can use the logic for mechanism design. Building a mechanism that imple- ments a social choice procedure satisfying the properties P1, P2, . . . Pn consists of finding a model for the formula ρP1 ∧ ρP2 ∧ . . . ∧ ρPn. We believe these are exciting possibilities for social choice theory and logic, and as the logic is decidable, they are in principle possible. 20 Acknowledgment An earlier abstract of this paper appeared as [14]. We thank the anonymous reviewers for their suggestions that helped to improve the paper. We are also grateful to the participants of TARK'09. This research is funded by the EPSRC grant EP/E061397/1 Logic for Automated Mechanism Design and Analysis (LAMDA). References [1] C. Areces and B. ten Cate. Hybrid Logics, volume Handbook of Modal Logic, chapter 14, pages 821 -- 868. Elsevier Science Inc., 2006. [2] K. Arrow. A difficulty in the concept of social welfare. Journal of Political Economy, 58(4):328346, 1950. [3] K. J. Arrow, A. K. Sen, and K. Suzumura, editors. Handbook of Social Choice and Welfare, volume 1. Elsevier, 2002. [4] E. Bonzon, M.-C. Lagasquie-Schiex, and J. Lang. Efficient coalitions in Boolean In K. Apt and R. van Rooij, editors, New Perspectives on Games and games. Interaction, volume 4 of Texts in Logic and Games, pages 283 -- 297. Amsterdam University Press, 2007. [5] B. F. Chellas. Modal Logic: an introduction. Cambridge University Press, 1980. [6] P. Dasgupta, P. Hammond, and E. Maskin. The Implementation of Social Choice Rules: Some General Results on Incentive Compatibiliy. Review of Economic Studies, 46:185 -- 216, 1979. [7] A. Gibbard. Manipulation of voting schemes: a general result. Econometrica, 41(4):587 -- 601, 1973. [8] K. May. A set of independent, necessary and sufficient conditions for simple majority decision. Econometrica, 20(4):680 -- 684, 1952. [9] H. Moulin. The Strategy of Social Choice. Advanced Textbooks in Economics. North Holland, 1983. [10] M. J. Osborne and A. Rubinstein. A Course in Game Theory. The MIT Press, 1994. [11] M. Pauly. Logic for Social Software. PhD thesis, University of Amsterdam, 2001. ILLC Dissertation Series 2001-10. [12] M. A. Satterthwaite. Strategy-proofness and arrow's conditions: Existence and correspondence theorems for voting procedures and social welfare functions. Journal of Economic Theory, 10:187 -- 217, 1975. 21 [13] N. Troquard, W. van der Hoek, and M. Wooldridge. A Logic of Games and Propositional Control. In Decker, Sichman, Sierra, and Castelfranchi, editors, 8th International Joint Conference on Autonomous Agents and Multi Agent Systems (AAMAS-09) , Budapest, Hungary, pages 961 -- 968. IFAAMAS, 2009. [14] N. Troquard, W. van der Hoek, and M. Wooldridge. A logic of propositional con- trol for truthful implementations. In TARK '09: Proceedings of the 12th confer- ence on Theoretical aspects of rationality and knowledge, pages 237 -- 246. ACM DL, 2009. [15] W. van der Hoek and M. Wooldridge. On the logic of cooperation and proposi- tional control. Artificial Intelligence, 164(1-2):81 -- 119, 2005. [16] M. Wooldridge, T. Ågotnes, P. E. Dunne, , and W. van der Hoek. Logic In Proceedings of the for automated mechanism design -- a progress report. Twenty-Second AAAI Conference on Artificial Intelligence (AAAI-2007), Vancou- ver, British Columbia, Canada, 2007. Proof of Theorem 2 Λscf [N, K] is sound and complete with respect to the class of models of social choice functions. Proof. It is routine to verify that all principles of Figure 6 are valid. We show that if a formula is consistent, it is provable in the system Λscf [N, K]. M = (cid:104)N, K, S, (Ri), (Pi), V(cid:105) such that: We first introduce the Kripke models of SCF. A Kripke model of SCF is a tuple • N and K are parameters; • S = {V ∈ ΘAt[N,K] ∀i ∈ N,∃Vi ∈ strategies[i, K] s.t. V = ∪i∈NVi}; • V is a valuation function of At[N, K] ∪ K where for every v ∈ S: -- p ∈ V(v) iff p ∈ v, p ∈ At[N, K]; -- there is a unique x ∈ K s.t. x ∈ V(v); [(cid:44)→ we say that the model is based on the outcome function outM when outM(v) = x iff x ∈ V(v)]. • Rivu iff vj = uj for all j (cid:44) i; • there is a <M ∈ L(K)N s.t. Pivu iff (if x ∈ V(v) and y ∈ V(u) then x <M i y); [(cid:44)→ we say that the model is based on <M]. Truth values of (cid:94)iϕ and (cid:95)iϕ in a Kripke model of SCF are obtained in the standard way from the relations Ri and Pi, respectively. Clearly, for every Kripke model M based on outM and <M, we can construct a model of social choice functions Mscf = (cid:104)N, K, outM, (<M By construction, there exists a bijection f : S −→ strategies[N, K] that associates a state s in M to a state v = (v1 . . . vn) in Mscf in such a way that for every p ∈ At[i, K], we have p ∈ V(s) iff p ∈ vi. i )(cid:105) and reciprocally. The following is easy to see. 22 Claim 1 M, s = ϕ iff Mscf , f (s) = ϕ. Hence, the proof of the theorem can be reduced to a proof of completeness of the logic wrt. to the class of Kripke models of SCF. Let Ξ be the set of maximally consistent sets (mcs.) of Λscf [N, K]. We define the proper canonical model Mcan = (cid:104)N, K, S, (Ri), (Pi), V(cid:105) as follows. N and K are the parameters of the logic. S = Ξ. RiΓ∆ iff ∀δ ∈ ∆, (cid:94)iδ ∈ Γ. PiΓ∆ iff ∀δ ∈ ∆, (cid:95)iδ ∈ Γ. p ∈ V(Γ) iff p ∈ ∆. x ∈ V(Γ) iff x ∈ ∆. Given an mcs. Γ0 we define the set of mcs. 'describing' the same SCF and where the players have the same true preferences (modulo the preferences concerning some outcome which is not feasible in the SCF): Cluster(Γ0) (cid:44) {Γ1 ∀ < ∈ L(K)N,∀x ∈ K, (cid:94)N(ballot(<) ∧ x) ∈ Γ1 iff (cid:94)N(ballot(<) ∧ x) ∈ Γ0} ∩ {Γ2 ∀i ∈ N,∀{x, y} ⊆ K, x (cid:74)i y ∈ Γ2 iff x (cid:74)i y ∈ Γ0} Let ϕ be a consistent formula of Lscf [N, K]. There is an mcs. Γϕ s.t. ϕ ∈ Γϕ. The proof consists in constructing a model from Γϕ such that it is indeed a Kripke model of SCF and there is a state satisfying ϕ. i , P(cid:48) i , V(cid:48)(cid:105) from Mcan as follows: We define Mϕ = (cid:104)N(cid:48), K(cid:48), S(cid:48), R(cid:48) • N(cid:48) = N and K(cid:48) = K; • S(cid:48) = ΞCluster(Γϕ); • R(cid:48) = RiCluster(Γϕ); • P(cid:48) = PiCluster(Γϕ); • p ∈ V(cid:48)(∆) iff p ∈ V(∆), ∆ ∈ S(cid:48). It is immediate that the truth lemma holds. i i Claim 2 Mϕ, Γ = δ iff δ ∈ Γ. Hence, Mϕ, Γϕ = ϕ. The set of states in Kripke models of SCF is defined as the set of valuations of At[N, K] encoding a preference profile. We prove that there exists a bijection between S(cid:48) and L(K)N. Claim 3 The following statements are true: 1. ∀∆ ∈ S(cid:48),∃! < ∈ L(K)N s.t. ballot(<) ∈ ∆; 2. ∀ < ∈ L(K)N,∃!∆ ∈ S(cid:48) s.t. ballot(<) ∈ ∆. the claim follows from the constraints of control The first part of (refl), (antisym-total) and (trans). We now argue that for every < ∈ L(K)N, there is exactly one ∆ ∈ S(cid:48) such that ballot(<) ∈ ∆. Let < ∈ L(K)N. We have (cid:96) (cid:94)iballoti(<) by (ballot). With (comp-At), we find that (cid:96) (cid:94)Nballot(<). Hence, (cid:94)Nballot(<) ∈ Γϕ, and there must be an mcs. ∆ s.t. ballot(<) ∈ ∆. Now suppose that ∆(cid:48) ∈ S(cid:48) also contains ballot(<). By (func2), ∆ and ∆(cid:48) contain the same formulae. Then ∆(cid:48) = ∆, which proves the second part of the claim. As a consequence we will be allowed to use the formulae of the form ballot(<) as world labels in Mϕ. We now prove the main claim of this proof. 23 Claim 4 Mϕ is a Kripke model of SCF. We first prove that for every mcs. Γ and ∆, we have that RiΓ∆ iff for all i (cid:44) j we have that ballotj(<) ∈ Γ iff ballotj(<) ∈ ∆. First, observe that for every i, Ri is an equivalence relation because by axioms (K(i)), (T(i)), (B(i)) and (comp∪) all (cid:3)i are S5 modalities. (⇒). Suppose RiΓ∆. Then by definition ∀δ ∈ ∆ we have (cid:94)iδ ∈ Γ. For any < ∈ L(K)N and j (cid:44) i, suppose also that ballotj(<) ∈ ∆. By (exclu), (cid:3)iballotj(<) ∈ ∆. Then by hypothesis (cid:94)i(cid:3)iballotj(<) ∈ Γ, which by (B(i)) entails that ballotj(<) ∈ Γ. Because RiΓ∆ is an equivalence relation, the same reasoning can be done to prove that if ballotj(<) ∈ Γ then ballotj(<) ∈ ∆. (⇐). Suppose ∀j (cid:44) i, ∀ < ∈ L(K)N we have ballotj(<) ∈ Γ iff ballotj(<) ∈ ∆. Suppose that balloti(<(cid:48)) ∈ ∆ and δ ∈ ∆. Let us note <∆ the preference profile i . . . <n). We hence have ballot(<∆) ∧ δ ∈ ∆. Which by (func2) means that By (ballot), we also have that (cid:86) (<1, . . . <(cid:48) (cid:3)N(ballot(<) → δ) ∈ ∆. From (exclu), (cid:3)i (cid:94)iballoti(<(cid:48)) ∈ Γ. Hence, by S5, (cid:94)iballot(<∆) ∈ Γ. We obtain that (cid:94)iδ ∈ Γ. We now prove that there is a linear order < ∈ L(K)N such that PiΓ∆ iff (if x ∈ V(Γ) and y ∈ V(∆) then x <i y). For every i ∈ N, we construct an order <◦ i over the set K◦ = {x ∈ K (cid:94)Nx ∈ Γϕ} such that x <◦ i y iff x (cid:74)i y ∈ Γϕ. (Note that the reason we restrict the preliminary construction of the preference order to the set of possible outcomes is because the language is not strong enough to talk about impossible outcomes. See Remark 4. A careless construction could lead to a relation over K that is not a linear order.) Capitalising on (unifPref ), it is immediate that <◦ i is transitive (4(<i)), antisymmet- ric (antisym(cid:48)) and total and reflexive (total(cid:48)). Then <◦ i is a linear order over K◦. It is now easy to obtain a linear order <i over K such that for all x and y in K◦ we have x <i y iff x <◦ j(cid:44)i ballotj(<) ∈ Γ. i y. This completes the proof that Mϕ is a Kripke model of SCF. Then, for every consistent formula ϕ, there is a Kripke model of SCF in which ϕ is satisfied. (cid:4) 24
cs/0701087	2	0701	2007-04-27T18:00:27	Artificiality in Social Sciences	[ "cs.MA" ]	This text provides with an introduction to the modern approach of artificiality and simulation in social sciences. It presents the relationship between complexity and artificiality, before introducing the field of artificial societies which greatly benefited from the computer power fast increase, gifting social sciences with formalization and experimentation tools previously owned by "hard" sciences alone. It shows that as "a new way of doing social sciences", artificial societies should undoubtedly contribute to a renewed approach in the study of sociality and should play a significant part in the elaboration of original theories of social phenomena.	cs.MA	cs	Artificiality in Social Sciences Jean-Philippe Rennard Grenoble Graduate School of Business Jp at rennard.org This a draft version of a paper to be published in: Rennard, J.P. (Ed.), Handbook of Research on Nature Inspired Computing for Economics and Management, Hershey, IGR, 2006. Abstract: This text provides with an introduction to the modern approach of artificiality and simulation in social sciences. It presents the relationship between complexity and artificiality, before introducing the field of artificial societies which greatly benefited from the computer power fast increase, gifting social sciences with formalization and experimentation tools previously owned by "hard" sciences alone. It shows that as "a new way of doing social sciences", artificial societies should undoubtedly contribute to a renewed approach in the study of sociality and should play a significant part in the elaboration of original theories of social phenomena. Introduction The "sciences of the artificial" deal with synthesized things which may imitate natural things; which have functions and goals and which are usually discussed in terms of imperatives as well as descriptives. Imitation with computer is now usually termed simulation and is used to understand the imitated system (Simon, 1996). Artificiality has invaded science over the last thirty years and physicists, chemists or biologists now daily use widespread computing tools for simulations. Social sciences did not set this trend aside (Halpin, 1999). This chapter will first introduce the essential link between complexity and artificiality before presenting the highly promising field of artificial societies. Complexity and artificiality Since the seminal book of Herbert Simon in 1969 (Simon, 1996), the sciences of the artificial knew a jerky evolution. In the field of artificial intelligence, the excessive ambitions of the sixties were considerably lowered in the seventies, before knowing a new wave of optimism in the mid eighties. The renewed interest toward artificiality originates in new approaches of artificial intelligence and in the success of the highly innovative related fields of artificial life (Langton, 1989) and artificial societies (Gilbert & Conte, 1995; Epstein & Axtell, 1996). Artificial life is at the crossroad of the rebirth of artificiality and offers lots of nice examples illustrating this revival, like this one: Many ant species tend to form piles of corpses (cemetery) in order to clean their nest. Experiments with different species showed that if corpses are randomly distributed, ants tend to gather them in some clusters within few hours. Deneubourg, Goss, Franks, et al. (1991) proposed a simple model of corpses gathering (see also Bonabeau, Dorigo, & Théraulaz, 1999). They designed virtual ants having the following behaviors: • f pp k k with The probability for an ant to pick up a corpse is )) ( ( + / = 1 1 threshold constant and f the fraction of perceived corpses in the neighborhood. 1k a 2 • 2k a dp f k f with The probability for an ant to deposit a corpse is: ( ( )) 2 = / + 2 f as the threshold constant. Deneubourg, Goss, Franks, et al. (1991) computed number of items perceived during the last t periods divided by the largest number of items that can be encountered during the last t periods. To put it simply, virtual ants tend to pick-up isolated corpses to drop them in dense zones. The result (see figure 1) is close to the real phenomenon. << FIGURE 1 >> Figure 1. Virtual ant cemetery Highly simple virtual individuals ("agents") without any knowledge of the global process, manage to carry out cemetery building. Furthermore, it "suffices" to define different types of objects to obtain sorting capabilities, like for example larval sorting observed in anthills. The gathering or the sorting process emerges from the interactions of simple agents. Emergence Emergence can be defined as the qualities or properties of a system which are new compared with the qualities or properties of the components isolated or differently organized (Morin, 1977). According to Gilbert (1995b): "Emergence occurs when interactions among objects at one level give rise to different types of objects at another level. More precisely, a phenomenon is emergent if it requires new categories to describe it that are not required to describe the behavior of the underlying components. For example, temperature is an emergent property of the motion of atoms. An individual atom has no temperature, but a collection of them does." Most authors consider that emergence relies on three conditions: 1. The global process is distributed, there is no central control and the result depends on the interactions between components. 2. The process is autonomous, there is no external controller. 3. The process is not at the same level as the components. The language or concepts used to describe the emergent process are different from the language or concepts used to describe the components. The "Test of emergence" thus relies on the 1L used to design surprise engendered by the difference between the language 2L , used to describe the resulting process (Ronald, components, and the language Sipper, & Capcarrère, 1999). According to Steels (1997), this change of language is sufficient to characterize emergence. We can clarify this concept with the classical and very general formalization proposed by (Baas, 1994), which is based on three elements: 1. 1J some index set finite or not. A set of first order structures J { }iS i , ∈ with 1 1 1 1 First order structures are primitive objects, abstract or physical; they can be organizations, machines as well as fields or concepts. 2. An observational mechanism Obs . 3. Interactions Int . The new kind of structure resulting from the observed interactions of first order structures is: , where R R S Obs Int S interaction process and the the result of is ) ( 2 1 1 1 = , , i 1 J ∈ 1 i 1 N N 1 − N 1 − 2S a second-order structure, and , ∈ families. Obs Obs S }iS i J . Baas calls 1( )i { 2 1 ≡ 2 2 1 2 2S can be measured with an The properties of the new unity resulting from the collection 2Obs . Then P is an emergent property of 2S iff observational mechanism P Obs S for all 1i . P Obs S 2( ) , but 1( )i 2 2 ∈ ∉ 1 The property P belongs to the emergent structure 2S , but is absent from the components. 2S ’s can interact to form a third order structure, and so on. A N-th order structure is The then: J i Int Obs S R S ) ( N 1 − ∈ , , = , N N i 1 1 − − N 1 − Baas calls it a hyperstructure and he considers that "complexity often takes the form of a hyperstructure." (Baas, 1994, p.525, original italics). According to Simon (1996), hierarchies are necessary to allow the evolution of complex structures. Baas distinguishes two different types of emergence: Deductible or computable emergence: A process or theory D exists which • P Obs S iS Obs Int allows to determine . That is typically the case of from ) 2( ) ( 2 1 1 1 ∈ , , 1 engineering constructions or "trivial emergence" like temperature evoked above. • Observational emergence: the emerging property P cannot be deduced (e.g. consequences of Gödel’s theorem). Bedau (1997) considers less drastically, that weak emergence characterizes emerging properties that can only be derived by simulation. Most of the recent modeling works deal with this type of weak emergence (Chalmers, 2002). Despite the thousands of pages published on emergence, or the recent emphasis on the reduction principle (the macrobehavior is reducible to the interactions of the components), (e.g. Holland, 1999; Kubik, 2003)), we are still far from an ontological concept of emergence (Emmeche, Koppe, & Stjernfelt, 1997), but, considering its success, emergence is undoubtedly epistemologically a fertile concept. Bottom-up modeling Emergence is a key feature of those famous non-linear systems which are said to be more than the sum of their parts (Waldrop, 1992). Non-linear systems do not obey the superposition principle —the linear combination of solutions is not a solution. Their dynamic cannot be reduced to the simple (linear) combination of their components ones. We have known since at least the end of the nineteenth century and Henri Poincaré (Poincaré, 1892), that the dynamic of such complex systems is unpredictable. The only way to know their state at a given step is to compute each step. The usual analytical method is of few help; the necessary mathematics are still to be invented. Even the small body of mathematics which directly deals with non-linearity depends upon linear approximations (Holland, 1999). Non- linearity thus challenges the traditional approach which tries to understand a system by analyzing its components: "The key feature of non-linear systems is that their primary behaviors of interest are properties of the interactions between parts, rather than being properties of the parts themselves, and these interactions-based properties necessarily disappear when the parts are studied independently." (Langton, 1989, p.41, original italics). How to deal with emergence? How to study processes which are "more than the sum of their parts"? How to analyze properties that cannot be forecasted? The solution proposed by computer scientists is termed bottom-up modeling. Bottom-up modeling is a very new way of building artificial systems. Since core properties disappear when the components are studied independently, bottom-up modeling is based on the gathering of interacting components. Corpses clustering or larval sorting models are then based on the building of rather simple agents (see below) which interact both with one another and with the environment. Such constructions and the study of the dynamic resulting from non-linear interactions of the simple components constitute the "bottom-up method". Instead of modeling the global dynamic of the studied system ("top-down method" usually based on differential equations) one merely models the components to study the potentially emerging regularities. This synthetic method is at the heart of the revival of artificiality. Commenting the first workshop on artificial life, C. Langton stated: "I think that many of us went away […] with a very similar vision, strongly based on themes such as bottom-up rather than top-down modeling, local rather than global control, simple rather than complex specifications, emergent rather than prespecified behavior, population rather than individual simulation, and so forth." (Langton, 1989, p.xvi, original italics). The 19th century ended with Poincaré’s discovery of the limits of the analytical method faced with non-linear systems. The 20th century ended with the unprecedented quick spread of a machine able to deal with these systems. Computers are in fact surprisingly adapted to the analysis of non-linear systems. Besides their ability to iteratively compute equations which do not have analytical solutions, computers —particularly since the development of object oriented programming— can easily deal with populations of interacting agents, so contributing to the study of Bedau’s weak emergence. "(…) Computer-based models offer a halfway house between theory and experiment [(…) and computer-based non-linear modeling] will certainly improve our understanding of emergence." (Holland, 1999, p.232). Bottom-up modeling is based on the interactions of (usually) simple virtual individuals. It massively uses multi-agent systems (MAS). Multi-Agent Systems MASs originate in Distributed Artificial Intelligence (DAI) and in artificial life. The basic idea of DAI is that intelligence is not only a matter of phenotype (brain) but also depends on the interactions with other individuals. Intelligence has a "social dimension" (Drogoul, 2005). The emergence of DAI is directly linked to the limits of the traditional symbolic AI (GOFAI) which tries to embed intelligence in a unique entity. The cognitive school of DAI associates a few complex agents to obtain some kind of group expertise (see e.g. Demazeau & Müller, 1991). The reactive school of DAI is more original. Strongly rooted in artificial life, it uses the insect (and animal) societies metaphor to try to obtain emergent intelligent behaviors by associating simple ("sub-cognitive") agents (Steels, 1990; Deneubourg, Goss, Beckers, & Sandini, 1991). We have seen that cemetery building was modeled with "virtual insects" i.e. some software processes that imitates insects’ behaviors. These virtual insects are agents. Jacques Ferber, one of the founders of the field, considers that an agent is a physical or virtual entity (Ferber, 1999): • • • goals. • having its own resources; but these resources depend on the environment. Agents are then open systems since they find resources in the environment, and close system, since they manage the use of these resources. • having a partial representation of their environment. An agent thus do not have to "fully understand" its environment; above all it does not have to perceive the global result of its actions. capable of acting. capable of communicating with other agents. driven by a set of tendencies. Autonomous agents act according to their own • possessing skills. • possibly able to reproduce itself. • tending to act according to its objectives. "The agent is thus a kind of ’living organism’, whose behavior, which can be summarized as communicating, acting and perhaps, reproducing, is aimed at satisfying its needs and attaining its objectives, on the basis of all the other elements (perception, representation, action, communication and resource) which are available to it." (Ferber, 1999, p.10). Ferber’s definition is restrictive and one can limit the characterization of agents to the following core properties (Wooldridge & Jennings, 1995): • autonomy: agents operate according to their own control. • social ability: agents can interact with one another through some kind of language. • reactivity: agents can perceive their environment and react according to its change. • pro-activness: agents act according to their own goals. Figure 2 summarizes the structure of an agent. << FIGURE 2 >> Figure 2. An agent in its environment Bottom-up modeling uses interacting agents by building multi-agent systems (MAS). A MAS contains the following elements (Ferber, 1999): An environment E ; a set of objects O having a specific position in the environment; a set of agents A with A O⊆ ; a set of relations R linking the objects to each other; a set of operations Op allowing the agent to "perceive, produce, consume, transform and manipulate" objects; operators able to apply the operations and to process the reaction of the environment. MASs and Agent Based Modeling (ABM) are the base of social simulation (see e.g. the Iterated Prisoners Dilemma—IPD (Axelrod, 1984, 1997)) and artificial societies (Conte, Gilbert, & Sichman, 1998). Artificial Societies How to connect virtual agents with human societies? Humans are quite different from ants and despite real progress —thanks to the quick growth of computer power— the intelligence of the most sophisticated agent ever programmed cannot be compared to human intelligence. The 2005 Nobel Prize in Economics was attributed to Thomas C. Schelling (along with Robert J. Aumann) who proposed in 1971 (Schelling, 1971, 1978) a far ahead of one’s time experiment, which will help us understand the link between agents and human societies. The seminal model of Thomas Schelling Schelling wanted to understand the pre-eminence of geographical segregation between black and white in American cities despite the fact that when they are questioned, citizens refute any desire of segregation. He designed very simple agents of two distinct colors ("black and white"), having the following abilities: • Each agent can compute the fraction of neighbors having the same color. • If this fraction is below the agent preference, then the agent moves to an unoccupied place which satisfies its preference. Schelling used cellular automata to implement its experiment. Very briefly, cellular automata are lattice of sites whose states —belonging to a finite set— evolve in discrete time step according to rules depending on the states of the neighbors sites. In a two dimensions implementation, Schelling used a "Moore" neighborhood, i.e. neighbors are the eight closest squares. The rules were: • If an agent has two neighbors, it will not move if at least one is of the same color. • If an agent has three to five neighbors, it will not move if at least two are of the same color. • If an agent has six to eight neighbors, it will not move if at least three are of the same color. These rules are compatible with a fully integrated structure. The initial state of Schelling (see figure 3) is thus an attractor since no agent needs to move. Schelling showed that a slight perturbation of this initial state is sufficient to give rise to a dynamic quite inevitably leading to segregation (see figure 3). << FIGURE 3 >> Figure 3. Schelling’s model Schelling’s model clearly demonstrates that local interactions (micromotives) lead to global structures (macrobehavior, (Schelling, 1978)). More important, he showed that the macrobehavior can be different from the underlying micromotives, since segregation occurs even when preference rules are compatible with integrated structure. Nowak and Latané (1993) used an extended model to study Dynamic Social Impact i.e. the change of attitudes or beliefs resulting from the action of other individuals. They notably showed that the system achieved stable diversity. The minority survived, thanks to a clustering process of attitudes, not because individuals moved, but due to the attitude change process. (Latané, 1996). The observed macrobehaviors are very robust. Schelling’s and Latané’s models were tested under a wide range of parameters and quite always evolve towards the same type of attractors. Pancs and Vriend (2003) recently enlarged the study of segregation process showing that it tends to occur even if people are anxious that segregation should not occur. Both these examples show that some complex social dynamics can be modeled from simple basis: "(…) there is a spirit in the air which suggests that we should look for simple explanations of apparent complexity." (Gilbert, 1995b). Stephen Wolfram recently brought a strong justification to this quest for simplicity (Wolfram, 2002). Its Principle of Computational Equivalence states that: "(…) almost all processes that are not obviously simple can be viewed as computations of equivalent sophistication. (…) So this implies that from a computational point of view even systems with quite different underlying structures (…) can always exhibit the same level of computational sophistication. (…) And what it suggests is that a fundamental unity exists across a vast range of processes in nature and elsewhere: despite all their detailed differences every process can be viewed as corresponding to a computation that is ultimately equivalent in its sophistication." (Wolfram, 2002, pp.717- 719). Without going as far as Wolfram, it is now clear that at least some social phenomena can be modeled with interacting sub-cognitive agents. • The Newtonian model uses systems of differential equations to study equilibrium; the best example being equilibrium theory in economics —which is also a brilliant example of the consequences of oversimplification motivated by the will to obtain tractable equations; the results having few to do with reality. • Considering the difficulty to write the equations of the system, the statistical model tries to discover regularities; the best example being the study of "social forces" by Durkheim in 1897 (Durkheim, 2004). Schelling’s or Latané’s models are then quite a new way of doing social sciences based on virtual experiments inside artificial societies. Artificial Societies as a new way of doing social sciences The field of artificial societies is based on the strong assumption that human societies are complex systems (Goldspink, 2000). Analysis is unable to point the source of macro- properties since there is no localized source, but a distributed process which obliges to consider the system as a whole (Goldspink, 2002). Furthermore, they are complex adaptive systems (CAS) i.e. systems where agents can learn and modify their rules according to their previous success (that is of course also the case of animal or insect societies, but the specificity of human —cognitive— societies is that they can also learn from their failures). Schelling’s segregation process or Nowak and Latané’s clustering process of people sharing the same opinion are emergences or "regularities at the global level" (Gilbert, 1995a). "As the number of elements and interactions of a system is increased, we can observe an emergent complexity. But somehow, regularities arise and we can observe emergent simplicity (Gershenson, 2002, original italics). Artificial societies then try to obtain emergent regularities: "(…) the defining feature of an artificial society model is precisely that fundamental social structures and group behaviors emerge from the interaction of individual agents operating on artificial environments (…)."(Epstein & Axtell, 1996, p.6, original italics). Considering European contributions to social modeling, Gilbert wrote: "One of the major objectives of the approach being reviewed here is to generate through simulation, emergent phenomena and thus to understand and explain the observable macro-level characteristics of societies." (Gilbert, 2000). This is quite a new way of doing science; so new that simulation is said to be "a third way of doing sciences" (Axelrod, 2006) different from deduction and from induction. In the fields of artificial intelligence and artificial life, Luc Steels termed it the synthetic method (see figure 4) (Steels & Brook, 1994). << FIGURE 4 >> Figure 4. Inductive vs. synthetic method Induction starts from observed facts and uses inferences to build a theory potentially able to globally explain the observed facts. The theory is then validated through the test of predicted facts. The synthetic method starts like induction from the observed facts and the inferred theory (but it can also start like deduction from a set of assumptions). On this basis, the synthetic method engineers an artificial system, the objective being that, while operating, this system will behave like the real one, thus confirming the tested theory. In their seminal work, Epstein and Axtell (1996) considered that artificial societies models may change the way we think about explanation in the social sciences. "Clearly, agent-based social science does not seem to be either deductive or inductive in the usual senses. But then what is it? We think generative is an appropriate term. The aim is to provide initial microspecifications that are sufficient to generate the macrostructures of interest." (Epstein & Axtell, 1996, p.177). This generative interpretation is directly linked to the disjunction between determinism and predictability which is a huge epistemological consequence of complexity sciences. Even if we perfectly understand the concerned forces, we are unable to predict the evolution of the system (Croquette, 1997). A high potential to stimulate novelty Agent based modeling is potentially a highly powerful tool for social scientists. Axelrod and Tesfatsion (forthcoming) recently synthesized its goals with four forms: • Empirical understanding: why have regularities emerged? • Normative understanding: how can models help to define norms/design? How to know if a given decision is positive for the society? • Heuristic: How to attain greater insight about fundamental mechanisms in social systems? • Methodological advancement: How to give researchers the method and tools to rigorously study social systems? Practically, these four forms rely on three pillars: Formalization, experiments and ability to study the macro to micro problem. the good Formalization Apart from the verbal and mathematical symbol systems, computer simulation can be considered as the "third symbol system" (Ostrom, 1988). Any theory originating in the first two models can be expressed in the third one. Simulation can then be considered as formal models of theories (Sawyer, 2004). That is an important point since computer symbols are more adapted to social sciences than mathematical ones (Gilbert & Troitzsch, 2005, pp.5-6): • Programming languages are more expressive and less abstract than mathematical techniques. • Programs deal more easily with parallel processing. • Programs are modular. Major changes can easily be made, that is not the case of mathematical systems. Computer modeling thus helps social scientists to formalize their theories. The difficulty — not to say the impossibility— to mathematically formalize many social sciences theories is considered to be a great weakness by "hard" scientists. This inability is closely linked to the inability of mathematics to deal with distributed emergent processes. Computer modeling can thus contribute to give social sciences some of the scientific tools they need to rigorously express their theoretical models. Experiments Simulation can be considered as a new experimental methodology. Gilbert and Conte (1995) defined it as "exploratory simulation". Such explorations can contribute to social sciences notably in the following ways: • Modeling allows a culture-dish methodology. The modeler designs the agents and the initial state of its society and studies its temporal evolution (Tesfatsion, 2002). Any sort of experiments can be carried out since the modeler has a complete control on the model. It is then possible to study the consequences of any given modification. This will notably contribute to the analysis of the minimal set of parameters and system characteristics necessary to give rise to a given behavior as well as to the analysis of (Goldspink, 2002). the attractors of dynamic social systems The ability to carry out experiments is something very new for social scientist that usually cannot test their theory in the field. Like formalization this contributes to bring closer social and "hard" sciences methods. • Modeling is potentially able to contribute to original discoveries. The same way the classification of cellular automata permitted to propose an original analysis of complex systems (Wolfram, 1984; Langton, 1990), simulations can play a role in the discovery of general, yet unattainable, laws. Implicit unknown effects can be detected (Gilbert & Conte, 1995). This ability to stimulate discovery does not only stand on the possibility to carry out otherwise impossible experiments, but also on the capacity of emergent modeling to give rise to original cognitive processes. In the field of artificial life, Cariani (1992) emphasizing non-stochastic models like the Game of Life, pointed out the fact that emergence relies on a cognitive process; a process is emergent only according to its observer: "The interesting emergent events that involve artificial life simulations reside not in the simulations themselves, but in the way that they change the way we think and interact with the world. Rather than emergent devices on their own right, these computer simulations are catalyst for emergent processes in our minds; they help us create new ways of seeing the world." (Cariani, 1992, p.790, original italics). • Modeling can go beyond some of the limits of the statistical tools usually used by social scientists, e.g. qualitative changes can be analyzed through simulation (see (Pyka, 2006)). Simulation also helps the study of processes. Usual statistical analyses study the correlations between variables at a given time. Simulations embed the processes which lead to these correlations (Gilbert & Troitzsch, 2005). Since social systems are fundamentally dynamic, simulation allows formalizing processes beyond the scope of statistical analysis. Furthermore, statistic is based on linearity assumptions which oblige to over simplify the observed facts. Simulation does not suffer from this limit. • limits of mathematical technical the is not concerned by Modeling formalization. For example, mathematical formalization obliges to consider agents as equivalent whereas simulation is able to manage heterogeneous population. In the same vein, simulation allows to relax assumptions necessary to obtain tractable equations (e.g. the rationality of economic agents). In economics, the highly promising field of Agent-Based Computational Economics (ACE) (Tesfatsion, 2002) clearly illustrates the potential of simulations. • More generally, the same way artificial life allows the study of "Life as it could be" (Langton, 1989), artificial societies allow the study of "Societies as they could be" (Gilbert, 2000), thus giving social sciences an unprecedented tool to understand fundamental invariants (Rennard, 2004). Study of the macro to micro problem The macro to micro problem—how to describe the relationship between macro-phenomena characterizing the dynamic of a system as a whole and micro-phenomena characterizing the dynamic of the components of the system— is a central issue of social sciences, but also of DAI (Schillo, Fischer, & Klein, 2000). Simulation is a ground-breaking tool to study the core problem of the micro/macro relations. The relations between different levels (individual, organization, societal) and the potential associated lock-in can be studied. Artificial life with its widely studied concept of Dynamical hierarchy which "refers to a system that consists of multiple levels of organization having dynamics within and between the entities described at each of the different levels." (Lenaerts, Chu, & Watson, 2005, p.403), should contribute to this study. Simulation can be used to study both the micro to macro and the macro to micro problems (Sawyer, 2003). Schelling’s or Latané’s models thus show how regularities can arise from micro-interactions. But such models also show that these regularities then constraint the system and impact the behaviors of individual agents. More directly, it is possible to conceive simulations that specifically study the impact of macro-phenomena. For example, Axtell (2000) while studying retirement behaviors, showed that modifying the sole network connections between agents can lead to great changes of the overall society behavior. The study of the micro/macro problem through simulation remains nevertheless very difficult while studying societies. In fact, humans are not limited to basic behavior, they notably have the ability to grasp macro-level phenomena and they can adjust their behavior according to this. That is what Gilbert (2000) terms second order emergence, characterizing systems where agents can detect and react to emergent properties. Models should then embed both the emergence of macro-properties and the ability to deal with the effects of these macro- properties on self-aware individuals. This remains a challenge (Gilbert, 1995b). Limits Artificial societies is a very recent field in which huge problems still are to be solved that challenges these researches. A first set of problems relies on the cognitive dimension of human societies. Guided by the success of artificial life, many artificial societies are based on reactive DAI, one of the most famous example being the Sugarscape of Epstein and Axtell (1996). The complexity of human cognition has a deep impact on the structuring of societies. • Self-awareness and the related second order emergence should be modeled. • Interpretativism in sociology leads to the idea that meanings are parts of the actions. “(…) meanings and concepts describing both the physical and the social world are said to be socially constructed by members of society” (Gilbert, 2000). Simulations should then embed the corresponding social constructions. As a consequence, artificial societies must find a way to associate cognitive and reactive DAI. This remains both a theoretical (how to build cognitive agents) and a practical (how to have sufficient computing power) problem. A second set of problems is linked to the tools and methods used for modeling and simulation. First of all, simulation uses tools that may make implicit assumptions having nothing to do with the tested theory. For example, the use of cellular automata assumes that the world is a regular grid, which may have massive consequences on the global dynamic of the simulation (Troitzsch, 1997). Then simulation tends to develop its own finality, hence the importance to ground it in social theories in order to avoid the trend to develop simulations for themselves and to mistake them for reality. The balance is difficult to find: "If our ’toy models’ serve only to reify and naturalize the conventional social science wisdom, then they are indeed a Medusan mirror, freezing the victim by the monster’s glance" Lansing (2002, p.289). The gap between social sciences and computer sciences also challenges the field. Some social sciences theories are mainly descriptive and discursive and such approaches may be very difficult to formalize through simulation. Moreover, despite common issues, the discussion between computer scientists and social scientists remains very difficult. For computer scientists, non formalized discursive social theories often seem blurred and they have difficulties in understanding them. Social scientists are often reluctant facing computer programming and they usually consider that computer scientists do not understand the complexity of human societies. Finally, the core problem (which is not limited to artificial societies) of "how to obtain from local design and programming, and from local actions, interests, and views, some desirable and relatively predictable/stable emergent results" (Castelfranchi, 2000, original italics) still remains to be solved. Conclusion The field of artificial societies, despite old roots, is now only ten years old. Along with artificial life, it participates to an emerging way of doing science. This way still has to reach maturity, but will undoubtedly contribute to complement more traditional methods. The debate now is not to choose between usual methods and methods originating in artificiality, but to convince "traditional" scientist that artificiality is not limited to some kind of, possibly funny, computer game and to find ways of building stronger bridges between these practices of science. The growing easiness of computer programming and the quick spread of computer culture among young scientists is potentially a promise of quick evolution of artificiality in social sciences; no doubt this will contribute to renew the field. References Axelrod, R. (1984). The evolution of cooperation. New York: Basic Books. Axelrod, R. (1997). The complexity of cooperation: Agent-based models of competition and collaboration. London: University College London Press. Axelrod, R. (2006). Advancing the art of simulation in the social sciences. In J.-P. Rennard (Ed.), Handbook of research on nature inspired computing for economics and management. Hershey, PA: IGR. Axelrod, R., & Tesfatsion, L. (forthcoming). A guide for newcomers to agent-based modeling in the social sciences. In L. Tesfatsion & K. Judd (Eds.), Handbook of computational economics, vol 2: Agent-based computational economics. Amsterdam: North-Holland. Axtell, R. L. (2000). Effect of interaction topology and activation regime in several multi-agent systems (Tech. Rep. No. 00-07-039). Santa Fe Institute. Baas, N. A. (1994). Emergence, hierarchies, and hyperstructures. In C. Langton (Ed.), Artificial life 3 (p. 515-537). Reading, MA: Addison-Wesley. Bedau, M. A. (1997). Weak emergence. In T. J. (Ed.), Philosophical perspectives: Mind, causation and world (p. 375-399). Malden MA: Blackwell. Bonabeau, E., Dorigo, M., & Théraulaz, G. (1999). Swarm intelligence. from natural to artificial systems. Oxford: Oxford University Press. Cariani, P. (1992). Emergence and artificial life. In C. Langton (Ed.), Artificial life 2 (p. 775-797). Redwood city: Addison-Wesley. Castelfranchi, C. (2000). Engineering social order. In LNCS, proceedings of the first international workshop on engineering societies in the agent world (Vol. 1972, p. 1-18). London: Springer. Chalmers, D. J. (2002). Varieties of Emergence. Retrieved 12/15/2005, 2005, from http://consc.net/papers/granada.html Conte, R., Gilbert, N., & Sichman, J. (1998). MAS and social simulation: A suitable commitment. In J. Sichman, R. Conte, & N. Gilbert (Eds.), Proceedings 1st. international workshop on multi-agent based simulation (Vol. 1534, p. 1-9). Berlin: Springer Verlag. Croquette, V. (1997). Déterminisme et chaos. In Pour la Science (Ed.), L’ordre du chaos (p. 64-87). Paris: Belin. Demazeau, Y., & Müller, J.-P. (1991). Decentralized AI 2. Amsterdam: Elsevier Norh Holland. Deneubourg, J.-L., Goss, S., Beckers, R., & Sandini, G. (1991). Collectively self-solving problems. In A. Babloyantz (Ed.), Self-organization, emergent properties and learning. New York: Plenum. Deneubourg, J.-L., Goss, S., Franks, N., Sendova-Franks, A., Detrain, C., & Chretien, L. (1991). The dynamics of collective sorting: Robots-like ant and ant-like robots. In J. Meyer & S. Wilson (Eds.), Proceedings first conference on simulation of adaptive behavior: From animals to animats (p. 356- 365). Cambridge, MA: MIT Press. Drogoul, A. (2005). Les sytèmes multi-agents. In J. Lautrey & J.-F. Richard (Eds.), L’intelligence. Paris: Hermès. Durkheim, E. (2004). Le suicide. Paris: PUF. Emmeche, C., Koppe, S., & Stjernfelt, F. (1997). Explaining emergence. Journal of General Philosophy of Science, 28, 83-119. Epstein, J. M., & Axtell, R. L. (1996). Growing artificial societies. Social sciences from the bottom up. Cambridge, MA: MIT Press. Ferber, J. (1999). Multi-agent systems. London: Addison-Wesley. Gershenson, C. (2002). Philosophical ideas on the simulation of social behaviour. Journal of Artificial Societies and Social Simulation, 5 (3), http://jasss.soc.surrey.ac.uk/5/3/8.html. Gilbert, N. (1995a). Emergence in social simulation. In N. Gilbert & R. Conte (Eds.), Artificial societies. the computer simulation of social life (p. 144-156). London: UCL Press. Gilbert, N. (1995b). Simulation: An emergent perspective. Lecture given at the Conference on New Technologies in the Social Sciences. Bournemouth, UK, 27-29 October, 1995. Gilbert, N. (2000). Modelling sociality: The view from europe. In T. Kohler & G. Gumerman (Eds.), Dynamics in human and primates societies: Agent-based modelling of social and spatial process (p. 355-372). Oxford: Oxford University Press. Gilbert, N., & Conte, R. (1995). Artificial societies. the computer simulation of social life. London: UCL Press. Gilbert, N., & Troitzsch, K. (2005). Simulation for the social scientist. Maidenhead: Open University Press. Goldspink, C. (2000). Modelling social systems as complex: Towards a social simulation meta-model. Journal of Artificial Societies and Social Simulation, 3 (2), http://www.soc.surrey.ac.uk/JASSS/3/2/1.html. Goldspink, C. (2002). Methodological implications of complex systems approaches to sociality: Simulation as a foundation of knowledge. Journal of Artificial Societies and Social Simulation, 5 (1), http://www.soc.surrey.ac.uk/JASSS/5/1/3.html. Halpin, B. (1999). Simulation in society. American Behavioral Scientist, 42, 1488-1508. Holland, J. (1999). Emergence. Cambridge, MA: Perseus books. Kubik, A. (2003). Toward a formalization of emergence. Artificial Life, 9, 41-65. Langton, C. G. (1989). Artificial life. In C. G. Langton (Ed.), Artificial life 1. Redwood city:Addison- Wesley. Langton, C. G. (1990). Computation at the edge of chaos: Phase transition and emergent computation. Physica D, 42, 12-37. Lansing, J. (2002). "artificial societies" and the social sciences. Artificial Life, 8, 279-292. Latané, B. (1996). Dynamic social impact: Robust predictions from simple theory. In R. Hegselmann, U. Mueller, & K. Troitzsch (Eds.), Modelling and simulating in the social sciences from a philosophy science point of view (p. 287-310). Kluwer. Lenaerts, T., Chu, D., & Watson, R. (2005). Dynamical hierarchies. Artificial Life, 11, 403-405. Morin, E. (1977). La méthode 1i. Paris: 1977. Nowak, A., & Latané, B. (1993). Simulating the emergence of social order from individual behaviour. In N. Gilbert & J. Doran (Eds.), Simulating societies: The computer simulation of social phenomena. London: UCL Press. Ostrom, T. (1988). Computer simulation: The third symbol system. Journal of Experimental Social Psychology, 24, 381-392. Pancs, R., & Vriend, N. (2003). Schelling’s spatial proximity model of segregation revisited (Dpt. of Economics, WP 487). University of London. Poincaré, H. (1892). Les méthodes nouvelles de la mécanique céleste. Paris: Blanchard. Pyka, A. (2006). Modelling qualitative development - agent based approaches in economics. In J.-P. Rennard (Ed.), Handbook of research on nature inspired computing for economics and management. Hershey, PA: IGR. Rennard, J.-P. (2004). Perspectives for strong artificial life. In L. de Castro & F. von Zuben (Eds.), Recent developments in biologically inspired computing. Hershey: Idea Group Inc. Ronald, E., Sipper, M., & Capcarrère, M. (1999). Design, observation, surprise! a test of emergence. Artificial Life, 5 (3), 225-239. Sawyer, R. (2003). Artificial societies: Multiagent systems and micro-macro link in sociological theory. Sociological Methods and Research, 31 (3), 325-363. Sawyer, R. (2004). Social explanation and computational simulation. Philosophical Explorations, 7, 219-231. Schelling, T. C. (1971). Dynamic model of segregation. Journal of Mathematical Sociology, 1 (2), 143- 186. Schelling, T. C. (1978). Micromotives and macrobehavior. New York: Norton. Schillo, M., Fischer, K., & Klein, C. (2000). The micro-macro link in DAI and sociology. Lecture Notes in Artificial Intelligence, 1979. Simon, H. (1996). The sciences of the artificial (3rd ed.). Cambridge, MA: MIT, Press. Steels, L. (1990). Cooperation between distributed agents through self-organization. In Y. Demazeau & J.-P. Müller (Eds.), Decentralized AI (p. 175-196). Amsterdam: Elsevier. Steels, L. (1997). Artificial life roots of artificial intelligence. In C. Langton (Ed.), Artificial life. An overview (p. 75-110). Cambridge, MA: MIT Press. Steels, L., & Brook, R. (1994). The artificial life route to artificial intelligence. New Haven, CT: Lawrence Eribaum Ass. Tesfatsion, L. (2002). Agent-based computational economics: Growing economies from the bottom up. Artificial Life, 8, 55-82. Troitzsch, K. G. (1997). Social simulation – origins, prospects, purposes. In R. Conte, R. Hegselmann, & P. Terna (Eds.), Simulating social phenomena (Vol. 456, p. 41-54). Berlin: Springer-Verlag. Waldrop, M. (1992). Complexity: The emerging science at the edge of chaos. New York: Simon and Schuster. Wolfram, S. (1984). Universality and complexity in cellular automata. Physica D, 10, 1-35. Wolfram, S. (2002). A new kind of science. Champain: Wolfram Media Inc. Wooldridge, M., & Jennings, N. (1995). Intelligent agents: Theory and practice. Knowledge Engineering Review, 10 (2), 115-152. Figure 1. Virtual ant cemetery Figure 2. An agent in its environment Figure 3. Schelling’s model Adapted from (Steels & Brook, 1994). Figure 4. Inductive vs. synthetic method
1904.08315	1	1904	2019-03-22T17:57:44	Multi-Level Mesa	[ "cs.MA" ]	Multi-level Mesa is an extension to support the Python based Agents Based Model (ABM) library Mesa. Multi-level Mesa provides ABM infrastructure to allow for the inclusion of complex networks, which have modules (groups) and hierarchies (layers) of agents. This approach allows for users to define and simulate multi-layered adaptions of complex networks. This study reviews other multi-level libraries currently in the field, describes the main functions and classes of the Multi-level Mesa, and describes its implementation and impact in numerous varieties using the seminal ABM - Sugarscape. Multi-level Mesa and Sugarscape examples are available on GitHub at https://github.com/tpike3/multilevel_mesa and https://github.com/tpike3/SugarScape.	cs.MA	cs	Multi-Level Mesa By Thomas Pike PhD Candidate, Computational Social Science George Mason University [email protected] Abstract: Multi-level Mesa is an extension to support the Python based Agents Based Model (ABM) library Mesa. Multi-level Mesa provides ABM infrastructure to allow for the inclusion of complex networks, which have modules (groups) and hierarchies (layers) of agents. This approach allows for users to define and simulate multi-layered adaptions of complex networks. This study reviews other multi-level libraries currently in the field, describes the main functions and classes of the Multi-level Mesa, and describes its implementation and impact in numerous varieties using the seminal ABM - Sugarscape. Multi-level Mesa and Sugarscape examples are available on GitHub at https://github.com/tpike3/multilevel_mesa and https://github.com/tpike3/SugarScape. Multi-Level Mesa provides ABM infrastructure to support modules and hierarchies. Modularity is the concept that clusters of linked nodes within a network can effectively act as a single node, and can also be known as communities or building blocks (Barabasi, 2016; Holland, 1995). Hierarchies represent the layers of emergence which can occur within a complex system. For example, using an individual as the focal point, hierarchies go smaller from the individual, as a human body is comprised of organs which are comprised of cells, which are comprised of chemicals and so on. Or, larger from the individual, as humans make up families, which make up neighborhoods, which make up towns and so on (Miller & Page, 2007). Modules and hierarchies are a critical part of complex adaptive systems as they provide sub-assemblies to retain working systems while enabling adaption (Holland, 1995; Simon, 1997). As an essential aspect of complex systems modules and hierarchies are critical to analyze. Object Oriented Programming (OOP) inherently consists of modularity and hierarchies and exploiting this capability is what provides greatest advantage to OOP languages (Booch et al., 2007). OOP provides analysts the ability to capture specific modules, hierarchies and processes of complex adaptive systems. ABM platforms and coding libraries then exploit the properties of OOP by providing ABM infrastructure. This infrastructure reduces the cost of the modeler who can focus on simulating his or her phenomenon of interest and not on writing code which manages the interactions of the phenomenon. ABMs, however, typically stop at two levels of interaction. Agents produce the bottom-up emergent behavior of the next level, but no further hierarchies are produced (Haman Tchappi, Galland, Kamla, & Kamgang, 2018; Morvan, 2013). Multi- level Mesa seeks to further extend the typical ABM dynamic to enable more complex interactions where agents and groups of agents can interact across multiple hierarchies and have cascading effects across those hierarchies. The goal of Multi-level Mesa is to 1 provide methods to help manage the complex interactions of agents and modules of agents (e.g. groups) across multiple hierarchies (e.g. levels). Multi-level Mesa starts from the view of complex systems as adaptive networks and allows not only for the formation and dissolution of modules but also for active and resting modules (or neutral networks) which can interact across layers. In the taxonomy of efforts to facilitate multiple levels in ABMs, Multi-level Mesa falls in the category of generalizable coding libraries (Taillandier et al. 2012; Morvan, 2013) and is the only Python based library. Multi-level Mesa is intended to be a readily available coding library which can be employed to support models developed to understand adaptive networks. This chapter proceeds in four sections. First, a literature review of current multi- level ABM approaches. Second, a discussion of Multi-level Mesa's conceptual approach. Third, a discussion of Multi-level Mesa's methods. Fourth, an implementation of Multi- level Mesa using the Sugarscape construct developed by Rob Axtell and Joshua Epstein (1996). Literature Review As modules and hierarchies are an inherent feature of complex systems there is a rich body of literature examining them across multiple disciplines. This expansive body of literature can be broken down into three broad categories. First, are approaches which want to identify existing processes that produce emergent modules, which in turn reify and become agents at a higher hierarchy. Second, are approaches which provide computational infrastructure so others can dictate their own emergent and reification processes. Third, are attempts at linking different models together each of which is its own module and falls within certain hierarchies. Due to the wide breadth of research a complete review is impractical, instead this survey will provide a brief overview of the first category, as it has the largest amount of literature, and then focus on the existing computational infrastructure before discussing attempts to link models together. To begin, however, it is important to discuss some of the various terms. Due to the large amount of research across multiple disciplines on this subject a diverse terminology has emerged for the concept of modules and hierarchies within a complex system. Terminology The three main terms for deliberate inclusion of modules and hierarchies into ABMs are multi-level, multi-scale and holons. The term multi-scale is favored by the natural sciences but contentious to other disciplines who argue multi-scale is inaccurate (Gil-Quijano, Louail, and Hutzler 2012; Morvan, 2013). Two cities, for example, Tucson, Arizona and New York, New York are both at the same 'city-level', but are of two different population scales, 535,000 and 8.6 million, respectively ("Population in the U.S. - Google Public Data Explorer," 2018). Natural sciences may counter families and states are really just different scales of human organization in one level of the Earth's ecological hierarchy, so the proper terminology is really determined by one's perspective. The second term is holon, which has accompanying descriptors such as holarchy, for 2 discussing the hierarchies within the system, or holonic, to describe a system with modules and hierarchies (Haman Tchappi et al., 2018). Holon comes from Arthur Koestler's book Ghost in the Machine and was invented to specifically address the existence of subassemblies within complex systems based on Herbert Simon's parable of the two watchmakers (1967; 1997). The coding module adopts the term multi-level, as it is more descriptive for the user who is concerned about the levels within a specific field. Multi-level was then selected instead of holon for the simple reason the term multi-level makes its purpose more obvious to potential users. Despite these different terms, multi- scale, multi-level and holon each mean the deliberate inclusion of modules and hierarchies. There are two other terms worthy of discussion in the literature, which intersect with multi-level ABMs but also have models outside the ABM set. First is multi- modelling (also referred to as meta models). This effort can be seen as a separate but intersecting focus area. Multi-models are an effort to link two or more models of a similar phenomenon together to allow for numerous research efforts to be combined (Scerri, Drogoul, Hickmott, & Padgham, 2010; Soyez, Morvan, Dupont, & Merzouki, 2013). Some of these efforts fall under the third category discussed in this literature review, while some are wholly independent from ABMs, notably the Coupled Earth System Model which consists of four publicly available models to explore the Earth's weather system.1 The second term is hybrid ABMs, these are primarily system biology models and combine ABMs with systems dynamics where one or more levels is agent based and their actions parameterize differential equations at other levels, whose output provides inputs to the agents (Cilfone, Kirschner, & Linderman, 2015; Smallwood & Holcombe, 2006). Each of these research areas also examine complex systems at multiple hierarchies, but are specialized approaches which can be seen as overlapping. With the main terminology described, the next step is to review each of the three categories of research efforts. Category One: Processes The first category is research trying to discover or define processes for creating multiple hierarchies. This category consists of the largest amount of research and exists across multiple fields. This research tries to address the theoretical issue of generalizable mechanisms for identifying and reifying emergent phenomenon and cross-level communication (Haman Tchappi et al., 2018; Morvan, 2013; Seck & Honig, 2012). For example, when a group of bacteria form a microbial colony and begin to act as a singular entity, a group of cells form a functioning organ, or a population of people act as a single nation. The natural disciplines prefer the term multi-scale and have well developed and coordinated research efforts to try to identify emergence and reification processes, which includes government sponsored projects, working groups, tools, databases, webinars and competitions, coding platforms and modeling languages (Falcone, Chopard, and Hoekstra 2010; "Interagency Modeling and Analysis Group" 2018; Morvan, 2013; Smallwood and 1 The climate models are available at http://www.cesm.ucar.edu/models/ccsm4.0/ 3 Holcombe 2006).2 Ecology has a series of models which look at the dynamics of multiple levels within a trophic web using ABMs. Existing ecological models focus on scale (level) transfer or clustering methods to explore the dynamics of how agents coalesce and how their actions impact levels above and below them from the micro level (e.g. soil) to the macro level (e.g. an ecosystem) (Morvan, 2013). The natural sciences have extensive work examining multiple levels of complex systems with significant effort placed on understanding the emergence and reification of entities, and interdependencies between levels. Expanding beyond biology and ecology, multi-level models are represented in three areas. The first area is traffic and pedestrian models. For these models, levels are added to make the models more computationally efficient as pedestrians begin to move together as a type of flocking model (Haman Tchappi et al., 2018; Navarro, Corruble, Flacher, & Zucker, 2013). The second area overlaps with traffic and pedestrian models to examine city development. This area includes several research efforts which try and address different aspects of the multi-level problem. These aspects include identifying when new agents emerge (Camus, Bourjot, & Chevrier, 2013; Gil-Quijano et al., 2012) and how different levels and modules should interact with each other, which intersects with research into coupling models together (e.g. multi-models) (Navarro et al., 2013). The third area is organizational and has seen applications trying to manage intelligent autonomous intelligence vehicles. In this area multiple levels are used to find ways to deconflict layers within an organization, such as fleets of vehicles autonomously conducting port operations (Haman Tchappi et al., 2018; Soyez et al., 2013). These three areas show considerable cross-fertilization as they are looking at similar systems of flow and organization for different purposes. Due to the cross fertilization of the previous three areas there are general frameworks which are used and improved upon for their specific research problems. These frameworks are CRIO (Capacity, Role, Interaction, Organization) (Haman Tchappi et al., 2018) , IRM4MLS (Influence Reaction Model for Multi-Level Simulations) (Soyez et al., 2013), and AA4MM (Agents and Artifacts for Multi-Modelling) (Camus et al., 2013; Siebert, Ciarletta, & Chevrier, 2010). Interestingly, each of these approaches are proposed by French universities, who have the most research papers, outside the natural sciences, on this subject. Research into natural processes for the emergence and reification of new layers and the cross communication between layers represented the largest amount of research on multi-level ABMs. The natural sciences have the most developed research efforts to examine this problem. There is also substantial research examining these phenomena in population flow and organizational models. Although this review focused on ABMs, there are similar efforts in discrete event simulations, specifically, the DEVs models, whose evolution over time has made them more similar to ABMs (Haman Tchappi et al., 2018; Morvan, 2013; Seck & Honig, 2012) 2 A concise website containing multi-scale modelling efforts and links to models, tools and databases is located at https://www.imagwiki.nibib.nih.gov/. 4 Category Two: Computational Infrastructure The second category in the literature is computational infrastructure and is the category of Multi-level Mesa. This category has two sub-areas, ABM platforms and coding libraries. ABM platforms are characterized by their own simplified coding language to reduce the barrier of entry for non-programmers. For the ABM platforms there are three which have multi-level models. They are NetLogo3, SPARK (Simple Platform for Agent-based Representation of Knowledge)4 and GAMA5. NetLogo has an extension dedicated to multi-level models, more accurately meta-models, called LevelSpace. LevelSpace's approach is to link models together (e.g. multi-models) so the dynamics of one can update another. Examples include linking NetLogo's Wolf Sheep Predation model with its Climate Change model where climate impacts grass growth and animal flatulence impacts greenhouse gases and animals whose decision-making function is linked to neural net models (Hjorth A., Head, B., & Wilensky, U., 2015; Hjorth, Weintrop, Brady, & Wilensky, 2016). Based on the taxonomy of this chapter, LevelSpace is infrastructure for the third category, but with OOP this line between connecting models and models with hierarchies and modules is blurry at best. SPARK is a Java-based platform modelled on Netlogo, designed specifically for the use with cell biology. SPARK does not explicitly allow for the formation of hierarchies relying on the implicit nature of object-oriented programming transferred to their coding language to allow the modeler to specify their agents and meta-agents (Solovyev et al., 2010). GAMA, also a Java-based platform, is the only platform which has specific methods for the emergence of new agents formed from lower level agents (Taillandier et al., 2012). Due to applicability of this approach to Multi-level Mesa it is worth looking at GAMA and its methods in more detail. GAMA like SPARK, uses the object-oriented nature of Java to embed agents in larger groups. GAMA then proceeds further by providing explicit commands for group formation and algorithms to detect new agents. The commands for group formation include: capture, which adds agents to a group, release, which removes agents from a group, and migrate, which moves agents from one group to another ("Multi-level architecture," n.d.). GAMA also has clustering algorithms embedded within its platform which can be used to specify the use of the capture, release or migrate commands (Taillandier et al., 2012). In addition, GAMA passes properties of Java's objected- oriented language into its GAML language so users can specify different behaviors for groups and their sub-agents and access each agents' respective attributes regardless of the level. GAMA is the only ABM platform which explicitly allows for multi-level architecture within an ABM. The second area for computational infrastructure is coding libraries. As each coding library for ABMs (e.g. MASON, Repast, Mesa, FLAME, MaDKit) uses object- oriented programming, each has an implicit ability to have modules and hierarchies. Of 3 https://ccl.northwestern.edu/netlogo/ 4 http://www.pitt.edu/~cirm/spark/ 5 https://gama-platform.github.io/ 5 the existing coding libraries, three identified models consisting of multiple layers Repast6, FLAME7, and MaDKit8 (Haman Tchappi et al., 2018; Morvan, 2013; Smallwood & Holcombe, 2006). Of these three only MaDKit provides explicit infrastructure to support agents operating in multiple levels embedding what it calls the Agent, Group, Role organizational model. Within the MaDKit documentation, this manifests itself in two places, first in the agent who can be assigned to multiple groups and assigned a role in each of these groups. Second, in the network management which maintains the different groups and roles (Michel, Gutknecht, & Ferber, 2017). Although Mason9 and Repast do not have explicit infrastructure for multi-level models they have built in features which help enable multi-level models. For MASON this includes Steppable and anonymous wrappers which allow modelers to group agents together and iterate through them in a schedule and place an agent (or group of agents) in the schedule multiple times (Luke, Cioffi-Revilla, Panait, Sullivan, & Balan, 2005). For Repast it has three features to aid multiple levels, which are also based on scheduling. First, scheduling annotations where certain actions are scheduled if a trigger event occurs. Second, scheduling global behaviors in which the modeler creates a context which is filled with agents who then are scheduled to behave within that context. Third, schedule with watcher, which allows for dynamic scheduling by letting agents know if certain conditions are met so they can execute some action ("Repast Simphony Reference Manual," 2018). Of the existing platform and coding libraries only two, GAMA and MaDKit, have explicit architecture for developing multiple hierarchies and allowing interaction between them. Although other platforms and coding libraries do not have explicit methods for multiple layers and hierarchies' modelers are able to leverage their object-oriented foundation to develop their own. In addition, MASON and Repast have additional features with their respective scheduler classes which can reduce the cost of integrating modules and hierarchies. Category Three: Connecting Models The final category is the concept of linking models together to create multi-level ABMs. This category intersects with a much larger field of connecting models and simulations together and are governed by High Level Architecture (HLA) standard of the IEEE (2010). What is significant about coupling models which are part of the same complex system is they will share variables as the various modules in their respective hierarchy update. Unfortunately, this critical dynamic falls outside the IEEE standard (Scerri et al., 2010). Simulating such interdependencies is critical to understanding how these complex interactions may ripple across the entire system. Beyond NetLogo's LevelSpace, this literature review found one effort to deal with this challenge. The paper provides an architecture with two main features to overcome this difficulty, first is a time 6 https://repast.github.io/ 7 http://flame.ac.uk/ 8 http://www.madkit.org/ 9 https://cs.gmu.edu/~eclab/projects/mason/#Features 6 Multi-level ABMs covers a wide breadth of disciplines and approaches. The manager to ensure all the models are synchronized in their sequential management. The second feature is a conflict resolver to determine which model should get access to shared variables first (Scerri et al., 2010). These features go beyond the features Multi-level Mesa will add, but is a dynamic which at some point must be considered. Linking models to simulate the interactions of different modules and hierarchies of a complex system presents new problems not addressed by the common standards of model and simulation coupling. natural sciences who are trying to understand the interaction of hierarchies and modules which have evolved over millennia are understandably trying to discover the specifics of those complex interactions. Outside the natural sciences researchers are trying to determine if there are common interaction processes among diverse human societies or develop reliable interaction processes to control fleets of autonomous vehicles. For the majority of ABM libraries and platforms they have relied on the inherent inclusion of modules and hierarchies in object-oriented programming. Only GAMA and MaDKit have explicitly included functionalities for modules and hierarchies, while MASON and Repast have elements in their schedulers which can implicitly aid more complex interactions between modules and hierarchies of agents. The literature review showed that modules and hierarchies are an implicit part of the OOP languages on which ABMs are built. Although modules and hierarchies are an implicit part of OOP languages managing the complex interaction of agents impacting higher level agents and vice versa and the ability for agents to be active in different modules or change from one group to another presents significant management challenges which Multi-level Mesa seeks to mitigate for ABM practitioners. Multi-level Mesa Approach Multi-level Mesa's approach is driven by the concept of complex systems as adaptive networks. The core data structure of Multi-level Mesa is a network graph using Python's NetworkX library (Hagberg, Schult, & Swart, 2008). This approach extends existing multi-level approaches as well as exploiting the OOP nature of Python. The most similar approaches are GAMA and MaDKit, GAMA incorporates clustering algorithms as an additional method of determining if agents are in the same group (Taillandier et al., 2012). Multi-level Mesa, leveraging NetworkX clustering algorithms allows for the same dynamic, while also allowing users to specify when a module may form or activate based on link type or a value associated with a link type. In MaDKit, the user must specify the use of Agent, Group and Role to manage which agent is doing what in which group (Michel et al., 2017). Multi-level Mesa extends this approach by incorporating a dynamic network. Instead of specifying specific groups and roles, as connections between agents change through the dynamics of the ABMs, new modules (groups) can form or dissolve and new behaviors can activate or lay dormant. This approach allows for neutral networks to exist within any model where certain behaviors may only emerge under specific conditions and are not previously seen. Multi-level Mesa goes beyond existing approaches by creating a greater synergy between network science and ABMs, the 7 interaction of agents produces a dynamic network, which in turn alters the behavior of the agents. The remainder of this section will discuss the specifics of Multi-level Mesa's implementation prior to discussing the results of the Sugarscape ABM used to develop Mult-level Mesa. The Multi-Level Mesa Library Multi-level Mesa has three main components. First, a collection of managers which tracks the agents, the modules of agents (groups), the network of agents, agents who belong to an existing group, and the schedule. Second, a series of functions which provides the user different options to form groups or dissolve them. Third, a group class which allows for the inclusion of different group policies, manages the behavior and status of the group, and implicitly produces hierarchies within the complex system. (Figure 3-1) Multi-level Mesa is available on GitHub at Figure 3-1: Multi-level Mesa Schematic https://github.com/tpike3/multilevel_mesa and is also part of the Python Package Index and can be installed with the pip install multilevel_mesa. Creating an Multilevel_Mesa Instance and the Multilevel_Mesa Managers Creating an instance of Multi-level Mesa requires no parameters, and initiates one attribute and six managers (Box 3-1). The ML Mesa does have two keyword parameters, 8 MultiLevel_Mesa.min_for_group and MultiLevel_Mesa.group_net. MultiLevel_Mesa.min_for_group tells the instance the minimum number of agents which must be in a group. The min_for_group parameter has a default setting of two. The second key word parameter is MultiLevel_Mesa.group_to_net, this parameter takes a Boolean and is defaulted to False. If a group is added to the network, this tells the Multi- level Mesa instance that the group as an agent can link with other nodes. User specified behavior can then dictate the complexity of these interactions, groups to agents, groups to groups and groups forming groups of groups. The one attribute of Multi-level Mesa is id_counter, which allows for unique_ids to be generated for groups. The six managers are (1) MultiLevel_Mesa._agents which is an ordered dictionary (a hash-table consisting of a key:value pair) that holds every agent added to the instance. This manager is critical to maintain the most granular dictionary possible of all agents and mimics the _agents dictionary found in Mesa. (2) MultiLevel_Mesa.net is an instance of a NetworkX graph. This feature provides the critical structure for tracking and managing agents and groups. (3) MultiLevel_Mesa.agents_by_type uses a dictionary of dictionaries to track agents by type. This feature allows for faster reference of specific types of agents when manipulating groups or schedules. (4) MultiLevel_Mesa.schedule replaces the Mesa schedule and is an ordered dictionary which manages the agents and when they execute a step function. (5) MultiLevel_Mesa.groups is an ordered dictionary and tracks the groups within the model performing the same function of tracking groups as the agents ordered dictionary. (6) MultiLevel_Mesa.reverse_groups is a dictionary of dictionaries of sets. The first dictionary key is the agent id, while the second is group types (link and link values) and the set is the group ids to which the agent belongs in those group types. This structure is necessary to ensure duplicate groups are not created or that an agent is added to an existing group instead of creating a new one. The use of sets also helps expedite computation by using set operations to evaluate if a group should be formed or agents added to an existing group. 1. _agents: Ordered Dictionary 2. net : NetworkX Undirected Graph 3. agents_by_type : Dictionary of Dictionaries 4. schedule : Ordered Dictionary 5. groups : Ordered Dictionary 6. reverse_groups : Dictionary of Dictionaries of Sets Box 3-1: Multi-level Mesa Managers and Data Types The Mutli-level Mesa Functions As shown in figure 1, Multi-level Mesa has two primary approaches for facilitating a multi-level ABM, an explicit approach and a network approach. Within these two approaches, Multi-level Mesa turns the desired agents into a bilateral link list which form the groups. Each input of agents is transformed into a network edge which 9 forms the groups or adds agents to an existing group. The use of links is also used to disband groups or remove agents from the group. These functions then create a more dynamic schedule with modules of agent within hierarchies. Forming and Dissolving Groups User Defined Formation Process: MultiLevel_Mesa.form_group The formation function of the explicit approach is MultiLevel_Mesa.form_group and takes a user defined process which must generate a list of bilaterally connected agents (Box 3-2). This approach can be computationally expensive, but is necessary to allow for the accurate recreation of the network. As dictionaries (e.g. the schedule) cannot be manipulated during iteration users must use a yield versus the more common return operator to pass the list of agents to the MultiLevel_Mesa.form_group function. def form_group(self, process, args, determine_id = 'default', double = False,\ policy = None, group_type = None, kwargs): Box 3-2: MultiLevel_Mesa.form_meta function The MultiLevel_Mesa.form_group function requires one parameter which is the user specified process which determines whether or not an agent should be in a group with other agents. The args and *kwargs allows the user to pass in the parameters for this process. The determine_id parameters ensures each group gets a unique id. If default it will simply append a number based on the id_counter attribute to the string 'group'. For the user to pass in an id he or she must yield the id as the first element of a tuple generated from the yield operator from the user defined process. Users must choose this id carefully as the id is used in the set operations to merge groups. The double parameter takes a Boolean value and is defaulted to False. If True the agent will remain in the schedule as an independent entity and be added as part of the group, while if False the agent is removed. This feature is to provide users maximum flexibility for agent scheduling and group processes. The policy parameter passes in the step processes for the group, which can consist of only internal processes or can consist of group processes and then execute the individual agent processes. The group_type parameter takes a string and allows the user to specify different types of groups so an agent can belong to different types of group such as 'family' and 'firm'. User Defined Dissolution Process: MultiLevel_Mesa.reassess_group The dissolution function for the explicit approach (although it can be used interchangeably with the network approach) is MultiLevel_Mesa.reassess_group (Box 3- 3). This function iterates through each group and then uses the user defined process to assess whether or not an agent should still belong to the group. Similar to the MultiLevel_Mesa.form_group this function requires a yield to provide the list of agents which should be removed and then proceeds to remove those agents while updating the appropriate managers. This function also ensures if the group fails to have a certain 10 number of agents within the group that the group will be removed. This minimum number of agents is the min_for_group attribute of the Multi-level Mesa instance and has a default setting of two. def reassess_group(self, process, args, reintroduce = True, group_type = None, *kwargs): Box 3-3: MultiLevel_Mesa.reassess_group function The MultiLevel_Mesa.reassess_group function requires one parameter, which is the process defined by the user for assessing whether or not the agent should remain within the group. The function also has a reintroduce parameter which takes a Boolean value and is defaulted to True. This parameter tells the function whether or not to reintroduce the removed agents back into the schedule. Network Defined Formation: MultiLevel_Mesa.net_group The formation function of the network approach is MultiLevel_Mesa.net_group (Box 3-4) and uses an undirected NetworkX graph object to assess what agents should form groups. With an undirected graph and as indicated in figure one, there are three possibilities for assessing whether or not linked agents should be in the same group. First, by whether or not a link exists between the agents. Second, if a specific type of link exists (e.g. friend, enemy). Third, if a link exists which has reached a certain value. For example, in the Sugarscape model discussed in the next section, one version forms a group if an agent and landscape cell are linked, in another version, the agents form a group if they have 10 or more trades between them. Although, NetworkX also offers the possibility of directed graphs and multi- graphs these options were not used for simplicity sake and because the dynamics of ABMs can account for the main aspects of these features. As NetworkX uses a dictionary structure to capture nodes and links, a multi-graph can be easily simulated by adding more link types along the edge, so a link may have the dictionary keys {family, tribe, job...} allowing for a link with multiple types similar to a multi-graph. The directed graph dynamic can also be achieved through agent interactions as the link attributes can dictate the direction of flow based on agent attributes and behaviors. The one cost is users cannot use the multi-graph and directed graph network evaluation functions in NetworkX. Using an undirected graph provides a leaner, more easily understood approach without loss of network dynamics. def net_group(self, link_type = None, link_value = None,\ double = False, policy = None): Box 3-4: MultiLevel_Mesa.net_schedule function 11 The MultiLevel_Mesa.net_group function requires no parameters and will default to whether or not a link exists or not between agents. As the group is formed purely based on the links between agents, no args or **kwargs arguments are required. As the net_group function has no process passed in there is no way to specify a group id, the function uses the default "group" if groups are forming based on the presence of a link, the link_type is not the default None or the link_type_link_value, plus a number from the MultiLevel_Mesa.id_counter attribute. If users decided they would like to pass in processes to provide a unique id for groups this could be added in future versions, but was not included in this version as it did not add anything substantive to the Multi-level Mesa dynamics. The link_type function allows the user to pass in what link key value should link agents together. The link_type can then be further specified with the link_value criteria. These values are also used as the dictionary keys in the MultiLevel_Mesa.reverse_groups manager. The link_value can either be a string to further classify the type of link, for example family: friendly or family: angry_teenager or it can be a value such as will be seen in the Sugarscape model trades: 10 (number of trades between agents), which in this case tracks a type of interaction between agents. As net_group is an additive process the value is assumed to be a threshold of greater than or equal to a value. The network can then be updated and evaluated through the other processes in the ABM using NetworkX object manipulation functions. For convenience, MultiLevel_Mesa also has MultiLevel_Mesa.add_links and MultiLevel_Mesa.remove_links functions. These functions take a list of agents, combines them in to a list of fully connected tuples and then adds or removes the links. Network Defined Dissolution: MultiLevel_Mesa.reassess_net_group The MultiLevel_Mesa.reassess_net_group (Box 3-5) uses the same taxonomy of options as MultiLevel_Mesa.net_group. First, an agent can be removed based on the presence of a link, the presence of a specific link type and finally the presence of a specific link value. The function will also check to ensure the meta-agent still has the minimum number of agents to remain a group which is defaulted to two with the MultiLevel_Mesa.min_for_group attribute. def reassess_net_group(self, link_type = None, link_value = None) Box 3-5: MultiLevel_Mesa.reassess_net_group function The dissolution function similar to the formation function requires no parameters and will default to determining if there is a link or not. The user can also specify link types which cause agents to be removed or link values, which can again be either strings or numbers. However, as this function is not additive, the agent will be removed if the value is less than or equal to the user specified value. 12 A Note on Formation Precedence A critical point for users to understand is agents belong to the first group with which they form. If an agent is not part of a group and meets the user given criteria it will then be added to the first group evaluated by function based on the specified user dynamics or randomly ordered dictionary of the NetworkX link dictionary. If both agents belong to a group the link between them at the agent level will remain in place. This approach was adopted because the dynamics of how agents should be integrated into groups is specific to the user's model. This approach, therefore, defaults to the first group joined which is consistent with human biases (Heuer, 1999; Pratkanis & Aronson, 2002). Appreciating how this dynamic works will allow users to leverage the other functions to specify group precedence. Schedule Functions As MultiLevel_Mesa replaces the normal schedule function of Mesa, it must also have the basic scheduling functions (Box 3-6). These are the add and remove functions, which remain at the individual agent level but have a higher degree of complexity as agents must be kept in multiple managers to ensure agents are being properly 'stepped' in the schedule or removed if the agent 'dies'. Multi-level Mesa also replaces Mesa's step function. Its primary schedule is random activation, but this can be turned off for an ordered activation and a staged activation can be executed through the agent_type manager. A future extension of MultiLevel_Mesa would be to store different schedules based on different network configurations. This would save computation time so specific agent schedules would be created less often. For example, if one was recreating daily life of a population and the night and morning hours used one configuration, while the daytime hours would use a different configuration, each calling different behavior routines for the agents. def add(self, agent, schedule = True, net = True) def remove(self, agent): def step(self, shuffled = True, by_type = False, const_update = False) Box 3-5: MultiLevel_Mesa. Schedule functions Similar to Mesa, the MultiLevel_Mesa.add function requires an agent object. It also has two keyword parameters which take Boolean parameters each with a default value of True. Keyword parameter schedule adds the agent to the schedule. This is an option in case the user begins with a complex network and the agent is already part of a group. The net parameter similarly adds the agent to the NetworkX object. This is done in case the user has an agent he or she does not want to be part of the network. For instance, in a Sugarscape model, the grid cells may not need to be a part of the network as what is of concern is the agent's network. The MultiLevel.Mesa.remove function requires an 13 agent object. If invoked this will remove the agent from all managers as applicable. The MultiLevel_Mesa.step function works in a similar way to the Mesa step function, where it iterates through each agent in schedule and executes their step function. Random activation is the default as identified by the keyword parameter shuffled. If shuffled is False it will follow the order in the ordered dictionary (the order the agents were added). The keyword parameter by_type is set to False but can take a list of agent types to simulate staged activation. Constant update provides the ability to have specific agent types activated after the more dynamic schedule. For example, an environmental variable which changes at a steady rate for each time step, such as sugar or spice growth in the Sugarscape model. The Group Class The Group class introduces hierarchy into the ABM. The Group class performs similar functions to Multi-level Mesa or Mesa's time module. The Group class has three managers, which includes a dictionary of the agents which belong to the Group, a dictionary of dictionaries with the agents in the Group by type and a NetworkX graph object of the sub_agents. The Group then has three attributes to make it easier for users to employ the Group. The first attribute is Group.active which is a Boolean value to help users activate and deactivate Groups as necessary. The next two attributes are Group.type and Group.__str__ which both equal "group" and allow the user greater ease in identifying and performing functions on the groups. The final attribute of the Group is its policy object, this object is passed in by the user and provides the Group behavior. The behavior of the Groups and its internal agents is done with two step functions the Group.group_step which calls the policy function and the individual agent step functions, again using a random order, but with the same options of the MultiLevel_Mesa.step function to dictate schedule ordering processes. Group.sub_agents = dictionary Group.agents_by_type = dictionary Group.net = NetworkX graph Group.policy = object of group policies Group.active = status of Group Attributes: Main Functions: Group.meta_step() = policies to dictate sub_agent behavior Group.step() = sub_agent behaviors Box 3-6: Group attributes and functions The interaction of the schedule, formation and dissolution of modules of agents, and the ability for hierarchies to exist allows for the easier introduction of these key features of complex systems. The functions can be employed as part of the normal step 14 function, at specific events or at specific intervals. By using a network data structure as the main management structure, Multi-level Mesa is able to integrate the interdependencies and changing dynamics of those interdependencies into ABM management structure providing a new dynamic which goes beyond the current multi- level approaches. With an understanding how the main functions and dynamics of Multi- level Mesa, it is now time to verify and validate the Multi-level Mesa library. Multi-level Mesa and Sugarscape Sugarscape was used to verify and validate the functioning of the Multi-level Mesa library.10 Sugarscape was one of the first ABMs to demonstrate bottom up emergence of system behavior based on the decentralized action of many agents. The specific variation used for Multi-level Mesa is the trade variation in which the landscape has two commodities sugar and spice and the agents must acquire and consume both based on their unique sugar and spice metabolism in order to survive. The agents trade their sugar and spice accumulations based on the amount of sugar and spice they have acquired and their marginal rate of substitution due to their metabolisms (Axtell & Epstein, 1996). This variation offers a great test case for Multi-level Mesa because the results are well founded on economic theory providing clear verification and validation for Multi-level Mesa, as well as providing enough complexity that groups can be introduced in different configurations and with different policies to show the impact of this additional dynamic. The Multi-level Mesa model uses the base case of a trading environment outlined in the beginning of chapter four of Growing Artificial Societies. The landscape is a 50 by 50 torus with each cell given a quantity of sugar and or spice from zero to six (Figure 3- 2). There are four mounds, each with a gradient that increases in sugar or spice as one gets closer to a peak. Each grid will regrow one sugar and one spice unit per step until its maximum sugar and spice allotment is reached. There are 200 hundred agents, each instantiated with a vision attribute between one and six which determines how many cells they can see using a Von Neumann neighborhood (four cardinal directions). Each agent is instantiated with a sugar and spice metabolism between one and six, which indicates how much sugar or spice each agent consumes with each step. Each agent is also given an initial endowment of sugar and spice from 25 to 50. On each time step, the schedule iterates through a randomly ordered list of agents and each agent moves to collect more sugar and spice, consume sugar and spice, and trade with agents within their vision. The agents move and trade based on their marginal rate of substitution and in accordance with what their vision allows according to a Von Neumann neighborhood and as calculated in Growing Artificial Societies (Axtell & Epstein, 1996). With this model, the different configurations of Multi-level Mesa are tested to both verify and validate its use as a library. 10 All code for this instantiation of Sugarscape can be found at https://github.com/tpike3/SugarScape. Due to the size the results were not included, but the code used to analyze the results was included. This allows any interested parties to run and analyze the code. The results can also be provided upon request. 15 Testing Multi-level Mesa occurred in three phases, the first phase is showing equivalency between Multi-level Mesa's explicit and network approach and a standard Sugarscape configuration. The second phase is showing equivalency with the formation of groups and the third phase is showing the impact of different group policies on agent behavior. This provides both verification and validation of the Multi-level Mesa library as well as justifying its existence based on the impact of even simple group policies on emergent behavior. Phase I: Equivalency Between Multi-level Mesa Approaches and a Standard Approach The first phase recreates Sugarscape, specifically the sugar and spice variation described in chapter four of Growing Artificial Societies (Axtell & Epstein, 1996), and replicates the output of this standard approach using Multi-level Mesa's explicit and network approach. In the standard approach the schedule randomly orders each agent as they iterate through the movement, consumption and trade functions. This model replicates the key result of the sugar and spice landscape as the price of both sugar and spice moves toward one and the standard deviation of the logarithmic mean of the price moves toward zero, as predicted by economic theory (Axtell & Epstein, 1996). This instantiation of the sugar and spice landscape does not match the trade volume in Growing Artificial Societies as the volume total is much less and follows a heavily skewed distribution (Figure 3). This difference is acceptable as the metric for validation is not the amount of trade but rather the trade price (Axtell and Epstein 1996). To ensure the proper functioning of the Multi-level Mesa library these results then needed to be replicated using the network and explicit approaches. To replicate these results using the explicit approach, the model forms a group with each agent and the landscape cell on which the agent is located. The model then steps forward each group, producing the same set of dynamics as the standard approach. The model then reassesses each group and if the agent has moved disbands the group. Similarly, for the network approach, the model forms a link between the agent and the cell it is on, the model steps through the agent functions and then the link between the agent and cell is reevaluated and removed if the agent is no longer on the cell. These two approaches then replicate the results of price, standard deviation of the logarithmic mean, and trade volume (Figure 3-3). In addition, these three variations were run for 100 runs over 1000 steps mimicking Growing Artificial Societies (Axtell & Epstein, 1996). The output of their respective price distributions was not qualitatively different and their surviving number of agents was not statistically different demonstrating the approaches are equivalent versions of the same dynamic. The additional process of creating and destroying the groups added a time cost from the standard variation with a two second addition to the mean for the network approach and a 14 second addition to the mean for the explicit approach (Figure 3-4). These results provide the simplest possible comparison to ensure the Multi-level Mesa approach does not fundamentally alter ABM dynamics. These results then allowed for advancement to phase two introducing groups of trading agents. 16 Figure 3-2: Sugar and Spice Landscape, the tan peaks represent more spice while the green peaks represent more sugar, the dots are the agents. Phase II: Introducing Groups In phase two the use of the Multi-level Mesa library was varied so groups formed if two or more agents reached a threshold of trades. The primary question for this phase was whether or not grouping agents together in the schedule would have any impact on the results. As demonstrated in Who Goes First? An Examination of the Impact of Activation on Outcome Behavior in Agent-based Models, activation schemes in ABMs do matter (Comer, 2014). In this case however, agents grouping together and being randomly activated as a group had no significant outcome on the results. To test this, groups were formed at one, five and 10 trades over 1000 steps and 100 runs. The overall results were compared, as well as specific results of the price and trade volume. For each parameter the grouping of agents had no observable impact on the results. Figure 3-5 shows agent configuration, the price, and standard deviation of the logarithmic mean of the price. Figure 3-6 shows the overall price distribution of 100 runs for the explicit and network 17 Figure 3-3: Single Run Results of Price, Standard Deviation of Logarithmic Mean and Trade Volume 18 Figure 3-4: Standard Deviation of Logarithmic Mean Price Distribution, Survivor and Time Histograms for 100 runs of 1000 step 19 Figure 3-5: Agent Types for 100 runs, Mean Trade Price for One Run, Standard Deviation of Logarithmic Mean for One Run 20 Figure 3-6: Standard Deviation of the Logarithmic Mean, Survivors and Time Histogram Results of 100 Runs with 10 Trades Forming a Group 21 approach as compared to the standard approach. Although shown for 10 trades, groups which formed at one and five trades showed similar results. The network approach time per run was comparable with the standard approach and actually had a mean of three second less. While the explicit approach incurred a time cost of 15 seconds. The reason for the network approach is comparable is it is able to use NetworkX's dictionary structure to reference specific agent groups, while the explicit approach must iterate through the model's trade dictionary and assess the trade status. Although, grouping agents in the schedule did not have an impact on this particular model this does not mean these results are generalizable. For this study, however, these results provide further verification of the functioning Multi-level Mesa. Phase III: Introducing Group Policy The final phase introduced policy into groups. Policy for Multi-level Mesa is understood to be group behavior which alters behavior of the agents in their group. If the agent is not part of the group it will behave differently. To assess the impact of policy three variations were implemented. First, the policy of the group changed the individual agent's behavior. Second, the group resources were available to all, but the agents explored and traded based on their own situation. Third, the groups shared their resources and explored as in the second version, but the groups could form groups with other groups. This tested the ability of Multi-level Mesa to allow multiple levels to emerge endogenously as the group could trade with other groups and form a group consisting of sub-groups and individual agents. Phase three tested the impact of group policy of the emergent behavior of the system. For the first variation, agents reaching a specified number of trades (i.e. one, five and 10) formed a group and the group agent applied a policy to their behavior which changed how the group's agents moved through the landscape. Once a group agent is formed it is randomly assigned one of three policies. (1) Each agent within the group moves to a new cell as though it has the lowest sugar or spice accumulation in the group. (2) Each agent within the group moves to a new cell as though it has the highest sugar or spice accumulation in the group. (3) Each agent moves to a new cell as though it has the geometric mean of sugar and spice accumulations of the group. It is important to note, the agent's perception was only changed based on accumulation, their metabolism was not changed so the agent explored the environment with their view of which cell provided the best resources based on their metabolism and their respective groups accumulation policy. The agents then continued to consume and trade based on their actual accumulations. These policies had a substantial impact on the outcomes of model (Figure 3-7 and 3-8). As shown in figure 3-7, the policies prevented the price from moving towards one and the standard deviation of the logarithmic mean from moving towards zero. In addition, and somewhat surprisingly, the policies reduced the number of independent agents as compared with no policy (Figure 3-5). Reducing the number of trades required to form a group further reduced the number of independent agents. This occurred because the agents when reaching a price equilibrium were also reaching a movement 22 equilibrium. The policies of the groups then caused the agents to explore more of the environment increasing the number of agents in contact and trading with each other, and reducing the number of independent agents. The policies however, reduced the survival rate of the agents and so if they had a choice, it would not be in the agent's interest to be part of a group (Figure 3-8). Other variations in which the agents searched based on the total group accumulations and could trade with everyone in their group regardless of distance and vision, also resulted in lower survival rates. From these results, one can conclude in Sugarscape searching the landscape based on someone else's situation is sub- optimal. As will be seen in the next variation however, sharing one's resources, and searching and trading based one's own metabolism can lead to much greater survival rates. Comparing the times between model approaches the network approach was slightly faster, while the explicit approach incurring approximately 47 seconds per run (Figure 3-8). The reason for the explicit is the additional iterations the explicit approach must do in order to assess the groups agents each step, while the network approach is able to reference the dictionary structure of the NetworkX object and its links. Comparing the mean for Network and Standard variations with and without policy shows the these two versions are effectively comparable with time, while other processes in the computer are impacting the exact results. The next variation for the group agents consolidated the accumulations of each agents to create a common resource available to all group agents. Agents, however, would interpret this accumulation through their own metabolism for trade and movement. This variation was then further explored by the Multi-level_Mesa instance parameter for adding the group to the network to allow group agents to form links with other group agents, creating multiple levels. Agents trade and form a group, then these groups trade and form a group and so on. These variations were only instantiated using the network approach but can be done in the explicit approach as well. The results continued to demonstrate that group policy changes the emergent behavior of the system. For both one level and multiple levels of agents accessing and consuming groups resources, while exploring their environment based on their metabolism, there was a change in the qualitative shape of the price distribution curve and increases to the survival of the population (Figure 3-9). The inclusion of multiple levels had no impact on the results, which makes sense as the behavior is moving and trading is at the agent level and the groups only provide a common accumulation. Reducing the number of trades (e.g. from 10 to five) did not change the shape of price distribution but it did change the height, increasing the center peak by approximately 400. This increase in trade frequency can be attributed to the dynamic that the sooner the agents are able to join a group the higher their chance of survival. Faster group formations had higher survival rates. These survival rates then changed the time dynamic of each run, more surviving agents resulted in longer run times. The majority of computation cost being in the agent also produced the counter intuitive results that the multi-level time mean was less than the mean time for one level runs. As the multi-level survival mean was slightly less than that of the one level survival mean (92.83 vs. 94.35) it was therefore slightly faster but not statistically different (p-value of .115) (Figure 3-9). 23 Figure 3-7: Impact of Policies on Groups, Mean Price and Standard Deviation of the Logarithmic Mean 24 Figure 3-8: Standard Deviation of Logarithmic Mean, Survivors and Time Histograms for Groups with Policies for 100 Runs, 1000 Steps 25 Figure 3-9: Price Distribution, Survivor and Time Histograms for 100 Runs, with 1000 Steps of Agent Groups whose Accumulations are a Common Resource 26 These results continued to demonstrate group policy matters and demonstrated Multi-level Mesa's ability to recursively form multiple levels of groups. Additional variations were attempted to further verify the group formation of Multi-level Mesa. When vision or agent density was increased, as expected, the agents tend to coalesce towards one super group with levels of groups within them, for this model up to four (one super group, two levels of intermediate groups, and individual agents). In addition, the order of action in the group was varied. In one version, the agents within the group all moved and collected, before eating and trading. In another variation each agent moved, collected, ate and traded in a random order. This had no noticeable impact on the results. Summary The sugar and spice variation of Sugarscape served as an excellent dynamic to both verify and validate Multi-level Mesa. Through Sugarscape, Multi-level Mesa successfully showed that first it 'does no harm' and can successfully replicate the key dynamics of the sugar and spice trading model. Second, it showed that agents forming groups with a policy to direct their agent's behavior, does impact the emergent behavior of the system. Both the shape of the price distribution and survival rates of the agents were impacted based on the policy. Finally, Multi- level Mesa demonstrated that it can create multiple levels of groups through recursion. These results were shown with a new way to conceptualize agent-based scheduling, networks. Dynamically changing agent activity based on complex networks offers new modelling possibilities which are consistent with other complex adaptive systems such a brain activity and genetic networks. Multi-level Mesa will help reduce the barrier of entry to analysts, planners and decision makers, while increasing their ability to develop models of the complex societies they are trying to influence. These models will help them conduct virtual experiments with complex population networks in pursuit of more effective policy with less resources. 27 BIBLIOGRAPHY Axtell, R., & Epstein, J. (1996). Growing Artificial Societies: Social Science From the Bottom Up. Washington D.C.: Brooking Institution Press. Barabasi, A.-L. (2016). Network Science (Online). Cambridge: Cambridge University Press. Retrieved from http://networksciencebook.com/ Booch, G., Maksimchuk, R. A., Engle, M. W., Conallen, J., Houston, K. A., & Ph.D, B. J. Y. (2007). Object-Oriented Analysis and Design with Applications, Third Edition (3rd ed.). Addison-Wesley Professional. Retrieved from http://proquest.safaribooksonline.com/book/software-engineering-and- development/object/9780201895513 Camus, B., Bourjot, C., & Chevrier, V. (2013). Multi-level Modeling as a Society of Interacting Models (technical No. hal-00913038) (p. 17). Retrieved from https://hal.inria.fr/hal- 00913038 Cilfone, N. A., Kirschner, D. E., & Linderman, J. J. (2015). Strategies for Efficient Numerical Implementation of Hybrid Multi-scale Agent-Based Models to Describe Biological Systems. Cellular and Molecular Bioengineering, 8(1), 119 -- 136. https://doi.org/10.1007/s12195-014-0363-6 Comer, K. (2014). Who Goest First? An Examination of the Impact of Activation on Outcome Behavior in Agent-based Models. George Mason University. Falcone, J.-L., Chopard, B., & Hoekstra, A. (2010). MML: towards a Multiscale Modeling Language. Procedia Computer Science, 1(1), 819 -- 826. https://doi.org/10.1016/j.procs.2010.04.089 Gil-Quijano, J., Louail, T., & Hutzler, G. (2012). From Biological to Urban Cells: Lessons from Three Multilevel Agent-Based Models. In N. Desai, A. Liu, & M. Winikoff (Eds.), Principles and Practice of Multi-Agent Systems (Vol. 7057, pp. 620 -- 635). Berlin, Heidelberg: Springer Berlin Heidelberg. https://doi.org/10.1007/978-3-642-25920-3_45 Hagberg, A. A., Schult, D. A., & Swart, P. J. (2008). Exploring Network Structure, Dynamics, and Function using NetworkX. In Proceedings of the 7th Python in Science conference (p. 5). Pasadena, CA. Haman Tchappi, I., Galland, S., Kamla, V. C., & Kamgang, J. C. (2018). A Brief Review of Holonic Multi-Agent Models for Traffic and Transportation Systems. Procedia Computer Science, 134, 137 -- 144. https://doi.org/10.1016/j.procs.2018.07.154 Heuer, R. J. (1999). Psychology of Intelligence Analysis. Washington, D.C.: Center for the Study of Intelligence, Central Intelligence Agency. Hjorth A., Head, B., & Wilensky, U. (2015). LevelSpace NetLogo Extension. NetLogo, Evanston, IL: Center for Connected Learning. Retrieved from https://github.com/NetLogo/LevelSpace Hjorth, A., Weintrop, D., Brady, C., & Wilensky, U. (2016). LevelSpace: Constructing Models and Explanations across Levels, 2. Holland, J. H. (1995). Hidden Order: How Adaption Builds Complexity. New York: Helix Books. IEEE. (2010, August 8). IEEE 1516-2010 - IEEE Standard for Modeling and Simulation (M&S) High Level Architecture (HLA)-- Framework and Rules. Retrieved November 28, 2018, from https://standards.ieee.org/standard/1516-2010.html 28 Interagency Modeling and Analysis Group. (2018). Retrieved November 24, 2018, from https://www.imagwiki.nibib.nih.gov/ Kissinger, H. (2014). Word Order. New York: Penguin Books. Koestler, A. (1967). Ghost in the Machine. New York: Hutchinson and Company. Luke, S., Cioffi-Revilla, C., Panait, L., Sullivan, K., & Balan, G. (2005). MASON: A Multiagent Simulation Environment. SIMULATION, 81(7), 517 -- 527. https://doi.org/10.1177/0037549705058073 Michel, F., Gutknecht, O., & Ferber, J. (2017, July 18). MaDKit 5.2. Retrieved November 28, 2018, from http://www.madkit.org/ Miller, J. H., & Page, S. E. (2007). Complex Adaptive Systems: An Introduction to Computational Models of Social Life (Vol. 27). Retrieved from http://books.google.com/books?id=XQUHZC8wcdMC&pgis=1 Morvan, G. (2013). Multi-level agent-based modeling A literature survey. ArXiv:1205.0561v7 [Cs.MA], 27. Morvan, G. (n.d.). Multi-level agent-based modeling A literature survey, 27. Multi-level architecture. (n.d.). Retrieved November 27, 2018, from https://gama- platform.github.io/wiki/MultiLevelArchitecture Navarro, L., Corruble, V., Flacher, F., & Zucker, J.-D. (2013). A Flexible Approach to Multi- Level Agent-Based Simulation with the Mesoscopic Representation. In Proceedings of the 12th International Conference on Autonomous Agents and Multiagent Systems (p. 8). Saint Paul, Minnesota: International Foundation for Autonomous Agents and Multiagent Systems. Population in the U.S. - Google Public Data Explorer. (2018, September 19). Retrieved November 23, 2018, from https://www.google.com/publicdata/explore?ds=kf7tgg1uo9ude_&met_y=population&hl =en&dl=en#!ctype=l&strail=false&bcs=d&nselm=h&met_y=population&scale_y=lin&i nd_y=false&rdim=country&idim=place:0477000&ifdim=country&hl=en_US&dl=en&in d=false Pratkanis, A., & Aronson, E. (2002). The Age of Propoganda: The Everyday Use and Abuse of Persuasion. New York, NY: Owl Books. Repast Simphony Reference Manual. (2018, November 13). Retrieved November 28, 2018, from https://repast.github.io/docs/RepastReference/RepastReference.html#_scheduling Scerri, D., Drogoul, A., Hickmott, S., & Padgham, L. (2010). An Architecture for Modular Distributed Simulation with Agent-Based Models. In Proceedings of the 9th International Conference on Autonomous Agents and Multiagent Systems (AAMAS) (p. 8). Toroto, Canada: Springer. Seck, M. D., & Honig, H. J. (2012). Multi-perspective modelling of complex phenomena. Computational and Mathematical Organization Theory, 18(1), 128 -- 144. https://doi.org/10.1007/s10588-012-9119-9 Siebert, J., Ciarletta, L., & Chevrier, V. (2010). Agents and artefacts for multiple models co- evolution. Building complex system simulation as a set of interacting models. In Autonomous Agents and Multiagent Systems - AAMAS 2010 (Vol. 1, pp. 509 -- 516). Toronto, Canada: ACM. Retrieved from https://hal.archives-ouvertes.fr/hal-00452865 Simon, H. a. (1997). The sciences of the artificial, (third edition) (Vol. 33). Smallwood, R., & Holcombe, M. (2006). The Epitheliome Project: Multiscale Agent-Based Modeling of Epithelial Cells. In 3rd IEEE International Symposium on Biomedical 29 Imaging: Macro to Nano, 2006. (pp. 816 -- 819). Arlington, Virginia, USA: IEEE. https://doi.org/10.1109/ISBI.2006.1625043 Solovyev, A., Mikheev, M., Zhou, L., Dutta-Moscato, J., Ziraldo, C., An, G., … Mi, Q. (2010). SPARK: A Framework for Multi-Scale Agent-Based Biomedical Modeling. International Journal of Agent Technologies and Systems, 2(3), 18 -- 30. https://doi.org/10.4018/jats.2010070102 Soyez, J.-B., Morvan, G., Dupont, D., & Merzouki, R. (2013). A Methodology to Engineer and Validate Dynamic Multi-level Multi-agent Based Simulations. In F. Giardini & F. Amblard (Eds.), Multi-Agent-Based Simulation XIII (Vol. 7838, pp. 130 -- 142). Berlin, Heidelberg: Springer Berlin Heidelberg. https://doi.org/10.1007/978-3-642-38859-0_10 Taillandier, P., Vo, D.-A., Amouroux, E., & Drogoul, A. (2012). GAMA: A Simulation Platform That Integrates Geographical Information Data, Agent-Based Modeling and Multi-scale Control. In N. Desai, A. Liu, & M. Winikoff (Eds.), Principles and Practice of Multi- Agent Systems (Vol. 7057, pp. 242 -- 258). Berlin, Heidelberg: Springer Berlin Heidelberg. https://doi.org/10.1007/978-3-642-25920-3_17 30
1812.10202	1	1812	2018-12-26T02:20:24	Gliders2d: Source Code Base for RoboCup 2D Soccer Simulation League	[ "cs.MA" ]	We describe Gliders2d, a base code release for Gliders, a soccer simulation team which won the RoboCup Soccer 2D Simulation League in 2016. We trace six evolutionary steps, each of which is encapsulated in a sequential change of the released code, from v1.1 to v1.6, starting from agent2d-3.1.1 (set as the baseline v1.0). These changes improve performance by adjusting the agents' stamina management, their pressing behaviour and the action-selection mechanism, as well as their positional choice in both attack and defense, and enabling riskier passes. The resultant behaviour, which is sufficiently generic to be applicable to physical robot teams, increases the players' mobility and achieves a better control of the field. The last presented version, Gliders2d-v1.6, approaches the strength of Gliders2013, and outperforms agent2d-3.1.1 by four goals per game on average. The sequential improvements demonstrate how the methodology of human-based evolutionary computation can markedly boost the overall performance with even a small number of controlled steps.	cs.MA	cs	Gliders2d: Source Code Base for RoboCup 2D Soccer Simulation League Mikhail Prokopenko1,2 and Peter Wang2 1 Complex Systems Research Group, Faculty of Engineering and IT The University of Sydney, NSW 2006, Australia [email protected] 2 Data Mining, CSIRO Data61, PO Box 76, Epping, NSW 1710, Australia Abstract. We describe Gliders2d, a base code release for Gliders, a soccer simulation team which won the RoboCup Soccer 2D Simulation League in 2016. We trace six evolutionary steps, each of which is encapsulated in a sequential change of the released code, from v1.1 to v1.6, starting from agent2d- 3.1.1 (set as the baseline v1.0). These changes improve performance by adjusting the agents' stamina management, their pressing behaviour and the action-selection mechanism, as well as their positional choice in both attack and defense, and enabling riskier passes. The resultant behaviour, which is suffi- ciently generic to be applicable to physical robot teams, increases the players' mobility and achieves a better control of the field. The last presented version, Gliders2d-v1.6, approaches the strength of Glid- ers2013, and outperforms agent2d-3.1.1 by four goals per game on average. The sequential improve- ments demonstrate how the methodology of human-based evolutionary computation can markedly boost the overall performance with even a small number of controlled steps. 1 Introduction The RoboCup Soccer 2D Simulation League contributes to the overall RoboCup initiative, sharing its in- spirational Millennium challenge: producing a team of fully autonomous humanoid soccer players capable of winning a soccer game against the 2050 FIFA World Cup holder, while complying with the official FIFA rules [1]. Over the years, the 2D Simulation League made several important advances in autonomous decision-making under constraints, flexible tactical planning, collective behaviour and teamwork, communi- cation and coordination, as well as opponent modelling and adaptation [2,3,4,5,6,7,8,9,10]. These advances are to a large extent underpinned by the standardisation of many low-level behaviours, world model up- dates and debugging tools, captured by several notable base code releases, offered by "CMUnited" team from Carnegie Mellon University (USA) [11,12], "UvA Trilearn" team from University of Amsterdam (The Netherlands) [13], "MarliK" team from University of Guilan (Iran) [14], and "HELIOS" team from AIST Information Technology Research Institute (Japan) [15]. The latter release in 2010 included a number of components: -- librcsc-4.0.0: a base library for the RoboCup Soccer Simulator (RCSS); -- agent2d-3.0.0: a base source code for a team; -- soccerwindow2-5.0.0: a viewer and a visual debugger program for RCSS; -- fedit2-2.0.0: a team formation editor for agent2d. As a result, almost 80% of the League's teams eventually switched their code base to agent2d over the next few years [9]. The 2016 champion team, Gliders2016 [16,9], was also based on the well-developed code base of agent2d-3.1.1 [15], and fragments of MarliK source code [14], all written in C++. 2 Prokopenko and Wang The winning approach developed by Gliders combined human innovation and artificial evolution, fol- lowing the methodologies of guided self-organisation [17,18,19,20] and human-based evolutionary compu- tation (HBEC). The latter comprises a set of evolutionary computation techniques that incorporate human innovation [21,22]. This fusion allowed us to optimise several components, including an action-dependent evaluation function proposed in Gliders2012 [23], a particle-swarm based self-localisation method and tac- tical interaction networks introduced in Gliders2013 [24,25,26,27,28], a new communication scheme and dynamic tactics with Voronoi diagrams utilised by Gliders2014 [29], bio-inspired flocking behaviour in- corporated within Gliders2015 [30], and opponent modelling diversified in Gliders2016 [16]. The overall framework achieved a high level of tactical proficiency ensuring players' mobility and the overall control over the soccer field. In this paper, we describe a base code release for Gliders, called Gliders2d, version v1, with 6 sequential changes which correspond to 6 evolutionary HBEC steps, from v1.1 to v1.6. Since Gliders2d release is based on agent2d, the version Gliders2d-v1.0 is identical to agent2d-3.1.1 (apart from the team name), but every next step includes a new release. It is important to point out that Gliders2d is an evolutionary branch separate from the (Gliders2012 -- Gliders2016) branch. Thus, the final version of the presented release, Gliders2d-v1.6, is neither a subset not superset of any of Gliders2012 -- Gliders2016 teams. However, as a point of reference, we note that Gliders2d-v1.6 has a strength approaching that of Gliders2013 [24], and future releases will improve the performance further. Our objectives in making this first release are threefold: (a) it includes several important code com- ponents which explain and exemplify various approaches taken and integrated within the champion team Gliders2016; (b) it illustrates the HBEC methodology by showing some of the utilised primitives, while explicitly tracing the resultant performance (i.e., the fitness) for each sequential step from v1.1 to v1.6; (c) it demonstrates how one can make substantial advances, starting with the standard agent2d code, with only a small number of controlled steps. It may also serve as a brief tutorial that may help new teams in making the first steps within the league, using the available base code. 2 Methodology and Results The HBEC approach evolves performance across an artificial "generation", using an automated evaluation of the fitness landscape, while the team developers innovate and recombine various behaviours. The muta- tions are partially automated. On the one hand, the development effort translates human expertise into novel behaviours and tactics. On the other hand, the automated evaluation platform, utilised during the develop- ment of Gliders, and Gliders2d in particular, leverages the power of modern supercomputing in exploring the search-space. Each solution, represented as the team source code, can be interpreted as a "genotype", encoding the en- tire team behaviour in a set of "design points". A design point, in the context of a data-farming experiment, describes a specific combination of input parameters [31], defining either a single parameter (e.g., pressing level), complex multi-agent tactics (e.g., a set of conditional statements shaping a positioning scheme for several players), or multi-agent communication protocols [9,10,32]. While some design points are easy to vary, others may be harder to mutate and/or recombine due to their internal structure. For example, a specific tactic (design point), created by a team developer, may be implemented via several conditional statements each of which comprises a condition and an action, involving multiple parameters and primitives (see next subsections for examples). These components can then be mutated and recombined as part of the genotype. The solutions are evaluated against a specific opponent, over thousands of games played for each gen- eration. In order to maintain coherence of the resultant code, which evolves against different opponents in Gliders2d: Source Code Base 3 parallel, auxiliary conditions switch the corresponding parts of design points on and off for specific op- ponents [9], in an analogy to epigenetic programming [33]. The fitness function is primarily based on the average goal difference, with the average points as a tie-breaker, followed by the preference for a lower standard error. The main thread in the evolutionary branch described in this release aims to ensure a better control of the soccer field, by different means: (i) stamina management with higher dash power rates; (ii) more intense pressing of the ball possessing opponent; (iii) actions' evaluation aimed at delivering the ball to points stretching the opposition most; (iv) attacking players positioning to maximise their ball reachability potential; (v) defending players positioning to minimise the ball reachability potential of the opponents; (vi) risky passes. These improvements may in general be applied to robotic teams in physical RoboCup leagues. All the changes in Gliders2d are marked with // G2d: <brief comment> For example, setting the role of the agent based on its uniform number is done as follows: // G2d: role int role = Strategy::i().roleNumber( wm.self().unum() ); while retrieving the opponent name is achieved in this fashion: // G2d: to retrieve opp team name bool helios2018 = false; if (wm.opponentTeamName().find("HELIOS2018") != std::string::npos) helios2018 = true; In tracing the relative performance of Gliders2d from v1.1 to v1.6 we used three benchmark teams: agent2d-3.1.1 itself [15], Gliders2013 [24], and the current world champion team, HELIOS2018 [34]. For each sequential step, 1000 games were played against the benchmarks. Against agent2d, the goal difference achieved by Gliders2d-v1.6 improves from zero to 4.2. Against HELIOS2018, the goal difference improves from −12.73 to −4.34. Finally, against Gliders2013, the goal difference improves from −5.483 to −0.212, achieving near-parity. Tables 1, 2, and 3 summarise the performance dynamics, including the overall points for and against, goals scored and conceded, the goal difference, and the standard error of the mean. 2.1 Gliders2d v1.1: Stamina management The first step in improving upon agent2d performance, along the released evolutionary branch, is adding adjustments to the agents' stamina management (confined to a single source file strategy.cpp). Specif- ically, there are four additional assignments of the maximal dash power in certain situations: // G2d: run to offside line else if ( wm.ball().pos().x > 0.0 && wm.self().pos().x < wm.offsideLineX() && fabs(wm.ball().pos().x - wm.self().pos().x) < 25.0 ) dash_power = ServerParam::i().maxDashPower(); // G2d: defenders else if ( wm.ball().pos().x < 10.0 && (role == 4 role == 5 role == 2 role == 3) ) dash_power = ServerParam::i().maxDashPower(); 4 Prokopenko and Wang // G2d: midfielders else if ( wm.ball().pos().x < -10.0 && (role == 6 role == 7 role == 8) ) dash_power = ServerParam::i().maxDashPower(); // G2d: run in opp penalty area else if ( wm.ball().pos().x > 36.0 && wm.self().pos().x > 36.0 && mate_min < opp_min - 4 ) dash_power = ServerParam::i().maxDashPower(); This fragment of the source code demonstrates how these specific situations are described through conditions constraining the ball position, the agent position and its role, the offside line, and the minimal intercept cycles for the Gliders2d team (mate_min) and the opponent team (opp_min). Such constraints can be evolved by mutation or recombination of primitives (argument (op) X), where X is a constraint, wm.ball().pos().x is the argument, and (op) is a relational operator, e.g., <, >, ==, and so on. The action form may vary from a simple single assignment (the maximal dash power in this case), to a block of code. Adding these four conditions increased the goal difference against HELIOS2018 from -12.729 to -6.868, and against Gliders2013 from -5.483 to -2.684. 2.2 Gliders2d v1.2: Pressing The second step along this evolutionary branch is adding adjustments to the agents' pressing behaviour (confined to a single source file bhv_basic_move.cpp). The pressing level is expressed as the number of cycles which separate the minimal intercept cycles by the agent (self_min) and the fastest opponent (opp_min). More precisely, the intercept behaviour forcing the agent to press the opponent with the ball is triggered when self_min < opp_min + pressing. In agent2d the pressing level is not distinguished as a variable, being hard-coded as 3 cycles, and making it an evolvable variable is an example of a simple innovation. Specifically, there are four assignments of the pressing level, tailored to different opponent teams, agent roles and their positions on the field, as well as the ball location: // G2d: pressing int pressing = 13; if ( role >= 6 && role <= 8 && wm.ball().pos().x > -30.0 && wm.self().pos().x < 10.0 ) pressing = 7; if ( fabs(wm.ball().pos().y) > 22.0 && wm.ball().pos().x < 0.0 && wm.ball().pos().x > -36.5 && (role == 4 role == 5) ) pressing = 23; if (helios2018) pressing = 4; if ( ! wm.existKickableTeammate() && ( self_min <= 3 ( self_min <= mate_min && self_min < opp_min + pressing ) ) ) { } Body_Intercept().execute( agent ); ... Again, adding these four evolved conditions increased the goal difference against agent2d from near- zero to 1.288, against HELIOS2018 from -6.868 to -6.476 (this increase is within the standard error of the mean), and against Gliders2013 from -2.684 to -1.147. Gliders2d: Source Code Base 5 2.3 Gliders2d v1.3: Evaluator The third step modifies the action evaluator, following the approach introduced in Gliders2012 [23], which diversified the single evaluation metric of agent2d by considering multiple points as desirable states. The action-dependent evaluation mechanism is described in detail in [23,16], and the presented release includes its implementation (source files sample_field_evaluator.cpp and action_chain_graph). In particular, a new variable, opp_forward, is introduced, counting the number of non-goalie oppo- nents in a sector centred on the agent and extending to the points near the opponent's goal posts: // G2d: number of direct opponents int opp_forward = 0; Vector2D egl (52.5, -8.0); Vector2D egr (52.5, 8.0); Vector2D left = egl - wm.self().pos(); Vector2D right = egr - wm.self().pos(); Sector2D sector(wm.self().pos(), 0.0, 10000.0, left.th(), right.th()); for ( PlayerPtrCont::const_iterator of = wm.opponentsFromSelf().begin(); of != wm.opponentsFromSelf().end(); ++of ) { } if ( sector.contains( (of)->pos() ) && !((of)->goalie()) ) opp_forward++; The single evaluation metric of agent2d is invoked when there are no opponents in this sector, or when the ball is located within (or close to) the own half: if ( wm.ball().pos().x < depth opp_forward == 0 ) { // stay with best point = opp goal } Otherwise, the logic enters into a sequence of conditions (marked in the released code), identifying the "best" point out of several possible candidates offered by Voronoi diagrams. A Voronoi diagram is defined as the partitioning of a plane with n points into n convex polygons, so that each polygon contains exactly one point, while every point in the given polygon is closer to its central point than any other [35]. The best point is selected to be relatively close to the teammates' positions, and far from the opponents' positions. The distance between the identified best point and the future ball location, attainable by the action under consideration, is chosen as the evaluation result: double weight = 1.0; if (wm.ball().pos().x > 35.0) weight = 0.3; double point = state.ball().pos().x * weight; ... point += std::max( 0.0, 40.0 - best_point.dist( state.ball().pos() ) ); The condition scaling the initial assignment of the point's value, by weight, is another example of a simple mutation. 6 Prokopenko and Wang The action-dependent evaluation mechanism increased the goal difference against agent2d from 1.288 to 1.616, while not providing a notable improvements against the two other benchmarks, as it is applicable in attacking situations which are rare in these match-ups at this stage. 2.4 Gliders2d v1.4: Positioning To make a better use of the new field evaluator, the positioning scheme of the players is adjusted by selecting points according to suitably constructed Voronoi diagrams. For example, a Voronoi diagram may partition the field according to the positions of the opponent players; the candidate location points can be chosen among Voronoi vertices, as well as among the points located at intersections between Voronoi segments and specific lines, e.g., offside line; subject to certain constraints, as illustrated in [29]. All the constrained conditions are evolvable. A small fragment of the new code, fully contained in source file strategy.cpp, is below: // G2d: Voronoi diagram ... VoronoiDiagram vd; ... std::vector<Vector2D> OffsideSegm_tmpcont; for ( PlayerPtrCont::const_iterator o = wm.opponentsFromSelf().begin(); o != wm.opponentsFromSelf().end(); ++o ) { ... vd.addPoint((o)->pos()); } ... vd.compute(); ... for ( VoronoiDiagram::Segment2DCont::const_iterator p = vd.segments().begin(), end = vd.segments().end(); p != end; ++p ) Vector2D si = (p).intersection( offsideLine ); if (si.isValid() && fabs(si.y) < 34.0 && fabs(si.x) < 52.5) { OffsideSegm_tmpcont.push_back(si); } { } Once the container with the candidate points is filled, some of the players (three forwards) are assigned to the most promising points. The positioning based on Voronoi diagrams increased the goal difference against agent2d from 1.616 to 2.387, again maintaining the performance against the two other benchmarks. 2.5 Gliders2d v1.5: Formations This step did not change any of the source code files -- instead the formation files, specified in configura- tions such as defense-formation.conf, offense-formation.conf, etc. were modified with fedit2. This approach, pioneered in the Simulation League by [36,37], is based on Constrained Delaunay Triangulation (CDT) [38]. For a set of points in a plane, a Delaunay triangulation achieves an outcome such that no point from the set is inside the circumcircle of any triangle. Essentially, CDT divides the soccer field into a set of triangles, based on the set of predefined ball locations, each of which is mapped to the positions of each player. Moreover, when the ball takes any position within a triangle, each player's position is dy- namically adjusted during the runtime in a congruent way [36,37,9]. Overall, a formation defined via CDT is an ordered list of coordinates, and so, in terms of evolutionary computation, mutating and recombining such a list can be relatively easily automated and evaluated. Gliders2d: Source Code Base 7 Fig. 1. Example of a Delaunay triangulation, used by defense-formation.conf, produced by fedit2. The triangle formed by points 106, 108 and 110 is highlighted. When the ball is located at 110, the players are supposed to be located in the shown positions. Figure 1 shows a CDT fragment; for example, the point 110, where the ball is located, defines the following intended positions for the players: Ball -48.66 22.71 1 -50.72 6.07 2 -46.08 3.12 3 -47.6 10.53 4 -43.58 -3.75 5 -48.49 18.65 6 -44.3 13.29 7 -41.17 5.8 8 -40.32 17.03 9 -21.01 -17.44 10 -19.94 26.01 11 -22.62 5.8 8 Prokopenko and Wang The released changes in Gliders2d-1.5 formations are aimed at improving the defensive performance, placing the defenders and midfielders closer to the own goal. A notable performance gain was observed against all three benchmarks. The goal difference against agent2d increased from 2.387 to 3.210; against HELIOS2018: from -6.422 to -4.383; and against Gliders2013: from -1.039 to -0.344. 2.6 Gliders2d v1.6: Risky passes The final step presented in this release introduced risk level, expressed as the number of additional cycles "granted" to teammates receiving a pass, under a pressure from opponent players potentially intercepting the pass (strict_check_pass_generator.cpp). If risk level is set to zero, the default passing behaviour of agent2d is recovered. For positive values of risk the passes are considered as feasible even if an ideal opponent interceptor gets to the ball trajectory sooner than the intended recipient of the pass. // G2d: risk passes int risk = 0; if ( wm.ball().pos().x < wm.offsideLineX() && receive_point.x > wm.offsideLineX() + 3.0 && wm.offsideLineX() - receiver.player_->pos().x < 5.0 ) { } if (heliosbase) risk = 5; else if (helios2018) else risk = 0; risk = 2; if ( M_pass_type == 'T' ) { if ( o_step + risk <= step ) { ... failed = true; } ... } else { if ( o_step + risk <= step + ( kick_count - 1 ) ) { failed = true; } } // G2d: risk in opponent check int risk = 0; if ((receive_point.x < pass_max_x fabs(receive_point.y) > pass_min_y) && (M_pass_type == 'T' M_pass_type == 'L') && fabs(ball_move_angle.degree() - oppDir) > pass_cut && fabs(ball_move_angle.degree()) < pass_angle && wm.ball().pos().x < wm.offsideLineX() && receive_point.x > wm.offsideLineX() + pass_depth ) { } if (heliosbase) risk = 2; else risk = 1; int n_step = ( n_turn == 0 ? n_turn + n_dash + risk : n_turn + n_dash + 1 ); // 1 step penalty for observation delay The conditional statements in this fragment include several new variables, such as pass_max_x, pass_min_y, pass_cut, pass_angle, pass_depth, used in mutating and recombining the conditions. The addition of risky passes increased the goal difference against agent2d from 3.210 to 4.2; and against Gliders2013: from -0.344 to -0.212. Gliders2d: Source Code Base 9 Gliders2d Points for Points against Goals scored Goals conceded Goal diff. Std. error v0.0 (agent2d) v1.1 (stamina) v1.2 (pressing) v1.3 (evaluator) v1.4 (positioning) v1.5 (formations) v1.6 (risky passes) 1.384 1.345 2.161 2.252 2.515 2.785 2.840 1.414 1.468 0.691 0.607 0.367 0.154 0.116 2.287 2.254 2.642 2.997 3.849 3.995 5.214 2.289 2.290 1.355 1.381 1.461 0.785 1.014 -0.002 -0.036 1.288 1.616 2.387 3.210 4.200 0.040 0.049 0.051 0.063 0.086 0.181 0.172 Table 1. Performance evaluation for Gliders2d against agent2d, over ∼ 1000 games carried out for each version of Gliders2d against the opponent. The goal difference improves from zero to 4.2, while the average game score improves from (2.29 : 2.29) to (5.21 : 1.01). Gliders2d Points for Points against Goals scored Goals conceded v0.0 (agent2d) v1.1 (stamina) v1.2 (pressing) v1.3 (evaluator) v1.4 (positioning) v1.5 (formations) v1.6 (risky passes) 0.000 0.001 0.003 0.004 0.002 0.027 0.024 3.000 2.998 2.994 2.992 2.996 2.952 2.961 0.123 0.231 0.248 0.269 0.298 0.273 0.260 12.852 7.099 6.724 6.821 6.720 4.655 4.600 Goal diff. -12.729 -6.868 -6.476 -6.552 -6.422 -4.383 -4.337 Std. error 0.514 0.276 0.140 0.310 0.223 0.197 0.161 Table 2. Performance evaluation for Gliders2d against HELIOS2018, over ∼1000 games carried out for each version of Gliders2d against the opponent. The goal difference improves from −12.73 to −4.34, while the average game score improves from (0.12 : 12.85) to (0.26 : 4.60). Gliders2d Points for Points against Goals scored Goals conceded Goal diff. Std. error v0.0 (agent2d) v1.1 (stamina) v1.2 (pressing) v1.3 (evaluator) v1.4 (positioning) v1.5 (formations) v1.6 (risky passes) 0.022 0.183 0.539 0.657 0.603 1.039 1.111 2.968 2.730 2.230 2.109 2.160 1.607 1.527 0.569 0.596 0.613 0.770 0.700 0.700 0.776 6.052 3.280 1.760 1.800 1.739 1.044 0.988 -5.483 -2.684 -1.147 -1.030 -1.039 -0.344 -0.212 0.213 0.071 0.063 0.067 0.077 0.026 0.038 Table 3. Performance evaluation for Gliders2d against Gliders2013, over ∼1000 games carried out for each version of Gliders2d against the opponent. The goal difference improves from −5.48 to −0.21, while the average game score improves from (0.57 : 6.05) to (0.78 : 0.99). 10 Prokopenko and Wang 3 Conclusions In this paper, we described the first version of Gliders2d: a base code release for Gliders (based on agent2d- 3.1.1). We trace six sequential changes aligned with six evolutionary steps. These steps improve the overall control of the pitch by increasing the players' mobility through several means: less conservative usage of the available stamina balance (v1.1); more intense pressing of opponents (v1.2); selecting more diversified actions (v1.3); positioning forwards in open areas (v1.4); positioning defenders closer to own goal (v1.5); and considering riskier passes (v1.6). As has been argued in the past, the simulation leagues enable replicable and robust investigation of complex robotic systems [39,40]. We believe that the purpose of the RoboCup Soccer Simulation Leagues (both 2D and 3D) should be to simulate agents based on a futuristic robotic architecture which is not yet achievable in hardware. Aiming at such a general and abstract robot architecture may help to identify a standard for what humanoid robots may look like in 2050, the year of the RoboCup Millennium challenge. This is the reason for focussing, in this release, on the features which can also be used by simulated 3D, as well as robotic, teams competing in RoboCup, aiming at some of the most general questions: when to conserve energy (stamina), when to run (pressing), where to kick the ball (actions), where to be on the field (positioning in attack and defense), and when to take risks (passes). While the provided specific answers may or may not be widely acceptable, general reasoning along these lines may bring us closer to a new RoboCup Humanoid Simulation League (HSL). In HSL, the Simulated Humanoid should be defined in a standard and generalisable way, approaching human soccer-playing behavior [41], while the behavioural and tactical improvements can be evolved and/or adapted to this standardised architecture. The location of the released code: http://www.prokopenko.net/gliders2d.html. The last presented version, Gliders2d-v1.6, is comparable to Gliders2013, achieving the average score of (0.78 : 0.99) against this benchmark, and outperforms agent2d-3.1.1 with the average score (5.21 : 1.01). In tracing this evolutionary branch, we illustrated the methodology of human-based evolutionary com- putation, showing that even a small number of controlled steps can dramatically improve the overall team performance. 4 Acknowledgments We thank several members of Gliders team contributing during 2012 -- 2016: David Budden, Oliver Cliff, Victor Jauregui and Oliver Obst. We are also grateful to participants of the discussion on the future of the RoboCup Simulation Leagues, in particular to Peter Stone, Patrick MacAlpine, Nuno Lau, Klaus Dorer, and Daniel Polani. References 1. Burkhard, H.D., Duhaut, D., Fujita, M., Lima, P., Murphy, R., Rojas, R.: The road to RoboCup 2050. Robotics Automation Magazine 9(2) (Jun 2002) 31 -- 38 IEEE 2. Noda, I., Stone, P.: The RoboCup Soccer Server and CMUnited Clients: Implemented Infrastructure for MAS Research. Autonomous Agents and Multi-Agent Systems 7(1 -- 2) (July -- September 2003) 101 -- 120 3. Riley, P., Stone, P., Veloso, M.: Layered disclosure: Revealing agents' internals. In Castelfranchi, C., Lesperance, Y., eds.: Intelligent Agents VII. Agent Theories, Architectures, and Languages -- 7th. International Workshop, ATAL-2000, Boston, MA, USA, July 7 -- 9, 2000, Proceedings. Lecture Notes in Artificial Intelligence. Springer, Berlin, Berlin (2001) Gliders2d: Source Code Base 11 4. Stone, P., Riley, P., Veloso, M.: Defining and using ideal teammate and opponent models. In: Proceedings of the Twelfth Annual Conference on Innovative Applications of Artificial Intelligence. (2000) 5. Butler, M., Prokopenko, M., Howard, T.: Flexible synchronisation within RoboCup environment: A comparative analysis. In: RoboCup 2000: Robot Soccer World Cup IV, London, UK, Springer (2001) 119 -- 128 6. Reis, L.P., Lau, N., Oliveira, E.: Situation based strategic positioning for coordinating a team of homogeneous agents. In: Balancing Reactivity and Social Deliberation in Multi-Agent Systems, From RoboCup to Real-World Applications, London, UK, Springer (2001) 175 -- 197 7. Prokopenko, M., Wang, P.: Relating the entropy of joint beliefs to multi-agent coordination. In Kaminka, G.A., Lima, P.U., Rojas, R., eds.: RoboCup 2002: Robot Soccer World Cup VI. Volume 2752 of Lecture Notes in Com- puter Science., Springer (2003) 367 -- 374 8. Prokopenko, M., Wang, P.: Evaluating team performance at the edge of chaos. In Polani, D., Browning, B., Bonarini, A., Yoshida, K., eds.: RoboCup 2003: Robot Soccer World Cup VII. Volume 3020 of Lecture Notes in Computer Science., Springer (2004) 89 -- 101 9. Prokopenko, M., Wang, P.: Disruptive Innovations in RoboCup 2D Soccer Simulation League: From Cyberoos'98 to Gliders2016. In Behnke, S., Sheh, R., Sariel, S., Lee, D.D., eds.: RoboCup 2016: Robot World Cup XX [Leipzig, Germany, June 30 - July 4, 2016]. Volume 9776 of Lecture Notes in Computer Science., Springer (2017) 529 -- 541 10. Zuparic, M., Jauregui, V., Prokopenko, M., Yue, Y.: Quantifying the impact of communication on performance in multi-agent teams. Artificial Life and Robotics 22(3) (Sep 2017) 357 -- 373 11. Stone, P., Asada, M., Balch, T.R., Fujita, M., Kraetzschmar, G.K., Lund, H.H., Scerri, P., Tadokoro, S., Wyeth, G.: Overview of robocup-2000. In Stone, P., Balch, T.R., Kraetzschmar, G.K., eds.: RoboCup 2000: Robot Soccer World Cup IV. Volume 2019 of Lecture Notes in Computer Science., Springer (2000) 1 -- 28 12. Stone, P., Riley, P., Veloso, M.: The CMUnited-99 champion simulator team. In Veloso, M., Pagello, E., Kitano, H., eds.: RoboCup-99: Robot Soccer World Cup III. Volume 1856 of Lecture Notes in Artificial Intelligence. Springer Verlag, Berlin (2000) 35 -- 48 13. Kok, J.R., Vlassis, N., Groen, F.: UvA Trilearn 2003 team description. In Polani, D., Browning, B., Bonarini, A., Yoshida, K., eds.: Proceedings CD RoboCup 2003, Padua, Italy, Springer-Verlag (July 2003) 14. Tavafi, A., Nozari, N., Vatani, R., Yousefi, M.R., Rahmatinia, S., Pirdir, P.: MarliK 2012 Soccer 2D Simulation In: RoboCup 2012 Symposium and Competitions: Team Description Papers, Mexico Team Description Paper. City, Mexico, June 2012. (2012) 15. Akiyama, H.: Agent2D Base Code. http://www.rctools.sourceforge.jp (2010) 16. Prokopenko, M., Wang, P.and Obst, O., Jaurgeui, V.: Gliders2016: Integrating multi-agent approaches to tactical diversity. In: RoboCup 2016 Symposium and Competitions: Team Description Papers, Leipzig, Germany, July 2016. (2016) 17. Nehaniv, C., Polani, D., Olsson, L., Klyubin, A.: Evolutionary information-theoretic foundations of sensory ecol- ogy: Channels of organism-specific meaningful information. Modeling Biology: Structures, Behaviour, Evolution (2005) 9 -- 11 18. Prokopenko, M., Gerasimov, V., Tanev, I.: Measuring spatiotemporal coordination in a modular robotic system. In Rocha, L., Yaeger, L., Bedau, M., Floreano, D., Goldstone, R., Vespignani, A., eds.: Artificial Life X: Proceedings of The 10th International Conference on the Simulation and Synthesis of Living Systems, Bloomington IN, USA (2006) 185 -- 191 19. Prokopenko, M., Gerasimov, V., Tanev, I.: Evolving spatiotemporal coordination in a modular robotic system. In Nolfi, S., Baldassarre, G., Calabretta, R., Hallam, J.C.T., Marocco, D., Meyer, J.A., Miglino, O., Parisi, D., eds.: From Animals to Animats 9: 9th International Conference on the Simulation of Adaptive Behavior (SAB 2006), Rome, Italy, September 25-29 2006. Volume 4095 of Lecture notes in computer science. (2006) 558 -- 569 20. Prokopenko, M.: Guided self-organization: Inception. Volume 9. Springer Science & Business Media (2013) 21. Kosorukoff, A.: Human based genetic algorithm. In: Systems, Man, and Cybernetics, 2001 IEEE International Conference on. Volume 5., IEEE (2001) 3464 -- 3469 22. Cheng, C.D., Kosorukoff, A.: Interactive one-max problem allows to compare the performance of interactive and human-based genetic algorithms. In Deb, K., ed.: Genetic and Evolutionary Computation -- GECCO 2004: Genetic and Evolutionary Computation Conference, Seattle, USA, June 26-30, 2004. Springer (2004) 983 -- 993 12 Prokopenko and Wang 23. Prokopenko, M., Obst, O., Wang, P., Held, J.: Gliders2012: Tactics with action-dependent evaluation functions. In: RoboCup 2012 Symposium and Competitions: Team Description Papers, Mexico City, Mexico, June 2012. (2012) 24. Prokopenko, M., Obst, O., Wang, P., Budden, D., Cliff, O.M.: Gliders2013: Tactical analysis with information dy- namics. In: RoboCup 2013 Symposium and Competitions: Team Description Papers, Eindhoven, The Netherlands, June 2013. (2013) 25. Budden, D., Prokopenko, M.: Improved particle filtering for pseudo-uniform belief distributions in robot localisa- tion. In: RoboCup 2013: Robot Soccer World Cup XVII, Springer (2013) 26. Lizier, J.T., Prokopenko, M., Zomaya, A.Y.: Coherent information structure in complex computation. Theory in Biosciences 131 (2012) 193 -- 203 27. Cliff, O.M., Lizier, J., Wang, R., Wang, P., Obst, O., Prokopenko, M.: Towards quantifying interaction networks in a football match. In Behnke, S., Veloso, M., Visser, A., Xiong, R., eds.: RoboCup 2013: Robot Soccer World Cup XVII, Springer (2013) 1 -- 12 28. Cliff, O.M., Lizier, J.T., Wang, X.R., Wang, P., Obst, O., Prokopenko, M.: Quantifying long-range interactions and coherent structure in multi-agent dynamics. Artificial Life 23(1) (2017) 34 -- 57 29. Prokopenko, M., Obst, O., Wang, P.: Gliders2014: Dynamic Tactics with Voronoi Diagrams. In: RoboCup 2014 Symposium and Competitions: Team Description Papers, Joao Pessoa, Brazil, July 2014. (2014) 30. Prokopenko, M., Wang, P., Obst, O.: Gliders2015: Opponent avoidance with bio-inspired flocking behaviour. In: RoboCup 2015 Symposium and Competitions: Team Description Papers, Hefei, China, July 2015. (2015) 31. Cioppa, T.M., Lucas, T.W.: Efficient nearly orthogonal and space-filling latin hypercubes. Technometrics 49(1) (2007) 45 -- 55 32. Gabel, T., Kloppner, P., Godehardt, E., Tharwat, A.: Communication in soccer simulation: On the use of wiretap- ping opponent teams. In: RoboCup 2018: Robot Soccer World Cup XXII, Springer (2018) 33. Tanev, I., Yuta, K.: Epigenetic programming: Genetic programming incorporating epigenetic learning through modification of histones. Information Sciences 178(23) (2008) 4469 -- 4481 34. Nakashima, T., Akiyama, H., Suzuki, Y., Ohori, A., Fukushima, T.: HELIOS2018: Team Description Paper. In: RoboCup 2018 Symposium and Competitions: Team Description Papers, Montreal, Canada, July 2018. (2018) 35. Dylla, F., Ferrein, A., Lakemeyer, G., Murray, J., Obst, O., Rofer, T., Schiffer, S., Stolzenburg, F., Visser, U., Wagner, T.: Approaching a formal soccer theory from the behavior specification in robotic soccer. In Dabnicki, P., Baca, A., eds.: Computers in Sport. Bioengineering. WIT Press (2008) 161 -- 186 36. Akiyama, H., Noda, I.: Multi-agent positioning mechanism in the dynamic environment. In Visser, U., Ribeiro, F., Ohashi, T., Dellaert, F., eds.: RoboCup 2007: Robot Soccer World Cup XI. Springer, Berlin, Heidelberg (2008) 377 -- 384 37. Akiyama, H., Shimora, H.: Helios2010 team description. In: RoboCup 2010: Robot Soccer World Cup XIV. Volume 6556 of Lecture Notes in Computer Science., Springer (2011) 38. Chew, L.P.: Constrained Delaunay Triangulations. Algorithmica 4(1-4) (1989) 97 -- 108 39. Budden, D.M., Wang, P., Obst, O., Prokopenko, M.: Robocup simulation leagues: Enabling replicable and robust investigation of complex robotic systems. IEEE Robotics and Automation Magazine 22(3) (2015) 140 -- 146 40. Prokopenko, M., Wang, P., Marian, S., Bai, A., Li, X., Chen, X.: Robocup 2d soccer simulation league: Evaluation challenges. In Akiyama, H., Obst, O., Sammut, C., Tonidandel, F., eds.: RoboCup 2017: Robot World Cup XXI [Nagoya, Japan, July 27-31, 2017]. Volume 11175 of Lecture Notes in Computer Science., Springer (2018) 325 -- 337 41. Stone, P., Quinlan, M., Hester, T.: Can robots play soccer? In Richards, T., ed.: Soccer and Philosophy: Beautiful Thoughts on the Beautiful Game. Volume 51 of Popular Culture and Philosophy. Open Court Publishing Company (2010) 75 -- 88
1703.05626	1	1703	2017-03-16T14:04:38	Scalable Accelerated Decentralized Multi-Robot Policy Search in Continuous Observation Spaces	[ "cs.MA", "cs.RO" ]	This paper presents the first ever approach for solving \emph{continuous-observation} Decentralized Partially Observable Markov Decision Processes (Dec-POMDPs) and their semi-Markovian counterparts, Dec-POSMDPs. This contribution is especially important in robotics, where a vast number of sensors provide continuous observation data. A continuous-observation policy representation is introduced using Stochastic Kernel-based Finite State Automata (SK-FSAs). An SK-FSA search algorithm titled Entropy-based Policy Search using Continuous Kernel Observations (EPSCKO) is introduced and applied to the first ever continuous-observation Dec-POMDP/Dec-POSMDP domain, where it significantly outperforms state-of-the-art discrete approaches. This methodology is equally applicable to Dec-POMDPs and Dec-POSMDPs, though the empirical analysis presented focuses on Dec-POSMDPs due to their higher scalability. To improve convergence, an entropy injection policy search acceleration approach for both continuous and discrete observation cases is also developed and shown to improve convergence rates without degrading policy quality.	cs.MA	cs	Scalable Accelerated Decentralized Multi-Robot Policy Search in Continuous Observation Spaces Shayegan Omidshafiei1, Christopher Amato2, Miao Liu3, Michael Everett1, Jonathan P. How1, John Vian4 7 1 0 2 r a M 6 1 ] A M . s c [ 1 v 6 2 6 5 0 . 3 0 7 1 : v i X r a Abstract- This paper presents the first ever approach for solving continuous-observation Decentralized Partially Observ- able Markov Decision Processes (Dec-POMDPs) and their semi- Markovian counterparts, Dec-POSMDPs. This contribution is especially important in robotics, where a vast number of sensors provide continuous observation data. A continuous-observation policy representation is introduced using Stochastic Kernel- based Finite State Automata (SK-FSAs). An SK-FSA search algorithm titled Entropy-based Policy Search using Contin- uous Kernel Observations (EPSCKO) is introduced and ap- plied to the first ever continuous-observation Dec-POMDP/Dec- POSMDP domain, where it significantly outperforms state- of-the-art discrete approaches. This methodology is equally applicable to Dec-POMDPs and Dec-POSMDPs, though the empirical analysis presented focuses on Dec-POSMDPs due to their higher scalability. To improve convergence, an entropy in- jection policy search acceleration approach for both continuous and discrete observation cases is also developed and shown to improve convergence rates without degrading policy quality. I. INTRODUCTION Decision-making under uncertainty is a ubiquitous robotics problem wherein a robot collects data from its environment and decides subsequent tasks to execute. While low-cost robotics platforms and sensors have increased the affordability of multi-robot systems, derivation of policies dictating robot decisions remains a challenge. This decision-making problem is even more complex in noisy settings with imperfect com- munication, requiring a formal framework for its treatment. A general representation of the multi-agent planning under uncertainty problem is the Decentralized Partially Observable Markov Decision Process (Dec-POMDP) [1], which extends single-agent POMDPs to decentralized domains. Due to Dec-POMDPs' usage of primitive actions (atomic actions assumed to each take a single time unit to execute) they have exceedingly large policy spaces which severely limits planning scalability. Recent efforts have extended Dec-POMDPs to use macro-actions (temporally extended actions), resulting in the Decentralized Partially Observable Semi-Markov Decision Process (Dec-POSMDP) [2], [3]. The result is a scalable asynchronous multi-robot decision-making framework which plans over the space of high-level robot tasks (e.g., Open-the- valve or Find-the-key) with non-deterministic durations. *This work was supported by The Boeing Company. 1Laboratory for Information and Decision Systems (LIDS), MIT, Cam- bridge, MA 02139, USA {shayegan,jhow}@mit.edu 2College of Computer and Information Science (CCIS), Northeastern University, Boston, MA 02115, USA [email protected] 3IBM Thomas J. Watson Research Center, Yorktown Heights, NY 10598, USA [email protected] (work completed while the author was at MIT) 4Boeing Research & Technology, Seattle, WA 98108, USA [email protected] Despite the increased action-space scalability offered by Dec-POSMDPs, they have so far been limited to planning over the space of discrete observations. To date, no algorithms exist for continuous-observation Dec-POSMDPs (or Dec-POMDPs [4]). This is a major research gap, especially important in the context of robotics where a vast number of real-world sensors provide continuous observation data. Application of Dec- POSMDPs to continuous problems such as robot navigation currently mandates observation space discretization, resulting in loss of valuable sensor information which could otherwise be used to better inform the decision-making policy. Several approaches have targeted single-agent continuous-observation POMDPs. These include partitioning of continuous spaces into lossless discrete spaces [5], Gaussian mixtures for belief representation [6], use of continuous-observation classifiers [7], and learned discrete representations for continuous state spaces [8]. This paper expands this body of work beyond the single-agent case, targeting scalable treatment of continuous- observation Dec-POSMDPs. The methods presented are applicable to domains with continuous underlying state spaces, as shown in some of the experiments used for evaluation. In order to develop solvers for continuous-observation Dec- POSMDPs, we build on current state-of-the-art discrete policy search methods [2], [3], [9]. Unfortunately, these algorithms suffer from convergence speed limitations-an issue which was identified in prior work but remains untreated [9]. A major gap exists in addressing these issues before extending the foundations of these discrete algorithms to the continuous case, where such convergence issues are exacerbated. To resolve this, we first introduce a maximal entropy injection approach targeting convergence acceleration for both discrete and continuous algorithms, without degrading overall policy quality. The approach is shown to significantly outperform existing search acceleration methods. The paper's key contribution is a stochastic kernel-based policy representation and search algorithm, allowing direct mapping of continuous observations to robot decisions (with no discretization necessary). This algorithm leverages the pro- posed entropy injection acceleration method and is evaluated on a multi-robot nuclear contamination domain-the first ever continuous-observation Dec-POMDP/Dec-POSMDP domain- in which discrete policy search algorithms perform extremely poorly. Failure modes of discrete methods are analyzed and compared to the superior continuous policy behavior. The contributions introduced in this paper can be readily applied to Dec-POMDPs and Dec-POSMDPs. However, as we are mo- tivated by applications to extremely large action-observation spaces, the notation used and experiments conducted focus on the more scalable Dec-POSMDP framework. II. BACKGROUND A. Decentralized Planning using Macro-Actions This section summarizes the Dec-POSMDP, a multi-robot decentralized decision-making under uncertainty framework targeting action-space scalability. For a more detailed intro- duction to Dec-POSMDPs, we refer readers to [2], [3], [9]. The Dec-POSMDP is a belief-space framework in which agents execute macro-actions (temporally-extended actions) with non-deterministic completion times, and receive noisy high-level observations of their post-MA state. Macro-actions (MAs) are abstractions of low-level POMDPs involving prim- itive actions u(i) , allowing execution of t high-level tasks (e.g., Park-the-car)1. Each MA executes until an -neighborhood of its belief milestone bgoal is reached. This neighborhood defines the MA termination condition or goal belief node, denoted Bgoal ={b : (cid:107)b − bgoal(cid:107) ≤ } [3]. Upon completion of an MA, each robot makes a macro (or high-level) observation oe(i) of the underlying high- level system state xe ∈ Xe. It also calculates its own final belief state, bf (i). Thus far, both Dec-POMDPs and Dec-POSMDPs have only seen limited applications to finite discrete observation spaces. Due to its action-space scalability, let us focus on the Dec-POSMDP, defined as follows: and observations o(i) t • I = {1, 2, . . . , n} is the set of heterogeneous robots. • B(1) × B(2) × . . . × B(n) × Xe is the belief space, with local belief milestones B(i) and joint environment (or high-level) space Xe. • ¯T = T(1) ×T(2) . . .×T(n) is the joint MA space, where T(i) is the finite set of MAs for the i-th robot. • ¯Oe = {¯oe} is the space of all joint MA-observations. , k¯b, xe; ¯π) is the high-level transition proba- • P (¯b(cid:48), xe(cid:48) bility model under MAs ¯π from (¯b, xe) to (¯b(cid:48), xe(cid:48) ). MA ¯π at (¯b, xe). with joint observation ¯oe = {oe(1), oe(2), . . . , oe(n)}. • ¯Rτ(¯b, xe; ¯π) is the high-level reward of taking a joint • P (¯oe¯b, xe) is the joint observation likelihood model, • γ ∈ [0, 1) is the reward discount factor. Macro-observations and final beliefs are jointly denoted as MA-observation oe(i) = (oe(i), bf (i)). Trajectories of MAs and received MA-observations are denoted as the MA-history, k = {oe(i) ξ(i) 0 , π(i) 0 , oe(i) , π(i) Transition probability P (¯b(cid:48), xe(cid:48) (¯b(cid:48), xe(cid:48) (1) , k¯b, xe;¯π) from (¯b, xe) to ) under joint MA ¯π ={π(1),. . ., π(n)} in k timesteps is, 1 , . . . , oe(i) k−1, π(i) 1 k }. k−1, oe(i) P (¯b(cid:48), xe(cid:48) = , k¯b0, xe 0, oe (cid:88) (cid:104) k; ¯π) = P (xe k−1, oe P (xe kxe k, ¯bk¯b0, xe 0, oe k; ¯π(¯bk−1))× k; ¯π) k−1,¯bk−1 xe k−1, ¯bk−1; ¯π(¯bk−1))P (xe P (¯bkxe k−1, ¯bk−1xe 0, ¯b0; ¯π(¯b0)) . (2) (cid:105) The generalized high-level team reward for a discrete-time 1We denote a generic parameter p of the i-th robot as p(i), a joint team parameter as ¯p, and a joint team parameter at timestep k as ¯pk. Dec-POSMDP during execution of joint MA ¯π is defined [9], ¯Rτ(¯b, xe; ¯π) =E γt ¯R(¯xt, xe t, ¯ut)P (¯x0) =¯b, xe 0 = xe; ¯π (3) (cid:35) (cid:34)τ−1(cid:88) t=0 where τ = mini mint{t : b(i) timestep at which any robot completes its current MA. t ∈ B(i),goal} is the first The joint high-level policy, ¯φ = {φ(1), . . . , φ(n)}, dictates MA selection. High-level policy φ(i) maps the i-th robot's MA-history ξ(i) to the next MA π(i) to be executed. Joint k Dec-POSMDP value under policy ¯φ is then [9], ; ¯πtk )¯b0, xe ¯φ(¯b, xe) = E γtk ¯Rτ (¯btk , xe tk 0; ¯φ (cid:35) (4) ¯V (cid:34) ∞(cid:88) γtk(cid:88) k=0 ∞(cid:88) = ¯Rτ(¯b, xe; ¯π)+ P (¯b(cid:48), xe(cid:48) ,oe(cid:48) k=1 ¯b(cid:48),xe(cid:48) , oe(cid:48) , k¯b, xe; ¯π) ¯V ¯φ(¯b(cid:48), xe(cid:48) ). (5) The optimal joint high-level policy is, ¯φ∗ = argmax ¯V ¯φ(¯b, xe). ¯φ (6) Solving the Dec-POSMDP results in joint high-level decision-making policy ¯φ dictating the MA π(i) executed by each robot based on its MA-history. Each MA is, itself, a policy over low-level actions u(i) . Thus, t decision-making using the Dec-POSMDP allows abstraction of task-level actions from low-level actions, leading to sig- nificantly improved planning scalability over Dec-POMDPs. B. Dec-POSMDP Policy Search Algorithms and observations o(i) t So far, research efforts have focused on Dec-POSMDP policy search for discrete observation spaces, resulting in several algorithms: Masked Monte Carlo Search (MMCS) [3], MacDec-POMDP Heuristic Search (MDHS) [2], and Graph-based Direct Cross Entropy method (G-DICE) [9]. These algorithms use Finite State Automata (FSAs) for policy representation. FSA-based policy φ(i) for robot i consists of }. FSA-based decision-making Nn FSA nodes, {q(i) is two-fold: each robot begins execution in FSA node q(i), where MA output function π(i) = λ(i)(q(i)) assigns it an MA, π(i). Following MA execution, the robot receives a high-level observation and selects its next FSA node using transition function q(cid:48)(i) = δ(i)(q(i), oe(i)). The graph-based nature of FSAs allows their application to infinite-horizon domains. 1 , . . . , q(i) Nn Though Dec-POSMDPs have increased the size of solvable planning domains beyond Dec-POMDP counterparts, major algorithm limitations still exist. MMCS is a greedy algorithm which succumbs to local optimality issues [3]. MDHS uses lower and upper bound value heuristics to bias search towards promising policy regions, by initiating an empty (partial) FSA and incrementally assigning nodes actions λ(i) and transitions δ(i). Partial policies with high upper bounds are expanded incrementally. Yet, each expansion involves TN child policies, severely limiting usage for large observation spaces. G-DICE is a cross entropy-based algorithm which iter- atively updates policies using two sampling distributions at each FSA node: MA distribution f (π(i)q(i); θ(i)(πq)) ¯O n and node transition distribution f (q(i)(cid:48)q(i), oe(i); θ(i)(q(cid:48)q,oe)), where θ(i) are parameter vectors. Each iteration samples the distributions Ns times, resulting in Ns deterministic FSA policies. Maximum likelihood estimates (MLE) of parameters θ(i) are calculated using the Nb≤ Ns best policies. To prevent convergence to local optima, smooth parameter updates, θk+1 ← αθk+1 + (1 − α)θk, (7) are used, with iteration number k and learning rate α ∈ (0, 1]. For sufficiently small values of α, this process minimizes cross entropy between each sampling distribution and a unit mass centered at the optimal policy [10]. G-DICE is executed until convergence, after which the best deterministic policy from the history of samples is returned. Using smooth parameter updates and sampling distributions initiating from a uniform distribution allows G-DICE to tradeoff exploration and exploitation in the policy space, outperforming other Dec-POSMDP search approaches given a fixed computational budget. Yet, G-DICE suffers from sample degeneracy and convergence issues related to the sampling distributions, and in its current form only applies to discrete observation settings. The following sections resolve these issues, resulting in a scalable, accelerated continuous- observation search algorithm. III. ACCELERATED POLICY SEARCH Prior to extending to continuous observations, this section treats the sampling distribution degeneracy issue in sampling- based Dec-POSMDP approaches. It also introduces a maximal entropy injection scheme which is then embedded in the proposed continuous-observation Dec-POSMDP algorithm. A. Sampling Distribution Degeneracy Problem A major issue with sampling distribution-based approaches, such as G-DICE, occurs when a low enough learning rate α is not used, causing underlying sampling distributions to rapidly converge to degenerate distributions far from the optimum [11]. All subsequent search iterations return identical samples of the policy space, stifling exploration altogether. Yet, one benefit of a high learning rate is fast convergence, especially useful for complex Dec-POSMDPs with large observation spaces and computationally expensive trajectory sampling and evaluation. Sampling distribution-based approaches such as G-DICE often require hand-tuned selection of α for good performance, even after which convergence may be exces- sively slow and can hinder experimentation and analysis. This trade-off was noted in [9], where it was left as future work. Recall the motivation behind the Dec-POSMDP framework is scalability to very large multi-robot planning domains. Despite the fact that policy search is conducted offline, hindrance of human-in-the-loop analysis due to slow convergence is undesirable. A naıve solution is to set α arbitrarily low, but this implies arbitrarily high convergence time (on the order of many days for complex domains). These foundational issues must first be resolved before extending these algorithms to treat the more complex continuous observation case. Several works have targeted this degeneracy problem. One approach uses dynamic smoothing of learning rates [12], αk = α0 − α0(1 − k−1)β, (8) where α0 is the baseline rate (typically close to 1) and β is the drop-off rate (typically between 5 to 10). The result is a monotonically decreasing αk which initially starts high. Another approach involves the addition of a noise term ωk to the sampling distribution at each iteration k to prevent degeneration. Linearly decreasing noise injection, ωk = max(ωmax − rk, 0), (9) was investigated in [13]. In the above, ωmax is the maximum allowable noise and r is the noise drop-off rate. These approaches are not ideal as they are agnostic to Dec-POSMDP value function convergence, meaning they do not adapt to domain-specific behaviors. Thus, sub-parameters (α0, β, ωmax, r) typically need significant tuning to alleviate convergence issues for individual domains. B. Maximal Entropy Injection A principled approach combining policy exploration with fast convergence is desired, without reliance on sensitive dy- namic smoothing or noise terms. As degenerate distributions have minimal entropy [14], an intuitive idea is to simulta- neously monitor policy value convergence and underlying sampling distribution entropy to alleviate degeneracy issues. In the proposed acceleration approach, search is conducted as usual for iterations where policy value has not converged, allowing policy space exploration. Once convergence occurs, entropies of sampling distributions f (π(i)q(i); θ(i)(πq)) and f (q(i)(cid:48)q(i), oe(i); θ(i)(q(cid:48)q,oe)) are calculated. If a distribu- tion's entropy is significantly below the max entropy for its distribution family, degeneracy has likely occurred [14]. Max entropy distributions are well-studied and closed form results for many families and constraint sets are known [15]. For Dec-POSMDPs, these entropy calculations are computa- tionally cheap as sampling distributions are categorical, with corresponding discrete uniform maximal entropy distributions. In post-degeneracy iterations, each sampling distribution's entropy is increased by incrementally combining its parame- ters θ(i) with the max entropy distribution parameters θM E, θk+1 ← (1 − αEI )[αθk+1 + (1 − α)θk] + αEI θM E, where αEI is the entropy injection rate. This encourages policy space exploration while still allowing usage of high learning rates (e.g., α > 0.5) for fast convergence. In practice, entropy injection rate αEI has a low value (between 1% - 3% per iteration). As this process is repeated only in post-convergence iterations, there is low sensitivity to αEI as entropy is incrementally increased whenever necessary. Injection stops as soon as the policy value diverges, allowing unhindered exploration. This acceleration approach is eval- uated in Section V-A and also integrated into the proposed continuous-observation search algorithm in the next section. (10) IV. CONTINUOUS-OBSERVATION DEC-POSMDP SEARCH This section focuses on multi-robot policy search in continuous observation spaces. It first presents an extension of traditional discrete, deterministic FSAs to allow repre- sentation of continuous policies. A continuous-observation Dec-POSMDP search algorithm is then introduced. (b) Given a high-level continuous observation made following MA execution (e.g., oe = (2.5, 1.5) above), the transition function outputs a categorical transition distribution over next-nodes q(cid:48). The robot samples this distribution to select its next node, q(cid:48). (a) The robot samples an MA using its node's MA distribution. Fig. 1: Overview of continuous-observation decision-making using SK-FSAs. A given robot's policy is represented by a set of stochastic FSA nodes, each containing an MA sampling distribution and node transition function. A 4 node SK-FSA (Nn = 4) is illustrated above, with the robot starting policy execution at SK-FSA node q = 3 (on the left). (c) The robot repeats this decision-making pro- cess at the next SK-FSA node q(cid:48) (one of the 4 nodes above), conducting the stochastic MA and transition selection process indefinitely. A. Stochastic Kernel-Based Finite State Automata We first extend the notion of deterministic policies used in existing Dec-POSMDP algorithms to stochastic policies. In a stochastic FSA, MA output function λ(i) and node transition function δ(i) provide robots with a probability distribution over MAs and next-nodes q(cid:48) during policy execution, rather than deterministic MA and transition assignments. The result- ing stochastic decision-making scheme allows robots to escape cycles of incorrect decisions which may otherwise occur in deterministic FSAs [16]. While it has been shown that finite- horizon Dec-POMDPs have at least one optimal deterministic policy (i.e., guaranteed to at least equal performance of the optimal stochastic policy) [17], in approximate searches, stochastic FSAs often result in a higher joint value [16], [18]. One can readily modify cross entropy-based search to provide such a stochastic policy by simply using the underlying sampling distributions f (π(i)q(i); θ(i)(πq)) and f (q(i)(cid:48)q(i), oe(i); θ(i)(q(cid:48)q,oe)) to define the policy, rather than the best sampled deterministic policy (as done in G-DICE). A second issue is extension of FSAs to support continuous observations, a formidable task as continuous observation spaces are uncountably infinite. Existing Dec-POSMDP algorithms are, thus, inapplicable. To resolve this, we assume policy smoothness over the observation space, a characteristic which occurs naturally in many robotics domains. In other words, the controller structure should induce similar decisions from similar observation chains. This typical assumption is also made by the continuous state-action MDP and POMDP literature [7], [8], [19]. We exploit this smoothness assumption and introduce Stochastic Kernel-based Finite State Automata (SK-FSAs) for policy representation (Fig. 1), which have similar structure to the controllers used in [7]. Policy execution in SK- FSAs is similar to traditional FSAs. Each robot's SK- FSA node (e.g., node q = 3 in Fig. 1) outputs cat- egorical MA distribution f (π(i)q(i); θ(i)(πq)), which the robot samples to select its next MA (Fig. 1a). Following MA execution, the robot receives a continuous high-level observation, which the SK-FSA node transition function δ(i) uses to output a corresponding node transition distribution f (q(i)(cid:48)q(i), oe(i); θ(i)(q(cid:48)q,oe)). Note the distinction between transition function and transition distribution-the transition function maps continuous observations to the Nn-dimensional simplex. Given an observation, δ(i) outputs an infinitesimal 'slice' representing a categorical transition distribution over next-nodes q(cid:48). Fig. 1b illustrates such a slice, evaluated at high-level observation oe = (2.5, 1.5). The robot samples this categorical distribution, transitions to its next SK-FSA node q(cid:48), and repeats this process indefinitely. We propose use of kernel logistic regression (KLR) to represent node transition functions. KLR is a non-parametric multi-class classification model (i.e., model complexity grows with the number of kernel points). In SK-FSAs, node transition functions use KLR with high-level observation inputs, oe, and output probabilities over next-nodes q(cid:48). KLR is a natural model for stochastic policies as it is a probabilistic classifier (i.e., SK-FSA transition distributions correspond to KLR probabilities) [20]. Our approach uses KLR with radial basis function (RBF) kernels over the observation space, K(oe, oe(cid:48) ) = exp(−0.5σ−2oe − oe(cid:48)2), (11) where σ is the kernel radius. RBF kernels are preferred as they provide smooth classification outputs while allowing non- linear decision boundaries [20], in contrast to linear kernels. The next section discusses SK-FSA policy search, including details on kernel basis selection and kernel weight training. B. Entropy-based Policy Search over SK-FSAs This section introduces an SK-FSA search algorithm titled Entropy-based Policy Search using Continuous Kernel Observations (EPSCKO). EPSCKO consists of 3 steps: cross entropy search for MA distributions (as done in G-DICE), memory-bounded KLR training for SK-FSA node transition functions, and entropy injection for search acceleration (as in SK-FSA Node 𝑞=3MA Distribution𝑃(𝜋)𝜋1𝜋2𝜋3𝜋4𝜋5𝑃(𝑞′=2𝑜𝑒)𝑃(𝑞′=3𝑜𝑒)𝑃(𝑞′=4𝑜𝑒)𝑃(𝑞′=1𝑜𝑒)Node Transition Function𝑞′=1𝑞′=4𝑞′=2𝑞′=3Node Transition Distribution8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 Section III-B). In each EPSCKO iteration, decision trajectories are sampled from the SK-FSA policy. The Nb best trajectories (evaluated using (4)) are used for policy update. We first detail the KLR training approach and then present the overall algorithm. As transition function δ(i) uses a kernel-based representation over the observation space, it requires a set of observation kernel basis points and weights. In EPSCKO, kernel weights constitute the node transition parameter vector θ(i)(q(cid:48)q,oe). To simplify notation, references to θ(i) in this section refer to this transition parameter vector. q(cid:48) The computational cost of training KLR models is O(N 3 d ) [20], where Nd is the training input size. For a sustainable training time, EPSCKO uses a memory-bounded kernel basis consisting of continuous observations received during evaluation of the Nb best policies in each of the latest NKLR iterations. In each iteration, the bundle of observations in the Nb best decision trajectories is pushed to a first-in, first-out (FIFO) circular queue of length NKLR. KLR training outputs are the corresponding sampled node transitions taken along these same trajectories. The non-parametric nature of KLR ensures that node transition function complexity increases in regions with high observation density, so the policy naturally focuses on prominent observation space regions. The result is a compact yet informative policy representation. To counter convergence to locally optimal SK-FSAs, EPSKCO uses a weighted log-likelihood function to train the KLR model. Weights are discounted such that observations sampled in earlier algorithm iterations are given higher value. Given learning rate α, the following weight set is used, wb = (1 − α)NKLR−1 α(1 − α)NKLR−b b = 1 b ∈ {2,··· , NKLR} (12) (cid:40) where wb is the training weight for the b-th observation bundle in the FIFO kernel queue. This weighting is derived from recursive application of (7), and is analogous to the smoothing step used in G-DICE. For each robot i, the weighted log- likelihood function is maximized over θ(i) for KLR training, q(cid:48) NKLR(cid:88) b=1 l(i)(θ(i)) = wb log (cid:80)Nn q(cid:48) oe(i) exp(θ(i)T b k=1 exp(θ(i)T qk oe(i) ) b b (13) , ) (cid:35) , (cid:34) NKLR(cid:88) , q(cid:48) b b (i), and θ(i) q(cid:48) where oe(i) are transition function training inputs, outputs, and kernel weights for the b-th observation bundle. The partial derivative with respect to the j-th component of each parameter is, b (cid:80)Nn b ) b=1 I(q(cid:48) l(θ(i)) = wboe(i) b,j q(cid:48) oe(i) k=1 exp(θ(i)T b = q(cid:48))− exp(θ(i)T ∂ qk oe(i) ∂θ(i) ) q(cid:48),j (14) where I(·) is the indicator function. The log-likelihood can be maximized using a quasi-Newton method (our implementation uses the Broyden-Fletcher-Goldfarb-Shanno algorithm). To improve the generalization of the learned model, L2 regular- ization is used during weight training. b EPSCKO is outlined in Alg. 1. It begins by specifying an empty SK-FSA policy and NKLR-length FIFO circular kernel basis queue for each robot (Alg. 1, Lines 2-3). The Algorithm 1: EPSCKO 1 Procedure: ¯φb = ¯Oe, I, Nn, Nk, Ns, Nb, NKLR, α, αEI ) EPSCKO(¯T, 2 For each robot, initialize SK-FSA policy with Nn nodes; 3 ¯QKLR ← initFIFOQueue(NKLR); 4 ¯Vb, ¯Vw,0 ← −∞; 5 for i = 1 to n do 6 ∀q and θ(i)(q(cid:48)q,o) Initialize θ(i)(πq) ∀q, o; 0 7 for k = 0 to Nk − 1 do 0 allowEntropyInject, entropyInjected ← False; ¯πlist, KLRlist, ¯Vlist ← ∅; for s = 1 to Ns do ¯V ¯φ,{¯π}s,{¯oe, ¯q(cid:48)}s ← Evaluate( ¯φ); if ¯V ¯φ ≥ ¯Vw,k then ¯πlist ← ¯πlist ∪ {¯π}s; KLRlist ← KLRlist ∪ {¯oe, ¯q(cid:48)}s; ¯Vlist ← ¯Vlist ∪ ¯V ¯φ; if ¯V ¯φ > ¯Vb then ¯Vb, ¯φb ← ¯V ¯φ, ¯φ; ¯πb,list, KLRb,list, ¯Vb,list← best Nb policies in ¯Vlist; ¯QKLR.push(KLRb,list); ¯Vw,k+1 ← min( ¯Vb,list); if ValueConverged() then allowEntropyInject ← True; for i = 1 to n do k+1 + (1 − α)θ(i)(πq) θ(i)(πq) k+1 ← MLE of θ(i)(πq) using ¯πb,list ∀q; θ(i)(πq) k+1 ← αθ(i)(πq) θ(i)(q(cid:48)q,o) k+1 if allowEntropyInject then ← trainWeightedKLR(Q(i) entropyInjected←tryInject(θ(i)(πq) k+1 , θ(i)(q(cid:48)q,o) KLR, α); k+1 ; k ); if entropyInjected then ¯Vw,k+1 ← −∞;; 31 return ¯φb; best-value-so-far, ¯Vb, and worst-joint-value, ¯Vw,0, are set to −∞ (Alg. 1, Line 4). To encourage policy space exploration, SK-FSA parameter vectors are initialized such that associated distributions are uniform (Alg. 1, Line 6). The main algorithm loop updates the SK-FSA policy over Nk iterations, using the maximal entropy injection scheme detailed in Section III-B to accelerate search. Entropy injection is initially disabled and a flag indicating successful entropy injection in the current iteration is set to False (Alg. 1, Line 8). The team's SK-FSA policies are evaluated Ns times, with perceived continuous observation and node transition trajectories saved for KLR training (Alg. 1, Line 11). MA selections and node transitions from policies exceeding the previous iteration's worst joint value are tracked in KLRlist (Alg. 1, Lines 13-15). The best-value-so-far, ¯Vb, is saved (Alg. 1, Line 17). Trajectory lists are pruned to retain only the best Nb trajectories (Alg. 1, Line 18). Continuous observations and node transitions from this list are pushed to the FIFO queue, causing old trajectories to be popped (Alg. 1, Line 19). The iteration's worst joint value, ¯Vw,k+1, is then updated. At this point, the algorithm checks if the Dec-POSMDP joint value has converged. If so, entropy injection is enabled to counter convergence to a local optima (Alg. 1, Line 22). This does not imply entropy injection will occur, only that it is allowed to occur. Each robot subsequently updates its MA distribution parameter vector, θ(i)(πq), using a smoothed MLE approach (Alg. 1, Lines 24-25). As discussed earlier, weighted log-likelihood maximization is used to train the KLR model for each node transition function (Alg. 1, Line 26). Next, if maximal entropy injection is allowed, entropies of sampling distributions are calculated and (if necessary) injection occurs (Alg. 1, Line 28). As transition function δ(i) is continuous and non-linear, an approximate measure of its entropy is calculated using transition distributions sampled at its underlying set of observation kernels. This approximation was found to work well in practice (Section V-B) and is computationally efficient as it avoids domain re-sampling. To increase entropy of the node transition function, a continuous uniform distribution injection is done using update rule (10). If entropy injection is conducted for any robot, the current iteration's worst joint value, ¯Vw,k+1, is set to −∞ (Alg. 1, Line. 30). This critical step ensures trajectories sampled in the next iteration can actually be used for policy exploration. EPSKCO is an anytime algorithm applicable to continuous- observation Dec-POMDPs and Dec-POSMDPs. This approach also offers memory advantages to discretization as SK- FSA memory usage is O(NKLRNbNn ¯O), in contrast to O(d ¯ON 2 n) for FSAs with discretization resolution d. V. EXPERIMENTS This section first validates maximal entropy search acceler- ation, which resolves a long-standing convergence issue for sampling-based Dec-POSMDP algorithms. Then, EPSCKO is evaluated against discrete approaches in the first ever continuous-observation Dec-POMDP/Dec-POSMDP domain. A. Accelerated Policy Search We evaluate policy search acceleration approaches dis- cussed in Section III on the benchmark Navigation Among Movable Obstacles (NAMO) domain [21] with horizon h = 25 and a 6 × 6 grid. Fig. 2 shows convergence trends for all approaches. A low learning rate of 0.15 is needed in G-DICE [9] to find the optimal policy (taking k = 200 iterations). 50 policies are sampled per iteration, with 1000 trajectories used to approximate policy value in each iteration, so 1e7 total policy evaluations are conducted. This computationally expensive evaluation becomes prohibitively large as domain complexity grows. Increasing learning rate to α = 0.5 causes fast convergence to a sub-optimal solution, after which exploration stops due to sampling distribution degeneration. Existing search acceleration approaches are also evaluated. Dynamic smoothing with a moderate baseline rate (α0 = 0.5, β = 15) slightly improves value. However, decay rate β is static with no closed-loop feedback from underlying sampling distributions. The result is a sub-optimal policy (found around iteration k = 35) which then quickly converges to the same value as the baseline approach with α = 0.5. Fig. 2: Comparison of search acceleration approaches for NAMO domain, using a Nn = 5 node policy. Linearly decreasing noise injection with ωmax = 0.02 and r = 2000−1 performs similarly, with fast initial increase in value and subsequent degeneration to a sub-optimal policy. The proposed entropy injection method significantly out- performs the above approaches. The same baseline learning rate as previous methods (α = 0.5) is used with a 3% entropy injection rate, resulting in much faster convergence (around k = 20). Sensitivity to α and injection rate is low as value convergence monitoring is conducted in all iterations. While some initial tuning of entropy injection rate is necessary, the key insight is that post-tuning results converge much faster and are more conducive to additional experimentation and analysis (e.g., with domain/policy structure). Oscillations in plots are due to post-convergence injections, which reset underlying sampling distributions and forces further policy space exploration. In practice, the best policy found in a fixed number of iterations would be returned by the algorithm. B. Continuous Observation Domain To evaluate EPSCKO, a multi-robot continuous-observation nuclear contamination domain is considered (Fig. 4a). This first-ever continuous-observation Dec-POMDP/POSMDP do- main involves 3 robots cleaning up nuclear waste. MAs are Navigate to base, Navigate to waste zone, Correct position, and Collect nuclear contaminant. Following MA execution, each robot receives a noisy high-level observation oe of its 2D (x, y) state. The above MAs have non-deterministic durations and a 30% failure probability (due to nuclear contaminant degrading the robots). This causes poor performance of observation-agnostic policies which memorize chains of MAs, rather than make informed decisions using the observations. Robots are initially at the base and must first navigate to the waste zone prior to collection attempt. Robots which execute the Navigate to base MA terminate with a random continuous state in a region centered on the base (brown region marked 'B' in Fig. 4c). The Navigate to waste zone MA results in a random terminal state within two large regions surrounding the nuclear zone (everything interior of gray regions marked 'L' in Fig. 4c, including the green regions marked 'S'). Collection attempts are only possible if the robot is within the waste zone (green regions marked 'S' in Fig. 4c). Collections attempted outside these small contamination regions result in wasted time, which further discounts the team's future joint rewards. Robot can attempt a Correct position MA, which re-samples their state to be within these smaller regions. However, repeated attempts may be necessary due to the 30% MA failure probabilities. After successful collection, each robot must return to the base to deposit the waste before attempting another collection. Each collection results in +1 joint team reward (with discount factor γ = 0.9). This domain is particularly challenging due to the high failure rate of MAs, and the presence of a continuous, non-linear decision boundary in the nuclear zone center, where the trade-off between the correction and collection MAs must be considered by robots given their noisy observations. Fig. 3 compares best values obtained using continuous- observation and discrete-observation policy search (EPSCKO, G-DICE with maximal entropy injection, and MDHS). Time horizon h = 40 was used for evaluation, with each MA taking an average of 1-4 time units to complete. Nn = 6 nodes were used for both discrete and continuous policies. G-DICE and MDHS results are shown for observation discretization factors d ∈ {2, . . . , 10}, with uniform discretization in each observation dimension. EPSCKO significantly outperforms the discrete approaches, more than doubling the mean policy value of the best discrete-observation case (d = 4). MDHS faces the policy expansion issues discussed in Section II-B. G-DICE policy values initially increase with higher dis- cretization resolutions (d = 2 to d = 4), yet a drop-off occurs beyond d = 5. While initially counterintuitive, as higher discretization factors imply increased precision regarding important decision boundaries in the continuous domain, Figs. 4b and 4d reveal the underlying problem. These plots show the normalized count of observation samples used to compute Nn = 5 node discrete policies for the d = 10 and d = 5 cases, with discounting of old observation samples using (7). In other words, they provide a measure of discrete observation bins which have informed each G-DICE policy throughout its iterations. The core issue for discrete policies is that no correlation exists between decisions at nearby observation bins. Fine discretization meshes, as in Fig. 4b, result in cyclic processes where observation bins with no previous samples are encountered, therefore causing the robot to make a poor MA selection. Nearby observation bins do not inform the robot during this process, leading it to repeatedly make incorrect decisions. This issue is especially compounded in this domain due to delays caused by high MA failure probabilities, which reduce the overall number of observations received by robots. The result is a highly uninformative policy with no observations made in many bins, in contrast to policies with lower discretization factor (Fig. 4d). To build intuition on continuous-policy decision-making, Fig. 5 plots transition functions for a 6-node EPSCKO policy. For each node q, colored 3D manifolds represent probabilities of transitioning to next-nodes, q(cid:48), given a continuous observation. Circles plotted beneath transition functions indicate base and nuclear zone locations. Colorbars indicate the transition manifold color associated with each node and the highest-probability MA, πmax, executed in it. Consider a robot policy starting at node q = 1 (far left in Fig. 5) which has two major manifolds (beige and green). Observations under a prominent green manifold region indicate high probability of transitioning to node q(cid:48) = 3 (as Fig. 3: Comparison of discrete and continuous policy search approaches for nuclear contamination domain. its colorbar is green), which has πmax = Navigate to waste zone. For q = 1, this green manifold is centered on the base, which makes intuitive sense as the Navigate to waste zone MA should only be executed if the robot is confident it is at the base. Thus, the robot most likely transitions to node q = 3, and a complex transition function manifold is encountered. Two beige peaks are centered on the small inner regions of the nuclear zone, indicating transition to node q(cid:48) = 5, which has πmax = Collect nuclear contaminant. Thus, when the robot is in q = 3 and confident that it is in the center of the nuclear zone, it attempts a collection MA. Yet, for observations outside the inner nuclear zone, the red and blue manifolds are most prominent. These indicate high probabilities of transitioning to q(cid:48) = 1 and q(cid:48) = 2, which have πmax = Correct position. Thus, the robot most likely performs a heading correction before continuing policy execution and attempting waste collection. This process continues indefinitely or until the time horizon is reached. Recall that SK-FSA policies are stochastic, so these discussions provide an intuition of the 'most likely' continuous-policy behaviors. VI. CONCLUSION This paper presented an approach for solving continuous- observation multi-robot planning under uncertainty problems. Entropy injection for policy search acceleration was presented, targeting convergence issues of existing algorithms, which are exacerbated in the continuous case. Stochastic Kernel- based Finite State Automata (SK-FSAs) were introduced for policy representation in continuous domains, with the Entropy-based Policy Search using Continuous Kernel Ob- servations (EPSCKO) algorithm for continuous policy search. EPSCKO was shown to significantly outperform discrete search approaches for a complex multi-robot continuous- observation nuclear contamination mission-the first ever Dec-POMDP/Dec-POSMDP domain. Future work includes extending the framework to continuous-time planning. REFERENCES [1] D. S. Bernstein, R. Givan, N. Immerman, and S. Zilberstein, "The complexity of decentralized control of Markov decision processes," Math. of Oper. Research, vol. 27, no. 4, pp. 819–840, 2002. [2] C. Amato, G. Konidaris, A. Anders, G. Cruz, J. How, and L. Kaelbling, "Policy search for multi-robot coordination under uncertainty," in Robotics: Science and Systems XI (RSS), 2015. [3] S. Omidshafiei, A.-A. Agha-Mohammadi, C. Amato, and J. P. How, "Decentralized control of partially observable markov decision processes using belief space macro-actions," in Robotics and Automation (ICRA), 2015 IEEE International Conference on. IEEE, 2015, pp. 5962–5969. [4] F. A. Oliehoek and C. Amato, A Concise Introduction to Decentralized POMDPs. Springer, 2016. DiscretizationFactord246810Value024681012EPSCKOG-DICEMDHSDomain (a) (artist's conception). overview (b) Density of observations used to update discrete Dec-POSMDP policy. Nn = 5 node policy case with discretization factor d = 10. Numerous low density bins present due to fine discretization. (c) Domain overview (key continuous regions). (d) Density of observations used to update a discrete Dec-POSMDP policy. Nn = 5 node policy case with discretization factor d = 5. Observation density increases in the low discretization resolution case. Fig. 4: Continuous-observation nuclear contamination domain overview and corresponding discrete policy results. Fig. 5: Visualization of an Nn = 6 node SK-FSA policy transition functions for nuclear contamination domain. For each node q and observation oe, colored 3D manifolds represent probabilities of transitioning to next-nodes q(cid:48). Colorbars indicate the color associated with each node, as well highest-probability MA, πmax, executed in it. [13] C. Thiery and B. Scherrer, "Improvements on learning tetris with cross [5] J. Hoey and P. Poupart, "Solving POMDPs with continuous or large discrete observation spaces," in IJCAI, 2005, pp. 1332–1338. [6] J. M. Porta, N. Vlassis, M. T. Spaan, and P. Poupart, "Point-based value iteration for continuous POMDPs," Journal of Machine Learning Research, vol. 7, no. Nov, pp. 2329–2367, 2006. [7] H. Bai, D. Hsu, and W. S. Lee, "Integrated perception and planning in the continuous space: A POMDP approach," The International Journal of Robotics Research, vol. 33, no. 9, pp. 1288–1302, 2014. [8] S. Brechtel, T. Gindele, and R. Dillmann, "Solving continuous POMDPs: Value iteration with incremental learning of an efficient space representation." in ICML (3), 2013, pp. 370–378. [9] S. Omidshafiei, A.-A. Agha-Mohammadi, C. Amato, S.-Y. Liu, J. P. How, and J. Vian, "Graph-based cross entropy method for solving multi- robot decentralized pomdps," in Robotics and Automation (ICRA), 2016 IEEE International Conference on. IEEE, 2016, pp. 5395–5402. [10] A. Costa, O. D. Jones, and D. P. Kroese, "Convergence properties of the cross-entropy method for discrete optimization." Oper. Res. Lett., vol. 35, no. 5, pp. 573–580, 2007. [11] Z. I. Botev and D. P. Kroese, "Global likelihood optimization via the cross-entropy method, with an application to mixture models." in Winter Simulation Conference. WSC, 2004, pp. 529–535. [12] D. P. Kroese, S. Porotsky, and R. Y. Rubinstein, "The cross-entropy method for continuous multi-extremal optimization," Methodology and Computing in Applied Probability, vol. 8, no. 3, pp. 383–407, 2006. entropy." ICGA Journal, vol. 32, no. 1, pp. 23–33, 2009. [14] L. Devroye, L. Gyorfi, and G. Lugosi, A probabilistic theory of pattern recognition. Springer Science & Business Media, 2013, vol. 31. [15] C. E. Shannon, "A mathematical theory of communication," ACM SIGMOBILE Mob. Comp. and Comm. Rev., vol. 5, no. 1, 2001. [16] C. Amato, D. S. Bernstein, and S. Zilberstein, "Optimizing fixed-size stochastic controllers for POMDPs and decentralized POMDPs," Auton. Agents and Multi-Agent Sys., vol. 21, no. 3, pp. 293–320, 2010. [17] F. Oliehoek, Value-based planning for teams of agents in stochastic partially observable environments. Amsterdam University Press, 2010. [18] D. S. Bernstein, C. Amato, E. A. Hansen, and S. Zilberstein, "Policy iteration for decentralized control of markov decision processes," J. of Artif. Intell. Res., vol. 34, no. 1, p. 89, 2009. [19] S. W. Carden, "Convergence of a Q-learning variant for continuous states and actions," J. of Artif. Intell. Res., vol. 49, pp. 705–731, 2014. [20] J. Zhu and T. Hastie, "Kernel logistic regression and the import vector machine," Journal of Computational and Graphical Statistics, 2012. [21] M. Stilman and J. J. Kuffner, "Navigation among movable obstacles: Real-time reasoning in complex environments," International Journal of Humanoid Robotics, vol. 2, no. 04, pp. 479–503, 2005. oe1012345oe2012345q=1oe1012345oe2012345q=2oe1012345oe2012345q=3oe1012345oe2012345q=4oe1012345oe2012345q=500.20.40.60.81Density of Observations used for Policy Updateoe1012345oe2012345q=1oe1012345oe2012345q=2oe1012345oe2012345q=3oe1012345oe2012345q=4oe1012345oe2012345q=5
1406.7770	3	1406	2017-02-20T19:59:31	A Psychologically-Motivated Model of Opinion Change with Applications to American Politics	[ "cs.MA", "cs.SI", "physics.soc-ph" ]	Agent-based models are versatile tools for studying how societal opinion change, including political polarization and cultural diffusion, emerges from individual behavior. This study expands agents' psychological realism using empirically-motivated rules governing interpersonal influence, commitment to previous beliefs, and conformity in social contexts. Computational experiments establish that these extensions produce three novel results: (a) sustained strong diversity of opinions within the population, (b) opinion subcultures, and (c) pluralistic ignorance. These phenomena arise from a combination of agents' intolerance, susceptibility and conformity, with extremist agents and social networks playing important roles. The distribution and dynamics of simulated opinions reproduce two empirical datasets on Americans' political opinions.	cs.MA	cs	ArXiv A Psychologically-Motivated Model of Opinion Change with Applications to American Politics Peter Duggins Abstract Agent-based models are versatile tools for studying how societal opinion change, including political polarization and cultural diffusion, emerges from individual behavior. This study expands agents' psychological realism using empirically-motivated rules governing interpersonal influence, commitment to previous beliefs, and conformity in social contexts. Computational experiments establish that these extensions produce three novel results: (a) sustained "strong" diversity of opinions across society, (b) opinion subcultures, and (c) pluralistic ignorance. These phenomena arise from a combination of agents' intolerance, susceptibility and conformity, with extremist agents and social networks playing important roles. The distribution and dynamics of simulated opinions reproduce two empirical datasets on Americans' political opinions. Keywords Agent-Based Model, Opinion Dynamics, Social Networks, Conformity, Polarization, Extremism Computational Neuroscience Research Group, University of Waterloo [email protected] Contents Introduction Social Psychology of Opinion Change Model Description Results Validation: American Political Opinions Discussion Conclusion References 1 2 3 4 10 14 16 16 Introduction Opinions are mutable: individuals revise their beliefs through social interaction, personal experience, and reflection, while societal norms shift in response to global events and public opinion. Opinion change at the individual and societal scales interact to produce political polarization, cultural globaliza- tion, and other important social trends. To understand these phenomenon and design appropriate interventions, we need quantitative tools that simulate the psychological and social as- pects of opinion change. For example, models of interpersonal communication will help activists organize grassroots support, help leaders design effective campaigns, and help peacekeep- ers prevent the spread of extremism. Computational models of opinion change have studied the relationship between polariza- tion, social influence, and political intolerance [1, 2, 3, 4, 5], while models of cultural diffusion have improved our under- standing of cultural convergence [6, 7], subculture formation [8, 9], and cultural stability within organizations [10, 11]. Building multi-level, quantitative, predictive models of opinion change is challenging because opinions arise from a multitude of neurological, psychological, and social pro- cesses. Empirically, the extent to which people are persuaded by each others' subjective evaluations depends on numerous factors, including previous beliefs and a desire to minimize cognitive dissonance [12]; motivations to be accurate, self- consistent, and socially accepted [13, 14]; issue framing, emo- tional arousal, and cognitive elaboration [15]; self-esteem [16]; social norms [17]; and more. Mathematical and compu- tational models help formally investigate both the interplay of internal psychological forces and the feedback between opinion change and social influence among many individuals. Unfortunately, models have historically neglected important elements of social psychology, assuming that individuals be- have identically, rationally, or with perfect information. This raises questions about whether their results properly inform our understanding of human societies. Agent-based models (ABMs) seek to explain macroscopic outcomes by showing that artificial societies populated by psychologically-plausible software individuals can, when ini- tialized in a virtual environment and evolved through time, en- dogenously "grow" complex social phenomenon [18]. Three features of ABMs make them ideal for modeling opinion change. First, agents are autonomous and heterogeneous: each individual has distinct internal attributes, such as an in- tolerance of opposing views, a propensity to socially conform, or a tendency towards stubbornness. Second, agents can be psychologically and cognitively authentic, endowed with ra- tional, emotional, and social thinking of arbitrary complexity [19]. Third, agents interact locally in an explicitly defined space: individuals have incomplete information about the A Psychologically-Motivated Model of Opinion Change with Applications to American Politics - 2/18 world, and interact in social networks of plausible size and composition, causing influence to spread through society in a manner constrained by personal connections. Although a rich literature of opinion dynamics using ABMs already exists [1, 20, 2, 3, 21, 22, 4, 23, 6, 7], several impor- tant questions remain unanswered: 1. How do social groups maintain a diversity of opin- ions? Previous models have shown that when agents exchange interpersonal influence, their opinions either converge to a single value (consensus) or diverge to homogeneous opinion groups (polarization). Although consensus and polarization are important political and cultural trends, real societies never converge or diverge absolutely: diversity is always preserved. Surprisingly, models have not yet shown that such a distribution can persist. 2. Will subcultures of opinions survive in a well-connected society? Pockets of extreme opinions exist within mod- erate real-world societies. Although such subcultures have emerged in previous models, they survive only be- cause of psychologically-implausible rules that curtail any interpersonal influence. 3. Does pluralistic ignorance affect societal opinion change? The views we express in public often differ from those we hold privately. This situation undoubtedly affects in- dividual and societal opinion dynamics, yet falsification remains unstudied in computational models. 4. Can an opinion change model reproduce empirical data? Opinion dynamics models capture qualitative phenomenon like polarization and clustering, but are rarely validated with quantitative empirical data. Clos- ing the loop will increase the scientific credibility and predictive power of these models. In this study, I aim to answer these questions by studying the relationship between the social psychology of personal opinion change and the distributions, dynamics, and geog- raphy of opinions across society. In Section 2, I review the literature on the social and psychological forces that drive opinion change. In Section 3, I describe the model, explaining how it extends previous models by expanding the psychologi- cal realism of agents. In Section 4, I pose hypotheses about the relationship between psychological forces and societal opinion change, run computational experiments to test them, and describe the emergence of (a) strong societal diversity, (b) persistent subcultures of opinions, and (c) pluralistic ig- norance. In Section 5, I compare these results with empirical data on Americans' political opinions. I conclude by summa- rizing the major findings, suggesting extensions to the model, and proposing a research agenda for agent models in the social sciences. Social Psychology of Opinion Change Social influence is a process in which the social exchange of information causes individuals to reevaluate their own opin- ions on a subjective issue. Arguably the most important feature of social influence is homophily, the principle that contact between similar people occurs more frequently and has greater impact than contact between dissimilar people. Empirical evidence for homophily and its effects on social influence abounds: for an overview, see [24]. Interpersonal in- fluence among friends is known to engender common attitudes [25, 26, 27], while the strength of dyadic connections concur- rently increases with similarity [10, 28]. On the other hand, interactions can impart negative social influence if opinions differ greatly [29, 30], causing individuals to adopt more ex- treme attitudes when exposed to counterattitudinal arguments [31, 32, 33]. Homophily is a cornerstone of opinion dynamics mod- els: individuals exert social influence on each other propor- tional to their ideological similarity. In dyadic conversations, similarity encourages consensus, while dissimilarity fosters polarization. A lineage of models have shown that a society with high tolerance (a parameter governing the relationship between opinion similarity and the magnitude of influence) leads to consensus, while low tolerance leads to polarization [2, 3, 1, 20, 21, 22, 4, 23, 6, 7]. Weak diversity, defined as the convergence of opinions to n > 1 attractor states, can be maintained when opinion subcultures form and become isolated. This outcome is common in bounded confidence models when influence between dissimilar agents goes to zero. Generally, strong diversity, defined as a smooth distri- bution of opinions along a continuous ideological spectrum, disappears in these models whenever social networks are fully connected [11, 34, 35], even accounting for noise and other minor deviations [36, 37]. Social influence does not take place in a vacuum, but in an environment filled by people who seek social acceptance and who judge each other upon personality and beliefs. Confor- mity describes an individual's desire to gain social approval and avoid rejection by expressing normative beliefs. There is substantial empirical evidence of people misrepresenting their true beliefs [13, 14, 38], though some "anticonformists" will express non-normative beliefs so as to appear more dis- tinct [35, 39]. Together, conformity and distinctiveness lead to pluralistic ignorance [40], a condition in which the true distribution of opinions in society differs from what is spoken and heard in public. Pluralistic ignorance makes people un- aware of others' true beliefs; a lack of accurate information can, though the mechanisms of social influence, feedback to change people's true opinions. For example, after years government oppression, levels of popular dissent in author- itarian societies may become suddenly obvious, leading to political turmoil and violent tipping points [41, 42]. Despite current enthusiasm for studying the effects of conformity and distinctiveness on opinion change [43, 35, 44], ABMs have yet to investigate the repercussions of agents' explicit belief A Psychologically-Motivated Model of Opinion Change with Applications to American Politics - 3/18 falsification on public opinion. The way an individual receives and internalizes others' beliefs can be as important as the content and context of the influence. People who hold strong opinions are committed to their beliefs: they resist opinion change, because it would challenge their political worldview and induce cognitive disso- nance, and because they judge contrary information as invalid due to confirmation bias [45]. Strongly opinionated individu- als have been shown to reject opinions contrary to their own belief and even become more extreme. On the other hand, moderately opinionated individuals are susceptible to opin- ion change and will more readily internalize beliefs presented by others [31, 32, 33]. Surprisingly, few models of opinion change have looked into how susceptibility and commitment help sustain diversity and prevent homogenization of small cultural groups [5]. Finally, the social networks through which individuals in- teract determine how opinion change spreads through society. These networks can be characterized by statistical descriptions such as the degree of connectivity (average size of a social net- work); real-world networks have positive assortativity (people with large networks tend to know others with large networks), low whole-network density (most people don't know each other), and high but heterogeneous clustering. Though sim- ulations have confirmed that the size and composition of so- cial networks strongly affect opinion change, their outcomes vary widely with the models' assumptions about the network [46, 47], which rarely take these empirical regularities into account. One procedure which does effectively reproduce these statistics is the social circle model [48], which is easily incorporated into an ABM framework [49]. The Influence, Susceptibility, and Conformity Model (ISC) To summarize, agents are randomly placed within a two- dimensional space. Each agent has a unique initial opinion, three parameters for tolerance, conformity, and susceptibil- ity, and a social network. Each round, every agent initiates a dialogue with members of his social network. In the dia- logue, each agent expresses an opinion that reflects his true opinion, his conformity, and the opinions already expressed in the dialogue. Afterwards, the initiating agent updates his true opinion based on his tolerance, susceptibility, and the expressed opinions' weighted influence. The model records the true and expressed opinion of each agent after every round. The model, data, and figures are available on GitHub. Agents' opinions, interpreted as beliefs on a single subjec- tive issue, lie on a continuous 0− 100 scale. Initial opinion, tolerance, conformity, susceptibility, and social reach are all drawn from normal distributions whose means and variances are specified in each experiment. Agents are randomly as- signed a continuously-valued (x,y) location, then each agent creates a social network N with all agents within euclidean radius equal to his social reach r, as per the social circle model. Agents remain stationary. Agent i initiates a dialogue with all agents j in his social network. He is the first to express an opinion, and always voices his true opinion (Oi). Subsequently, each j distorts his opinion in order to conform or appear distinct. Specifically, j calculates the average of all opinions (Ek) expressed so far in the dialogue (D), then expresses an opinion (E j) that is between his true opinion (O j) and the dialogue's opinion norm (conformity), or that is distanced by some amount from the dialogue's norm (distinctiveness): E j = O j + c j k j ∗ 1 N k D ∑ (Ek − O j) (1) The agent parameter c j represents an agent's inherent willing- ness to misrepresent his beliefs in social contexts in order to appear either normal or distinct. The parameter captures both conformity (c j > 0) and distinctiveness (c j < 0). Greater magnitude c j produces greater belief falsification: c j = 0 causes the agent to speak truthfully, c j = 1 causes the agent to express the dialogue's "mean opinion", and c = −1 causes the agent to express an opinion that is more dislike the mean than his true opinion. In this model, conformity and distinctiveness are manifest in expression but not directly in opinion change: agents attempt to gain social favor by stating opinions that dif- fer from their true beliefs, but do not change their true beliefs to reflect this posturing. The extent of j's conformity is further mitigated by his cur- rent commitment k j, which is proportional to his susceptibility s j and the extremeness of his current opinion: k j = 1 + s j ∗ 50− O j 50 . (2) The susceptibility parameter s j represents an agent's inherent commitment to strong beliefs; it causes him to be less affected by social context and social influence. Its magnitude governs how a departure from a neutral opinion (Oi = 50) translates to a shrinking of influence: higher values result in less opinion change. After each j has expressed E j once in the dialogue, i updates his true opinion according to the dialogue's influence (Ii), which is proportional to each E j and the weight that i assigns to that expression (wi j): Ii = j wi j ∗ (E j − Oi) ∑N j wi j ∑N (3) Conceptually, the dialogue's influence Ii results from i being pulled towards (or pushed away from) each opinion expressed in the dialogue, E j, by an amount proportional to the intera- gent weight, wi j. The weight, in turn, is calculated according to homophily: the greater the absolute distance between i's opinion and j's expression, the more negative the weight, and the less influence j's expression will exert on i's opinion: wi j = 1−ti E j − Oi 50 (4) A Psychologically-Motivated Model of Opinion Change with Applications to American Politics - 4/18 where ti represents i's inherent intolerance of dissimilar opin- ions. Its magnitude dictates how strongly a given opinion difference translates to a loss of interagent weight. A high value implies that an agent will only assign positive weight to opinions that are similar to his own beliefs; a low value implies the agent will be positively influenced by a wider range of opinions. Mathematically, ti is the slope of i's weight vs. ∆ opinion curve, which is continuous and linear. This is a departure from the canonical bounded confidence approach, in which weight is a threshold function of an agent's intolerance εi. I believe continuous weighting better reflects the subtleties of opinion appraisal and social influence than a binary "full acceptance vs. complete disregard" judgment. This approach has also been adopted by [35]. Weights are bounded from −1 to +1. Finally, i updates his true opinion based on his previous opinion and the dialogue's influence, scaled by his commit- ment: Oi,t+1 = Oi,t + Ii ki (5) This process is repeated for each i in the population, conclud- ing one timestep. I use four metrics to investigate the diversity, dynamics, and geography of opinions within the population. Opinion histograms plot the frequency of opinions across the ideolog- ical spectrum at particular times, and are the most complete measure of strong vs. weak diversity. Opinion trajectories plot each agent's history as a line on a opinion vs. time graph, and are used to study dynamics towards or away from diversity. To distinguish different regions of opinion space, I use the terms centrist to describe agents who hold (33 < Oi < 66), moderate to describe agents with moderately-strong opin- ions (16 < Oi < 33 or 66 < Oi < 83), and extremist to de- scribe agents with the strongest opinions (0 < Oi < 16 or 83 < Oi < 100). Spatial maps plot each agent as a circle in (x,y) space with color representing the agent's opinion, and can help identify subgroup formation and regions of ideologi- cal mixing. Finally, the Jensen-Shannon Divergence (JSD) is a measure of the similarity of true opinions and expressed opinions across society. The JSD quantifies pluralistic igno- rance and is used to study how agents' falsifications affects the diversity of opinions within society. It is calculated from the Kullback–Leibler divergence D(PQ), a standard entropy metric for probability distributions: D(PM) + JSD(PQ) = (6) 1 2 D(QM) 1 2 where P and Q are the true and expressed opinion distributions, M = 1 2 (P + Q), and D(PQ) = ∑ i P(i)log P(i) Q(i) . (7) The JSD ranges from 0 (identity) to 1 (minimum mutual in- formation) Results Experiment 1: Social Influence and Intolerance To begin, I reproduce a classical experiment in opinion dy- namics, in which the final distribution of opinions is examined as a function of intolerance. In this model, intolerance is an agent-level parameter ti which is initially drawn from a normal distribution of mean µt and variance σt. For these prelimi- nary experiments I assume no heterogeneity of intolerance, susceptibility, or conformity: σt ,σs,σc = 0. Hypothesis 1: low intolerance promotes societal opinion convergence, while high intolerance produces opinion polarization and weak diversity. In a society with low intolerance, µt = 0.7, most agents assign positive weight to each others' opinions during dialogues, and are consequently pulled towards the mean opinion in that dia- logue. Figure 1 (left) shows that an initial normal distribution of opinions, µO = 50,σO = 20, rapidly converges to a single, centrist opinion: given enough time, diversity will completely disappear, and all agents will believe Oi = 50. Conversely, in a society with high intolerance, µt = 1.0, many agents assign negative weight to each others' opinions and are pushed away from the dialogue mean. As agents adopt stronger opinions, they assign stronger negative weights, resulting in polarizing feedback. Figure 1 (right) shows this society rapidly diverges to two extremists opinions at either end of the opinion spec- trum. As t → ∞ only weak diversity remains: all agents either hold Oi = 0 or Oi = 100. These base-case results confirm the classical finding that, in the absence of other psychologi- cal forces, the degree of individuals' intolerance determines whether society homogenizes or polarizes. Experiment 2: Conformity and Distinctiveness Next I introduce social context into the simulation by allow- ing agents to misrepresent their true opinions in dialogues. Though opinion falsification does not directly affect agents' true opinion update, it does affect the information available to those agents. If falsification is significant, agents will perceive an unrepresentative distribution of opinions (compared to each others' true beliefs) and change their beliefs accordingly. Hypothesis 2: a conformist society will homogenize under conditions that otherwise cause polarization, while a society driven by distinctiveness will polarize under conditions that otherwise favor consensus. First, I simulate a society whose high intolerance would normally cause polarization, µt = 1.0, but introduce a moder- ate tendency towards social conformity, µc = 0.5. Conform- ing agents now express opinions that are close enough to the dialogue mean that almost nobody assigns these (falsified) opinions a negative weight. Agents adopt opinions closer to the norms expressed in the dialogue, and opinions converge, eventually resulting in societal consensus as shown in Figure 2 (right). Second, I simulate the opposite conditions: a society A Psychologically-Motivated Model of Opinion Change with Applications to American Politics - 5/18 Figure 1. A society populated by agents who initially hold normally distributed opinions converges to a single centrist opinion if agents' intolerance of dissimilar opinions is low, and diverges to two extremists opinion if agents' intolerance is high. This result reproduces findings in classical opinion dynamics and represents the base case of the simulation, in which social influence is the only active psychological force and all agents are identically intolerant. Additional runs show that societies with intolerance below µt = 0.7 always converge, societies with intolerance above µt = 1.0 always diverge, and societies in between can either converge or diverge, depending on initial conditions. Strong diversity doesn't emerge. whose low intolerance would normally cause convergence, µt = 0.7, but filled with agents who possess a strong desire to be distinct, µc = −1.5. When agents converse, they ex- press opinions that differ radically from centrist norms. As expressions become more extreme, social influence causes agents to adopt and retain extreme beliefs. Figure 2 (left) shows that opinions initially converge, then diverge towards two homogeneous extremist parties. These results indicate that contextual opinion falsification can reverse the effects of intolerance, but cannot sustain strong diversity or pluralistic ignorance. Experiment 3: Commitment to Strong Beliefs I conclude the preliminary experiments by investigating whether, when agents' susceptibility to influence decreases with their belief extremity, different patterns of societal opinion change emerge. Hypothesis 3: when extremist agents undergo less opinion change than moderate or centrist agents, their persistent influence will prevent centrist homogenization and produce weak diversity. Beginning with the simple case of a tolerant society with no opinion falsification, I test whether strong commitment, µs = 10.0, can reverse trends towards convergence. The mean opinion expressed in dialogues is still ¯O (cid:39) 50, but because agents are tolerant and truthful, they assign positive weights to all opinions they hear. Extreme agents undergo little opinion change due to their commitment, but without a repelling force to push them away from social norms (i.e. intolerance or dis- tinctiveness), they are still pulled slowly towards this centrist opinion. Although they remain steadfast in their views for longer periods of time than in Experiment 1, they eventually converge to a single centrist opinion like the rest of society (not shown). This result contradicts Hypothesis 3, showing that commitment by itself cannot reverse homophilous opin- ion convergence. However, personal susceptibility can affect a society that is intolerant and conformist, which normally homogenizes as in Experiment 2. Opinions initially converge due to the strong centrist norms perceived in conformist social dialogues, but extremists are slow to change. By t = 300, most agents have adopted moderate or centrist opinions and expressions, but about 2% of agents have, through intolerant repelling, adopted maximally extreme opinions, as shown in Figure 3. These extreme agents are now so committed to their beliefs that they barely soften their expressions to socially conform, Oi = 100 → Ei = 95, and their strongly opinionated vocaliza- tions polarize their social networks. Over time, this influence bifurcates society, as can be seen by the divergence of opin- ions past t = 500. This experiment indicates that personal commitment fosters pockets of extremism whose long-term influence significantly alters societal opinion dynamics. Experiment 4: Strong Diversity Equipped with a basic understanding of model behavior and the independent effects of intolerance, susceptibility, and com- mitment, I now simulate societies in which all three psychoso- cial forces interact. In this experiment, all agent parameters are drawn from normal distributions with nonzero means and variances, creating an artificial society with greater hetero- geneity and psychological realism than previous opinion dy- namics models. Hypothesis 4: when agents are simultaneously motivated by social influence, personal susceptibility, and social context of varying degrees, society will (a) maintain a strong diversity of opinions and (b) exhibit and pluralistic ignorance. A Psychologically-Motivated Model of Opinion Change with Applications to American Politics - 6/18 Figure 2. When agents with low intolerance wish to appear distinct in social contexts, they express extremists beliefs in dialogues, which eventually polarizes society and leaves weak, bimodal diversity. When agents with high intolerance are motivated to socially conform, they express normative centrist views, causing society to converge to centrism. First I examine a society which is on average tolerant and conformist (µt = 0.7, µc = 0.4), but with enough diver- sity to create some intolerant and distinction-seeking agents (σt = 0.3,σc = 0.1). These forces tend to pull society towards convergence. However, an opposing commitment to strong beliefs (µs = 5.0,σs = 0.4) helps initially-extreme agents re- sist normalization and locally exert polarizing influence. The opinion trajectory plot in Figure 4 shows that society initially converges, but scattered extremists retain their strong views. Unlike in Experiment 3, where strong commitment kept ex- tremists from softening their expressions, extremists in this society moderate their expressions and few extreme opinions are publicly voiced. This, combined with low extremist den- sity, prevents radicals from attracting many followers: by t = 1,000, less than 5% of the population holds or expresses extreme views. However, these extremists exert enough influ- ence that they keep centrist norms from completely homog- enizing society. By t = 10,000, the distribution of opinions and expressions have settled into a diverse centrist group, a moderate fringe, and scattered extremists. This strongly diverse distribution of opinions persists past t = 100,000 de- spite small opinion fluctuations. A spatial map of opinions shows that the diversity within the centrist party arises from the minor influence exerted by extremists, which keeps the surrounding neighborhoods to the ideological left and right of ¯O (cid:39) 50. This result is, to my knowledge, the first evidence of indefinitely-sustained strong diversity in a continuous-opinion model. Experiment 5: Opinion Subcultures Next, I increase agents' average intolerance and conformity (µt = 1.0, µc = 0.5, µs = 5.0), then tweak their psychological diversity (σt = 0.3, σs = 0.3, σc = 0.3). The results are shown in Figure 5. After an initial period of convergence, several ex- tremists neighborhoods develop, affecting partial polarization. Society quickly self-organizes into distinct, geographically- clustered opinion subcultures, as can be seen in the spatial map. These subcultures are stable and coherent, but continue to influence each other through persuadable agents on their mutual border. Eventually, society settles into two extremist groups and a centrist group. The spatial orientation of the extremist parties is such that the centrist party receives ap- proximately equal influence from both sides of the opinion spectrum, and acts as a relatively stable buffer between the two extremes. Experiment 6: Pluralistic Ignorance Pluralistic ignorance and unpredictable dynamics are also possible under various conditions, such as when agents have intermediate intolerance (µt = 0.8, σt = 0.3), low commit- ment (µs = 0.1, σs = 0.1), and highly variable conformity (µc = 0.3, σc = 0.5). Opinions converge early on, and soci- ety is sufficiently tolerant and uncommitted that only a few agents retain extreme opinions. Through some combination of the extremists' social influence, conformity of their neigh- bors, and distinction of agents from centrist norms, opinions throughout society begin drifting towards the extreme. How- ever, unlike in Experiment 4, the extremists abruptly convert to centrism, causing a dramatic turn towards convergence, Fig- ure 6. Before these conversions, centrist or extremist agents express moderate opinions in dialogues, and pluralistic igno- rance spikes. The perceived moderate norm pulls centrists towards extremism, causing the slow drift before t = 500 in the opinion trajectories, but also pulls extremists towards centrism, causing the occasional conversion of an extremists agent. If the former trend dominates, society bifurcates; if the latter dominates, society homogenizes. Although this ex- periment shows that strong diversity does not always persist in the model, it suggests that pluralistic ignorance precedes dramatic and sometimes nonlinear changes in societal opinion dynamics. A Psychologically-Motivated Model of Opinion Change with Applications to American Politics - 7/18 Figure 3. In a society filled with individuals who are intolerant, conformist, and committed to extreme beliefs, a minority of agents will distance themselves from centrist norms and stubbornly express extremist views. The influence of these agents polarizes their neighbors, spreading extremism spatially outward until society bifurcates into weak diversity. Spatial maps at t = 200 and t = 2000 show that polarization originates from neighborhoods that contain extremist agents. A Psychologically-Motivated Model of Opinion Change with Applications to American Politics - 8/18 Figure 4. When agents are motivated by social influence, personal susceptibility, and social context of heterogeneous strength, novel opinion distributions emerge at the societal scale. The opinion trajectory shows that society settles into a stable opinion configuration, while the opinion histogram confirms that the final distribution of opinions is strongly diverse. As confirmed by the spatial map, most agents have adopted a centrist opinion, but a small minority of extremists counterbalance homogenizing norms, preventing total convergence but exerting too little influence to bifurcate society . A Psychologically-Motivated Model of Opinion Change with Applications to American Politics - 9/18 Figure 5. Extremists counteract initial trends towards convergence and form neighborhoods of strongly (but uniquely) opinionated agents. These groups compete on the border: when one group exerts greater influence, they persuade moderate agents to become extremists; and when both groups exert equal influence, a buffer zone of centrist forms between them. These strongly diverse subcultures persist through time without artificial geographic or social barriers to prevent communication. A Psychologically-Motivated Model of Opinion Change with Applications to American Politics - 10/18 Figure 6. The conversion of influential extremists may shift the course of societal opinion change from polarization to convergence. These events occur rapidly and are frequently preceded by spikes in pluralistic ignorance, such as at t (cid:39) 500,1000,5800. This suggests that when sustained levels of opinion falsification are finally revealed, tipping-point phenomenon may occur, leading to nonlinear opinion dynamics. Experiment 7: Social Reach Finally, I investigate whether strong diversity, opinion subcul- tures, and pluralistic ignorance are robust to changes in the size of agents' social networks: Hypothesis 5: small social networks will promote geographically-distinct opinion subcultures while large social networks will dissolve subgroups; strong diversity and pluralistic ignorance will not be affected. I reproduce Experiment 5 with smaller social networks, ob- tained by reducing agents' social reach (from µr = 22, σr = 4 to µr = 11, σr = 2). Opinions rapidly become clustered in geographically-constrained networks, producing discrete sub- cultures, Figure 7. These subcultures continue to receive influence from surrounding networks, often through persuad- able agents on the border who continually oppose consensus and promote strong diversity within each subculture. The partial isolation of subcultures prevents both homogenization and polarization globally, which is reflected in a wide, mul- timodal opinion distribution. Both opinion subcultures and strong diversity remain stable past t = 5,000. Similar results were obtained by reducing social reach in Experiments 4 and 6. Conversely, increasing agents' social reach promotes ho- mogenization. Initially, most agent converge to centrism, while the large size and strong centrist norms in dialogues en- courage the remaining extremists to express moderate views. Two outcomes are possible: either all extremists convert and society converges to centrism; or, as shown in Figure 8, an imbalance of extremists remains, and society drifts towards the most vocal group. Unlike in Experiment 5, where a buffer zone between extremist groups prevented takeover, large net- works decrease the likelihood that centrists participate in dia- logues that are well-balanced between the two extremes. This increases the probability that the dominant group will exert the strongest influence in all geographic regions, and that centrists and moderates will turn towards that extreme. In either case, strong diversity vanishes, in contradiction with Hypothesis 5. Validation: American Political Opinions Although the ISC model is grounded in social psychology and reproduces features of real-world opinion dynamics like strong diversity, opinion subcultures, and pluralistic ignorance, I have not shown that it quantitatively captures real-world data. In this section, I validate the model by reproducing empirical data on the distributions and dynamics of political opinions in American society. As a proof-of-concept for strong diversity, I compare the expressed opinion distributions produced by the ISC model with a survey that assessed people's opinions on each of twelve issues in contemporary American politics [50]. Each respondent was asked which of seven idealized positions, ranging from extremely liberal to extremely conservative statements about that issue, best described his or her belief, creating a seven-point opinion scale. Using several parameter- space exploration strategies, including an evolutionary algo- rithm and hyperopt[51], I found values for mean intoler- ance, susceptibility, and conformity that produced the dis- tributions shown in Figure 9. These parameters, optimized to reduce the root-mean-square-error between the model dis- tribution and Broockman's data over n = 4 realizations, lie within the bounds of the values used in the above experiments (except for σO = 50). The model captures distributions with a variety of different shapes, including: normal distributions around a centrist opinion (gun control) and a moderate opin- ion (affirmative action); centrist dominance with an extreme group (healthcare and contraception); and other strange shapes (abortion and immigration, with less accuracy). This result quantitatively demonstrates that real-world opinion distribu- tions are within the output-space of the model. A Psychologically-Motivated Model of Opinion Change with Applications to American Politics - 11/18 Figure 7. Shrinking agents' social networks encourages the formation of geographically-organized opinion subcultures. Reduced interaction between these groups prevents both centrist and extremist takeover, but continuing dialogues with intermediary agents keeps these groups, and society as a whole, strongly diversity. Histograms and maps show opinions at t = 10,000. A Psychologically-Motivated Model of Opinion Change with Applications to American Politics - 12/18 Figure 8. Expanding agents' social networks dilutes the influence of extremist agents over larger network, while globalized influence prevents the formation of opinion subcultures and eventually destroys strong diversity. The spatial maps at t = 4,000 and t = 5,000 show that the buffer zone that previously preserved diversity no longer prevents the takeover of the dominant extremist group. A Psychologically-Motivated Model of Opinion Change with Applications to American Politics - 13/18 Figure 9. The ISC model produces expressed opinion distributions that align with American's opinions on contemporary political issues ranging from gun control to healthcare to immigration. Data reproduced with permission from [50], parameters for each realization available on GitHub A Psychologically-Motivated Model of Opinion Change with Applications to American Politics - 14/18 I also compare the model's opinion dynamics with a large- N, multi-year survey of American's ideological consistency conducted by the Pew Research Center [52]. The survey consisted of ten questions assessing individuals' attitudes about current political issues such as "[the] size and scope of government, the social safety net, immigration, homosex- uality, business, the environment, foreign policy and racial discrimination," with each response coded −1 (liberal), +1 (conservative) , or 0 (don't know/refused). These values were summed for each individual, creating an "ideological consis- tency" scale ranging from −10 (liberal responses to every question) to +10 (conservative responses to every question). The study found that Americans have become increasingly polarized from 1994 to 2014: individuals who previously held mixed liberal and conservative positions on different issues are increasing partisan and ideologically uniform. As shown in Figure 10, this trend manifests a spreading of the empirical distribution over time. Using the same parameter-space explo- ration tools as above, I found the ISC model produced similar patterns of polarization: a normal-like opinion distribution midway through the simulation gradually spread as extremists on both sides pulled centrists towards the periphery. It can also reproduce more subtle dynamics, such as the leftward shift of the kernel density estimate's central peak from 1994 to 1999, then back to the right as a sharper peak from 1999 to 2004. Discussion How do social groups maintain a strong diversity of opinions? Psychological forces such as a commitment to strong opinions or a drive to distinctiveness may oppose homogenizing social influence and preserve strong diversity. When society contains agents with heterogeneous intolerance, susceptibility, and conformity, each individual is simultaneously pulled towards centrism and extremism. If these forces are balanced, a strong diversity of opinions emerges and remains stable through time. Figures 4, 5, and 7 showed this diversity can take the form of (a) a centrist party diversified by influence from a few extremist, (b) two extremist parties with undecided agents on the borders, and (c) geographically-isolated opinion subcultures. The maintenance of strong diversity is a novel result in opinion change and cultural diffusion models based on bounded confidence, which assume that agents influence one another only if their opinion similarity is above an interaction threshold [53, 7, 2, 3, 21, 22, 4]. Though this approximation of intolerance has proved a useful first step in understanding convergence vs. polarization, I argue that it is overly rigid: people do not classify each others' trustworthiness according to a binary scheme. The ISC model assumes that social in- fluence changes continuously with intolerance, commitment, and context, and produces sustained, strong diversity under multiple psychological and network conditions. This result is intuitive, since societies do not converge to a single opinion or diverge to two polar opposites, and is also quantitatively plausible, as shown through empirical validation. Whenever possible, agent-based modelers should move away from psy- chologically and socially implausible assumptions and adopt empirically-motivated cognitive heuristics: doing so will so- lidify the model's foundations and, as exemplified by this study, produce more complex and realistic results. Maintaining a diversity of opinions is important outside the modeling community. Indigenous cultures dissolve in the face of globalization as people substitute traditional languages and practices for the norms of modern society. Corporations fall prey to groupthink when individuals with original ideas choose not to voice them. Political and religions groups be- come polarized due to intolerance of dissimilar beliefs. To promote cultural and ideological diversity, leaders must rec- ognize that social influence is not the only force that drives single-mindedness. They must recognize, not just concep- tually but with the quantitative precision afforded by com- putational models, the role of psychological heterogeneity, personal commitment, and social context in destroying the valuable resource of diversity. Will subcultures of opinions survive in a well con- nected society? In the ISC model, communities of dissenters can survive among globalized centrism or extremist competition in two circumstances. In a tolerant or conformist society, the push towards centrism rapidly homogenizes most agents, but leaves a few intolerant or committed agents on the ideological pe- riphery. When intolerant agents who hold opposing beliefs live together, they reject the opposite perspective so strongly that they become extreme despite centrist influence, as in Experiments 3 and 5. Polarization spreads outward, leav- ing cohesive extremist parties and undecided agents on the neighborhoods' borders. Densely clustered extremists resist moderate influence; it seems that neither a centrist majority nor an opposing extreme minority effectively moderates their speech or prevents their polarizing influence. If left undis- turbed, these individuals will either settle into small conflicted communities or radicalize society. In the second scenario, agents' small social networks limit communication, producing a larger number of semi-isolated, cohesive, persistent communities. Communication is still possible between such communities, but must travel through bridging individuals whose influence is often overcome by the group consensus. These communities cannot coalesce when agents' social networks are large; social influence, when distributed over a large network, may cause either centrist convergence or extremist takeover. These results imply that (a) a lack of communication within society can encourage ide- ological splintering in the same way that geographic barriers facilitate speciation and genetic diversification, and (b) when advances in communication technology put isolated cultures into contact with the outside world, the inflow of globalized ideas can overwhelm the distinct features of their culture. In A Psychologically-Motivated Model of Opinion Change with Applications to American Politics - 15/18 Figure 10. The ISC model produces opinion dynamics that are consistent with the polarization of Americans' political opinions from 1994 to 2014. Blue and green lines are Gaussian kernel density estimates for the respective distributions. Data reproduced from the Pew Research Center [52] A Psychologically-Motivated Model of Opinion Change with Applications to American Politics - 16/18 reality, extremism often emerges in locations with limited communication and access to external information. Although networking extremists with new individuals has the potential to spread radicalization, it also increases the probability that extremists will find a bridge to more moderate attitudes that, over time, persuades them to soften their beliefs, as occurred in Experiment 6. Does pluralistic ignorance affect societal opinion change? We assess others' opinions through their expressions and use that assessment to reevaluate our own beliefs. Experiment 2 showed that agents' desire to conform can lead others to mistakenly think that agreement exists in society, reverse the process of opinion polarization, and bring an intolerant society back to consensus. It also showed that when agents express opinions with the goal of appearing distinct, no observable consensus exists on any issue, and the false atmosphere of extremism causes societal polarization. Social norms like the desire to be distinct from the previous generation or to conform to the community's religious beliefs do have far reaching effects on opinion change at the societal level; any simulation which assumes agents have perfect information is missing an important aspect of social communication. Pluralistic ignorance appeared in simulations with nonlin- ear opinion dynamics, spiking during the critical periods of change in that society's history, such as before stubborn ex- tremist converted to centrism and when centrist experimented with moderate expressions. One interpretation of these re- sults is that long-term history is relatively predictable when everyone communicates perfectly, but when social context en- courages belief falsification, tensions between what is heard and what is felt build until they are suddenly released. This interpretation agrees with Kuran's work on the role of pref- erence falsification in authoritarian revolutions [41], and was likely a contributing factor in the unexpected and rapid nature of the arab spring [42]. Can the ISC model reproduce empirical opinion dis- tributions and dynamics? All models should be treated with skepticism until they have been credibly validated with empirical data. The agreement between Broockman's data and the simulated opinion distribu- tions shows that the model reproduces strong diversity. These political opinion distributions are sometimes far from normal, and may be non-symmetric or have few agents at the ideo- logical center. Furthermore, the similarities between political polarization in the Pew dataset and in the simulation shows the model also captures certain features of opinion change, in- cluding short-term centrist fluctuations and long-term societal polarization. Overall, the validation experiments should be seen as an existence proof of plausible diversity and dynam- ics in the model, not as evidence of a calibrated simulation capable of precisely predicting opinion change. Conclusion In this study, I examined the relationship between the psy- chosocial forces driving opinion change and the resulting distributions, dynamics, and topologies of opinions across so- ciety. This research extends previous studies in computational opinion dynamics by expanding the psychological depth of agents to include previously unstudied forces. Through a se- ries of computational experiments, I showed that networks of heterogeneous agents will interact to produce (a) distribu- tions of opinions that match political opinion data (b) opinion subcultures, and (c) trend-setting pluralistic ignorance. These results are significant advances in the study of macroscopic opinion change and suggest that modest increases in the com- plexity of agent models can produce opinion dynamics that align better with reality. Many extensions of the ISC model are possible. People ac- tively promote their opinions at rallies or online, while others join organizations that enforce their beliefs through coercion and punishment. Introducing social mechanisms for these behaviors would permit the study of collective action prob- lems and suggest more specific strategies that leaders could take to achieve desired patterns of opinion change. Another extension would allow for dynamic social networks. Though the social reach procedure captures important statistics of so- cial networks, the people with whom we converse change constantly. Introducing dynamic networking, possibly in an expanded virtual environment, would permit a more complete study of how opinions change in a society dominated by social media. I would also like to compare opinion geography and pluralistic ignorance to empirical data. I contend that empirically-accurate patterns of opinion change only emerge when agents act according to plausible rules, and that modelers must expand the depth of agents' so- cial cognition to explain complex social phenomenon. This is best achieved by endowing agents with human-like cognitive architectures capable of affecting perception, memory, emo- tion, attention, and communication. Several opinion change models have already incorporated neurally-inspired mecha- nisms to great effect [54, 55]. Recent advances in neural engineering suggest that building agents with artificial brains may soon be possible [56]. In future work, I plan to incorpo- rate such artificial intelligences into social simulations. References [1] Bibb Latan´e. The psychology of social impact. American psychologist, 36(4):343, 1981. [2] Rainer Hegselmann, Ulrich Krause, et al. Opinion dy- namics and bounded confidence models, analysis, and simulation. Journal of Artificial Societies and Social Simulation, 5(3), 2002. [3] Guillaume Deffuant, Fr´ed´eric Amblard, G´erard Weis- buch, and Thierry Faure. How can extremism prevail? a study based on the relative agreement interaction model. A Psychologically-Motivated Model of Opinion Change with Applications to American Politics - 17/18 Journal of artificial societies and social simulation, 5(4), 2002. [4] Wander Jager and Fr´ed´eric Amblard. Uniformity, bipo- larization and pluriformity captured as generic stylized behavior with an agent-based simulation model of attitude change. Computational & Mathematical Organization Theory, 10(4):295–303, 2005. [5] Delia Baldassarri and Peter Bearman. Dynamics of political polarization. American sociological review, 72(5):784–811, 2007. [6] Kathleen Carley. A theory of group stability. American Sociological Review, pages 331–354, 1991. [7] Noah Mark. Beyond individual differences: Social dif- ferentiation from first principles. American Sociological Review, pages 309–330, 1998. [8] Jenna Bednar, Aaron Bramson, Andrea Jones-Rooy, and Scott Page. Emergent cultural signatures and persistent diversity: A model of conformity and consistency. Ratio- nality and Society, 22(4):407–444, 2010. [9] James A Kitts. Social influence and the emergence of norms amid ties of amity and enmity. Simulation Mod- elling Practice and Theory, 14(4):407–422, 2006. [10] Glenn R Carroll and J Richard Harrison. Come together? the organizational dynamics of post-merger cultural in- tegration. Simulation Modelling Practice and Theory, 10(5):349–368, 2002. [11] James A Kitts and Paul T Trowbridge. Shape up or ship out: social networks, turnover, and organizational culture. Computational and Mathematical Organization Theory, 13(4):333–353, 2007. [12] Richard E Petty, Duane T Wegener, and Leandre R Fab- rigar. Attitudes and attitude change. Annual review of psychology, 48(1):609–647, 1997. [13] Wendy Wood. Attitude change: Persuasion and social influence. Annual review of psychology, 51(1):539–570, 2000. [14] Robert B Cialdini and Noah J Goldstein. Social influence: Compliance and conformity. Annu. Rev. Psychol., 55:591– 621, 2004. [15] Bertram Gawronski and Galen V Bodenhausen. Asso- ciative and propositional processes in evaluation: an in- tegrative review of implicit and explicit attitude change. Psychological bulletin, 132(5):692, 2006. [16] Gregory J Pool, Wendy Wood, and Kira Leck. The self- esteem motive in social influence: agreement with valued majorities and disagreement with derogated minorities. Journal of personality and social psychology, 75(4):967, 1998. [17] Maria Knight Lapinski and Rajiv N Rimal. An explication of social norms. Communication Theory, 15(2):127–147, 2005. [18] Joshua M Epstein. Generative social science: Studies in agent-based computational modeling. Princeton Univer- sity Press, 2006. [19] Joshua M Epstein. Agent Zero: Toward Neurocognitive Foundations for Generative Social Science. Princeton University Press, 2014. [20] Morris H DeGroot. Reaching a consensus. Journal of the American Statistical Association, 69(345):118–121, 1974. [21] Pranav Dandekar, Ashish Goel, and David T Lee. Biased assimilation, homophily, and the dynamics of polariza- tion. Proceedings of the National Academy of Sciences, 110(15):5791–5796, 2013. [22] Laurent Salzarulo. A continuous opinion dynamics model based on the principle of meta-contrast. Journal of Artifi- cial Societies and Social Simulation, 9(1), 2006. [23] Robert Axelrod. The dissemination of culture a model with local convergence and global polarization. Journal of conflict resolution, 41(2):203–226, 1997. [24] Miller McPherson, Lynn Smith-Lovin, and James M Cook. Birds of a feather: Homophily in social networks. Annual review of sociology, pages 415–444, 2001. [25] Noah E Friedkin. Structural cohesion and equivalence ex- planations of social homogeneity. Sociological Methods & Research, 12(3):235–261, 1984. [26] Peter V Marsden. Homogeneity in confiding relations. Social networks, 10(1):57–76, 1988. [27] Noah E Friedkin. Structural bases of interpersonal in- fluence in groups: A longitudinal case study. American Sociological Review, pages 861–872, 1993. [28] James A Kitts, Michael W Macy, and Andreas Flache. Structural learning: Attraction and conformity in task- oriented groups. Computational & Mathematical Orga- nization Theory, 5(2):129–145, 1999. [29] Milton E Rosenbaum. The repulsion hypothesis: On the nondevelopment of relationships. Journal of Personality and Social Psychology, 51(6):1156, 1986. [30] George Smeaton, Donn Byrne, and Sarah K Murnen. The repulsion hypothesis revisited: Similarity irrelevance or Journal of Personality and Social dissimilarity bias? Psychology, 56(1):54, 1989. [31] Charles G Lord, Lee Ross, and Mark R Lepper. Biased assimilation and attitude polarization: The effects of prior theories on subsequently considered evidence. Journal of personality and social psychology, 37(11):2098, 1979. [32] Arthur G Miller, John W McHoskey, Cynthia M Bane, and Timothy G Dowd. The attitude polarization phe- nomenon: Role of response measure, attitude extremity, and behavioral consequences of reported attitude change. Journal of Personality and Social Psychology, 64(4):561, 1993. A Psychologically-Motivated Model of Opinion Change with Applications to American Politics - 18/18 [47] Damon Centola, Robb Willer, and Michael Macy. The emperor's dilemma: A computational model of self- enforcing norms1. American Journal of Sociology, 110(4):1009–1040, 2005. [48] Lynne Hamill and Nigel Gilbert. Social circles: A simple structure for agent-based social network models. Jour- nal of Artificial Societies and Social Simulation, 12(2):3, 2009. [49] Elisabeth zu Erbach-Schoenberg, Seth Bullock, and Sally Brailsford. A model of spatially constrained social net- work dynamics. Social Science Computer Review, page 0894439313511934, 2013. [50] David E Broockman. Approaches to studying policy representation. Legislative Studies Quarterly, 41(1):181– 215, 2016. [51] James Bergstra, Brent Komer, Chris Eliasmith, Dan Yamins, and David D Cox. Hyperopt: a python library for model selection and hyperparameter optimization. Com- putational Science & Discovery, 8(1):014008, 2015. [52] Michael Dimock, Jocelyn Kiley, Scott Keeter, and Carroll Doherty. Political polarization in the american public. Technical report, Pew Research Center, 2014. [53] Michael W Macy, James A Kitts, Andreas Flache, and Steve Benard. Polarization in dynamic networks: A hop- field model of emergent structure. Dynamic social net- work modeling and analysis, pages 162–173, 2003. [54] Tobias Schroder and Paul Thagard. The affective mean- ings of automatic social behaviors: Three mechanisms that explain priming. Psychological Review, 120(1):255, 2013. [55] Ingo Wolf, Tobias Schroder, Jochen Neumann, and Ger- hard de Haan. Changing minds about electric cars: An empirically grounded agent-based modeling approach. Technological forecasting and social change, 94:269–285, 2015. [56] Chris Eliasmith, Terrence C Stewart, Xuan Choo, Trevor Bekolay, Travis DeWolf, Yichuan Tang, and Daniel Ras- mussen. A large-scale model of the functioning brain. science, 338(6111):1202–1205, 2012. [33] Charles S Taber and Milton Lodge. Motivated skepticism in the evaluation of political beliefs. American Journal of Political Science, 50(3):755–769, 2006. [34] Andreas Flache and Michael W Macy. Local convergence and global diversity from interpersonal to social influence. Journal of Conflict Resolution, 55(6):970–995, 2011. [35] Michael Mas, Andreas Flache, and James A Kitts. Cul- tural integration and differentiation in groups and orga- nizations. In Perspectives on Culture and Agent-based Simulations, pages 71–90. Springer, 2014. [36] Konstantin Klemm, V´ıctor M Egu´ıluz, Ra´ul Toral, and Maxi San Miguel. Global culture: A noise-induced tran- sition in finite systems. Physical Review E, 67(4):045101, 2003. [37] Luca De Sanctis and Tobias Galla. Effects of noise and confidence thresholds in nominal and metric axel- rod dynamics of social influence. Physical Review E, 79(4):046108, 2009. [38] Solomon E Asch. Effects of group pressure upon the mod- ification and distortion of judgments. Groups, leadership, and men, pages 222–236, 1951. [39] Roland Imhoff and Hans-Peter Erb. What motivates nonconformity? uniqueness seeking blocks majority in- fluence. Personality and Social Psychology Bulletin, 35(3):309–320, 2009. [40] Deborah A Prentice and Dale T Miller. Pluralistic igno- rance and alcohol use on campus: some consequences of misperceiving the social norm. Journal of personality and social psychology, 64(2):243, 1993. [41] Timur Kuran. Sparks and prairie fires: A theory of unan- ticipated political revolution. Public choice, 61(1):41–74, 1989. [42] Jeff Goodwin. Why we were surprised (again) by the arab spring. Swiss Political Science Review, 17(4):452–456, 2011. [43] Matthew Jarman, Andrzej Nowak, Wojciech Borkowski, David Serfass, Alexander Wong, and Robin Vallacher. The critical few: Anticonformists at the crossroads of mi- nority opinion survival and collapse. Journal of Artificial Societies and Social Simulation, 18(1):6, 2015. [44] Paul E Smaldino and Joshua M Epstein. Social confor- mity despite individual preferences for distinctiveness. Royal Society open science, 2(3):140437, 2015. [45] Icek Ajzen. Nature and operation of attitudes. Annual review of psychology, 52(1):27–58, 2001. [46] Fr´ed´eric Amblard and Guillaume Deffuant. The role of network topology on extremism propagation with the rel- ative agreement opinion dynamics. Physica A: Statistical Mechanics and its Applications, 343:725–738, 2004.
cs/0211030	1	0211	2002-11-22T12:18:11	Integration of Computational Techniques for the Modelling of Signal Transduction	[ "cs.MA", "q-bio.CB" ]	A cell can be seen as an adaptive autonomous agent or as a society of adaptive autonomous agents, where each can exhibit a particular behaviour depending on its cognitive capabilities. We present an intracellular signalling model obtained by integrating several computational techniques into an agent-based paradigm. Cellulat, the model, takes into account two essential aspects of the intracellular signalling networks: cognitive capacities and a spatial organization. Exemplifying the functionality of the system by modelling the EGFR signalling pathway, we discuss the methodology as well as the purposes of an intracellular signalling virtual laboratory, presently under development.	cs.MA	cs	Integration of Computational Techniques for the Modelling of Signal Transduction Pedro Pablo González Pérez Maura Cárdenas García Carlos Gershenson García Jaime Lagúnez-Otero Instituto de Química Universidad Nacional Autónoma de México Ciudad Universitari a, 04510, MÉXICO, D.F. [email protected] http://132.248.11.4 Abstract: A cell can be seen as an adaptive autonomous agent or as a society of adaptive autonomous agents, where each can exhibit a particular behaviour depending on its cognitive capabilities. We present an intracellular signalling model obta ined by integrat ing several computational techniques into an agent-based paradigm. Cellulat, the model, takes into account two essential aspects of the intracellular signalling networks: cognitive capacities and a spatial o rganizatio n. Exempl ifying the functionality of the system by modelling the EGFR signalling pat hwa y, we discuss the method olo gy as w ell as t he p urp ose s of an i ntr acel lul ar s igna lli ng vi rtu al l abo rat ory, presently under development. Key words: in tra cel lul ar s igna lli ng pa thw ays, aut onomou s age nts , bl ackb oard arch ite ctu re, v irt ual lab orat ory. 1. Introduction Each cell in a multicellular organism receives specific combinations of chemical signals generated by other cells or from their internal milieu. The final effect of the signals received by a cell can be translated in the regulation of the cell metabolism, in cellular division or in its death. Once the extracellular signals bind to the receptors, different signalling processes are activated, genera t ing complex information transm iss ion networks. The more experimental data about cellular function we obtain, the important the more computational models become. The models allow for the visualization of the network components and permit the prediction of the effects of perturbations on com pon ent s or sect ion s of the sign all ing p athway. Within the computer sciences, the artificial intelligence is one of the main areas to model biological systems. This is due to the great variety of models, techniques and methods that support this research area, many of which are inherited from d isc ip l ines such as cognitive sciences and neuroscience. Among the main techniques of artificial intelligence and computer sciences commonly used to model cellular signalling networks are artificial neural networks (Bray and Lay, 1994; Pritchard and Dufton, 2000), Boolean networks (Edwards, 1995), petri nets (Fuss, 1987), rule-based systems (Cárdenas-García, 2001), cellular automata (Edwards, 1995; Wurthner, Mukhopadhyay and Peimann, 2000), and multi-agent systems (Fisher, Paton and Matsuno, 1999; Paton, Staniford and Kendall, 1995; Schwab and Pienta, 1997). The high complexity level of intracellular communication networks makes them difficult to model with any isolated technique. However, when integrating the most relevant features of these techniques in a single computational system, it should be possible to obtain a more robust model of signal transduction. This would permit a better visualization, understanding of the processes and components that integrate the networks. The theory of behaviour-based systems constitutes an useful approach for the modelling of in t race l lu la r (Gonzá lez, s igna l l ing ne two rk s Gershenson, Cárdenas and Lagúnez-Otero, 2000). The model permits to take into account communication between agents via a shared data structure, in which other cellular compartments and elements of the signalling pathw ays can b e expl icitl y repres ented . In this sense, the blackboard architecture (Nii, 1989) becomes appropriate. In this paper, we demonstrate that an effective and robust model of intracellular signalling can be obtained when the main structural and functional characteristics of behaviour-based systems are joined with the blackboard architecture. That is, a cell can be seen as a society of autonomous agents, where each agent communicates with the others through the creation or modification of signals on a shared data structure, named “blackboard”. The autonomous agents model determinated functional components of intracellular signallin g pathways, such as signalling proteins and other mechanisms. The blackboard levels represent different cellular compartments related to the signalling pathways, whereas the different objects created on the blackboard represent signal molecules, activation or inhibition signals or others elements belonging to the intracellular medium. In this way, when the autonomous agents are used in an intracellular signalling model, the cognitive capabilities of the signalling pathway can be taken into account. On the other hand, the use of blackboard architecture permits to model the high level of spatial organization exhibited by the intracellular signalling networks and structural characteristics of the intracellular medium. The rest of the paper is structured as follows: the next section presents an overview of the information processes of signalling intracellular networks. In section 3, some computational models for intracellular signalling are discussed; section 4 presents the intracellular signalling model proposed; section 5 describes the methodology steps for the creation of an intracellular signalling model using Cellulat. In section 6 an instantiation of the paradigm is made with the EGF receptor signalling pathway. Section 7 is focussed on the final goal of this work: the creation of an int race llu lar sign all ing v irt ual lab orat ory. 2. An overview of signalling intracellular netw orks In multicellular organisms, cell decisions about survival, growth, gene expression, differentiation and senescence or death, are made on the basis of external signals. These stimuli include cell-cell adhesion, growth factors, hormones, cytokines, neuropeptides and ions. The skill to integrate information from multiple sources is essential for the ability of the cell to respond appropriately to a wide range of conditions, and therefore enhances the adaptability and survival of the organism. Signal transduction networks allow cells to perceive changes in the extracellular environment in order to produce an appropriate response. One of the most exciting recent development in molecular and cell biology has been the step by step construction of signaling cascades that trace the path of the effects of an extracellular stimulus from the external membrane to the cell nucleus. A cellular process network mediates the transmission of extracellular signals to their intracellular targets. In general, the external signals are transmitted to the interior of the cells through receptors activating diverse signaling pathways. They can follow a single way and generate an answer or a specify cellular final process, or branch out and give rise to others. These pathways considered as a whole form an interconnected network, because pathways corresponding to different stimuli cross and generate alternative trajectories. The intracellular signalling implies several molecular processes. The signals can be as simple as the direct introduction of the signal to the nucleus and the activation of the transcription of proteins involved in the specific cellular function, which is expected. On the other hand, they can be very complicated and include multiple stages. For example, the receptor activates effector proteins like second messengers, kinases or phosphatases. They, in turn, activate transcription factor proteins, which determine the transcription of genes codifying for proteins involved in the specified cellular function. Several mechanisms of signal transduction systems have been described for the extracellular signal membrane receptors; some of them include: • • • • • • Adenilate cyclase system Phosphoinositides-calcium system. Mitogen activated protein kinase system (MAPK). JAK/STAT activation system. The esfingomyeline-ceramide system. The ionic channel function receptor system. The MAPK cascade is a central signal transduction pathway that is activated by growth factors, and is known to be involved in diverse cellular functions. In particular, we modeled the pathway activated by EGF since this pathway has crosstalk with other mentioned systems. The systems related to aging and cellular proliferation constitute our main goal. We are not modelling a specific cellular type. Once we have the functional model with all the elements involved in one pluripotential cell, it is relatively easy to extend it to cellular types like epidermic cells or hepatocites, etc. 3. Computational models for intracellular signalling Computational models in signal transduction pat hwa ys have been made using different points of view. Each research group chose the approach which seemed best for them and applied the most adequate computational tool for their purpose. The different models have been proposed according the to perspective that each research group has of the pathway they want to model. Thi s perspecti ve involve s a range from the types of information processing present at cellular level, such as sequential, parallel, distributed, concurrent and emergent; to the cognitive capabilities exhibited by certain signal transduction pathway component, such as memory, learning, pattern recognition and handling fuzzy data. In this sense, several computational approaches have been proposed to model the cellular signalling pathways, such as artificial neural networks (Bray and Lay, 1994; Pritchard and Dufton, 2000), Boolean networks (Edwards, 1995), petri nets (Fuss, 1987), rule-based systems (Cárdenas-García, 2001), cellular automata (Edwards, 1995; Wurthner, Mukhopadhyay and Peimann, 2000), and multi-agent systems (Fisher, Paton and Matsuno, 1999; Paton, Staniford and Kendall, 1995; Schwab and Pienta, 1997). Table 1 summarizes the main characteristics of these computational approaches, taking into account the idea behind the approach, the cognitive capabilities that can be modelled, types of presen t informat ion process ing, and the part of the cellular signalling to be modelled. From each model we took some points that we considered suitable for our work. Thus, our approach sees the dynamic of the cellular signal transduction in terms of a collection of autonomous agents communicating between them through a shared data structure, where each agent is implemented as an artificial neural network, a package of production rules, a Boolean network or a molecular automata, depending of the complexity of the task carried out by the agent and the knowledge degree or cognitive capabilities required by it. 4. Modelling intracellular signalling pathways using the adaptive autonomous agents approach Idea behind the approach Cognitive capabilities Cellular signalling pathway modelled References Types of information processing Parallel B o o l e a n computation l o g i c Intracellular signalling Genetic networks Intracellular signalling M i t o g e n a c t i v a t ed protein kin ase system K a u f f m a n , 1 9 9 1 ; Edwards, 1995; Karp and Paley, 1994; Armas et. al., 2000 Lagúnez -Otero, 1998; Cárdenas, 2001; Takai- Ig a r a s h i , T . , a n d Kaminuma,T.1998 Marijuan, 1994; Levy, 1 9 9 2 ; W u r t h n e r , Mukhopadhyay and Peimann, 2000 Intracellular signalling Holcombe, 1994 Extracellular signalling Intracellular signalling k n o w l e d g e - b a s e d inference and deduction Sequential Parallel None None Parallel Sequential Concurrent Computational approach Boolean networks Expert systems Cellular automata Petri nets Artificial neural networks (ANN) The cell can be modelled as a network of two- sate components interacting between them. The state of each component in the network depends of a particular boolean function. The in terac t ions (activation, phosphorylation, etc.) between signalling network components are modelled using production rules The interaction between cells or molecules is modelled as a matrix, where the state of an element of the matrix dep ends on the states of the neighbouring elements. The cell is seen as a connected graph with two types of nodes. One type represents elements, such as signalling molecules and proteins, whereas the other type represen ts transitions, such as activations. The proteins in signalling networks are seen as artificial neurons in ANN. Like an artificial neuron, a protein receives weighted inputs, produces an output, and has an activation value. D i s t r i b u t e d systems (agents) The cell is seen as a collection of agents working in parallel. The agents commun icate between them through messages. Memory Learning Pattern recognition Distributed Parallel Emergent Memory Learning Pattern recognition Handling fuzzy data Adaptive action selection Distributed Parallel Emergent Glucagon S ign a l l ing Pathway Intracellular Signalling Pho spho inos i tes /Ca2+ pathway Intracellular signalling Bray, 1990; Bray and Lay, 1994; Bray, 1995; Pritchard and Dufton, 2000; Paton, 1993 P a ton , S tan iford and Kendall, 1995; Fisher, Paton and Matsuno, 1999 T ab l e 1 . Summ a ry o f d i f f e r en t compu t a t ion a l app ro a ch e s fo r th e mod e l l ing o f c e l lu la r s ign a l l ing . opportunistically to changes on the blackboard, some mechanism is necessary to control these changes and to decide, at each relevant moment, which actions should be the taken. The control mechanism handles interaction between the blackboard , the knowl edge sources, and the outsourcing; such as users and control or data acquisition subsystems. The knowledge sources can be built as adaptive autonomous agents when an adaptive action selection mechanism has been allowed to control the interactions that the first ones execute on blackboard (González, 2000). In this way, these agents can model intracellular signalling pathway components such as receptors, proteins and enzymes. On the other hand, the blackboard structure allows to model the intracellular medium structure through its levels. That is, different cellular structures involved in the intracellular signalling could be mapped to different blackboard levels. In this sense, cellular membrane, juxta- membrane region, cytoplasm and nucleus will constitute different blackboard levels. The blackboard structure also provides a continuous trace of all interactions occurred between the agents. This trace can be seen as a topologic map distributed between the blackboard levels and its elements reflect the different inac t iva t ion , phosphory la t ion and ac t iva t ion , desphosphorylation degree that characterize the intracellular signalling in a given time. These previous considerations constitute the functional and structural essence of the intracellular signalling model described below. 4.3. The model Our proposal consists in modelling the cell as an autonomous agent, which in turn is c omp ose d by a society of autonomous agents, where each agent communicates through a blackboard with others. The model proposed here constitutes a refinement and adaptation of an action selection mechanism structured on a blackboard architecture previously developed by us, called Internal Behaviours Network (IBeNet) (Gershenson, Gonzá lez and Negrete, 2000a; Gershenson, González and Negrete, 2000b; González, 2000; González, Negrete, Barreiro and Gershenson, 2000). Although the IBeNet was initially built to control the action selection in autonomous agents (physical robots, animats, or artificial creatures simulated on a computer), it constitutes a working environment for the bottom-up modelling of information processing systems characterized by: (1) mod ula rit y, (2) parallel, distributed and emergent processing, (3) coordination and opportunistic integration of several tasks in real time, (4) use of several abstraction or context levels for the different types of information that participate in the processing network, (5) decision making, and (6) cognitive capabilities such as adaptive action selection, memory and learnin g. It is known that the information processing at a cellular level is characterized by many of the properties mentioned before. In this sense, the model proposed must constitute a good approach to model the intracellular signalling pathways. As mentioned above, the intracellular signalling model has been named Cellulat (a kind of animat (artificial animat) which behaves as a cell). In Figure 2, the structure of Cellulat can be appreciated. Three main components define the the Cellulat structure: blackboard, the internal autonomous agents and interface autonomous agents. the cellular The blackboard represents compartments. Different levels in the blackboard correspond to differen t cellular c ompartmen ts through which the signal transduction take place. I n th is w ay, the cellular membrane, the cytoplasm and the nucleus could be represented as different blackboard levels. The solution elements recorded on the blackboard represent two main types of intracellular signals: signalling molecules and activation/inactivation signals. Both types of signals are synthesized or created by internal autonomous agents and these, either directly or ind ire ctl y, promote the activation/inactivation of other internal aut onomou s age nts . Ot her t ypes of cellular elements or structures can be represented on the blackboard as well. The term “internal autonomous agents” has been used to ide ntify the auto nomous age nts whose tasks deal with the creation or modification of signals on the blackboard. An internal autonomous agent obtains a signal or combination or signals from a determinated blackboard level and transduces these into other signals on the same or other blackboard level. The way in which a signal is transduced depends of the cognitive capabilities of the internal autonomous agent. Internal autonomous agents model components of the intracellular signalling network such as proteins and other molecules necessary to carry out the signal transduction On the other hand, the function of an interface autonomous agent is to establish the communication between the blackboard and the external medium (they are similar to sensors or actuators in BBS). Not all external signals or combinations of these are recognized by an interface autonomous agent; this recognition depends both of the signal characteristics and the cognitive capabilities of the interface autonomous agent. Interface autonomous agents model the cell surface receptors and the mechanisms for the production of signalling molecules. directed. This is, each intracellular signal registered on the blackboard or each extracellular signal perceived constitutes an event, which could activate or inactivate one or more autonomous agents. When an internal autonomous agent is activated then it executes its action, consisting in the creation or modification of a signal on the blackboard. In Table 2 and Table 3 the characteristics and functionality of two intracellular signalling components in Cellulat are shown. Table 2 shows an interface autonomous agent whose identity, attributes and behaviour repertoire correspond to a type I receptor, whereas Table 3 represents an internal autonomous agent which models an adapter protein. In both cases, the behaviours of these components have been modelled using production rules, as a first approach. R e c e p t o r Iden t i ty ( I ) S t r u c t u r a l s t a t e ( S S ) P h o s p h o r y l a t io n S t a t e (P S ) A c t i v a t i o n S ta t e (A S ) L i g a nd v e c to r ( L V ) P h o s ph o r y la t i on S i te V e c to r ( P SV ) I n t er a c ti o n P ro t e in V e c to r ( IP V ) C e l lu l a r C om p a rtm e nt ( C C ) C r ea t e -R e c ep to r ( ) Ex t e rna l - In t e ra c t ion ( ) Pho spho ry la t ion ( ) A c t iva t ion ( ) In t e rna l - In t e ra c t ion ( ) D e s t roy -R e c ep to r ( ) Tab le 2 . Recep to r mode l in C e l lu la t . Adap ter Pro te in Iden t i ty ( I ) D om a in V e c to r ( D V ) I n t er a c ti o n P ro t e in V e c to r ( IP V ) C e l lu l a r C om p a rtm e nt ( C C ) I n it i a l I n a ct i v e C o n c e nt r a ti o n (I I C ) A c t u al A c ti v e C on c e n tr a t io n ( A A C ) I n a ct i v e A c t u al C o n ce n t ra t i on ( I A C) C r ea t e -P ro t e in ( ) P ro t e in -R e c ep to r - In t e ra c t ion ( ) P ro t e in -P ro t e in - In t e ra c t ion ( ) Chang e -S ta t e -D oma in ( ) Tab le 3 . Adap te r p ro te in mo de l in Ce l lu la t . 4.4. Cognitive capabilities of Cellulat’s agents It is known that certain types of proteins exhibit F igu re 2 A rch i tec tu re o f the Ce l lu la t . Each agent, independently of its type, has a condition part and an action part. The way in which both parts are linked depends on the complexity of the intracellular component modelled by the agent. For this reason, agents which model complex components could use more advanced techniques, such as neural networks, or any combination of other techniques, to link both parts. Agents which model less complex components could use less sophisticated but useful techniques, such as production rules, Boolean networks or others. The work of both types of agents is event- several cognitive capabilities such as pattern recognition, memory, and handling fuzzy data (Fisher, Paton and Matsuno, 1999). The evolution, learning and emergence of properties have also been suggested in protein netwo rks (P ritch ard an d Duft on, 20 00). In Cellulat, both types of autonomous agents exhibit several cognitive capabilities including pattern recognition, handling uncertainty, and fuzzy data, adaptive act ion sel ect ion , memor y, and learning. These capabilities allow the autonomous agents to exhibit adaptive behaviour. These cognitive capabilities are supported by several artificial intelligence paradigms including the following approaches: rule-based reasoning, probabilistic reasoning, artificial neural networks, Boolean networks, fuzzy logic systems, and multidimensional logic systems. 4.5. Modelling spatial organization in Cellulat Another important aspect to consider in the modelling of intracellular signalling networks is their spatial organization. Recent experimental data clearly demonstrate that many intracellular signalling networks exhibit a high level of spatial organization. Cellular functions could depend on the spatial organization of the cell’s components. New intracellular signalling models take this into account (Fisher, Malcom and Paton, 2000). Cellulat allows to model the spatial organization taking into account two organizational criteria of the blackboard architecture. The first is horizontal organization, given by the different abstraction levels of the blackboard, which allow intra level signal processing. The other is vertical organization, given by columns that vertically cross different blackboard levels. These columns arise as result of the adjoining work of several internal autonomous agents that operate at a same section of blackboard, which cover different blackboard levels. We have named these columns “agency columns”. These columns represent signalling pathways. Convergence and divergence of agency columns could occur, and these processes could be related with evolution and learning of the signalling networks (González, 2000). In this way, the model proposed allows to model intracellular signalling pathways by taking into account their spatial organization. This is, the two information processing levels present in Cellulat (horizontal and vertical) allow us to establish what may be called a “topology preserving map”. 5. Methodology Once Cellulat has been implemented, it constitutes a type of shell to model particular signal transduction pathways. The creation of an intracellular signalling network can be seen as an incremental process of definition and refinement of the pathway elements such as receptors, enzymes and secondary messengers, which are expressed as interface autonomous agents, internal autonomous agents or objects recorded on blackboard levels in Cellulat. In this sense, the creation of an intracellular signalling network involves the application of the following methods: b. 1. Select an individual signalling pathway to model. a. Divide the signalling pathway in sections to model and test. We have consider the following three sections: i. First Section: from ligand to the activation of the receptor. ii. Second Section: from the activated receptor to the activation of cytosolic kinases. iii. Third Section: from the activated cytosolic kinases to the activation of transcription factors. Identify and define th e component s belonging to the pathway section. That is, internal autonomous agents such as proteins, interface autonomous agents such as receptors, and objects recorded on blackboard levels such as activation/ phosphorylation signals and signalling molecules. c. Test the signalling pathway section, displaying the resultant behaviour through activity-time and concentration-time curves and activity maps. If the resultant behaviour of the model is not desirable, adjust and refine the signalling pathway section. If there are still signalling pathway sections to model then go to step 1.b, otherwise go to next step. 2. Join the different signalling pathway sections to form the complete pathway and to test this one. a. Adjust and refine the description of the individual signalling pathway, displaying the resultant behaviour through activity-time and concentration-time curves and activity maps. If there is another intracellular signalling pathway to model then go to step 1. Else, go to d. e. b. next step. 3. Define connections of two or more individual pathways, taking into account the following criterion: • Second messengers produced by one pathway are used as inputs to other pathways. • Enzymes whose activation was regulated by one pathway are connected to substrates belonging to other pathways. a. Test the signalling network behaviour against published experimental data. If the network resultant behaviour is not expected then refine the connections between pat hwa ys and return t o step 3.a. Else, go to next step. 4. The model of the signalling network has been completed. b. The definition of the signalling pathway components consists in the creation of Cellulat’s agents. This process can be seen as the creation of new instances from templates declared and contained in Cellulat. A new instance is created defining the appropriate values for the attributes specified as part of an instance state. On the other hand, the adjustment and refinement of an individual signalling pat hwa y, signalling pathway sections or interaction between pat hwa ys is reduced to the modification of different attribute values belonging to certain agents in Cellulat. 6. Modelling EGF receptor signalling pathway: an initial approach Applying part of the methodology previously presented we modelled a MAPK signalling pathway activated by the epidermal growth factor (EGF). We began with this pathway because some of its elements participate in other pathways that we will model afterwards. Figure 3 shows an interaction model of the MAPK signalling pathway. The identity of proteins belonging to MAPK signalling pathway are shown in Table 4. In this section we will present only a few steps of t he p revi ous ly de scr ibe d me tho dol ogy. The EGF receptor (EGFR) signal pathway is one of the main regulators of cell proliferation, differentiation, senescence and apoptosis. For the modelling of this pathway we divided it in three parts. The elements included in these parts are explained below. F igu re 3 . M APK s inga l l ing pa thw ay in te rac t ion m ode l . N um b e r s a r e in d i v i d u al p r o t e in s . A r r ows r e p r e s en t r e l at i o n b e tw e en two p ro t e in s . Two o r mo r e a r row s m e an coop e r a t iv e o r a l t e rn a t i v e a c t i o n . S e c t i o n s a r e i n d ic a t e d . 1 EGF 7 p120 2 EGFR 8 p190 13 Rho 14 Ras 19 RasGRp 25 Pyk2 20 RasGRF 26 PI3K 3 Grb2 9 p115 15 Rap 21 Src 31 PKC 32 PDK 33 AKT 34 MEK 37 ERK1 38 ERK2 39 JNKK 40 JNK 43 cJun 49 Elk1 44 JunD 50 CREB 45 FosB 51 cAbl 46 JunB 52 STAT 27 PAK 28 Raf 16 cdc42 22 PLCg 4 Shc 5 Gab2 10 cdc42 GAP 11 RapG AP 17 Rac 23 SHP2 29 JAK 35 MEKK 41 MKP1 47 AP1 53 ATF2 6 Sos 12 Dbl 18 KSR 24 SHP1 30 MLK 36 RSK 42 cFos 48 IP1 54 SRF T ab l e 4 . MAP K s ign a l t r an sdu c t ion e l em en t s . Th e numb e r s c o r r e s p o n d to F i g u r e 3 . Polypeptides such as growth factors, differentiation factors and hormones are crucial components of the regulatory systems that coordinate development of multicellular organisms. Many of these factors mediate their pleiotropic actions by binding to and activating the cell surface receptors with an intrinsic protein tyrosine kinase activity. It appears that ligand-induced activation of the kinase domain and its signalling potential are mediated by receptor dimerization. Ligand binding and the subsequent conformational alteration of the extracellular domain induce receptor dimerization, which stabilizes interaction of kinase function by molecular interaction. Receptor dimerization permits the transmission of a conformational change from the extracellular domain to the cytoplasmic domain and it allows its interaction with the following proteins of the signalling cascade. These interactions give rise to the first section of the model. The identified components belonging to this section are the follow ing: Signalling molecules: EGF • Receptors: EGFR • Table 5 presents the value assignment to the attributes defining one of the previously identified components. The elements of the Ligand Vector represents the lig and ide nti ty and affinity of the receptor for the ligand. The elements of the Phosphorylation Site Vector represent the identity of the molecule, identity of the phosphorylation protein, and the site state. When this identity is equal to the receptor identity (in this case, EGFR), it is a autophosphorylation site. The elements of the Interaction Protein Vector represent the protein ide nti ty, the domain identity and the affinity of the receptor for the pro tei n. A n aff ini ty value equals to 1 corresponds to the greater affinity value. E p i d e rm a l G r ow t h F a c t o r R e ce p t o r ( E G F R ) A t t r ibu t e S t ruc tu ra l s ta te ( S S ) P h o s p h o r y la t i o n S t a t e ( P S ) Ac t iva t ion S ta te (A S ) V a l u e 0 /mono me r / 0 /non pho spho ry la ted / 0 /* inac t iva ted * / L i g an d v ec t o r ( L V ) ( (EGF , 2 ) ) P h o s p h o r y la t i o n S i te V e c to r ( PS V ) ( (Y920 ,S r c ,1 ) , (Y891 ,S r c ,2 ) , (Y1173 ,EG FR ,1 ) , (Y1148 ,EG FR ,2 ) , (Y1086 ,EG FR ,2 ) , (Y1068 ,EG FR ,2 ) , (Y992 ,EGF R ,2 ) , (T654 ,MA PK ,2 ) , (T669 ,MA PK ,1 ) ) In te rac t ion P ro te in V e c to r ( IP V ) ( (G rb2 , SH2 ,3 ) , (Sh c , SH2 ,3 ) , (PLCg , SH2 ,1 ) , (P I3K , p8 5SH2 ,2 ) ) C e l l u l a r C om p a r tm en t (CC ) m emb r an e T a b l e 5 . I n i ti a l va l u e s o f a n in t e r fa c e a g e nt EG FR . Controlling the state of phosphorylation is an important mechanism by which signalling molecules regulate the activity of other proteins. A common theme in molecular signalling is the kinase cascade, in which a linear series of kinases is activated by phosphorylation by a kinase. The first kinase is specifying growth factor receptor. In this phase of the signalling cascade, phosphatases that regulates the state of activation and inactivation, and adapter proteins that facilitate two proteins also interaction between participate. This phase gives rise to the second section of the model. The identified components belonging to the second sect ion are the f ollowing: Adapter proteins: GRB2, SHC, Gab2 • Guanine interchanging proteins: Sos, p120, • p190, p115, cdc42GAP, RapGAP G proteins: Rho, Ras, Rap, cdc42, Rac Enzymes: KSR, RasGRP, RasGRF, Src, P LCg, SHP-2 • • Table 6 presents the value assignment to the attributes defining a previously identified component in second section. The elements of the Interaction Protein Vector represent the protein identity, the domain identity and the affinity. G R B 2 A t t r ibu t e V a l u e D o m a in V e c to r ( D V ) In te rac t ion P ro te in V e c to r ( IP V ) C e l l u l a r C om p a r tm en t (CC ) In i t i a l In a c t iv e C o n c e n tr a t io n ( I IC ) A c tu a l A c t iv e C o n c e n tr a t io n (AAC ) I n a c t i v e A c t ua l C o n c e n tr a t io n ( IAC ) ( (SH2 , f r e e ) , (SH3A , f r e e ) , (SH3B , f r e e ) ) ( (So s ,SH3 ,1 ) , (G ab2 ,SH3 ,2 ) ) cy top la sm 40 nM 0 nM 40 nM T a b l e 6 . I n it i a l va l u e s o f a n i n t er n a l a g en t GRB 2 . The third section of the model corresponds to the activation of transcription factors by kinase proteins. The transcription factors activate the transcription of genes codifying for proteins involved, in this case, in cell growth. Some of these formed proteins belong to this signalling pathway, and others give rise to transcription factors. This is a reason why a positive or negative feedback can be observed. This third section includes the following components: Enzymes: Pyk2, PI3K, PAK, Raf, JAK, PAK, • MLK, PKA, PKC, PDK, AKT, MEK, MEKK, RSK, ERK1, ERK2, JNKK, JNK, SEK, S6K, MKP1 Transcription factors: c-Fos, c-Jun, JunD, FosB, JunB, IP-1, Elk-1, CREB, c-Abl, STAT, ATF2 In Table 7 the initial values assigned to a component in this last section are presented. • J N K K A t t r ibu t e I n t e r a ct i o n P r o te i n V e c t o r ( I P V ) V a l u e ( ( JNK ,Y17 ,1 ) ) C e l l ul a r C om p ar tm e n t (CC ) cy top la sm I n i t ia l I n a c ti v e C o n c e n tr a t i o n ( I IC ) A c t u a l A c ti v e C o n c e n tr a t i o n (AAC ) I n a c t i v e A c t ua l C o n c e n tr a t io n ( IAC ) 1nM 0 nM 1 nM P h o s p h o r y la t i o n S ta t e ( P S ) 0 /non pho spho ry la ted / P h o s p h o r y la t i o n S it e V e c t o r ( P S V ) ( (TX , MEK K ,1 ) ) T a b l e 7 . In i ti a l v a lu e s of a n i n te r n al a ge n t J N K K . All the components of the signalling pathway MAPK identified in each of the three sections are shown in Table 8. S igna l l ing mo lecu le R e c e p t o r S e c t i o n I E G F EG FR Sec t ion I I Adap ter pro te in G r b 2 , S h c , G a b 2 G u a n i n e i n t e rc h a n g i n g pro te in S o s , p 1 2 0 , p 1 9 0 , p 1 1 5 , c d c 4 2GA P , R a pGA P , D b l G Pro te in En z ym e s En z ym e s T r a s c r ip t i on F a c t o r s R h o , R a s , R a p , C d c 4 2 , R a c K SR , R a sGR P , R a sGR F , S r c , P L C g , SH P 2 , S H P 1 Sec t ion I I I P y k 2 , P I 3 K , P A K , R af , JA K , M L K , P K A , P K C , P D K , A K T , M E K , M E K K , R S K , E R K 1 , E RK 2 , J N K K , J N K , M K P 1 c F o s, c Ju n , J u n D , F o s B , J u nB , A P 1 , I P 1 , E lk 1 , CREB , cA b l , STAT , AT F 2 T ab l e 8 . Id en t i f i ed compon en t s o f th e MAPK s ign a l l ing pa thway . Figure 4 shows the semantics assigned to the different blackboard levels and types of agent in Cellulat for this pathway. For example, the agents of the “Adapter protein” type would be GRB2, SHC, and Gab2; and some of the agents of the “G-protein” type would be Rho, Ras, and Rap. F igu re 4 . M ode l o f E GFR s ig na l l ing pa thwa y in Ce l lu la t . Once defined the elements belonging to each of the sections o f the MAPK s ignalling pathw ay, according to our methodology, each section is tested, for then assembling the sections to form the complete pathway. The individual pathway is also tested. 7. Towards an intracellular signalling virtual laboratory Virtual laboratories have been developed in different areas to reproduce experiments, that in most cases, were made in real laboratories. They are used for understanding the systems they simulate, and they can also be used for teaching. Virtual labs do not replace physical labs, but they have advantages over them. They have relatively low costs, there are no inconveniences in failing experiments, and the most important thing: the researchers can control and reproduce easily situations. We will not find new structures in virtual labs, because all structures were programmed. But they are quite useful for understanding the processes and emergent properties of the systems they simulate. Thus, people can go back to the physical world, and control a system after understanding its mechanisms in a virtual lab. Virtual labs have been developed for different areas, such as physics, ele ctr oni cs, rob oti cs, phys iol ogy, c hem ist ry, engineering, economics, ecology, and eth ology (Gershenson, González and Negrete, 2000). One of the reasons for our interest in the analysis and understanding of the signalling pathways, is the possibility of regulating them. In principle, it is possible to observe this process in a virtual laboratory based on our paradigm. In particular, we would like to see the effects of perturbations on the systems such as adding elements or taking them out as knock-outs. The expected effects would be directly on the cognitive capacity of the network and ultimately on the decisions taken by the cell in order to differentiate, proliferate or become senescent. Pathologies and natural processes can be followed in the computation of the interactions made by the components in the modelled network. The paradigm presented here is the backbone of the virtual lab orat ory. With the virtual laboratory we hope that etiologies and expected results of putative therapeutic strategies can be visualized. 8. Conclusion We have constructed an agent-based system where cognitive capabilities are coded using behaviour- based paradigms and the blackboard architecture, combined with other artificial intelligence techniques. Recruiting these techniques, the complexity of the topology and cognitive capacities of intracellular signalling system can be studied. In our group we are currently following them in the context of pathological states in particular, senescence and cancer. 9. Acknowledgements We would like to thank E. St. James for excellent discussions and proofreading. Part of this work was supported by DGAPA/UNAM. 10. References A rma s , M . , A rm a s , O . , C á r d e n a s , M . , V a s c o n c e lo s , J . and L a g ú n e z -O te r o , J . , 2000 . S igna l T ran sduc t ion A na lys is U s ing Log ic P r og ramm ing Pa rad igms . In N . M as to rak is ed . , pp 165 -16 8 , WSE S P re s s . B ray , D . , 1990 . In t r a c e l lu l a r s i g n a l l in g a s a p a r a l l el d i s t r ib u t e d p r o c e s s . J o u r n a l o f T h e o r e t i c al B i o l o g y, 1 4 3 , 2 1 5 - 2 3 1 . B ray , D . and La y , S . , 1994 . Compu te r s imu la ted e vo lu t ion o f a n e two rk o f ce l l s igna l l ing m o lecu les . B i o p h y s i c a l J o u r n a l, V o l . 6 6 , N o . 4 , p p . 9 7 2 -9 7 7 . B ray , D . , 199 5 . P ro te in m o lecu les as c ompu ta t io na l e lemen ts in l i v i ng c e l l s. N a t u r e , 3 7 6 , 3 0 7 - 3 1 2 . B rook s , R . A . , 1986 . A rob u s t laye red co n t ro l sys tem fo r a mob i le robo t , IEEE J o u r n a l o f R o b o t i c s a n d A u t om a t i o n, RA - 2 , p p . 1 4 - 2 3 . Cá rden as -Ga rc ía , M . , 2 0 0 1 . B ú s q u e d a d e b l a n c o s te r a p é u t ic o s em p l e a n d o un mode lo d e s e ñ a l e s i nt r a c e l u l a re s m e d i a d a s po r Ras , T e s i s D o c to r a l, UNAM , M é x i c o . Edwa rd s , C .F . , 1 9 9 5 . C om p u t a t i o n al m o d e l s f o r c e ll u l a r s y s t e m s , u n p u b l i s h ed i n f o r m a ti o n p r o c e s s i n g underg radua te thes is , Un ive rs i ty o f L ive rpoo l . F i sh e r , M . J . , P a t o n, R .C . a n d M a t s u n o , K . , 1 9 9 9 . I n t r a c e l l ul a r s ign a l l ing p ro t e in s as 'sma r t ' a g e n t s i n p a ra l l e l d i st r i b u t e d p rocesse s , B i o S ys t em s , 5 0 , 1 5 9 - 1 7 1 . F i sh e r , M . J . , M a l c om , G . a n d Pa ton , R .C . , 2 0 0 0 . S p a t i o - lo g i c a l p r o c e s s e s in in t race l lu la r s igna l l ing , B i o S ys t em s , 5 5 , 8 3 - 9 2 . Fu s s , H . , 1 9 8 7 . S im u l a t io n o f B i o l o g i c a l Sy s t em s w i th P e t r i Ne t s - I n t r o d u c ti o n t o M o d e l l ing o f D i s t r ibu t ed Sy s t em s . In A d v a n c e s in Sys tem Ana lys is , Mo l l e r D . ( ed . ) , B r a u n s c hw e ig , W i e s b a d e n : V i ew e g , 1 -1 2 . G e r sh en son , C . , G o n z á l ez , P . P . a n d N eg r e t e , J . , 2 0 0 0 a. A c t i o n S e l e c t io n P r o p e r t i es i n a S o ftw a r e S im u l a t e d Ag e n t , Lec tu re N o te s i n A r t i f ic i a l I n t e l li g e n c e , V o l . 1 7 9 3 , p p 6 3 4 - 6 4 8 . G e r sh en son , C . , G o nz á l ez , P . P . and N eg r e t e , J . ,2000b . Th ink ing A d a p t i v e : Tow a rd s a B eh av iou r s V i r tu a l L abo r a to ry . P r o c e e d i n g o f SAB ’ 2 0 0 0 , P a r i s, F r a n c e , p p . 8 9 - 9 7 . G o n z á l e z , P .P . and J . N eg r e t e . , 1997 . REDS IE X : A coop e r a t iv e n e two rk o f e x p er t s ys t em s w i th b lackboa rd a rch i tec tu re s , E x p e rt S y st e m s, V o l . 1 4 , N o . 4 , p p . 1 8 0 -1 8 9 . G o n z á l e z , P . P . , 2 0 0 0 . R ed e s d e Conduc ta s In te rna s co mo Nod o s - P i z a r rón : S e l e c c ión d e A c c ion e s y Ap r end i z a j e en un R o b o t R e a c t ivo , T e s i s D o c t o r a l, I n s ti t u t o de I n v e s t ig a c i o n e s B i om é d i c a s /UNAM , M é x i c o . G o n z á l e z , P .P . , N e g r e te , J . , Ba r re i ro , A .J . and Ge rshen son , C . , 2 0 0 0 . A M o d e l f o r C om b i n a ti o n o f E x t e r n a l a n d I n te r n a l S t imu l i in th e A c t ion S e l e c t ion o f an Au tonomou s Agen t , Lec tu re N o t e s i n A r t i fi c i a l I n t e ll i g e n c e , V o l . 1 7 93 , p p 6 2 1 - 6 3 3 . G o n zá le z , P . P . , G e r s he n s o n , C ., C á r d e n a s, M . a n d L a g ú ne z -O t e r o , J . , 2000 . Mod e l l ing in t r a c e l lu l a r s ign a l ling n e two rk s u s ing behav iou r -based sys tem s and the b la c kbo a rd a r c h i t e c t ur e . In N . M as to rak is ed . M a th em a t i c s and Compu t e r s i n M o d e r n S c i en c e . p p 1 6 5 - 1 68 , W SES P re s s . H o l c omb e , M . , 199 4 . F rom V LS I th roug h Mach ine Mo de ls to C e l lu l a r M e t a b o l i sm . C om p u t i n g W i t h B i o lo g i c a l M e t a p h o r s , P a to n , R . e d ., C h a pm a n a n d H a ll , 1 1 - 2 5 . K a r p , P .D . and Pa ley , S .M . , 1994 . R ep resen ta t io ns o f Me tabo l ic K n ow l e d g e : P a t hw a y s . P r o c ee d i n g s o f S e c o nd I n t e r n a t io n a l C o n f e r e n c e o n I n t e ll i g e n t S y s t em s f o r M o l e c u l a r B i o lo g y . p p 2 0 3 - 2 1 1 K au f fm an , S . A . , 1 9 9 1 . A nt i c haos and a dap ta t ion . Sc ien t i f ic Am e r i c a n 2 6 5 , 2 , p p . 6 4 - 7 0 . L a g ú n e z -O te r o , J . , 1998 . T he Ce l l a s a n E xp e r t S ys t em , D ag s tuh l S em i n a r R e p o rt : 2 1 5 , p .6 0 . I S SN 0 9 4 0 - 1 1 2 1 . Levy , S . , 1992 . A r t i f ic ia l l ive : the que s t f o r a n ew c r e a t ion . Jon a th an C ap e , London . Mae s , P . , 1 9 94 . A dap t ive A u tonomo us Agen ts , J o u r n a l o f Ar t i f ic ia l L i fe , V o l . 1 , N o . 1 a n d 2 . M a r i ju an , P . C . , 1994 . Enz yme s , au toma ta , and a r t i f ic ia l ce l l s . Compu t ing w i th B io log i c a l M e t apho r s , P a ton , R . ( ed . ) , C h a pm a n a n d H a l l, p p . 5 0 - 6 8 . N e g r e t e , J . and Gon z á l e z , P .P . , 1998 . N e t o f mu l t i - ag en t exp e r t s y s tem s w i th eme rge n t con t ro l , Exper t S ys tems w i th A p p l i c a t io n s , V o l . 1 4 , N o . 1 , 1 0 9 - 1 1 6 . N i i , H . P . , 1989 . B l a ckbo a rd s y s t em s . I n A . B a r r , P . R . C o h e n and E . A . F e i ge n b a um (e d . ) , T h e H a n d b o o k o f A r t if i c i a l I n t e l l i g e n c e, vo lum en IV , Add i son -W e s l ey Pub l i sh ing Comp any . P a ton , R .C . , 1 9 9 3 . S om e C om p u t a t i o n a l M o d e l s a t t h e C e l l u la r Leve l , B i o S ys t em s , 2 9 , 6 3 - 7 5 . P a ton , R .C . , S ta n i f o r d , G . a n d K end a l l , G . , 1995 . Sp e c i fy ing L o g i c a l Agen ts in Ce l lu la r H ie ra rc h ies , P r o c e e d i n gs o f I P C A T ( I n f o rm a t io n P r o c e ss i n g in C e l l s and T i s su e s ) , P a ton , R .C . , H o l c om b e , M . y S t a n if o r d , G . (E d s . ) , p p 3 0 2 - 3 1 7 . P r i t c h a rd , L . and Du f ton , M . J . , 2000 . D o P ro te in s Lea rn to E v o l v e ? T h e H o p f i e l d N e tw o r k a s a Ba s i s fo r th e Und e r s t and ing o f P ro te in Evo lu t ion , J o u r n a l o f T h e o r e t i ca l B i o l o g y, 2 0 2 , 7 7 - 8 6 . S c hw a b , E .D . a n d P i e nt a , K . J ., 1 9 9 7 . M o d el in g s i g n a l t r a n s d u ct i o n in n o rm a l a n d c a n c e r c e l ls u s i n g c om p l e x adap t ive sys te ms , M e d i c a l H y p o t h e s e s, 4 8 , 1 1 1 - 1 2 3 . Taka i - Iga rash i , T . and K am inuma , Ts . , 199 8 . A p a t hw a y f ind ing s y s t em f o r t h e c e l l si g n a l i n g ne tw o r k d a t a b a s e . J o u r n a l o f Compu t a t ion a l B io logy , 5 ( 4 ) 7 4 7 - 7 5 4 . Wu r thn e r , J .U . , M u k h o p a d h ya y , A . K . and P iemann C .J . , 2000 . A c e l l u l a r a u t om a t o n m o d e l o f c e l lu l a r s ign a l tr an sdu c t ion . C om p u t e r s i n B io l o g y a n d M e d i c i n e, 3 0 , 1 - 2 1 .
1205.7025	1	1205	2012-05-31T15:50:27	Engineering hierarchical complex systems: an agent-based approach. The case of flexible manufacturing systems	[ "cs.MA" ]	This article introduces a formal model to specify, model and validate hierarchical complex systems described at different levels of analysis. It relies on concepts that have been developed in the multi-agent-based simulation (MABS) literature: level, influence and reaction. One application of such model is the specification of hierarchical complex systems, in which decisional capacities are dynamically adapted at each level with respect to the emergences/constraints paradigm. In the conclusion, we discuss the main perspective of this work: the definition of a generic meta-model for holonic multi-agent systems (HMAS).	cs.MA	cs	Engineering hierarchical complex systems: an agent-based approach The case of flexible manufacturing systems Gildas Morvan1,2, Daniel Dupont1,3, Jean-Baptiste Soyez1,4, and Rochdi Merzouki1,4 1 Univ. Lille Nord de France, 1bis rue Georges Lefèvre 59044 Lille cedex, France 2 LGI2A, U. Artois, Technoparc Futura 62400 Béthune, France. email: first 3 HEI, 13 rue de Toul 59046 Lille Cedex, France. email: first [email protected] 4 LAGIS, EC-Lille, Avenue Paul Langevin BP 48 59651 Villeneuve D'ascq cedex, [email protected] France Abstract. This article introduces a formal model to specify, model and validate hierarchical complex systems described at different levels of anal- ysis. It relies on concepts that have been developed in the multi-agent- based simulation (MABS) literature: level, influence and reaction. One application of such model is the specification of hierarchical complex systems, in which decisional capacities are dynamically adapted at each level with respect to the emergences/constraints paradigm. In the con- clusion, we discuss the main perspective of this work: the definition of a generic meta-model for holonic multi-agent systems (HMAS). Keywords: multi-level multi-agent based simulations, formal models, hierarchical systems 1 Introduction Engineering a complex system such as a flexible manufacturing system (FMS) is a challenging problem. The target system is complex, holonic, relies on distributed decisional processes, and must be adaptive, i.e., robust to perturbations and easily reconfigurable. To solve these problems, proposed solutions1 take advantage of system -- complexity, distributing the control in system components that embody primitive cognitive capacities, e.g., be able to be identified, to communicate, to react to environmental changes, -- holonic structure, using dedicated meta-models and conception methodolo- gies. 1 E.g., heterarchical [3] or semi-heterarchical [25] control, holonic multi-agent systems (HMAS) [2,6,35,34] or intelligent product based concepts [24]. An important tool in the design, simulation and validation of such solutions has been multi-agent-based simulation (MABS). This article introduces a formal model to specify, model and and validate hierarchical complex systems. It takes inspiration from two trends in MABS research: -- the formalization of interaction models, -- multi-level modeling, where interacting agents are ontologically distributed among multiple layers of organization. The article is organized as follows: -- in the section 2, the two trends of MABS research cited above, multi-level modeling and formal modeling2, are introduced, -- the section 3 presents a generic formal model for multi-level MABS, -- an abstract implementation of this model, focusing on the specification of hierarchical multi-agent systems (MAS), in which decisional capacities are dynamically adapted at each level with respect to the emergences/constraints paradigm, is proposed in the section 4, -- the conclusion (section 5) summarizes our contributions and perspectives. 2 Two trends in MABS research 2.1 Multi-level modeling A level represents a point of view on the system, and its relations to other points of view [16]. While this concept seems important to understand complex systems3, it generally remains abstract: implementations tend to constraint this definition, in particular the relations between levels. Therefore, a multi-level model integrates knowledge on different levels and their relations. Multi-scale are multi-level models characterized by hierarchical relations in levels [7,12,19,23]. A level may represent, according to the context, a spatio-temporal extent, a position in a decision hierarchy, etc. Let consider these two examples. 1. The system is characterized by processes that have different spatio-temporal extents. Two types of relations can be commonly found in such models: -- scaling, i.e., computing macroscopic (resp. microscopic) variables from microscopic (resp. macroscopic) processes, -- grouping and degrouping (or aggregation and disaggregation) [6,20,26], i.e., defining a process at a level as a group (resp. part) of processes (resp. a process) at an other level. 2. Levels are characterized by decisional capacities; relations represent the emer- gence of new capacities and the constraint over existing capacities [15,17]. A level is often viewed as a level of organization. This concept is closely related to the notion of holon [6]. This aspect is discussed in the section 5. 2 The presentation focuses on the influences → reaction model (IRM). Other ap- proaches such as IODA [10] or based on DEVS [18] are not described. 3 Multi-level approaches have proven useful in many domain such as statistics [8], chemistry [9,11], physics [28], hydrology [26] or biology [33]. 2.2 The influences → reaction model The influences → reaction model (IRM) has been developed to address issues raised by the classical vision of action in Artificial Intelligence as the transfor- mation of a global state [5]: -- simultaneous actions cannot be easily handled, -- the result of an action depends on the agent that performs it but not on other actions, -- the autonomy of agents is not respected. Basically, it decomposes action in two phases: agents and environment (mi- cro level) produce a set of influences, then the system (at macro level) reacts to influences; e.g., detects and solves influence conflicts such as in the platform Jaak4. As [13] notes, "the influences [produced by an agent] do not directly change the environment, but rather represent the desire of an agent to see it changed in some way". Thus, reaction computes the consequences of agent de- sires and environment dynamics. In recent years, variants of IRM have been developed to handle specific situations [13,16,36,37]. This presentation focuses on the influence reaction model for simulation (IRM4S) [13]. Let δ(t) ∈ ∆ be the dynamic state of the system at time t: δ(t) =< σ(t), γ(t) >, (1) where σ(t) ∈ Σ is the set of environmental properties and γ(t) ∈ Γ the set of influences, representing system dynamics. The state of an agent a ∈ A is characterized by its physical state φa ∈ Φa with Φa ∈ Σ (e.g., its position) and its internal state sa ∈ Sa (e.g., its beliefs). The evolution of the system from t to t + dt is a two-step process: 1. agents and environment produce a set of influences5 γ ′(t) ∈ Γ ′, 2. the reaction to influences produces the new dynamic state of the system. An agent a ∈ A produces influences through a function Behaviora : ∆ 7→ Γ ′. This function is decomposed into three functions executed sequentially: pa(t) = P erceptiona(δ(t)), sa(t + dt) = M emorizationa(pa(t), sa(t)), γ ′ a(t) = Decisiona(sa(t + dt)). (2) (3) (4) The environment produces influences through a function N aturalω : ∆ 7→ Γ ′: γ ′ ω(t) = N aturalω(δ(t)). (5) 4 http://www.janus-project.org/Jaak 5 the sets of producible influence sets and influences produced at t are denoted re- spectively Γ ′ and γ ′(t) to point out that the latter is temporary and will be used to compute the dynamic state of the system at t + dt. Then the set of influences produced in the system at t is: γ ′(t) = {γ(t) ∪ γ ′ ω(t) ∪ [ γ ′ a(t)}. a∈A (6) After influences have been produced, the new dynamic state of the system is computed by a function Reaction : Σ × Γ ′ 7→ ∆ such as: δ(t + dt) = Reaction(σ(t), γ ′(t)). (7) 3 A generic meta-model for multi-level MABS In this section, a generic meta-model for multi-level MABS, called IRM4MLS, is presented6. This model has the following interesting properties: -- any valid instance can be simulated [27], -- simulation scheduling is logically distributed by level, -- complexity of simulation algorithm can be optimized according to model structure. 3.1 Specification of the levels and their interactions A multi-level model is defined by a set of levels L and a specification of the relations between levels7. Two kinds of relations are specified in IRM4MLS: an influence relation (agents in a level l are able to produce influences in a level l′ 6= l) and a perception relation (agents in a level l are able to perceive the dynamic state of a level l′ 6= l), represented by directed graphs denoted respectively < L, EI > and < L, EP >, where EI and EP are two sets of edges, i.e., ordered pairs of elements of L. Influence and perception relations in a level are systematic and thus not specified in EI and EP (cf. eq. 8 and 9). E.g.,∀l, l′ ∈ L2, if EP = {ll′} then the agents of l are able to perceive the dynamic states of l and l′ while the agents of l′ are able to perceive the dynamic state of l′. The in and out neighborhood in < L, EI > (respectively < L, EP >) are denoted N − I and N + I (resp. N − P and N + P ) and are defined as follows: ∀l ∈ L, N − I (l) (resp. N − P (l)) = {l} ∪ {l′ ∈ L : l′l ∈ EI (resp. EP )}, (8) ∀l ∈ L, N + I (l) (resp. N − E.g., ∀l, l′ ∈ L2 if l′ ∈ N + P (l)) = {l} ∪ {l′ ∈ L : ll′ ∈ EI (resp. EP )}, I (l) then the environment and the agents of l are I (l′), i.e., l′ (9) able to produce influences in the level l′; conversely we have l ∈ N − is influenced by l. 6 The dynamic aspects of the meta-model, i.e., simulation algorithms, are not de- scribed here. An exhaustive presentation can be found in [16]. 7 The notion of level is here similar to the notion of brute space in the MASQ meta- model [29]. 0..n agent 0..n 0..n environment 1 1..n 0..n level Fig. 1. Main concepts of IRM4MLS (cardinalities are specified in the UML fashion) 3.2 Agent population and environments The set of agents in the system at time t is denoted A(t). ∀l ∈ L, the set of agents belonging to l at t is denoted Al(t) ⊆ A(t). An agent belongs to a level iff a subset of its physical state φa belongs to the state of the level: ∀a ∈ A(t), ∀l ∈ L, a ∈ Al(t) iff ∃φl a(t) ⊆ φa(t)φl a(t) ⊆ σl(t). (10) Thus, an agent belongs to zero, one, or more levels. As notes [29, p. 815], the physical state of an agent in a level, i.e., its body, is "the manifestation of an agent in the environment and allows others to perceive it." An environment can also belong to multiple levels (cf. fig. 1). 3.3 Action modeling The dynamic state of a level l ∈ L at time t, denoted δl(t) ∈ ∆l, is a tuple < σl(t), γl(t) >, where σl(t) ∈ Σl and γl(t) ∈ Γ l are the sets of environmental properties and influences of l. The influence production step takes into account the influence and perception relations between levels: ∀a ∈ Al, Behaviorl a : Y lP ∈N + P (l) ∆lP 7→ Y lI ∈N + I (l) Γ lI ′. (11) Once influences have been produced, interactions between levels do not mat- ter anymore. Thus, the reaction function defined in IRM4S can be re-used: Reactionl : Σl × Γ l ′ 7→ ∆l, (12) where Reactionl is the reaction function proper to each level. 4 Engineering hierarchical complex systems with IRM4MLS 4.1 The emergence/constraint paradigm In many MABS, processes are considered on the following 2-level relative hier- archy: micro macro. Arrows represent causality relations between levels. Dashing suggests that they are generally not explicitly defined but emerge from interactions between entities. A contrario, a multi-level approach considers these relations explicitly. In engineering applications, a level may rather represents a position in a decision hierarchy (cf. section 2.1). Two kinds of relation may be distinguished in such systems: emergence of new capacities and constraint over existing capacities [14]. Let consider an example in the domain of FMS engineering. In a case study on automated guided vehicle (AGV) control presented in [17] (cf. section 4.4), the model relies on the following relations: emergences AGV deadlock solving. constraints Macro agents (representing a set of "trapped" AGVs) emerge from micro agent interactions when an interaction pattern defined as a deadlock is detected, and then constraint their behaviors to solve it. While the notions of emergence and constraint were informally defined in [17], formal definitions in the context of IRM4MLS are given in the following. 4.2 IRM4MLS implementation Let L be a hierarchy and {µ, M } ⊆ L two hierarchically coupled levels, µ refer- ring to the micro level and M to the macro level. Thus, Aµ (respectively AM ) de- notes the agents of the micro-level (resp. macro-level). The emergence/constraint paradigm supposes that EI ⊇ {µM, M µ}. ∀l ∈ L, γl ′(t) = {γl(t), γM ω , γµ ω, [ a∈AM γM a ′(t), [ a∈Aµ γµ a ′(t)}. (13) An emergence e at the level M is an influence that has the following proper- ties: -- e belongs to the macro-level but not to the micro-level: e ∈ Γ M but e /∈ Γ µ, (14) -- e cannot be produced by the behavior of an agent or the environment of M : ∀t, e /∈ [ a∈AM BehaviorM a (δ(t)) ∪ N aturalM ω (δ(t)), (15) with δ(t) =< δM (t), δµ(t) >. Emergent influences generally determine the life-cycle (creation, evolution, de- struction) of agents at the macro-level. A constraint over an influence i, denoted ¬i, is the special kind of influence that has the following properties: -- {i, ¬i} belongs to the micro-level but not to the macro-level: {i, ¬i} ⊆ Γ µ but {i, ¬i} * Γ M , (16) -- ¬i cannot be produced by the behavior of an agent or the environment of µ: ∀t, ¬i /∈ [ a∈Aµ Behaviorµ a (δ(t)) ∪ N aturalµ ω(δ(t)), with δ(t) =< δM (t), δµ(t) >, -- ¬i inhibits i: if {i, ¬i} ⊆ γµ′(t) then Reactionµ(σµ(t), γµ ′(t)) = Reactionµ(σµ(t), γµ ′(t)\{i}). (17) (18) 4.3 Conception of hierarchical systems The approach described below can be viewed as a semi-heterarchical control one and takes advantage of complexity and hierarchical (not yet holarchical) orga- nization of the system, distributing the control by level. Heterarchical control methods rely on self-organization principles8 and therefore assume that the sys- tem is able to achieve its goals and is easily reconfigurable, i.e., that the normal functioning mode emerges from the interactions between system components (products, machines, simulated entities, etc.) that embody limited cognitive ca- pabilities (cf. introduction). However, the trajectory of such systems may lead to non desired attractors. The proposed methodology is presented in the fig. 2. The system is designed iteratively in a two-step process. 1. From an initial specification of the system, a model of the system in normal functioning mode, is defined and verified, i.e., that system components have the necessary cognitive capacities to perform their tasks. 2. From non desired attractors exhibited by the simulation of the model, the control strategy may be designed and validated. However, it is likely that the specification of the system has be modified to do so, e.g., because a new decisional level is needed. The notion of influence is very general and therefore, may have many possible meanings. In this case, let 8 Self-organized systems are generally characterized by the use of environment as a communication medium to carry local informations as well as positive and negative feedbacks. modification Specification Model design Control design Implementation verification validation Fig. 2. Engineering methodology -- γl ′(t) be the capacities of each agent of a level l at time t, i.e., the tasks they can perform at the moment, -- γl(t) the actual affectation of tasks to agents; the only cognitive capacity required for agents is to expose services they may provide. Thus, Reactionl is a task assignment algorithm that computes < σl(t), γl(t) > from < σl(t), γl ′(t) >. Note that the hierarchical nature of the system allows to decompose the specification of the system S by level: S = {γl(δl) : ∀l ∈ L, ∀δl ∈ ∆l}, (19) i.e., task assignments for all functioning modes. That design should lead to the definition of reaction functions that control goal affectations. If such a function cannot be defined, then the system design is not valid and must be redefined. This process is iterated until a solution is found (cf. fig. 2). 4.4 Case study: AGV deadlocks in gradient field-based FMS The main functionalities of an intelligent transportation system (ITS) are: (1) transport assignment, (2) routing, (3) gathering traffic information, (4) collision avoidance, (5) deadlock avoidance [38]. Gradient field-based approaches, where AGV trajectories are computed from gradient fields, allow to implement efficient ITS in FMS [31,32]. A dedicated task assignment algorithm is generally used to ensure functionality 1, while function- alities 2 -- 4 rely on AGV and shop self-organization properties. Thus, an AGV has two cognitives capabilities: sense attractive or repulsive force fields and emit a repulsive force field. Similarly, a shop is able to emit attractive fields to require products to process and give back the result to the system. A known problem of gradient field-based approaches is that a group of AGVs may be trapped in local minima that lead to a system deadlock [30,32,39]. However, this issue can be easily addressed by hierarchical control methods that compute explicit trajectories9. 9 Readers interested in general, i.e., not gradient-field based approaches, deadlock avoidance techniques in FMS may refer e.g., to [1,4,40]. Task assignment Deadlock solving Task assignment ¬i e AGV (a) Initial design ¬i Shop e ¬i ¬i e AGV (b) Final design ¬i Shop mediated interactions Fig. 3. Decisional levels in the case study on AGV control The first design of the system is presented in fig. 3(a): a task assignment algorithm affects goals to AGVs (statically, a signal to maximize) and shops (dinamically, products to process). The deadlock avoidance functionality is not explicitly programmed but is supposed to emerge from mediated interactions between AGVs and shops. Various researches have shown that such a solution may reduce the number of deadlock occurrences but not eliminate it: routing is not deadlock avoidance [39]. A new system architecture is then designed (cf. fig. 3(b)): if a deadlock (reified by an emergence) is observed by a deadlock solving algorithm, constraints over signal sensing and emission are computed to solve it 10. 5 Conclusion In this article, we have presented a formal model for MABS and its implementa- tion to engineer hierarchical complex systems. Two types of influences have been distinguished in this approach : emergence, that basically triggers a new system behavior when a specific pattern is detected (in the previous short example of gradient field-based FMS, the detection of a deadlock triggers the modification of AGV repulsive signal emission) and constraint, that, as its name suggests, constraints decisional capacities of system entities to solve a situation. The main advantage of this approach lies in the multi-level and simulation capabilities of IRM4MLS, to model a system in which decisional capacities are distributed in its components and evolve along time to meet user's goals and to simulate a model whiteout bias and temporal deadlocks11. Its main drawback is the strict hierarchical organization in levels. Holonic multi-agent systems (HMAS) can be viewed as a specific case of multi-level multi-agent-systems (MAS), the most obvious aspect being the loosely 10 Practical aspects of this approach are discussed in [17]. E.g., AGVs embody the deadlock solving algorithm, becoming multi-level agents. This problem has been an important motivation in the development of IRM4MLS. 11 Simulation properties of IRM4MLS may be exploited to explore model behavior us- ing, e.g., the polyagent concept [22,21]. Such an approach may be used to determine fail probabilities of system components or control strategies. hierarchical organization of levels. However, from a methodological perspective, differences remain: thus, most of holonic meta-models focus on organizational aspects (cf. e.g., [2,6,35,34]). An important issue towards a generic meta-model for HMAS would be to define a holon with respect to IRM4MLS concepts: a holon cannot be defined with IRM4MLS first class abstractions (level, agent or environment), as it represents a multi-level entity. This situation is the main perspective of this work. Acknowledgments Authors would like to thank Daniel Jolly (LGI2A, Université d'Artois, Béthune France) and Alexandre Veremme (HEI, pôle recherche Ingénierie et Sciences du Vivant, Lille France) for their help and support. Jean-Baptiste Soyez is funded by the InTrade project12. References 1. Banaszak, Z., Krogh, B.: Deadlock avoidance in flexible manufacturing systems with concurrently competing process flows. IEEE Transactions on Robotics and Automation 6(6), 724 -- 734 (1990) 2. Bendriss, S., Benabdelhafid, A., J.Boukachour, Boudebous, D.: Meta-modèle de référence holonique pour la gestion de la traçabilité du produit dans la chaîne logistique. In: 5ème Colloque International Conception et Production Intégrées - CPI'2007 (2007) 3. Duffie, N.: Heterarchical control of highly distributed manufacturing systems. In- ternational Journal of Computer Integrated Manufacturing 9(4), 270 -- 281 (1996) 4. Ezpeleta, J., Tricas, F., f. Garcia-Valles, Colom, J.: A banker's solution for deadlock avoidance in fms with flexible routing and multiresource states. IEEE Transactions on Robotics and Automation 18(4), 621 -- 625 (2002) 5. Ferber, J., Müller, J.P.: Influences and reaction: a model of situated multiagent systems. In: 2nd International Conference on Multi-agent systems (ICMAS-96). pp. 72 -- 79 (1996) 6. Gaud, N., Galland, S., Gechter, F., Hilaire, V., Koukam, A.: Holonic multilevel simulation of complex systems : Application to real-time pedestrians simulation in virtual urban environment. Simulation Modelling Practice and Theory 16, 1659 -- 1676 (2008) 7. Gil Quijano, J., Hutzler, G., Louail, T.: Accroche-toi au niveau, j'enlève l'échelle: Éléments d'analyse des aspects multiniveaux dans la simulation à base d'agents. Revue d'Intelligence Artificielle 24(5), 625 -- 648 (2010) 8. Goldstein, H.: Multilevel Statistical Models. Wiley Series in Probability and Statis- tics, Wiley-Blackwell, 4th revised edition edn. (2010) 9. Horstemeyer, M.: Practical Aspects of Computational Chemistry Methods, Con- cepts and Applications, chap. Multiscale Modeling: A Review, pp. 87 -- 135. Springer (2010) 12 http://www.intrade-nwe.eu 10. Kubera, Y., Mathieu, P., Picault, S.: Interaction-oriented agent simulations: From theory to implementation. In: Proceeding of the 2008 conference on ECAI 2008: 18th European Conference on Artificial Intelligence. pp. 383 -- 387. IOS Press (2008) 11. Lucia, A.: Multi-scale methods and complex processes: A survey and look ahead. Computers & Chemical Engineering 34(9), 1467 -- 1475 (2010), selected papers from the 7th International Conference on the Foundations of Computer-Aided Process Design (FOCAPD, 2009, Breckenridge, Colorado, USA. 12. McGregor, S., Fernando, C.: Levels of description: A novel approach to dynamical hierarchies. Artificial Life 11(4), 459 -- 472 (2005) 13. Michel, F.: The IRM4S model: the influence/reaction principle for multiagent based simulation. In: AAMAS '07: Proceedings of the 6th international joint conference on Autonomous agents and multiagent systems. pp. 1 -- 3. ACM, New York, NY, USA (2007) 14. Morin, E.: Method: Towards a Study of Humankind, vol. 1. Peter Lang Pub Inc (1992) 15. Morvan, G., Jolly, D., Veremme, A., Dupont, D., Charabidze, D.: Vers une méthode de modélisation multi-niveaux. In: Actes de la 7ème Conférence de Modélisation et Simulation MOSIM, Paris, France. vol. 1, pp. 167 -- 174 (2008) 16. Morvan, G., Veremme, A., Dupont, D.: IRM4MLS: the influence reaction model for multi-level simulation. In: Multi-Agent-Based Simulation XI. No. 6532 in Lecture Notes in Artificial Intelligence, Springer (2011) 17. Morvan, G., Veremme, A., Dupont, D., Jolly, D.: Modélisation et conception multi- niveau de systèmes complexes : stratégie d'agentification des organisations. Journal Européen des Systèmes Automatisés 43, 381 -- 406 (2009) 18. Müller, J.P.: Towards a formal semantics of event-based multi-agent simulations. In: David, N., Sichman, J. (eds.) Multi-Agent-Based Simulation IX, Lecture Notes in Computer Science, vol. 5269, pp. 110 -- 126. Springer Berlin / Heidelberg (2009) 19. Müller, J.P., Ratzé, C., Gillet, F., Stoffel, K.: Modeling and simulating hierarchies using an agent-based approach. In: Proceedings of the MODSIM 2005 International Congress on Modelling and Simulation. pp. 1631 -- 1638 (2005) 20. Navarro, L., Flacher, F., Corruble, V.: Dynamic level of detail for large scale agent- based urban simulations. In: Proc. of 10th Int. Conf. on Autonomous Agents and Multiagent Systems (AAMAS 2011). pp. 701 -- 708 (2011) 21. Parunak, H.: Pheromones, probabilities and multiple futures. In: Multi-Agent- Based Simulation XI, Lecture Notes in Artificial Intelligence, vol. 6532, pp. 44 -- 60. Springer (2011) 22. Parunak, H., Brueckner, S.: Concurrent modeling of alternative worlds with polya- gents. In: Multi-Agent-Based Simulation VII, pp. 128 -- 141. Lecture Notes in Com- puter Science, Springer (2007) 23. Ratzé, C., Gillet, F., Müller, J.P., Stoffelb, K.: Simulation modelling of ecological hierarchies in constructive dynamical systems. Ecological Complexity 4(1 -- 2), 13 -- 25 (2007) 24. Sallez, Y., Berger, T., Deneux, D., Trentesaux, D.: The lifecycle of active and intelligent products: The augmentation concept. International Journal of Computer Integrated Manufacturing 23(10), 905 -- 924 (2010) 25. Sallez, Y., Berger, T., Raileanu, S., Chaabane, S., Trentesaux, D.: Semi- heterarchical control of FMS: From theory to application. Eng. Appl. Artif. Intell. 23, 1314 -- 1326 (2010) 26. Servat, D., Pierrer, E., Treuil, J., Drogoul, A.: Virtual Worlds, chap. Towards Virtual Experiment Laboratories : How Multi-Agent Simulations Can Cope with Multiple Scales of Analysis and Viewpoints, pp. 205 -- 217. No. 1434 in Lecture Notes in Computer Science, Springer-Verlag Berlin Heidelberg (1998) 27. Soyez, J.B., Morvan, G., Merzouki, R., Dupont, D., Kubiak, P.: Modélisation et simulation multi-agents multi-niveaux. Submitted to Studia Informatica Univer- salis (2011) 28. Steinhauser, M.: Computational Multiscale Modeling of Fluids and Solids Compu- tational Multiscale Modeling of Fluids and Solids. Springer (2008) 29. Stratulat, T., Ferber, J., Tranier, J.: Masq: towards an integral approach to in- teraction. In: AAMAS '09: Proceedings of The 8th International Conference on Autonomous Agents and Multiagent Systems. pp. 813 -- 820. International Founda- tion for Autonomous Agents and Multiagent Systems, Richland, SC (2009) 30. Ueda, K., Kitob, T., Fujii, N.: Modeling biological manufacturing systems with bounded-rational agents. CIRP Annals - Manufacturing Technology 55(1), 469 -- 472 (2006) 31. Ueda, K., Markusb, A., Monostorib, L., Kalsc, H., Arai, T.: Emergent synthesis methodologies for manufacturing. CIRP Annals - Manufacturing Technology 50(2), 535 -- 551 (2001) 32. Ueda, K., Vaario, J., Ohkura, K.: Modelling of biological manufacturing systems for dynamic reconfiguration. CIRP Annals - Manufacturing Technology 46(1), 343 -- 346 (1997) 33. Uhrmacher, A.M., Ewald, R., John, M., Maus, C., Jeschke, M., Biermann, S.: Combining micro and macro-modeling in devs for computational biology. In: Pro- ceedings of the 39th conference on Winter simulation: 40 years! The best is yet to come. pp. 871 -- 880. WSC '07, IEEE Press, Piscataway, NJ, USA (2007) 34. Van Brussel, H.: Holonic manufacturing systems, the vision matching the problem. In: in Proc. of First European Conf. on Holonic Manufacturing Systems (2007) 35. Van Brussel, H., Wyns, J., Valckenaers, P., Bongaerts, L., Peeters, P.: Reference ar- chitecture for holonic manufacturing systems: Prosa. Computers in Industry 37(3), 255 -- 274 (1998) 36. Weyns, D., Holvoet, T.: Multi-Agent System Technologies, chap. Model for Simul- taneous Actions in Situated Multi-agent Systems, pp. 105 -- 118. No. 2831 in Lecture Notes in Artificial Intelligence, Springer-Verlag Berlin Heidelberg (2003) 37. Weyns, D., Holvoet, T.: A formal model for situated multi-agent systems. Funda- menta Informaticae 63(2 -- 3), 125 -- 158 (2004) 38. Weyns, D., Holvoet, T., Schelfthout, K., Wielemans, J.: Decentralized control of au- tomatic guided vehicles: applying multi-agent systems in practice. In: Companion to the 23rd ACM SIGPLAN conference on Object-oriented programming systems languages and applications. pp. 663 -- 674. OOPSLA Companion '08, ACM, New York, NY, USA (2008) 39. Weyns, D., N.Boucké, Holvoet, T.: Gradient field-based task assignment in an agv transportation system. In: Proceedings of the fifth international joint conference on Autonomous agents and multiagent systems. pp. 842 -- 849. AAMAS '06, ACM, New York, NY, USA (2006) 40. Yoo, J.W., Sim, E., Cao, C., Park, J.W.: An algorithm for deadlock avoidance in an agv system. The International Journal of Advanced Manufacturing Technology 26, 659 -- 668 (2005)
1505.06012	3	1505	2016-11-07T16:39:38	The Specification of Sugarscape	[ "cs.MA" ]	Sugarscape is a well known and influential Agent Based Social Simulation (ABSS). Various parts of Sugarscape are supplied as examples in almost all Agent Based Model (ABM) toolkits. It has been used for demonstrating the applicability of different approaches to ABM. However a lack of agreement on the precise definition of the rules within Sugarscape has curtailed its usefulness. We provide a formal specification of Sugarscape using the Z specification language. This demonstrates the ability of formal specification to capture the definition of an ABM in a precise manner. It shows that formal specifications could be used as an approach to tackle the replication problem in the field of ABM. It also provides the first clear interpretation of Sugarscape identifying areas where information is missing and/or ambiguous. This enables researchers to make proper comparisons between different implementations of this model.	cs.MA	cs	THE SPECIFICATION OF SUGARSCAPE JOSEPH KEHOE Abstract. Sugarscape is a well known and influential Agent Based Social Sim- ulation (ABSS). Various parts of Sugarscape are supplied as examples in almost all Agent Based Model (ABM) toolkits. It has been used for demonstrating the applicability of different approaches to ABM. However a lack of agreement on the precise definition of the rules within Sugarscape has curtailed its useful- ness. We provide a formal specification of Sugarscape using the Z specification language. This demonstrates the ability of formal specification to capture the definition of an ABM in a precise manner. It shows that formal specifications could be used as an approach to tackle the replication problem in the field of ABM. It also provides the first clear interpretation of Sugarscape identifying areas where information is missing and/or ambiguous. This enables researchers to make proper comparisons between different implementations of this model. 1. Introduction 1.1. Overview. First we give an short informal summary of Sugarscape. We follow with a brief introduction to formal specification. We then specify the single resource simulation. Following the standard Z patterns of development we list the basic types and constants first, followed by the specification of the basic state attributes and invariant properties. Then the rules are presented in order. After this we present the extended specification, that is the specification with two resources. Here we highlight the differences between the single resource and two resource specifications by presenting the specifications in bold face whenever the specification changes from the original. 2. Sugarscape 2.1. Agent Based social Simulations. Sugarscape was the first large scale Agent Based Social Simulation (ABSS). It was developed by Epstein and Axtell and presented in their book Growing Artificial Societies [Epstein and Axtell, 1996]. The release of this simulation is considered an important event in the emerging field of Agent Based Social Simulation. The Sugarscape ABSS was used to investigate how individual behaviour can influence and cause different social dynamics within large populations. It has been used to show how, for example, inheritance of wealth affects resource distribution Date: 9/03/2015. 1 2 JOSEPH KEHOE in populations and how disease can spread through a population. It remains influ- ential today and every major simulation toolkit (Swarm, Repast, Mason and Net- Logo) [Railsback et al., 2006, Berryman, 2008, Inchiosa and Parker, 2002] comes with a partial implementation of Sugarscape that demonstrates that toolkit's ap- proach to simulation. Since Sugarscape first appeared ABSSs have been applied to fields as diverse as Anthropology[Campillo et al., 2012], Biomedical Science, Ecology, Social Science [Axtell and Axtell, 2000], Epidemic modelling and Market Analysis[Macal and North, 2009, Troitzsch, 2009, Gilbert, 2004]. ABSS's employs a bottom-up approach to modelling populations. Instead of pre- computing the overall population behaviour, as done in equation based models, individual agents and their local interactions within the population are modelled. The behaviour of the overall population is left to emerge from these local inter- actions. This approach allows us to address failings in the top-down approach and demonstrates the causal factors behind the emergence of group dynamics. In cases where we do not know what the overall behaviour will be or where we are trying to find out the causes of this behaviour, bottom-up based ABSSs are the only possible approach. 2.2. Issues with Sugarscape. Currently, social science simulations are start- ing to embrace concurrency in an effort to allow for bigger, more complete and faster implementations of ABMs. Different concurrency researchers have used the Sugarscape model as a testbed for benchmarking different approaches to parallelis- ing ABMs [Lysenko and D'Souza, 2008, Perumalla, 2006, Richmond et al., 2010]. However although the rules of Sugarscape have been defined there is no general agreement on their exact meaning. These difficulties hamper the ability of re- searchers both to properly compare their approaches, provide complete implemen- tations of Sugarscape or replicate their results. Most of the rules require some form of conflict resolution. We have specified the rules in a manner consistent with the original intention (agents acting con- currently) but independent of any particular approach to how this concurrency is implemented. That is, we have refrained from imposing any specific conflict resolution rules. By formalising Sugarscape and providing a single precisely defined reference for the rules we can produce a standard definition of Sugarscape. Compliance with this single reference will allow proper comparisons to be made between different approaches. It also leaves it open to the implementer to decide what approach to conflict resolution they wish to take. We detected ambiguities present in the cur- rent rule definitions, provided precise interpretations, where possible, and flagged irresolvable problems where not. We made the decision to restrict the initial specification to one pollution type and one resource type in an effort to guarantee clarity. While the rules were designed so that they could be extended to arbitrary numbers of resources and pollutants, explicitly specifying for an arbitrary number of resources and pollutants THE SPECIFICATION OF SUGARSCAPE 3 would make the specification even more difficult to understand and thus more likely to either contain or cause mistakes. Once we had a specification for the single resource scenario we extended the specification to a two resource situation. This allowed us to specify the final rule, T rade, as that rule requires two resources to function. This allowed for: (1) A simpler and easier to understand specification of the rules that use only one resource (trading clarity against completeness); (2) A complete (but separate) specification for simulations that use two re- sources. We do not provide specifications for multiple pollutants as multiple pollutants were never actually implemented in Sugarscape1. Similarly we did not provide a specification for more than two resources as we deem the benefits of doing so counterbalanced by both the complexity of the result- ing specification and the lack of any requirement to use such a complex simulation for benchmarking purposes. Sugarscape has only ever been implemented with two resources types, known respectively as sugar and spice. Anyone wishing to extend Sugarscape further can use the two resource specification for guidance. 2.3. Synchronous and Asynchronous Updating. Originally the rules were stated with an explicit assumption that the underlying implementation would be sequential. Concurrency was simulated through randomisation of the order of each rule application on the individual agents, and models that follow this regime are termed asynchronous. All results reported here have been produced by running the model on a serial computer; therefore only one agent is"active" at any instant. In principle the model could be run on parallel hardware, permitting agents to move simultaneously (although M would have to be supplemented with a conflict resolution rule to handle cases in which two or more agents simultaneously decide to inhabit the same site). [Footnote 12, Chapter II] The alternative to asynchronous updating is synchronous updating. Synchro- nous updating assumes that all updates occur concurrently. While it is clear that the original authors have no objection to employing synchronous updating on sug- arscape it is well known that asynchronous and synchronous updating produce different results. What is not known is how divergent these results are in the case of complex ABMs such as Sugarscape or indeed how to apply synchronous updating to all the complex interaction types in Sugarscape. 1We leave this open as an exercise for the reader. 4 JOSEPH KEHOE In order to answer these questions we present initially a specification that as- sumes a synchronous updating regime, as this is the most novel approach. Follow- ing this we give the equivalent Asynchronous updating version for comparison. 3. Single Resource Sugarscape Sugarscape is a discrete turn based simulation composed of a set of interacting agents that move across an environment. The environment, or simulation space, is modelled as a two dimensional M by M grid or matrix of discrete locations known as the lattice. This lattice is toroidal in nature, that is, it wraps around on all four edges. Every lattice location has a position denoted by its x and y coordinates. For any lattice location [i, j] there are four direct (von Neumann) neighbours (up, down, left and right) at positions [i, (j + 1)%M], [i, (j − 1)%M], [(i− 1)%M, j] and [(i + 1)%M, j]. We denote this set of von Neumann neighbours as N1(i, j), and further use Nk(i, j) to denote the set of von Neumann neighbours where each is a maximum distance of k locations from the location [i,j]. Each location can hold a number of resource and pollution types. While there is no limit placed on how many resource or pollution types can exist in a Sugarscape simulation we are unaware of any Sugarscape derived simulation that uses more than two resource types and one pollution type. When there is only one resource type it is called sugar and if there are two then the second resource is known as spice. These amounts are measured as natural numbers (≥ 0). Each individual location has limits placed on the maximum amount of resources of any type it may carry at any one time. These limits are defined at simulation startup and remain fixed during a simulation run. Agents consume the resources at their current location. Locations replenish their resources by some defined amount during each time step. Each location can also hold at most one agent at a time. Agents reside at locations within the lattice but are mobile and can change location at most once per step. At a minimum each agent has the following attributes: Metabolism Rate (one per resource type): The rate at which an agents resource stores decrease during each simulation step. Different resource types have independent metabolism rates. Once an agent runs out of re- sources it dies (is removed from the simulation); Age: The number of steps that the agent has been present in the simulation; Maximum Age: The maximum number of steps that an agent is allowed to exist during the simulation run. Once an agent reaches its maximum age it is removed from the simulation; Resource Store (one for each resource type): The amount of each re- source that the agent currently has; Vision: How far in each of the cardinal directions that the agent can see. An agent can only interact with locations and agents that are in its neigh- bourhood Nvision. To ensure locality all agent values for vision will be less THE SPECIFICATION OF SUGARSCAPE 5 than some predefined maximum and this maximum will be much smaller than the lattice dimension size (M). In the more complex versions of Sugarscape agents can also have a "culture" identifier (identifying which tribe the agent belongs to), a set of outstanding loans of resources that the agent has given to (or received from) other agents, a set of diseases that the agent has contracted and, an immunity system that gives each agent immunity from certain diseases. A simulation run consist of a series of turns or steps during which certain rules are applied to each location and agent. Each rule is applied concurrently and instantaneously to each agent and/or location. The rules are generally fairly simple and the only information that an agent (or location) can use when deciding how to apply a rule is local information, that is an agent or location at position [i,j] can only access information from locations and/or agents that are within the set Nk, where k≤ vision (in most cases k=1). The rules for locations decide how resources are replenished and how pollution is created or spread. The rules for agents are more varied and determine agent movement and interaction. Agent interaction can range from spreading disease, trading, entering financial agreements and even combat. There are a large number of rules but not all rules need to be (or indeed can be) applied in the same simula- tion run. The rules are chosen based on what we wish to model. A simulation that wishes to see the effect of trading on wealth distribution would have no need for the combat or culture rules while one modelling disease transmission would only require the movement and disease transmission rules. 3.1. Basic Types and Constants. First we identify the basic types and any required constants. Many are self explanatory or will become clear when their associated rules are specified. A simulation is defined by the values given to these constants and the combination of rules employed. 6 JOSEPH KEHOE M : N1 (1) CULT URECOUNT : N1 (2) MAXV ISION : N1 (3) MINMET ABOLISM, MAXMET ABOLISM : N (4) (5) SUGARGROW T H : N1 MAXAGE, MINAGE : N1 (6) MAXSUGAR : N1 (7) (8a) DURAT ION : N1 RAT E : A (8b) INIT IALSUGARMIN, INIT IALSUGARMAX : N(9) (10) W INT ERRAT E, SEASONLENGT H : N1 (11) P RODUCT ION, CONSUMP T ION : N COMBAT LIMIT : N (12) (13) IMMUNIT Y LENGT H : N (14) INIT IALP OP ULAT IONSIZE : N (15) P OLLUT IONRAT E : N CHILDAMT : N (16) CULT URECOUNT mod 2 = 1 MINMET ABOLISM < MAXMET ABOLISM MAXAGE < MINAGE MAXV ISION < M INIT IALSUGARMIN < INIT IALSUGARMAX INIT IALP OP ULAT IONSIZE ≤ M ∗ M (1) The simulation space is represented by a two dimensional M by M matrix of locations. Each location in the simulation space is referenced by two indices representing its position in this matrix; (2) CULT URECOUNT determines the size of the bit sequence used to rep- resent cultural allegiances. This is always equal to an odd number so that the number of 1's in the sequence is never equal to the number of 0's; (3) Agents can only "see" in the four cardinal directions, that is the locations to the north, south, east and west. Agents are endowed with a random vision strength that indicates how many locations the can "see" in each direction. This endowment is always less than MAXV ISION and MAXV ISION is always less than M; (4) Agents consume an amount of sugar (resources) during each turn. This sugar represents the amount of energy required to live. Each agent is en- dowed, on creation, with a random metabolism between MINMET ABOLISM and MAXMET ABOLISM; (5) Agents consume sugar (resources) from the location they occupy. Each location can renew its sugar at a rate determined by SUGARGROW T H. THE SPECIFICATION OF SUGARSCAPE 7 After each turn up to a maximum of SUGARGROW T H units of sugar are added to each location (in accordance with the Growback rule); (6) MAXAGE and MINAGE are, respectively, the maximum and minimum allowable lifespan for any agent; (7) MAXSUGAR is the maximum amount of sugar that any location can possibly hold. This is known as the carrying capacity of a location; (8) RAT E and DURAT ION are used for determining the rate of interest charged for loans and the duration of a loan; (9) INIT IALSUGARMIN and INIT IALSUGARMAX are the lower and upper limits for initial endowment of sugar given to a newly created agent; (10) If seasons are enabled then two seasons, winter and summer are allowed with a duration of SEASONLENGT H turns (ticks) and a new separate lower seasonal grow back rate calculated using W INT ERRAT E (as de- termined by the SeasonalGrowback rule); (11) Pollution can occur at a rate determined by the production and consump- tion of resources determined by the P RODUCT ION and CONSUMP T ION constants respectively; (12) The combat rule posits the maximum reward COMBAT LIMIT that can be given to an agent through killing another agent; (13) Immunity in agents is represented using a fixed size sequence of bits of length IMMUNIT Y LENGT H; (14) We have some predetermined initial population size INIT IALP OP ULAT IONSIZE that is used to initialise the simulation; (15) P OLLUT IONRAT E determines the number of steps that elapse before pollution levels diffuse to their neighbours; (16) A certain amount of sugar reserves, CHILDAMT , are required for an agent to have children. [AGENT ] (1) P OSIT ION == 0 . . M − 1 × 0 . . M − 1 (2) (3) SEX ::= male f emale (4) BIT ::= 0 1 af f iliation ::= red blue (5) boolean ::= true f alse (1) AGENT is used as a unique identifier for agents; (2) P OSIT ION is also used to make specifying indices within the grid so as to make the schemas easier to read and more compact; (3) All agents have a sex attribute; (4) BIT s are used to encode both culture preferences and diseases of agents; 8 JOSEPH KEHOE (5) Every agent has a cultural affiliation of either belonging to the blue tribe or red tribe. Agents can, using the Mating rule, have offspring if they are fertile. Fertility is determined by the age of the agent, where fertility starts at some predefined age and ends at another. These boundaries are defined for all agents. The numbers are set out by Epstein and Axtelland although there appears to be no special sig- nificance attached to these numbers we will stick with the originals. Male fertility ends 10 turns later than female fertility. F EMALEF ERT ILIT Y ST ART, F EMALEF ERT ILIT Y END : N MALEF ERT ILIT Y ST ART, MALEF ERT ILIT Y END : N 12 ≤ F EMALEF ERT ILIT Y ST ART ≤ 15 40 ≤ F EMALEF ERT ILIT Y END ≤ 50 12 ≤ MALEF ERT ILIT Y ST ART ≤ 15 50 ≤ MALEF ERT ILIT Y END ≤ 60 MALEF ERT ILIT Y END = F EMALEF ERT ILIT Y END + 10 The replacement rule requires a sugar allocation be given to new agents set between 5 and 25. Again there appears to be no special significance attached to these numbers. ST ART SUGARMIN, ST ART SUGARMAX : N ST ART SUGARMIN = 5 ST ART SUGARMAX = 25 3.2. The Sugarspace Lattice. The simulation space in Sugarscape consists of a finite discrete two-dimensional array of locations. Each location is identified its row and column value. Each location contains a number of resources. While only two resources are ever used it is clear that the intention of the original authors was that the simulation could be extended so that any number of different resources can be present. Similarly each location can contain a number of pollutant levels. In practice, although the rule is explicitly defined for an arbitrary number of pollution types only one is ever used. Again, in line with actual Sugarscape usage and to make the specification more readable we assume only one pollution type. Pollution fluxes are used in the rules to help calculate how pollution levels change over time. Although explicitly referenced in the Pollution rule these do not need to be explicitly modelled in the specification. THE SPECIFICATION OF SUGARSCAPE 9 Lattice sugar : P OSIT ION 7→ N maxSugar : P OSIT ION 7→ N pollution : P OSIT ION 7→ N dom sugar = dom maxSugar = dom pollution = P OSIT ION ∀ x : P OSIT ION • sugar(x) ≤ maxSugar(x) ≤ MAXSUGAR (1) sugar is a mapping that stores the amount of sugar stored at each position; (2) maxSugar is a mapping that records the maximum amount of sugar that (1) (2) (3) (4) (5) can be stored (carried) in each position; (3) pollution records the amount of pollution at each location; (4) Every position has a sugar level, a maximum allowed sugar level (or carry- ing load) and a pollution level; (5) Every position's sugar level is less than or equal to the maximum al- lowed amount for that position which is in turn less than or equal to the MAXSUGAR constant; We need to track the number of turns that have occurred in the simulation. Each turn consists of the application of all rules that form part of the simulation. Step step : N 3.3. Agents. Every agent is situated on a location within the grid and each lo- cation is capable of containing only one agent at a time (putting an upper limit on the number of possible agents). Agents are mobile, that is they can move to a new location if a suitable unoccupied location is available. Movement is both discrete and instantaneous, it is possible for an agent to move to a new location instantly while skipping over all intermediate locations. The attributes that every agent has are: Vision: How far in the four cardinal directions that an agent can see; Age: Number of turns of the simulation that an agent has been alive; Maximum Age: Age at which an agent dies; Sex: Agents are either male or female; Sugar Level: The amount of sugar that an agent currently holds. There is no limit to how much sugar an agent can hold; Initial Sugar: The amount of sugar the agent was initialised with on cre- ation; Metabolism: The amount of energy, defined by sugar (or resource) con- sumption, used during every turn of the simulation; Culture Tags: A sequence of bits that represents the culture of an agent; Children: For each agent we track its children (if any). To apply the In- heritance rule the full list of an agents children is required. 10 JOSEPH KEHOE Loans: Under the credit rule agents are allowed lend and/or borrow sugar for set durations and interest rates so we need to track these loans. For each loan we need to know the lender, the borrower, the loan principal and the due date (represented as the step number); Diseases: Diseases are sequences of bits that can be passed between agents. An agent may carry more than one disease; Immunity: Each agent has an associated bit sequence that confers immu- nity against certain diseases. If the bit sequence representing a disease is a subsequence of an agents immunity bit sequence then that agent is considered immune to that disease. THE SPECIFICATION OF SUGARSCAPE 11 Agents population : P AGENT position : AGENT ֌ P OSIT ION sex : AGENT 7→ SEX vision : AGENT 7→ N1 age : AGENT 7→ N maxAge : AGENT 7→ N1 metabolism : AGENT 7→ N agentSugar : AGENT 7→ N initialSugar : AGENT 7→ N agentCulture : AGENT 7→ seq BIT children : AGENT 7→ P AGENT loanBook : AGENT ↔ (AGENT × (N, N)) agentImmunity : AGENT 7→ seq BIT diseases : AGENT 7→ P seq BIT population = dom position = dom sex = dom vision = dom maxAge = dom agentSugar = dom children = dom agentCulture = dom metabolism = dom age = dom agentImmunity = dom diseases dom loanBook ⊆ population dom(ran loanBook) ⊆ population ∀ x, y : AGENT ; d : seq BIT • x, y ∈ population ∧ x 6= y ⇒ (1) (2) (3) (4) ((age(x) ≤ maxAge(x) ∧ MINAGE ≤ maxAge(x) ≤ MAXAGE ∧ # agentCulture(x) = CULT URECOUNT ∧ # agentImmunity(x) = IMMUNIT Y LENGT H ∧ vision(x) ≤ MAXV ISION ∧ MINMET ABOLISM ≤ metabolism(x) ≤ MAXMET ABOLISM ∧ position(x) = position(y) ⇔ x = y) d ∈ ran diseases(x) ⇒ # d < IMUNIT Y LENGT H (1) Every existing agent has an associated age, sex, vision, etc. Note that the population holds only the currently existing agent IDs; (2) Only current members of the population can be lenders; (3) Only current members of the population can be borrowers (4) Every agent in the population is guaranteed to have a current age less than the maximum allowed age for that agent, a maximum age less than or equal to the global MAXAGE, a metabolism between the allowed limits and vision less than or equal to the maximum vision. The sequence of bits representing its culture tags is CULT URECOUNT in size while those representing immunity is IMMUNIT Y LENGT H in size. All diseases 12 JOSEPH KEHOE are represented by sequences of bits that are shorter than the immunity sequence. The entire simulation consists of locations, agents and a counter holding the tick count. We combine them all in the schema SugarScape. SugarScape Agents Lattice Step The initial state of the schema when the simulation begins must also be stated. InitialSugarScape Sugarscape′ step′ = 0 # population′ = INIT IALP OP ULAT IONSIZE loanBook′ = ∅ ∀ a : AGENT • a ∈ population′ ⇒ (1) (2) (3) (4) (age(a) = 0 ∧ diseases′(a) = ∅ ∧ children′(a) = ∅ ∧ INIT IALSUGARMIN ≤ agentSugar′(a) ≤ INIT IALSUGARMAX) ∧ initialSugar′(a) = agentSugar′(a) (1) step is set to zero; (2) The population is set to some initial size; (3) There are no loans as yet; (4) Every agent in the starting population has an age of zero, no diseases or children and some initial sugar level within the agreed limits. The other attributes have random values restricted only by the invariants; 3.4. Rules. There are a number of rules that can be employed in different com- binations to give different simulations. We will quote each rule as laid out in the appendix of [Epstein and Axtell, 1996] and follow, where necessary, with a more detailed explanation of the rule. In many cases the simple rule definitions are not complete. Extra information, embedded in the original text, has been extracted where necessary to help complete these rules. The majority of rule definitions assume only one resource (sugar) and it is these that are specified in this section. The simulation is discrete with each time interval representing one complete set of rule applications. We use the step variable in the SugarScape schema to keep track of the current time interval number. Where there exist ambiguities in the rule definitions we will identify them and propose one or more possible interpretations consistent with what we believe to THE SPECIFICATION OF SUGARSCAPE 13 be the authors intentions. Throughout the rule definitions constants such as α, β are used but they have different meanings in each rule. For the sake of clarity we will give each constant a meaningful and globally unique name. 3.5. Tracking Steps. While not defined explicitly as a rule, we must ensure that we record the current step number. We increment the Step variable before every sequence of rule applications that compose a single turn of the simulation. There is an issue with metabolism in that every turn of the simulation requires that agents use up their sugar reserves at a rate determined by their metabolism. It is not explicitly stated when or where this sugar deduction occurs within the rules. It could be placed, for example, in the movement rule but it can also be placed, just as validly, within any rule that is guaranteed to be applied during every turn. Since there is no obvious reason why one is superior to the other, as long as it is consistently applied, we choose to place the metabolism deduction within the T ick schema. This new rule can be stated simply as follows: Tick: At the start of every time interval increase every agents age by one and decrease every agents sugar level by their metabolism rate. T ick ∆Agents ∆Step population′ = population position′ = position sex′ = sex vision′ = vision maxAge′ = maxAge metabolism′ = metabolism initialSugar′ = initialSugar agentCulture′ = agentCulture children′ = children loanBook′ = loanBook agentImmunity ′ = agentImmunity diseases′ = diseases step′ = step + 1 ∀ x : AGENT • x ∈ population ⇒ age′(x) = age(x) + 1 ∧ agentSugar′(x) = agentSugar(x) − metabolism(x) (1) Add one to the step count; (2) Increase everyone's age by one; (3) Decrease everyone's agentSugar by their metabolism. (1) (2) (3) 14 JOSEPH KEHOE 3.6. Sugarscape Growbackα. Sugarscape Growbackα: At each Lattice position, sugar grows back at a rate of α units per time interval up to the capacity at that position. Growback determines the rate at which location resources are replenished. The integer constant α indicates the amount by which resources grow during a single step or time interval. If α = ∞ then each resource returns to its maximum value during each turn, i.e. it is instantly fully replenished after each step. The rule only refers to a single resource, sugar, but the book explicitly defines one other resource spice and it is clear that generalisations allowing an arbitrary number of resource types to be held at each Lattice position are acceptable. Since we are dealing only with one resource, sugar, we only need to define α for this resource . The constant SUGARGROW T H represents alpha in this rule and we use this to update the sugar level of each position. Since the maximum carrying level of each resource cannot be exceeded we will set the resource levels to its maximum value if application of the replenishment rate would result in a value greater than this maximum. With these definitions we can express the Growback rule in a simple manner. The last line in the schema (see below) does the work of updating the resource levels of every location. Growback ∆Lattice pollution′ = pollution maxSugar′ = maxSugar sugar′ = {x : P OSIT ION • x 7→ min({sugar(x) + SUGARGROW T H, maxSugar(x)})} (1) (1) The new sugar levels are calculated using a simple formula to either, the maximum possible level for that location or the old level plus the SUGARGROW T H whichever is the smaller. 3.7. Seasonal Growback Sα,β,γ. Seasonal Growback Sα,β,γ: Initially it is summer in the top, half of the Sugarscape and winter in the bottom half. Then every γ time periods the seasons flip - in the region where it was summer it becomes winter and vice versa. For each site, if the season is summer then sugar grows back at a rate of α units per time interval; if the season is winter then the grow back rate is α units per β time intervals. Seasonal growback is an alternative to the previous grow back rule. Which rule is chosen will depend on what the simulation is trying to demonstrate. Seasonal grow back allow us to introduce seasonal factors into the original Growback rule. There are two seasons (representing summer and winter) and each lasts γ turns THE SPECIFICATION OF SUGARSCAPE 15 before switching. We rename γ to SEASONLENGT H. α is the summer season SUGARGROW T H rate and α/β is the winter season rate. We use the existing SUGARGROW T H to hold the summer rate and introduce W INT ERRAT E as β. Determining what season it is during a turn is fairly trivial. When seasonLength divides into the Step variable evenly it is summer in the top half and winter in the bottom half (and vice versa). SeasonalGrowback ∆Lattice ΞStep pollution′ = pollution maxSugar′ = maxSugar ∀ x : P OSIT ION • (step div SEASONLENGT H) mod 2 = 0 ⇒ sugar′ = (step div SEASONLENGT H) mod 2 6= 0 ⇒ sugar′ = {x : P OSIT ION f irst(x) < M div 2 • x 7→ min({sugar(x) + SUGARGROW T H, maxSugar(x)})} ∪ {x : P OSIT ION f irst(x) ≥ M div 2 • x 7→ min({sugar(x) + SUGARGROW T H div W INT ERRAT E, maxSugar(x)})} {x : P OSIT ION f irst(x) < M div 2 • x 7→ min({sugar(x) + SUGARGROW T H div W INT ERRAT E, maxSugar(x)})}(2a) ∪ {x : P OSIT ION; y : N f irst(x) ≥ M div 2 • x 7→ min({sugar(x) + SUGARGROW T H, maxSugar(x)})} (1a) (2b) (1) (1b) (2) (1) If the season is summer then: a) Top half of grid is updated as normal; b) Bottom half is updated at winter rate. (2) Otherwise if it is winter: a) Top half of grid is updated at winter rate; b) Bottom half is updated as normal. 3.8. Movement - M. Movement - M: north, south, east and west; • Look out as far as vision permits in each of the four lattice directions, • Considering only unoccupied lattice positions, find the nearest position • Move to the new position • Collect all resources at that location producing maximum welfare; 16 JOSEPH KEHOE The previous rules affected only the locations but the remaining rules affect agents as well as locations. The Movement rule determines how agents select their next location. There are a number of different versions of this rule. We will specify the simplest rule first as it is the only movement rule explicitly defined in the appendix but we will also specify the other movement rules defined in the text. We add a subscript to the rule title (Mbasic) to distinguish between the different movement rule specifications. Not explicitly stated within the rule but stated as a footnote to the rule is the restriction that the order in which the lattice directions are searched should be random. This comes into play when two or more available sites exist with the same welfare score. This rule does not guarantee that an agent will move to the best location. To see why this is the case consider what happens if two agents both try to move to the same location. Only one can succeed and the other will have to move to a less advantageous location. How we decide which agent succeeds is not defined. We assume that either a conflict resolution or conflict avoidance rule is available to make this decision but it is not stated what this rule should be. The original implementation is sequential with agents assumed to be moving in a random order thus enforcing collision avoidance. No guidance is provided for concurrent implementations. To help make the specification clear we define some simple helper functions. The distance between two positions is only defined for positions that are directly horizontal or vertical to each other. This function must take into account the torus-like (wrap around) structure of the simulation. distance : P OSIT ION ×P OSIT ION →N ∀ x1, x2, y1, y2 : N • distance((x1, y1), (x1, y2)) = distance((x1, y1), (x2, y2)) = (1) min({ y2 − y1 , M− y2 − y1 }) (2) min({ x1 − x2 , M− x1 − x2 }) (3) x1 6= x2 ∧ y1 6= y2 distance((x1, y1), (x2, y2)) = ∞ ⇔ (1) If two agents are vertically aligned we calculate distance based on the horizontal distance; (2) If two agents are horizontally aligned we calculate distance based on the vertical distance; (3) Otherwise the distance is defined as infinity. THE SPECIFICATION OF SUGARSCAPE 17 We use this to define the adjacent function that lets us know if two agents are directly beside each other. adjacent : P OSIT ION ×P OSIT ION →boolean ∀ a, b : P OSIT ION • adjacent(a, b) ⇔ distance(a, b) = 1 visibleAgents takes an agent, a function mapping agents to positions and the vision range of the agent and returns the set of agents that are within that agent's neighbourhood. visibleAgents : AGENT ×(AGENT 7֌ P OSIT ION) ×N 7→ F AGENT ∀ agent : AGENT ; pos : AGENT 7֌ P OSIT ION; range : N • visibleAgents(agent, pos, range) = {ag : AGENT ag ∈ dom pos ∧ 1 ≤ distance(pos(ag), pos(agent)) ≤ range} 18 JOSEPH KEHOE Movementbasic ∆SugarScape step′ = step population′ = population maxSugar′ = maxSugar pollution′ = pollution sex′ = sex vision′ = vision age′ = age maxAge′ = maxAge agentCulture′ = agentCulture loanBook′ = loanBook diseases′ = diseases agentImmunity ′ = agentImmunity children′ = children metabolism′ = metabolism initialSugar′ = initialSugar ∀ a : AGENT ; l : P OSIT ION • a ∈ population′ ⇒ distance(position′(a), position(a)) ≤ vision(a) (distance(position(a), l) ≤ vision(a) ∧ (l 6∈ ran position′)) ⇒ sugar(l) ≤ sugar(position′(a)) ∧ (distance(l, position(a)) < distance(position′(a), position(a))) ⇒ sugar(l) < sugar(position′(a)) agentSugar′ = {∀ a : AGENT a ∈ population′ • sugar′ = sugar ⊕ {∀ l : P OSIT ION l ∈ ran position′ • l 7→ 0} a 7→ agentSugar(a) + sugar(position′(a))} (1) (2) (2a) (2b) (3) (4) After the rule is applied the following will be the case for every agent: (1) They will be located within one of the locations in their original neighbour- hood (possibly the same position as before); (2) After every agent has moved: a) There will exist no remaining available locations from the original neigh- bourhood of an agent that would have given a better welfare score than the location that agent now inhabits (we picked the maximum reward); b) If there was more than one location with maximum reward then the agent moved to the closest location. (3) Agent sugar levels increase because they consume all the sugar at their new location (even if the new location is the same as their old location); THE SPECIFICATION OF SUGARSCAPE 19 (4) Location sugar levels are set to zero everywhere there is an agent present. The specification states what is true after the application of the rule but not how we achieve that state. In any implementation some conflict resolution strategy will be needed but in this specification we remain agnostic as to what it should be. The rule is well stated but requires that we precisely define welf are. For a single resource simulation welfare is precisely equal to the amount of sugar available at a location. We will define welfare for multiple resource simulations later. 3.9. Pollution Formation PΠ,χ. Pollution Formation Pα,β: When sugar quantity s is gathered from the Sugarscape, an amount of production pollution is gathered in quantity αs. When sugar amount m is consumed (metabolised), consumption pollution is generated according to βm. The total pollution on a site at time t, pt, is the sum of the pollution present at the previous time, plus the pollution resulting from production and consumption activities, that is, pt = pt−1 + αs + βm. This single resource pollution rule is easiest to understand and the most common form of the pollution rule. When pollution is incorporated into the Sugarscape the movement rule is changed so that the welfare of a location is now defined using the sugar to pollution ratio - the greater the ratio the greater the welfare. This ratio is defined as sugar/(1 + pollution) where the "plus one" prevents division by zero. As the pollution rule requires that we know both the sugar consumed and sugar metabolised during the last move of an agent to that location it is simpler to incorporate the P ollutionF ormation rule into the movement rule. The alternative is to track the sugar consumed during each move which would require another attribute defined in the Agent schema. 20 JOSEPH KEHOE Movementpollution ∆SugarScape step′ = step maxSugar′ = maxSugar sex′ = sex population′ = population vision′ = vision age′ = age maxAge′ = maxAge agentCulture′ = agentCulture loanBook′ = loanBook children′ = children agentImmunity ′ = agentImmunity diseases′ = diseases metabolism′ = metabolism initialSugar′ = initialSugar ∀ a : AGENT ; l : P OSIT ION • a ∈ population′ ⇒ distance(position′(a), position(a)) ≤ vision(a) (distance(position(a), l) ≤ vision(a) ∧ (l 6∈ ran position′)) ⇒ [sugar(l)/(1 + pollution(l))] ≤ [sugar(position′(a))/(1 + pollution(position′(a)))] ∧ (distance(l, position(a)) < distance(position′(a), position(a))) (1) ⇒ sugar(l)/(1 + pollution(l)) < sugar(position′(a))/(1 + pollution(position′(a))) sugar′ = sugar ⊕ {∀ l : P OSIT ION l ∈ ran position′ • l 7→ 0} agentSugar′ = {∀ a : AGENT a ∈ population • a 7→ agentSugar(a) + sugar(position′(a))} pollution′ = pollution ⊕ {∀ l : P OSIT ION; x : AGENT position′(x) = l • l 7→ (P RODUCT ION ∗ sugar(l) + CONSUMP T ION ∗ metabolism(x))}(2) (1) We use our new formula to calculate the desirability of a location; (2) The new pollution value for any location that an agent is present at is calculated as per rule definition. The rule as stated in the appendix is the generalised rule defined for an arbitrary number of pollutants and resources. We have specified the simpler version as it is easier to grasp. The more complex version has not been used in any of the Sugarscape simulations. We state the generalised rule below for completeness but do not present a formal specification of it. THE SPECIFICATION OF SUGARSCAPE 21 Pollution Formation PΠ,χ: For n resources and m pollutants, when n- dimensional resource vector r is gathered from the Sugarscape the m- dimensional pollution production vector p is produced according to p = Πr, where Π is an m× n matrix; when n-dimensional (metabolism) vector m is consumed then m-dimensional consumption pollution vector c is produced according to c = χm, where χ is an m × n matrix. 3.10. Pollution Diffusion Dα. Pollution Diffusion Dα: • Each α time periods and at each site, compute the pollution flux the average pollution level over all its von Neumann neighbouring sites; • Each site's flux becomes its new pollution level. This rule determines how pollution diffuses over grid. Pollution diffusion is calculated every α turns and is computed as the average pollution level of all the locations von Neumann neighbours. We use the constant P OLLUT IONRAT E in place of alpha. The von Neumann neighbours of a location are those immediately above, below, left and right of the current locations (aka north, south, east and west). We define the four cardinal directions taking into account to fact that the grid wraps around at its edges (i.e. it is a torus). north : P OSIT ION ֌→ P OSIT ION south : P OSIT ION ֌→ P OSIT ION east : P OSIT ION ֌→ P OSIT ION west : P OSIT ION ֌→ P OSIT ION ∀ x, y : N • west((x, y)) = ((x − 1) mod M, y) east((x, y)) = ((x + 1) mod M, y) south((x, y)) = (x, (y − 1) mod M) north((x, y)) = (x, (y + 1) mod M) We use this to define a function that returns true if two agents are von Neumann neighbours. It takes as parameters the two agents and a function that maps each agent onto their location in the simulation. 22 JOSEPH KEHOE vonNeumanNeighbour : (AGENT × AGENT × (AGENT ֌ P OSIT ION)) → boolean ∀ a, b : AGENT ; position : AGENT ֌ P OSIT ION • vonNeumanNeighbour(a, b, position) ⇔ position(a) = north(position(b)) ∨ position(a) = south(position(b)) ∨ position(a) = east(position(b)) ∨ position(a) = west(position(b)) P ollutionDif f usion ∆Lattice ΞStep maxSugar′ = maxSugar sugar′ = sugar (step mod P OLLUT IONRAT E 6= 0) ⇒ pollution′ = pollution (step mod P OLLUT IONRAT E = 0) ⇒ pollution′ = {∀ l : P OSIT ION • l 7→ (pollution(north(l)) + pollution(south(l)) +pollution(east(l)) + pollution(west(l))) div 4} The P ollution Dif f usion Rule can be simplified slightly, by removing the re- dundant mention of f lux. Pollution Diffusion Dα: After every α time periods and at each location, the average pollution level over all a site's von Neumann neighbouring locations becomes its new pollution level. 3.11. Replacement - R[a,b]. Replacement - R[a,b]: When an agent dies it is replaced by an agent of age 0 having random genetic attributes, random position on the Sugarscape, random initial endowment, and a maximum age selected from the range [a,b]. The two constants a and b we have defined already as LOW ERAGELIMIT and UP P ERAGELIMIT and we assume that the range is inclusive. It is not stated whether the new agents immediately consume the resources at the location they are placed in. We assume they do not, but accept that the alternative interpretation is equally valid. Although not part of the rule definition in the appendix it is stated elsewhere in the book that new agents will have initial resource levels set between 5 and 25. We have defined ST ART SUGARMIN and ST ART SUGARMAX for this purpose. Although the simulation can be run without employing the replacement rule (in an effort, for example, to determine the total carrying load - maximum tolerable THE SPECIFICATION OF SUGARSCAPE 23 population of agents - of a simulation space) there is no stated separate death rule. We will first add a schema that defines "death" explicitly to ensure consistency. Death : When an agent reaches its maximum allowed age or runs out of resources it is removed from the simulation and all its associated loans (either as borrower or lender) are considered void. Death ∆Agents population′ = population\ loanBook′ = population′ {a : AGENT age(a) = maxAge(a) ∨ agentSugar(a) = 0} (1) (2) ⊳ loanBook ⊲ {x : AGENT × (N × N) f irst(x) ∈ population′} (3) ∀ a : AGENT • a ∈ population′ ⇒ (sex(a) = sex′(a) ∧ vision(a) = vision′(a) ∧ maxAge(a) = maxAge′(a) ∧ agentCulture(a) = agentCulture′(a) ∧ position(a) = position′(a) ∧ age(a) = age′(a) ∧ agentSugar(a) = agentSugar′(a) ∧ metabolism′(a) = metabolism(a) ∧ diseases′(a) = diseases(a) ∧ agentImmunity ′(a) = agentImmunity(a) ∧ children′(a) = children(a)) ∧ initialSugar′(a) = initialSugar(a) (1) We remove from the population all agents who have reached their maximum age or who have no sugar reserves; (2) We remove all loans owed by or owing to these dying agents; (3) Any agent not being removed still has all attributes completely unchanged. The replacement rule follows readily from this rule, the only addition being the generation of new agents to replace the agents being removed. In effect we have broken the replacement rule into two parts Death followed by Replacement; although the Death rule may be used in isolation the Replacement rule must always be preceded by the application of the Death rule. 24 JOSEPH KEHOE Replacement ∆Agents # population′ = INIT IALP OP ULAT IONSIZE loanBook′ = loanBook ∀ a : AGENT • a ∈ population ⇒ (a ∈ population′ ∧ sex(a) = sex′(a) ∧ vision(a) = vision′(a) ∧ maxAge(a) = maxAge′(a) ∧ agentCulture(a) = agentCulture′(a) ∧ position(a) = position′(a) ∧ age(a) = age′(a) ∧ agentSugar′(a) = agentSugar(a) ∧ metabolism′(a) = metabolism(a) ∧ diseases′(a) = diseases(a) ∧ agentImmunity ′(a) = agentImmunity(a) ∧ children′(a) = children(a)) ∧ initialSugar′(a) = initialSugar(a) (age′(a) = 0 ∧ ST ART SUGARMIN ≤ agentSugar′(a) ≤ ST ART SUGARMAX ∧ initialSugar′(a) = agentSugar′(a) ∧ diseases′(a) = ∅ ∧ children′(a) = ∅) a ∈ population′ \ population ⇒ (1) (2) (3) (1) The new population has the correct number of members; (2) The existing agents remain unchanged and part of the new population; (3) All new agents have new values initialised within the allowed limits (those not stated explicitly are random values within the ranges set by the speci- fication invariants. We do not state the positions of any new agents because they are chosen randomly. Our schema invariants ensure that they are on the grid in a location not occupied by any other agent. We need to add some extra information to this rule definition to ensure that: (1) Newly created agents have no diseases, children or loans; (2) Their initial endowment of resources is within a set range. 3.12. Agent Mating S. Agent Mating S: • Select a neighboring agent at random; • If the neighboring agent is of the opposite sex and if both agents are fertile and at least one of the agents has an empty neighboring site then a newborn is produced by crossing over the parents' genetic and cultural characteristics; THE SPECIFICATION OF SUGARSCAPE 25 • Repeat for all neighbors. This rule determines how mating takes place amongst agents to produce off- spring. An agent is fertile if its age is within preset boundaries. This is represented by the simple isF ertile function below. isF ertile : (N × SEX) ֌ boolean ∀ age : N • isF ertile(age, male) ⇔ MALEF ERT ILIT Y ST ART ≤ age ≤ MALEF ERT ILIT Y END isF ertile(age, f emale) ⇔ F EMALEF ERT ILIT Y ST ART ≤ age ≤ F EMALEF ERT ILIT Y END We define two functions that take in an agent and a mapping from parents to offspring and returns the father or mother of the agent. f ather : AGENT × ((AGENT × AGENT ) 7֌ AGENT ) 7֌ AGENT mother : AGENT × ((AGENT × AGENT ) 7֌ AGENT ) 7֌ AGENT ∀ x, m, f : AGENT ; Of f spring : (AGENT × AGENT ) 7֌ AGENT • f ather(x, Of f spring) = m ⇔ Of f spring((m, f )) = x mother(x, Of f spring) = f ⇔ Of f spring((m, f )) = x The issues encountered with the mating rule are similar to those with movement. If two sets of parent try to produce offspring in the same vacant location only one can succeed. As there is no preferred conflict resolution rule we cannot state any preference for which agents succeed in producing children and which do not. All we can state is that the maximum number of offspring will be produced given the space constraints but we cannot always be sure which offspring make it into this set. Neighbours in this rule refers to von Neumann Neighbours only. Mating although proceeding concurrently throughout the population is an ex- clusive event. That is, if agent A is mating with agent B then A cannot be mating with any other agent at the same time: you can only ate with one partner at a time. The rule itself specifies that each agent will mate with all available partners so the execution of the rule can involve a sequence of mating events for specific agents. Although it is not stated in the rule definition the accompanying book mentions that each parent should gift half of its sugar to its offspring and will only mate if it has a sugar level equal to or greater than its initial sugar level (that is its sugar level on creation). This significantly complicates the rule and dramatically changes its definition and characteristics. However we will assume that this information was inadvertently omitted from the rule definition as the rule makes more sense if we include these extra factors. 26 JOSEPH KEHOE Since each individual agent can involve itself in a sequence of up to four mating events during rule execution we require a specification that retains global concur- rency while still imposing a sequential ordering based on these constraints. We do this by collecting all possible potential mating partners into a set and then di- viding this set into a sequence of maximally sized sets where each subset contains only mating events that can occur concurrently. These sets are produced using a conflict resolution rule that ensures that only pairing that can occur simultane- ously appear within each such subset. The rule then proceeds by executing mating events within each subset concurrently while the sets are evaluated in sequence. AgentMating ΞLattice ∆Agents loanBook′ = loanBook ∃ potentialMatingP airs : P(AGENT × AGENT ) potentialMatingP airs = {(a : AGENT, b : AGENT ) sex(a) 6= sex(b) ∧ isF ertile(age(a), sex(a)) ∧ isF ertile(age(head), sex(head)) ∧ adjacent(position(a), position(head))} , children′, diseases′, agentImmunity ′, age′, sex′, initialSugar′) = (population′, position′, vision′, agentSugar′, agentCulture′, metabolism′ concurrentMating(getConf ictF reeP airs(potentialMatingP airs), population, position, vision, agentSugar, agentCulture, metabolism, children, diseases, agentImmunity, age, maxAge, sex, initialSugar) (1) Generate the set of all possible mating pairs; (2) Recursively proceed with concurrent mating within the conflict free subsets. THE SPECIFICATION OF SUGARSCAPE 27 getConf ictF reeP airs : P(AGENT × AGENT ) →seq(P(AGENT × AGENT )) ∀ AllP airs : P(AGENT × AGENT ); a : AGENT ; • ∃ conf lictF reeSet : P(AGENT × AGENT ) conf lictF reeSet ⊆ AllP airs ∧ a ∈ ran conf lictF reeSet ⇒ a 6∈ dom conf lictF reeSet ∧ a ∈ dom conf lictF reeSet ⇒ a 6∈ ran conf lictF reeSet ∀ otherSet : P(AGENT × AGENT ) otherSet ⊆ AllP airs ∧ a ∈ dom otherSet ⇒ a 6∈ ran AllP airs ∧ a ∈ ran otherSet ⇒ a 6∈ dom AllP airs • # otherSet ≤ # conf listF reeSet getConf ictF reeP airs(∅) = ∅ getConf ictF reeP airs(AllP airs) = (3) hconf lictF reeSeti a getConf ictF reeP airs(AllP airs \ conf lictF reeSet) (1) (2) (1) Generate a collision free (conflict resolved) set where each agent can only once within the set; (2) Ensure this set is as large as possible; (3) Recurse through the remaining pairs dividing them into more conflict free sets. 28 JOSEPH KEHOE concurrentMating : seq P(AGENT × AGENT ) × P AGENT ×AGENT ֌ P OSIT ION ×AGENT 7→ N1 ×AGENT 7→ N ×AGENT 7→ seq BIT ×AGENT 7→ N ×AGENT 7→ P AGENT ×AGENT 7→ P seq BIT ×AGENT 7→ seq BIT ×AGENT 7→ N ×AGENT 7→ N1 ×AGENT 7→ SEX ×AGENT 7→ N ↔P AGENT ×AGENT ֌ P OSIT ION ×AGENT 7→ N1 ×AGENT 7→ N ×AGENT 7→ seq BIT ×AGENT 7→ N ×AGENT 7→ P AGENT ×AGENT 7→ P seq BIT ×AGENT 7→ seq BIT ×AGENT 7→ N ×AGENT 7→ N1 ×AGENT 7→ SEX ×AGENT 7→ N ∀ tail : seq P(AGENT × AGENT ); head : P(AGENT × AGENT ); population : P AGENT ; position : AGENT ֌ P OSIT ION; vision : AGENT 7→ N1; agentSugar : AGENT 7→ N; agentCulture : AGENT 7→ seq BIT ; metabolism : AGENT 7→ N; children : AGENT 7→ P AGENT ; diseases : AGENT 7→ P seq BIT ; agentImmunity : AGENT 7→ seq BIT ; age : AGENT 7→ N; maxAge : AGENT 7→ N1; sex : AGENT 7→ SEX; initialSugar : AGENT 7→ N; THE SPECIFICATION OF SUGARSCAPE 29 ∃ newpopulation : P AGENT ; newposition : AGENT ֌ P OSIT ION; newvision : AGENT 7→ N1; newagentSugar : AGENT 7→ N; newagentCulture : AGENT 7→ seq BIT ; newmetabolism : AGENT 7→ N; newchildren : AGENT 7→ P AGENT ; newdiseases : AGENT 7→ P seq BIT ; newagentImmunity : AGENT 7→ seq BIT ; newage : AGENT 7→ N; newmaxAge : AGENT 7→ N1; newsex : AGENT 7→ SEX; newinitialSugar : AGENT 7→ N; (newpopulation, newposition, newvision, newagentSugar, newagentCulture, newmetabolism, newchildren, newdiseases, newagentImmunity, newage, newmaxAge, newsex, newinitialSugar) = applyMating(asSeq(head), population, position, vision, agentSugar, agentCulture, metabolism, children, diseases, agentImmunity, age, maxAge, sex, initialSugar) • agentSugar, agentCulture, metabolism, children, diseases, agentImmunity, age, maxAge, sex, initialSugar) = concurrentMating(hi, population, position, vision, (population, position, vision, agentSugar, agentCulture, metabolism, children, diseases, agentImmunity, age, maxAge, sex, initialSugar) concurrentMating(hheadi a tail, population, position, vision, agentSugar, agentCulture, metabolism, children, diseases, agentImmunity, age, maxAge, sex, initialSugar) = concurrentMating(tail, newpopulation, newposition, newvision, newagentSugar, newagentCulture, newmetabolism, newchildren, newdiseases, newagentImmunity, newage, newmaxAge, newsex, newinitialSugar) 30 JOSEPH KEHOE applyMating : seq(AGENT × AGENT ) × P AGENT ×AGENT ֌ P OSIT ION ×AGENT 7→ N1 ×AGENT 7→ N ×AGENT 7→ seq BIT ×AGENT 7→ N ×AGENT 7→ P AGENT ×AGENT 7→ P seq BIT ×AGENT 7→ seq BIT ×AGENT 7→ N ×AGENT 7→ N1 ×AGENT 7→ SEX ×AGENT 7→ N ↔P AGENT ×AGENT ֌ P OSIT ION ×AGENT 7→ N1 ×AGENT 7→ N ×AGENT 7→ seq BIT ×AGENT 7→ N ×AGENT 7→ P AGENT ×AGENT 7→ P seq BIT ×AGENT 7→ seq BIT ×AGENT 7→ N ×AGENT 7→ N1 ×AGENT 7→ SEX ×AGENT 7→ N ∀ population : P AGENT ; position : AGENT ֌ P OSIT ION; sex : AGENT 7→ SEX; vision : AGENT 7→ N1; age : AGENT 7→ N; initialSugar : AGENT 7→ N; maxAge : AGENT 7→ N1; metabolism : AGENT 7→ N; agentSugar : AGENT 7→ N; agentCulture : AGENT 7→ seq BIT ; children : AGENT 7→ P AGENT ; agentImmunity : AGENT 7→ seq BIT ; diseases : AGENT 7→ P seq BIT ; head : AGENT × AGENT ; tail : seq(AGENT × AGENT ); • THE SPECIFICATION OF SUGARSCAPE 31 ∃ of f spring, a, b : AGENT ; newsex : AGENT 7→ SEX; newvision : AGENT 7→ N1; newmetabolism, newagentSugar, newinitialSugar : AGENT 7→ N; newmaxAge : AGENT 7→ N1; newagentCulture : AGENT 7→ seq BIT ; newchildren : AGENT 7→ P AGENT ; newagentImmunity : AGENT 7→ seq BIT ; inheritedImmunity : seq BIT ; inheritedCulture : seq BIT ; of f spring 6∈ population a = f irst(head) ∧ b = second(head) newchildren : children ∪ {of f spring 7→ ∅ a 7→ children(a) ∪ {of f spring}, b 7→ children(b) ∪ {of f spring}} newsex = sex ∪ {of f spring 7→ male} ∨ newsex = sex ∪ {of f spring 7→ f emale} newvision = vision ∪ {of f spring 7→ vision(a)} ∨ newvision = vision ∪ {of f spring 7→ vision(b)} newmaxAge = maxAge ∪ {of f spring 7→ maxAge(a)} ∨ newmaxAge = maxAge ∪ {of f spring 7→ maxAge(b)} newmetabolism = metabolism ∪ {of f spring 7→ metabolism(a)} ∨ newmetabolism = metabolism ∪ {of f spring 7→ metabolism(b)} newinitialSugar = initialSugar⊕ {of f spring 7→ initialSugar(a)/2 + initialSugar(b)/2, a 7→ initialSugar(a)/2, b 7→ initialSugar newagentSugar = agentSugar ∪ {of f spring 7→ initialSugar} ∧ ∀ n : 1 . . IMMUNIT Y LENGT H • (inheritedImmunity(n) = agentImmunity(a)(n) ∨ inheritedImmunity(n) = agentImmunity(b)(n)) newagentImmunity : agentImmunity ∪ {of f spring 7→ inheritedImmunity} ∧ ∀ n : 1 . . CULT URECOUNT • (inheritedCulture(n) = agentCulture(a)(n) ∨ inheritedCulture(n) = agentCulture(b)(n)) newagentCulture : agentCulture ∪ {of f spring 7→ inheritedCulture} 32 JOSEPH KEHOE applyMating(hi, population, position, vision, agentSugar, agentCulture, (population, position, vision, agentSugar, agentCulture, metabolism, metabolism, children, diseases, agentImmunity, age, maxAge, sex, initialSugar) = children, diseases, agentImmunity, age, maxAge, sex, initialSugar) applyMating(hheadi a tail, population, position, vision, agentSugar, agentCulture, if ((∃ loc : P OSIT ION (adjacent(loc, position(ag))) metabolism, children, diseases, agentImmunity, age, maxAge, sex, initialSugar) = ∨ adjacent(loc, position(head)) ∧ loc 6∈ dom position) ∧ (agentSugar(head) > initialSugar(head)) ∧ (agentSugar(ag) > initialSugar(ag applyMating(tail, population ∪ {of f spring}, position ∪ {of f spring 7→ loc}, newvision, newagentSugar, newagentCulture, newmetabolism, newchildren, diseases ∪ {of f spring 7→ ∅}, newagentImmunity, age ∪ {of f spring 7→ 0}, newmaxAge, newsex, initialSugar) applyMating(tail, population, position, vision, agentSugar, agentCulture, else metabolism, children, diseases, agentImmunity, age, maxAge, sex, initialSugar) 3.13. Agent Inheritance I. Agent Inheritance I: When an agent dies its wealth is equally distributed among all its living children. The rule definition is deceptively simple but some assumptions must be made in order to give it a precise definition. These assumptions are required because of the discrete nature of the simulation. Only living children can inherit from a parent. If a child is alive but scheduled to die at the same time as their parent then (because all agents who are due to die will die simultaneously) this child should not inherit from their parent. If we were to allow them to inherit we would either have to impose an ordering on the allocation of inheritance making the rule more complex or accept than the ordering will sometimes result in part of an inheritance disappearing. This extra complexity brings no real benefit to the simulation so we discount it. The second assumption is that we allow for rounding errors. Resources (sugar) come in discrete amounts so division between children requires integer division. This is also true of division of the loans amongst an agents children. We just accept any rounding errors as part of the discrete nature of the simulation. Finally we note that inheritance is separate from the actual death or replacement rule, it reallocates the resources of agents due to die but it does not remove those agents from the simulation. We leave that to the actual Replacement or Death rule and assume that one of these rules is applied after the inheritance rule. This simplifies the Inheritance schema. To enable inheritance to handle the loan book (when an agent dies its loans are passed on to its children) we introduce some helper functions. The asSeq function THE SPECIFICATION OF SUGARSCAPE 33 turns a set of items into a sequence of items. It does not specify the ordering in the sequence. [X] asSeq : P X → seq X ∀ x : P X; y : seq X • asSeq(x) = y ⇔ (ran y = x ∧ # y = # x) The second function disperseLoans takes in the loan book, a sequence contain- ing all the dying agents and the children of the agents and produces an updated loan book with the loans of the dying agents now dispersed amongst their children. To do this it employs a third function oneAgentLoans that takes in a single agent (who is marked for removal) the loans (in a sequence) held by that agent and the set containing its children. It outputs a new set of loans generated by dispersing all this agents loans amongst its children. In both cases we use sequences for the parameter we are recursing over as it makes the recursion easier to specify. ×(AGENT 7→ P AGENT )) disperseLoans : (P(AGENT × (AGENT × (N × N))) × seq AGENT → P(AGENT × (AGENT × (N × N))) ∀ Loans : P(AGENT × (AGENT × (N × N))); a : AGENT ; tail : seq AGENT ; Children : AGENT 7→ P AGENT • disperseLoans(Loans,hi, Children) = Loans disperseLoans(Loans,hai a tail, Children) = disperseLoans(({a} −⊳ Loans) ∪ oneAgentLoans(a, asSeq(ran({a} ⊳ Loans)), Children(a)), tail, Children) oneAgentLoans : AGENT × seq(AGENT × (N × N)) × P AGENT ֌ P(AGENT × (AGENT × (N × N))) tail : seq(AGENT × (N × N)) • ∀ a, borrower, inheritor : AGENT ; Children : P AGENT ; amt, dur, newAmt : N; oneAgentLoans(a,hi, Children) = ∅ oneAgentLoans(a,h(borrower, (amt, dur))i a tail, Children) = {x : AGENT x ∈ Children • (x, (borrower, (amt div # Children, dur)))} ∪ oneAgentLoans(a, tail, Children) The getMother and getFather functions simply take in an agent and the children set and finds the mother (father) of the agent from this set. 34 JOSEPH KEHOE getMother : AGENT × (AGENT 7→ P AGENT ) × AGENT 7→ SEX 7→ AGENT getF ather : AGENT × (AGENT 7→ P AGENT ) × AGENT 7→ SEX 7→ AGENT ∀ child, parent : AGENT ; children : AGENT 7→ P AGENT • getMother(child, children, sex) = parent ⇔ child ∈ children(parent) ∧ sex(parent) = f emal getF ather(child, children, sex) = parent ⇔ child ∈ children(parent) ∧ sex(parent) = male An agent can inherit from at most two different agents, one male and one female. We use this to facilitate the specification by treating each sex separately. Inheritance ∆Agents population′ = population ∧ sex′ = sex position′ = position ∧ vision′ = vision age′ = age ∧ maxAge′ = maxAge agentCulture′ = agentCulture ∧ children′ = children metabolism′ = metabolism ∧ diseases′ = diseases agentImmunity ′ = agentImmunity initialSugar′ = initialSugar ∃ dying : P AGENT ; inheritF romF emale, inheritF romMale : AGENT 7→ N; (1) ∀ x : AGENT ; ∃ p : AGENT x ∈ population \ dying • dom inheritF romF emale = dom inheritF romMale = population \ dying dying = {x : AGENT x ∈ population ∧ getMother(x, children, sex) 6∈ dying ⇒ p = getMother(x, children, sex) ∧ p ∈ dying ⇒ ∧ (age(x) = maxAge(x) ∨ agentSugar(x) = 0)} inheritF romF emale(x) = 0 (1a) inheritF romF emale(x) = agentSugar(p) div #(population ∩ children(p) \ dying)) inheritF romMale(x) = 0 getF ather(x, children, sex) 6∈ dying ⇒ p = getF ather(x, children, sex) ∧ p ∈ dying ⇒ inheritF romMale(x) = agentSugar(p) div #(population ∩ children(p) \ dying)) +inheritF romMale(x) + inheritF romF emale(x) loanBook′ = disperseLoans(loanBook, asSeq(dying), children) x ∈ dying x 6∈ dying ⇒ agentSugar′(x) = 0 ⇒ agentSugar′(x) = agentSugar(x) (1b) (3) (4) (5) THE SPECIFICATION OF SUGARSCAPE 35 (1) First we construct the set of dying agents. Then using this set of dying agents we can construct two functions, one mapping amounts inherited from a female parent and one mapping amounts inherited from a male parent. These sets are then used to update the sugar of each agent; a) The function giving the amount each inheriting agent gets from its fe- male parent is constructed by finding all healthy agents who have a dying mother and determining their share of their dying mother's resources; b) The function listing amounts each agent gets from a male parent is constructed in an almost identical manner. (2) If an agent is dying its sugar level is set to zero (because it is being real- located to its children); (3) Otherwise the agents sugar level is its old level plus whatever it inherits from both dying parents; (4) Finally we update the loanBook using our disperseLoans function. 3.14. Agent Culture K. Agent cultural transmission: • Select a neighboring agent at random; • Select a tag randomly; • If the neighbor agrees with the agent at that tag position, no change is made; if they disagree, the neighbor's tag is flipped to agree with the agent's tag; • Repeat for all neighbors. Group membership: Agents are defined to be members of the Blue group when 0s outnumber 1s on their tag strings, and members of the Red group in the opposite case. Agent Culture K: Combination of the "agent cultural transmission" and "agent group membership" rules given immediately above. Group membership is defined with the assumption that there are always an odd number of tags. tribe returns the affiliation of an agent based on the number of bits of each type in its culture sequence. If the majority of bits in a sequence are 0 then it belongs to the blue tribe, otherwise it belongs to the red tribe. This is used by the culture rule. tribe : seq BIT → af f iliation ∀ aSeq : seq BIT • tribe(aSeq) = blue ⇔ #(aSeq ⊲ {0}) > #(aSeq ⊲ {1}) tribe(aSeq) = red ⇔ #(aSeq ⊲ {0}) < #(aSeq ⊲ {1}) f lipT ags is a recursive function that takes in a culture tag sequence belonging to an agent, a sequence of neighbouring agents and the mapping containing all agent's culture tag sequences. It returns a new tag sequence generated by each 36 JOSEPH KEHOE neighbouring agent flipping one bit chosen at random of the original agent's tag sequence. It is aided in this by the function f lipBit that takes in two bit sequences and returns a new sequence equal to the first bit sequence with one bit changed at random to match the other sequence at that position. f lipBit : seq BIT × seq BIT ↔ seq BIT ∀ original, other, new : seq BIT • f lipBit(original, other) = new ⇔ # original = # other = # new ∧ ∃ i : 0 . . # original • ∀ j : 0 . . # original • (i 6= j ⇒ new(j) = original(j)) ∧ new(i) = other(i) f lipT ags : seq BIT × seq AGENT × (AGENT 7→ seq BIT ) ↔ seq BIT ∀ aSeq : seq BIT ; ag : AGENT ; tail : seq AGENT ; culturalResources : AGENT 7→ seq BIT • f lipT ags(aSeq,hi, culturalResources) = aSeq f lipT ags(aSeq,hagi a tail, culturalResources) = f lipT ags(f lipBit(aSeq, culturalResources(ag)), tail, culturalResources) The sequence of neighbours is provided by the Culture scheme which employs the asSeq function to convert a set of neighbours into a sequence. THE SPECIFICATION OF SUGARSCAPE 37 Culture ∆Agents population′ = population sex′ = sex position′ = position vision′ = vision age′ = age maxAge′ = maxAge agentSugar′ = agentSugar children′ = children loanBook′ = loanBook diseases′ = diseases metabolism′ = metabolism agentImmunity ′ = agentImmunity initialSugar′ = initialSugar ∀ a : AGENT • a ∈ population ⇒ agentCulture′(a) = f lipT ags(agentCulture(a), (1) asSeq({b : AGENT adjacent(position(a), position(b))}), agentCulture) (1) For every agent a in the population we allow each other agent that counts a as a neighbour to flip one bit at random of a's culture bit sequence. 3.15. Combat Cα. Agent Combat Cα: • Look out as far as vision permits in the four principle lattice directions; • Throw out all sites occupied by members of the agent's own tribe; • Throw out all sites occupied by members of different tribes who are • The reward of each remaining site is given by the resource level at the site plus, if it is occupied, the minimum of α and the occupant's wealth; wealthier then the agent; • Throw out all sites that are vulnerable to retaliation; • Select the nearest position having maximum reward and go there; • Gather the resources at the site plus the minimum of α and the occu- • If the site was occupied then the former occupant is considered "killed" pants wealth if the site was occupied; - permanently removed from play. reward is used by the combat rule and values a position based on its sugar content and the sugar reserves held by any agent at that position. The combat rule is really an extension of the movement rule where we are now allowed to move to locations occupied by other agents under certain predefined conditions. 38 JOSEPH KEHOE ×(AGENT 7→ N) × N reward : P OSIT ION × (P OSIT ION 7֌ N) × (AGENT 7֌ P OSIT ION) →N ∀ l : P OSIT ION; sugar : P OSIT ION 7֌ N; agentSugar : AGENT 7֌ N; if l ∈ ran positions then positions : AGENT 7֌ P OSIT ION • reward(l, sugar, positions, agentSugar, COMBAT LIMIT ) = sugar(l) + min({COMBAT LIMIT, agentSugar(positions ∼(l))}) else reward(l, sugar, positions, agentSugar, COMBAT LIMIT ) = sugar(l) availMoves returns the set of all safe moves that an agent can make. availMoves : AGENT × (AGENT 7֌ P OSIT ION) × (P OSIT ION 7֌ N)× (AGENT 7֌ N) × (AGENT 7֌ seq BIT ) × N 7→ P P OSIT ION ∀ agent : AGENT ; positions : AGENT 7֌ P OSIT ION; vision : N; sugar : P OSIT ION 7֌ N; agentSugar : AGENT 7֌ N; culture : AGENT 7֌ seq BIT • availMoves(agent, positions, sugar, agentSugar, culture, vision) = {l : P OSIT ION; x : AGENT distance(l, positions(agent)) ≤ vision ∧ positions(x) = l ⇒ (agentSugar(x) < agentSugar(agent) ∧ ((distance(positions(x), l) ≤ vision) ∧ tribe(culture(x)) 6= tribe(culture(agent))) ∧ tribe(culture(x)) 6= tribe(culture(agent))) ⇒ (1) (2) (3) agentSugar(x) < agentSugar(agent) +reward(l, sugar, positions, agentSugar, COMBAT LIMIT )) • l} (1) Only locations within an agents neighbourhood are considered; (2) If a location is occupied it must be occupied by an agent belonging to a different tribe who has lower sugar levels; (3) We only consider a position already containing an agent from another tribe if there are no other agents from a different tribe within the neighbourhood of that location who are stronger than we will be once we have consumed the resources of the new location (that is agents who may retaliate against us for killing an agent belonging to their own tribe). We note that the rule as stated means we consider retaliation under all conditions even if we are just moving to an empty location. It is unclear from the definition given as to how exactly we check for retaliation. Do we base our check on agents visible from our current position or from the proposed position. We have assumed that it is based on the proposed position but it could easily be otherwise. We THE SPECIFICATION OF SUGARSCAPE 39 also assume that the range used is based on the vision of the moving agent as this seems logical. The synchronous version of the combat rule assumes that all combat occurs instantaneously (concurrently). We note that it is simpler to specify in that we just state the before and after states and make no mention of orderings of combat. CombatSynchronous ∆SugarScape step′ = step maxSugar′ = maxSugar pollution′ = pollution loanBook′ = population′ population′ ⊆ population sugar′ = sugar ⊕ {p : P OSIT ION p ∈ ran position′ • p 7→ 0} ∀ ag : AGENT ; l : P OSIT ION • ag ∈ population′ ⇒ ⊳ loanBook ⊲ (population′ ⊳ (ran loanBook)) (sex′(ag) = sex(ag) ∧ vision′(ag) = vision(ag) ∧ age′(ag) = age(ag) ∧ maxAge′(ag) = maxAge(ag) ∧ children′(ag) = children(ag) ∧ agentCulture′(ag) = agentCulture(ag) ∧ agentImmunity ′(ag) = agentImmunity(ag) ∧ metabolism′(ag) = metabolism(ag) ∧ diseases′(ag) = diseases(ag) ∧ initialSugar′(a) = initialSugar(a) ∧ agentSugar′(ag) = agentSugar(ag) ∧ position′(ag) ∈ ) (1) (2) (3) (4) (5) (6) +reward(position′(ag), sugar, position, agentSugar, COMBAT LIMIT ) availMoves(ag, position, sugar, agentSugar, agentCulture, vision(ag))) ag ∈ population \ population′ ⇒ (7) ∃ x : AGENT • position′(x) = position(ag) ∧ tribe(culture(x)) 6= tribe(culture(ag)) (8) (l ∈ availMoves(ag, position, sugar, agentSugar, agentCulture, vision(ag)) ≥ reward(position′(ag), sugar, position, agentSugar, COMBAT LIMIT ) ∧ reward(l, sugar, position, agentSugar, COMBAT LIMIT ) ∧ distance(position(ag), l) < distance(position(ag), position′(ag))) ⇒ ∃ x : AGENT • position′ ∼(l) = x ∧ position(x) 6= l 40 JOSEPH KEHOE (1) Every agent that is removed from the simulation is also removed from the loanBook; (2) No new agents are introduced; (3) Location sugar levels are updated; (4) Every agent that remains in the population has all its attributes unchanged apart from (possibly) position and sugar; (5) We update the sugar levels of each agent using the reward function; (6) Every agent has moved somewhere within their old neighbourhood; (7) Every agent that is no longer part of the population was removed by com- bat, that is, there is another agent (the agent that killed them) now situated in their old position; (8) If a location available to an agent and the reward of that location is better or equal to that agent's new position and it was closer than that agents new position to its old position then it must be the case that some other agent has just moved to that location (otherwise we would have moved there); We have had to make some assumptions here. It is not stated what happens when there are no available moves, for example if all sites are subject to retaliation. We have assumed that a move is preferable to staying still and that the only time that an agent stays in the same position is when there are no available moves. That is, if every site, including our current one, is subject to retaliation then we do not move anywhere. A more complex interpretation would be to for an agent that cannot escape retaliation to attack another agent anyway and hope for the best but purely in the interests of simplicity we have agents remain where they are. 3.16. Credit Ld r. Credit Ld r: • An agent is a potential lender if it is too old to have children, in which case the maximum amount it may lend is one-half of its current wealth; • An agent is a potential lender if it is of childbearing age and has wealth in excess of the amount necessary to have children, in which case the maximum amount it may lend is the excess wealth; • An agent is a potential borrower if it is of childbearing age and has insufficient wealth to have a child and has income (resources gathered, minus metabolism, minus other loan obligations) in the present period making it credit-worthy for a loan written at terms specified by the lender; • If a potential borrower and a potential lender are neighbors then a loan is originated with a duration of d years at the rate of r percent, and the face value of the loan is transferred from the lender to the borrower; THE SPECIFICATION OF SUGARSCAPE 41 • At the time of the loan due date, if the borrower has sufficient wealth to repay the loan then a transfer from the borrower to the lender is made; else the borrower is required to pay back half of its wealth and a new loan is originated for the remaining sum; lender simply takes a loss; • If the borrower on an active loan dies before the due date then the • If the lender on an active loan dies before the due date then the bor- rower is not required to pay back the loan, unless inheritance rule I is active, in which case the lender's children now become the borrower's creditors. totalOwed calculates the total amount owed from a given sequence of loans. We have assumed that interest is simple interest and not compound. totalOwed : seq(AGENT × (N × N)) → N ∀ a : AGENT ; amt, dur : N; tail : seq(AGENT × (N × N)) • totalOwed(hi) = 0 totalOwed(h(a, (amt, dur))i a tail) = (amt + amt ∗ RAT E ∗ DURAT ION) +totalOwed(tail) canLend and willBorrow are simple rules. The definition of what determines credit-worthiness is missing so we have assumed it means an agent has enough money to pay all their outstanding loans. canLend : N × SEX × N → boolean ∀ age, sugar : N • canLend(age, male, sugar) ⇔ ∧ sugar > CHILDAMT ) canLend(age, f emale, sugar) ⇔ age > MALEF ERT ILIT Y END ∨ (MALEF ERT ILIT Y ST ART ≤ age ≤ MALEF ERT ILIT Y END age > F EMALEF ERT ILIT Y END ∨ (F EMALEF ERT ILIT Y ST ART ≤ age ≤ F EMALEF ERT ILIT Y END ∧ sugar > CHILDAMT ) ∧ sugar < CHILDAMT ) ∧ sugar > totalOwed(asSeq(loans)) willBorrow(age, f emale, sugar, loans) ⇔ ∧ sugar < CHILDAMT ) ∧ sugar > totalOwed(asSeq(loans)) (F EMALEF ERT ILIT Y ST ART ≤ age ≤ F EMALEF ERT ILIT Y END 42 JOSEPH KEHOE willBorrow : N × SEX × N × P(AGENT × (N × N)) → boolean ∀ age, sugar : N; loans : P(AGENT × (N × N)) • willBorrow(age, male, sugar, loans) ⇔ (MALEF ERT ILIT Y ST ART ≤ age ≤ MALEF ERT ILIT Y END amtAvail depends on whether an agent can still have children. If they are no longer fertile then they can loan out half their available sugar. If that are still fertile then they have to retain enough sugar to have children. amtAvail : N × SEX × N → N ∀ age, sugar : N • amtAvail(age, male, sugar) = if (age > MALEF ERT ILIT Y END)then sugar div 2 else if(isF ertile(age, male) ∧ sugar > CHILDAMT )then else sugar − CHILDAMT 0 amtAvail(age, f emale, sugar) = if (age > F EMALEF ERT ILIT Y END)then sugar div 2 else if(isF ertile(age, f emale) ∧ sugar > CHILDAMT )then else sugar − CHILDAMT 0 amtReq is the amount that a lender requires. This is not defined so we can only use a best guess as to what it is. We assume that the amount required is that which gives the borrower enough sugar to have children. This is the simplest sensible definition we can think of. amtReq : N → N ∀ sugar : N • amtReq(sugar) = CHILDAMT − sugar We supply some simple helper functions that extract the borrower and lender from a loanBook entry, calculate the amount due from a loan, the principal and the due date (defined as the step when payment is due). THE SPECIFICATION OF SUGARSCAPE 43 lender : AGENT × (AGENT × (N × N)) ֌ AGENT borrower : AGENT × (AGENT × (N × N)) ֌ AGENT amtDue : AGENT × (AGENT × (N × N)) ֌ N principal : AGENT × (AGENT × (N × N)) ֌ N due : AGENT × (AGENT × (N × N)) ֌ N ∀ l, b : AGENT ; p, d : N • lender(l, (b, (p, d))) = l borrower(l, (b, (p, d))) = b amtDue(l, (b, (p, d))) = p + p ∗ RAT E ∗ DURAT ION principal(l, (b, (p, d))) = p due(l, (b, (p, d))) = d Finally, using these functions we can present the P ayLoans schema. P ayLoans ∆Agents ΞStep population′ = population sex′ = sex position′ = position vision′ = vision age′ = age maxAge′ = maxAge agentCulture′ = agentCulture agentImmunity ′ = agentImmunity children′ = children diseases′ = diseases metabolism′ = metabolism initialSugar′ = initialSugar ∃ dueLoans, newLoans : (AGENT ↔ (AGENT × (N × N))) • dueLoans = loanBook ⊲ (ran(loanBook) ⊲ {a : (N × N) second(a) = step}) (loanBook′, agentSugar′) = payExclusiveLoans(chooseConf lictF reeSets(dueLoans), agentSugar, loanBook) This schema is complicated by the fact that it is possible that an agent has a loan due and cannot pay this loan off. In this case, according to the rule defini- tion, the borrower must pay half of its sugar to the lender and renegotiate another loan to cover the remainder of its debt. Under this rule some issues will arise if the borrower has more than one due loan and cannot pay these loans off. The lender must pay each borrower in sequence the amount of half its sugar. This 44 JOSEPH KEHOE cannot be performed simultaneously (for example if we owe three loans we can- not give each lender half our sugar as this would mean giving out more sugar than we actually have). In order to remain true to the rule definition we must, when we have more than one loan due, pay each loan in some sequence (defined using a conflict resolution rule e.g. pay biggest loan first). The helper function chooseConf lictF reeLoans returns a sequence of groups of loans that are conflict free (i.e. a borrower can only appear once in each group). The function payExclusiveLoans takes in this sequence of loan sets and pro- cesses each set concurrently in the same manner as the Mating rule. chooseConf lictF reeLoans : (AGENT ↔ (AGENT × (N × N))) ↔seq(AGENT ↔ (AGENT × (N × N))) ∀ a : AGENT ; dueLoans : (AGENT ↔ (AGENT × (N × N))) • chooseConf lictF reeLoans(∅) = hi chooseConf lictF reeLoans(dueLoans) = ∃ maxSet : (AGENT ↔ (AGENT × (N × N))) maxSet ⊆ dueLoans (1) ∧ #({a} ⊳ (ran dueLoans)) > 0 ⇒ #({a} ⊳ (ran maxSet)) = 1 hmaxSeti a chooseConf lictF reeLoans(dueLoans \ maxSet) (1) We choose the largest convict free set possible where a set is deemed conflict free if all borrowers only appear in that set at most once. payExclusiveLoans : seq(AGENT ↔ (AGENT × (N × N))) ×AGENT 7→ N ×(AGENT ↔ (AGENT × (N × N))) ↔((AGENT ↔ (AGENT × (N × N))) ×AGENT 7→ N) ∀ tail : seq(AGENT ↔ (AGENT × (N × N))); head, loanBook : (AGENT × (AGENT × (N × N))); agentSugar : AGENT 7→ N • payExclusiveLoans(hi, agentSugar, loanBook) = (loanBook, agentSugar) payExclusiveLoans(hheadi a tail, agentSugar, loanBook) = ∃ newAgentSugar : AGENT 7→ N; newLoans : (AGENT ↔ (AGENT × (N × N))) (newLoans, newAgentSugar) = makeP ayments(asSeq(head), ∅, agentSugar) • payExclusiveLoans(tail, newAgentSugar, (LoanBook \ head) ∪ newLoans) makeP ayments is a recursive function that goes through a sequence of loans and makes the final payment on each one. It is used in the P ayLoans schema THE SPECIFICATION OF SUGARSCAPE 45 where it takes in a sequence of the due loans and the agents current sugar levels and returns a set of renegotiated loans, where payment is unable to be made, and the new agent sugar levels. makeP ayments : seq(AGENT × (AGENT × (N × N)))× P(AGENT × (AGENT × (N × N))) × (AGENT 7→ N) →(P(AGENT × (AGENT × (N × N))) × (AGENT 7֌ N)) ∀ renegotiatedLoans, new : P(AGENT × (AGENT × (N × N))); updatedSugar, agentSugar : AGENT 7→ N; loan : (AGENT × (AGENT × (N × N))); tail : seq(AGENT × (AGENT × (N × N))) • (renegotiatedLoans, updatedSugar) makeP ayments(hi, renegotiatedLoans, updatedSugar) = makeP ayments(hloani a tail, new, agentSugar) = if amtDue(loan) ≤ agentSugar(borrower(loan)) then makeP ayments(tail, new, agentSugar (1) (2) (2a) ⊕{lender(loan) 7→ agentSugar(lender(loan)) + amtDue(loan), borrower(loan) 7→ agentSugar(borrower(loan)) − amtDue(loan)}) (2b) else makeP ayments(tail, new ∪ {(lender(loan), (borrower(loan), (amtDue(loan) − agentSugar(borrower(loan)) div 2, due(loan) + DURAT ION)))}, agentSugar ⊕ {lender(loan) 7→ agentSugar(lender(loan)) +agentSugar(borrower(loan)) div 2, borrower(loan) 7→ agentSugar(borrower(loan)) div 2}) For the final part of the Credit rule we need to be able to work out the total owed by an agent over all loans. First we define two helper functions: sumLoans and totalOwed. sumLoans : seq(AGENT × (AGENT × (N × N))) ֌ N ∀ tail : seq(AGENT × (AGENT × (N × N))); top : AGENT × (AGENT × (N × N)) • sumLoans(hi) = 0 sumLoans(htopi a tail) = sumLoans(tail) + amtDue(top) 46 JOSEPH KEHOE totalOwed : AGENT × (AGENT × (AGENT × (N × N))) ֌ N totalLoaned : AGENT × (AGENT × (AGENT × (N × N))) ֌ N ∀ agent : AGENT ; loans : AGENT × (AGENT × (N × N)) • totalOwed(agent, loans) = totalLoaned(agent, loans) = sumLoans(asSeq(loans ⊲ ({agent} ⊳ (ran loans)))) sumLoans(asSeq({agent} ⊳ loans)) THE SPECIFICATION OF SUGARSCAPE 47 MakeLoans ∆Agents ΞStep population′ = population ∧ sex′ = sex position′ = position ∧ vision′ = vision age′ = age ∧ maxAge′ = maxAge initialSugar′ = initialSugar agentCulture′ = agentCulture ∧ agentImmunity ′ = agentImmunity diseases′ = diseases ∧ children′ = children ∧ metabolism′ = metabolism ∃ newLoans : P(AGENT × (AGENT × (N × N))); ∀ ag, lender, borrower : AGENT ; amt, due : N • loanBook′ = loanBook ∪ newLoans ag ∈ dom newLoans ⇒ ag ∈ dom(ran newLoans) ⇒ ag 6∈ dom(newLoans) ∪ dom(ran newLoans) ⇒ willBorrow(age(ag), sex(ag), agentSugar′(ag), agentSugar′(ag) = agentSugar(ag) − totalLoaned(ag, newLoans) agentSugar′(ag) = agentSugar(ag) + totalOwed(ag, newLoans) agentSugar′(ag) = agentSugar(ag) (1) (2a) (2b) (2c) ran(loanBook′ ∩ {a : AGENT × (AGENT × (N × N)) (2d) borrower(a) = borrower(loan)})) ⇒ ¬ ∃ ag2 : AGENT • canLend(age(ag2), sex(ag2), agentSugar′(ag2)) ∧ adjacent(position(ag2), position(ag)) totalLoaned(ag, newLoans) ≤ amtAvail(age(ag), sex(ag), agentSugar(ag))(3) (4) totalOwed(ag, newLoans) ≤ amtReq(agentSugar(ag)) (5) (lender, (borrower, (amt, due))) ∈ newLoans ⇒ (5a) {borrower} ⊳ (ran loanBook)) (canLend(age(lender), sex(lender), agentSugar(lender)) ∧ willBorrow(age(borrower), sex(borrower), agentSugar(borrower), ∧ amt ≤ min({amtAvail(age(lender), sex(lender), agentSugar(lender)), (5b) (5c) ∧ due = step + DURAT ION ∧ adjacent(position(lender), position(borrower))) (5d) amtReq(agentSugar(borrower))}) (1) The new loan book is the old book plus the new loans; (2) The following properties ensure sugar is updated correctly and that the correct amount of borrowing has taken place: a) If an agent is a lender then their new sugar levels decrease by the amount the have lent; b) If an agent is a borrower then their sugar has increased by the amount they have borrowed; 48 JOSEPH KEHOE c) Any agent that neither borrowed or lent has the same sugar levels as before; d) If there remain any agents who still need to borrow then it is because there are no agents in their neighbourhood who are still in a position to borrow. (3) The total amount loaned by any agent is no greater than the amount that agent had available; (4) The total amount borrowed is less than or equal to the amount required by the borrower; (5) Every loan in this set must have the following properties: a) The lender must be in a position to lend; b) The borrower must need to borrow; c) The amount is less than or equal to the minimum of (i) the amount required by the borrower and (ii) the maximum amount available from the lender; d) The due date of the loan is set by the DURAT ION constant; e) the borrower and lender must be neighbours. 3.17. Agent Disease E. Agent immune response: is immune), else (the agent is infected) go to the following step; • If the disease is a substring of the immune system then end (the agent • The substring in the agent immune system having the smallest Ham- ming distance from the disease is selected and the first bit at which it is different from the disease string is changed to match the disease. Disease transmission: For each neighbor, a disease that currently afflicts the agent is selected at random and given to the neighbor. Agent disease processes E: Combination of "agent immune response" and "agent disease transmission" rules given immediately above subseq is a function for determining whether one sequence is a subsequence of another. hammingDist determines the number of bit differences in two sequences of the same size. subseq : seq BIT × seq BIT → boolean ∀ mid, aSequence : seq BIT • subseq(mid, aSequence) ⇔ ∃ pref ix, suf f ix : seq BIT • pref ix a mid a suf f ix = aSequence THE SPECIFICATION OF SUGARSCAPE 49 hammingDist : seq BIT × seq BIT → N ∀ tail, rest : seq BIT • hammingDist(hi,hi) = 0 hammingDist(h1i a tail,h1i a rest) = hammingDist(tail, rest) hammingDist(h0i a tail,h0i a rest) = hammingDist(tail, rest) hammingDist(h0i a tail,h1i a rest) = 1 + hammingDist(tail, rest) hammingDist(h1i a tail,h0i a rest) = 1 + hammingDist(tail, rest) applyDiseases takes in a bit sequence representing the immunity of an agent and a list of diseases that affect the agent and produces a new immunity bit sequence that is updated by the disease list. More precisely, for every disease not in the immunity sequence a single bit in the closest subsequence that matches the disease is flipped to make the sequence more closely match the disease. It uses another function processInf ection to process each disease in the disease set. applyDiseases : seq BIT × seq seq BIT → seq BIT ∀ I, d : seq BIT ; tail : seq seq BIT • applyDiseases(I,hi) = I applyDiseases(I,hdi a tail) = applyDiseases(processInf ection(I, d), tail) processInf ection : seq BIT × seq BIT → seq BIT ∀ I, d : seq BIT • if subseq(d, I)then processInf ection(I, d) = I else ∃ a, b, c : seq BIT ; ∀ x : seq BIT • a a b a c = I (# b = # x ∧ subseq(x, I)) ⇒ hammingDist(b, d) ≤ hammingDist(x, d) ∃ i : 1 . . # I • (y(i) 6= d(i) ∧ ∀ j : N • j < i ⇒ d(j) = y(j)) processInf ection(I, d) = I ⊕ {(i + # a) 7→ b(i)}) ImmuneResponse is the simplest part of this rule to specify. The recursive function applyDiseases does all the work. 50 JOSEPH KEHOE ImmuneResponse ∆Agents loanBook′ = loanBook population′ = population sex′ = sex position′ = position vision′ = vision age′ = age maxAge′ = maxAge agentCulture′ = agentCulture diseases′ = diseases children′ = children agentSugar′ = agentSugar metabolism′ = metabolism initialSugar′ = initialSugar agentImmunity ′ = {a : AGENT a ∈ population • a 7→ applyDiseases(agentImmunity(a), asSeq(diseases(a)))} ∀ x : AGENT • x ∈ population ⇒ agentSugar′(x) = agentSugar(x)− #{d : seq BIT d ∈ diseases(a) ∧ ¬ subseq(d, agentImmunity(a)} (1) Although not stated in the rule definition careful reading of the accompanying text [Epstein and Axtell, 1996] shows that there is a penalty that is applied to each agent carrying diseases that it has no immunity to. The text states that for every disease carried by an agent that it has no immunity to, sugar metabolism is increased by one. So if an agent carried two diseases that it has no immu- nity to then its metabolism rate increases by two. This extra cost can equally be deducted by the metabolism rule or the disease rule. Purely for the sake of narrative it is placed in the ImmuneResponse rule where it is first referenced in the original Sugarscape book. This is implemented by the final two lines (1) of the ImmuneResponse schema. The transmission of diseases is the more complex part of this rule. We will again use a recursive helper function newDiseases to construct a set of diseases that an agent can catch from its neighbours. It takes the set of neighbours and their current diseases as input and constructs a set of diseases where one disease is chosen from each neighbour. THE SPECIFICATION OF SUGARSCAPE 51 newDiseases : seq AGENT × (AGENT 7→ P(seq BIT )) → P seq BIT ∀ a : AGENT ; tail : seq AGENT ; diseases : AGENT 7→ P(seq BIT ) • newDiseases(hi, diseases) = ∅ newDiseases(hai a tail, diseases) = newDiseases(tail, diseases) if diseases(a) = ∅then else ∃ d : seq BIT d ∈ diseases(a) • {d} ∪ newDiseases(tail, diseases)) T ransmission ∆Agents loanBook′ = loanBook population′ = population sex′ = sex position′ = position vision′ = vision age′ = age maxAge′ = maxAge agentCulture′ = agentCulture agentImmunity ′ = agentImmunity children′ = children agentSugar′ = agentSugar metabolism′ = metabolism initialSugar′ = initialSugar ∀ a : AGENT • a ∈ population ⇒ diseases′(a) = diseases(a) ∪ newDiseases(asSeq(visibleAgents(a, position, 1)), diseases) (1) (1) visibleAgents returns the set of neighbours of an agent and this set is then passed to the newDisease function which returns a set of diseases, one chosen from each agent in the neighbour set. 3.18. Rule Application Sequence. Each tick of simulation time consists of the application of a set sequence of rules. Not all rules can be used together so we identify the allowable sequences of rules. We note that it is not stated in the book what order the rules are to be applied. In the absence of this information we will pick one ordering and restrict ourselves to this ordering. The display the different allowable combinations of rules in any given simulation we use the following terminology. 52 JOSEPH KEHOE {Rule}: The indicates that Rule is optional. We can choose to include it or not in a simulation; RuleA RuleB: This indicates that there is a choice of which rule to apply - either one or the other but not both. This rule ordering is for simulations using only a single resource and so omits the T rade rule. T ick [# Growback # SeasonalGrowback] [# Movementbasic (# Movementpollution # P ollutionDif f usion) # Combat] {# Inheritance}{# Death{[# Replacement # AgentMating]}} {# Culture}{# P ayLoans # MakeLoans} {# T ransmission # ImmuneResponse} 4. Asynchronous Sugarscape Specification AU is the sequential application of rules to agents during a simulation step. If, for example, all agents move during a single step then a sequential ordering is imposed on all of the agents and they will move one at a time (that is, sequentially) based on that ordering. This is in contrast to SU where all agents will attempt to move simultaneously (concurrently). AU is easier to implement that SU as it maps directly onto the current standard sequential programming practice. AU requires no collision detection and resolution (as for example when two agents try to simultaneously move to the same location) because concurrency is excluded - only one agent can move at any one time. It is well know that the AU and SU approaches can deliver different simulation results. Although AU and SU are both commonly used in CA based simulations, where agent interactions are simple in nature, AU is prevelent in ABM. This is due to the lack of any good SU algorithms that can handle the complex interactions such as Movement, Combat or T rade that can appear in ABM based simulations. We have provided a specification of Sugarscape that assumes SU. For the sake of completeness and to allow us to make comparisons between synchronous and asyn- chronous updating in Sugarscape we will now present an AU based specification of the rules of Sugarscape. 4.1. Variants of Asynchronous Updating. There are a number of varieties of AU. These variations differ in how they sequentially order agents for updating. The best known variations are [?]: Fixed Direction Line-By-Line: The locations in the lattice representing the simulation space are updated in the order they appear in the lattice (usually left to right, top-down); THE SPECIFICATION OF SUGARSCAPE 53 Fixed Random Sweep: The order that is used is determined randomly at the start of the simulation and this order is used for every step in the simulation; Random New Sweep: The order that the agents are updated in is deter- mined randomly at the start of each step (each step uses a different order); Uniform Choice: Each agent has an equal probability of being chosen. If there are n agents, then n agents are chosen randomly during a step. During any single step an agent may not be picked at all or may be picked more than once (in contrast Random New Sweep guarantees every agent is picked exactly once per step); Exponential Waiting Time: This is a Time Driven method, all the oth- ers are step driven. Every agent has its own clock which rings when the agent is to be updated. The waiting times for the clock are exponentially distributed (with mean 1). The probability that an event occurs at time t follows e−t where t is a real number, t ≥ 0. This is most similar to Uniform Choice. We will provide a specification for each variation in turn. Fixed Direction Line-by-Line takes in a set of agents and their positions on the lattice. It produces a sequence of agents where every agent appears once and only once in the sequence and the order of the sequence is determined by the agents position on the lattice. lineByLine : AGENT ֌ P OSIT ION →seqAGENT ∀ theP ositions : AGENT ֌ P OSIT ION; theSequence : seq AGENT • lineByLine(theSet) = theSequence ⇔ ran theSequence = dom theP ositions ∧ # theSequence = # theP ositions (1) (n, a) ∈ theSequence ⇔ n = f irst(theP ositions(a)) ∗ DIM + second(theP ositions(a)) (2) (1) Each agent in the population appears in the sequence once and only once; (2) If one agent appears before another in the sequence then it also appears before that agent on the lattice. Fixed Random Sweep returns a sequence of the agents in some fixed random ordering. This random ordering is chosen once at the start of the simulation and is fixed for the entire simulation run. RANDOMORDER : seq P OSIT ION # RANDOMORDER = # P OSIT ION (1) ∀ n, m : N • RANDOMORDER(n) = RANDOMORDER(m) (2) ⇔ n = m 54 JOSEPH KEHOE (1) RANDOMORDER is a globally defined sequence containing an ordering of positions on the lattice; (2) Each position on the lattice appears once and only once in this sequence. Any ordering that satisfies these constrains is allowable according to our specifi- cation. This introduces the randomness into the sequence. f ixedRandom : AGENT ֌ P OSIT ION ↔seqAGENT ∀ theP ositions : AGENT ֌ P OSIT ION; theSequence : seq AGENT • f ixedRandom(theP ositions) = theSequence ∀ i : 0 . . # theSequence − 2; a1, a2 : AGENT • ⇔ ran theSequence = dom theP ositions ∧ # theSequence = # theP ositions (1) (i, a1) ∈ theSequence ∧ (i + 1, a1) ∈ theSequence ⇒ (∃ x1, x2 : N (x1, a1), (x2, a2) ∈ RANDOMORDER ∧ x1 < x2 (2) (1) Every agent in the population appears once and only once in the resulting sequence; (2) The ordering of agents in the sequence is based on the ordering defined in RANDOMORDERING. Random New sweep is simpler to specify. We return a random ordering of agents after each call. We only need to ensure that every agent appears in this sequence exactly once. rndNewSweep : AGENT ֌ P OSIT ION ↔seqAGENT ∀ theP ositions : AGENT ֌ P OSIT ION; theSequence : seq AGENT • rndNewSweep(theP ositions) = theSequence ⇔ ran theSequence = dom theP ositions ∧ # theSequence = # theP ositions(1) (1) Every agent in the population appears once and only once in the resulting sequence; Uniform Choice allows for an agent to be picked multiple times. The only constraints are that the sequence returned contains only agents in the population and that the size of the sequence equals the number of agents. THE SPECIFICATION OF SUGARSCAPE 55 unif ormChoice : AGENT ֌ P OSIT ION ↔ seq AGENT ∀ theP ositions : AGENT ֌ P OSIT ION; n : N; theSequence : seq AGENT 0 ≤ n < # theSequence unif ormChoice(theP ositions) = theSequence ⇔ theSequence(n) ∈ dom theP ositions ∧ # theSequence = # theP ositions (1) Every agent in the sequence is an agent from the simulation population; (2) the size of the sequence equals the total number of individual agents in the population. Each variation of asynchronous updating can now be covered by the simple matter of swapping in the appropriate ordering function within the specifications. 4.2. Growback, Seasonal Growback and Replacement. Replacement, Growback and SeasonalGrowback belong to the category of rules we term independent. This category includes all rules where the agent involved in the update (or rule execu- tion) does not interact with any other agent - the update result is independent of any outside factor. It follows then that the order in which these rules are executed will have no bearing on their outcome. Given this we need make no changes to any of these rules. 4.3. Pollution Diffusion. P ollutionDif f usion is defined specifically as a syn- chronous rule. There is no asynchronous alternative to this rule as imposing AU would redefine the rule entirely. For this reason we do not produce a AU specifi- cation of this rule. 4.4. Movement. The specification of rules under an AU regime follows a standard pattern. First we impose an ordering on all the agents subject to the rule and then we recursively apply the update to each agent in the defined order. Each individual agent update can affect the global state and these changes must be passed forward to the next sequence of agent updates. This is in contrast to SU where all updates occur simultaneously. We always define the application of the rule to agents in a sequence recursively. While the rules themselves can be quite simple the Z notation forces us to pass to each update all parts of the global state that can be changed. This can result in large function signatures. 56 JOSEPH KEHOE AsyncMovementbasic ∆SugarScape step′ = step loc′ = loc maxSugar′ = maxSugar pollution′ = pollution sex′ = sex vision′ = vision age′ = age maxAge′ = maxAge agentCulture′ = agentCulture loanBook′ = loanBook diseases′ = diseases agentImmunity ′ = agentImmunity children′ = children metabolism′ = metabolism population′ = population initialSugar′ = initialSugar (sugar′, agentSugar′, position′) = applyMove(rndNewSweep(position), vision, sugar, agentSugar, position) Movement is a typical example of this structure. The main specification AsyncMovementbasic simply passes the relevant state information alongside the ordering of agents (ac- cording to whatever AU variant we are using) to the recursive function applyMove. This recursive function applies the move rule to each agent in turn and returns the final updated agent position, agent sugar levels and lattice sugar levels. THE SPECIFICATION OF SUGARSCAPE 57 applyMove : seq AGENT ×AGENT 7→ N ×(P OSIT ION 7֌ N) ×AGENT 7֌ N ×AGENT 7֌ P OSIT ION ↔((P OSIT ION 7֌ N) ×AGENT 7֌ N ×AGENT 7֌ P OSIT ION) ∀ head : AGENT ; tail : seq AGENT ; population : P AGENT ; positions : AGENT 7֌ P OSIT ION; sugar : P OSIT ION 7֌ N; agentSugar : AGENT 7֌ N; vision : AGENT 7→ N; • applyMove(hi, vision, sugar, agentSugar, positions) = (sugar, agentSugar, positions) (1) applyMove(hheadi a tail, vision, sugar, agentSugar, positions) = (2) ∃ newLoc : P OSIT ION newLoc ∈ neighbourhood(position(head), vision(head)) (3) ∧ ∀ otherLoc : P OSIT ION otherLoc ∈ neighbourhood(position(head), vision(head)) ⇒ sugar(otherLoc) ≤ sugar(newLoc) • applyMove(tail, vision, sugar ⊕ {newLoc 7→ 0}, agentSugar ⊕ {head 7→ agentSugar(head) + sugar(newLoc)}, positions ⊕ {head 7→ newLoc}) (1) The base case. If there are no agents left to update then we simply return the current state; (2) The recursive case. If we have agents left to process then we move the first agent in the list and apply the rule to the remaining agents; (3) Find the best location for the agent to move to based on sugar levels at each location. 4.5. Pollution Diffusion. The movement rule for pollution is almost identical to the simpler basic movement rule. It only differs in that it takes pollution into account when selecting the best new position for an agent to move to. 58 JOSEPH KEHOE AsyncMovementpollution ∆SugarScape step′ = step loc′ = loc maxSugar′ = maxSugar sex′ = sex pollution′ = pollution vision′ = vision age′ = age maxAge′ = maxAge agentCulture′ = agentCulture loanBook′ = loanBook children′ = children agentImmunity ′ = agentImmunity diseases′ = diseases metabolism′ = metabolism population′ = population (sugar′, agentSugar′, position′) = (1) applyMovepollution(rndNewSweep(position), vision, sugar, agentSugar, position, pollution) pollution′ = pollution ⊕ {∀ l : P OSIT ION; x : AGENT position′(x) = l • (2) l 7→ (P RODUCT ION ∗ sugar(l) + CONSUMP T ION ∗ metabolism(x))} (1) Call the recursive applyMovepollution to apply movement rule to each agent in turn; (2) Update location pollution levels based on agent movement. THE SPECIFICATION OF SUGARSCAPE 59 applyMovepollution : seq AGENT ×AGENT 7→ N ×(P OSIT ION 7֌ N) ×AGENT 7֌ N ×AGENT 7֌ P OSIT ION ×P OSIT ION 7֌ N ↔((P OSIT ION 7֌ N) ×AGENT 7֌ N ×AGENT 7֌ P OSIT ION) ∀ head : AGENT ; tail : seq AGENT ; population : P AGENT ; positions : AGENT 7֌ P OSIT ION; sugar : P OSIT ION 7֌ N; agentSugar : AGENT 7֌ N; vision : AGENT 7→ N; pollution : P OSIT ION 7֌ N • applyMovepollution(hi, vision, sugar, agentSugar, positions, pollution) = (sugar, agentSugar, positions) applyMovepollution(hheadi a tail, vision, sugar, agentSugar, positions, pollution) = ∃ newLoc : P OSIT ION newLoc ∈ neighbourhood(position(head), vision(head)) ∧ ∀ otherLoc : P OSIT ION otherLoc ∈ neighbourhood(position(head), vision(head)) ⇒ sugar(otherLoc) div (1 + pollution(otherLoc)) ≤ applyMovepollution(tail, vision, sugar ⊕ {newLoc 7→ 0}, sugar(newLoc) div (1 + pollution(position′(newLoc))) • agentSugar ⊕ {head 7→ agentSugar(head) + sugar(newLoc)}, positions ⊕ {head 7→ newLoc}, pollution) 4.6. Combat. Asynchronous Combat is undertaken with the applyAllCombat function which applies the combat rule to each agent in a random order using the singleF ight function. We note in passing that the synchronous specification seems to us to be simpler than the asynchronous one (even if the implementation is not). 60 JOSEPH KEHOE Combatasync ∆SugarScape step′ = step maxSugar′ = maxSugar pollution′ = pollution loanBook′ = population′ ⊳ loanBook ⊲ (population′ ⊳ (ran loanBook)) ∀ ag : AGENT ; l : P OSIT ION • ag ∈ population′ ⇒ (sex′(ag) = sex(ag) ∧ vision′(ag) = vision(ag) ∧ age′(ag) = age(ag) ∧ maxAge′(ag) = maxAge(ag) ∧ children′(ag) = children(ag) ∧ agentCulture′(ag) = agentCulture(ag) ∧ agentImmunity ′(ag) = agentImmunity(ag) ∧ metabolism′(ag) = metabolism(ag) ∧ diseases′(ag) = diseases(ag)) ∧ initialSugar′(ag) = initialSugar(ag) (population′, position′, sugar′, agentSugar′) = applyAllCombat(rndNewSweep(position), population, position, sugar, agentSugar, vision, agentCulture) THE SPECIFICATION OF SUGARSCAPE 61 applyAllCombat : seq AGENT × P AGENT ×(AGENT 7֌ P OSIT ION) ×(P OSIT ION 7֌ N) ×(AGENT 7֌ N) ×(AGENT 7→ N) ×(AGENT 7→ seq BIT ) →(P AGENT ×(AGENT 7֌ P OSIT ION) ×(P OSIT ION 7֌ N) ×(AGENT 7֌ N)) ∀ head : AGENT ; tail : seq AGENT ; population : P AGENT ; positions : AGENT 7֌ P OSIT ION; sugar : P OSIT ION 7֌ N; agentSugar : AGENT 7֌ N; vision : AGENT 7→ N; culture : AGENT 7→ seq BIT • applyAllCombat(hi, population, positions, sugar, agentSugar, vision, culture) = (population, positions, sugar, agentSugar, vision, culture) applyAllCombat(hheadi a tail, population, positions, sugar, agentSugar, vision, culture) = if(head ∈ population)then applyAllCombat(tail, singleF ight(head, population, positions, sugar, agentSugar, vision, culture)) else applyAllCombat(tail, population, positions, sugar, agentSugar, vision, culture) 62 JOSEPH KEHOE singleF ight : AGENT× P AGENT ×(AGENT 7֌ P OSIT ION) ×(P OSIT ION 7֌ N) ×(AGENT 7֌ N) ×(AGENT 7→ N) ×(AGENT 7→ seq BIT ) →(P AGENT ×(AGENT 7֌ P OSIT ION) ×(P OSIT ION 7֌ N) ×(AGENT 7֌ N)) ∀ agent : AGENT ; population : P AGENT ; positions : AGENT 7֌ P OSIT ION; sugar : P OSIT ION 7֌ NagentSugar : AGENT 7֌ N; vision : AGENT 7→ N; culture : AGENT 7→ seq BIT • singleF ight(agent, population, positions, sugar, agentSugar, vision, culture) = if (availMoves(agent, positions, sugar, agentSugar, culture, vision(agent)) = ∅) then (population, positions, sugar, agentSugar) else ∃ loc : P OSIT ION; available : P P OSIT ION; ∀ otherLoc : P OSIT ION loc, otherLoc ∈ availMoves(agent, positions, sugar, agentSugar, culture, vision(agent)) ∧ otherLocation 6= location • (distance(position(agent), loc) > distance(position(otherLoc), position′(agent)) ⇒ reward(loc, sugar, position, agentSugar, COMBAT LIMIT ) ≥ reward(otherLoc, sugar, position, agentSugar, COMBAT LIMIT ) reward(loc, sugar, position, agentSugar, COMBAT LIMIT ) > reward(otherLoc, sugar, position, agentSugar, COMBAT LIMIT )) • (population \ {positions ∼(loc)}, (positions −⊲ {loc}) ⊕ {agent 7→ loc}, sugar ⊕ {loc 7→ 0}, agentSugar ⊕ {agent 7→ agentSugar(agent)+ reward(position′(agent), sugar, position, agentSugar, COMBAT LIMIT )}) 4.7. Disease. Disease is a simple rule that follows the standard pattern for AU specification. We place all agents into a sequence, ordered according to the vari- ation of AU we are using, and apply the rule to each agent in turn updating the state as we go along. THE SPECIFICATION OF SUGARSCAPE 63 T ransmission ∆Agents loanBook′ = loanBook population′ = population sex′ = sex position′ = position vision′ = vision age′ = age maxAge′ = maxAge agentCulture′ = agentCulture agentImmunity ′ = agentImmunity children′ = children agentSugar′ = agentSugar metabolism′ = metabolism initialSugar′ = initialSugar diseases′ = applyT ransmission(rndNewSweep(position), diseases, position) (1) (1) Call recursive applyT ransmission on each agent in population in deter- mined order. applyT ransmission : seq AGENT ×AGENT 7→ P seq BIT ×AGENT ֌ P OSIT ION ↔AGENT 7→ P seq BIT ∀ head : AGENT ; tail : seq AGENT ; diseases : AGENT 7→ P seq BIT ; position : AGENT ֌ P OSIT ION ∃ newInf ections : P seq BIT newInf ections applyT ransmission(hi, diseases, position) = applyT ransmission(hheadi a tail, diseases, position) = = newDiseases(asSeq(visibleAgents(head, position, 1)), diseases) • diseases (1) (2) (3) applyT ransmission(tail, diseases ⊕ {head 7→ (diseases(head) ∪ newInf ections)} , position) (1) Construct a set of new infections for an agent using the previously defined newDiseases; (2) Base case: Noting to do, return new disease mapping; 64 JOSEPH KEHOE (3) Recursive case: Add new diseases to the first agent in the list (according to the rule definition) and then recursively apply the rule to the rest of the list. 4.8. Culture. Culture is specified in an identical manner to Disease. AsyncCulture ∆Agents population′ = population sex′ = sex position′ = position vision′ = vision age′ = age maxAge′ = maxAge agentSugar′ = agentSugar children′ = children loanBook′ = loanBook diseases′ = diseases metabolism′ = metabolism agentImmunity ′ = agentImmunity initialSugar′ = initialSugar agentCulture′ = applyCulture(rndNewSweep(position), agentCulture, position) (1) Call recursive applyCulture on each agent in population in determined order. applyCulture : seq AGENT ×AGENT 7→ seq BIT ×AGENT ֌ P OSIT ION ↔AGENT 7→ seq BIT ∀ head : AGENT ; tail : seq AGENT ; culture : AGENT 7→ seq BIT ; position : AGENT ֌ P OSIT ION; ∃ n : P AGENT n = neighbours(head, position(head), 1) • applyCulture(hi, culture, position) = applyCulture(hheadi a tail, culture, position) = applyCulture(tail, culture ⊕ {head 7→ f lipT ags(culture(head), asSeq(n), culture)} , position) culture (1) (2) (1) Base case: return new values for culture tags; THE SPECIFICATION OF SUGARSCAPE 65 (2) Recursive case: Flip the tags of the first agent in the list and repeat (re- cursively) for the remaining agents in the list (sequence). 4.9. Inheritance. Inheritance also follows the same pattern as Culture and Disease. AsyncInheritance ∆Agents population′ = population sex′ = sex position′ = position vision′ = vision age′ = age maxAge′ = maxAge agentCulture′ = agentCulture children′ = children metabolism′ = metabolism diseases′ = diseases agentImmunity ′ = agentImmunity initialSugar′ = initialSugar (loanBook′, agentSugar′) = (1) applyInheritance(rndNewSweep(position), children, loanBook, agentSugar) (1) Use recursive applyInheritance function to calculate inheritance based on each agent in turn. 66 JOSEPH KEHOE applyInheritance : seq AGENT ×AGENT 7→ P AGENT ×(AGENT ↔ (AGENT × (N × N))) ×AGENT 7→ N ×AGENT 7→ N ×AGENT 7→ N 7→(AGENT ↔ (AGENT × (N × N))) ×AGENT 7→ N ∀ head : AGENT ; tail : seq AGENT ; children : AGENT 7→ P AGENT ; loans : AGENT ↔ (AGENT × (N × N)); agentSugar : AGENT 7→ N; age, maxAge : AGENT 7→ N ∃ newLoans : AGENT ↔ (AGENT × (N × N)); newAgentSugar : AGENT 7→ N (1) newAgentSugar = agentSugar ⊕ ({head 7→ 0} ∪ {(a, amt) a ∈ children(head) ∧ amt = agentSugar(a) +agentSugar(head)/ # children(head)}) Children(head)) • ∧ newLoans = ({head} −⊳ loans) ∪ oneAgentLoans(a, asSeq(ran({a} ⊳ loans)), applyInheritance(hi, children, loans, agentSugar, age, maxAge) = applyInheritance(hheadi a tail, children, loans, agentSugar, age, maxAge) = if (age(head) = maxAge(head) ∨ agentSugar(head) = 0) then else applyInheritance(tail, children, newLoans, newAgentSugar, age, maxAge) (loans, agentSugar) (2) (3) (4) applyInheritance(tail, children, loans, agentSugar, age, maxAge) (1) Distribute the dying agents sugar equally amongst its children; (2) Distribute any loans where the dying agent is the lender equally amongst its children; (3) Base case of recursion. Nothing to do but return results; (4) Recursive case: If the first agent in the list is dying then handle that agents inheritance and recurse through the rest of the agents otherwise just ignore it and apply the rule to rest of agents. 4.10. Mating. The AU specification of Mating is simpler than the SU version as it does not have to construct conflict free sets. It just puts all of the potential pairs in a sequence ordered according to the variant of AU we are using and applies the rule to each in turn. THE SPECIFICATION OF SUGARSCAPE 67 AsyncAgentMating ΞLattice ∆Agents loanBook′ = loanBook ∃ potentialMatingP airs : P(AGENT × AGENT ) potentialMatingP airs = {(a : AGENT, b : AGENT ) sex(a) 6= sex(b) ∧ isF ertile(age(a), sex(a)) ∧ isF ertile(age(head), sex(head)) ∧ adjacent(position(a), position(head))} metabolism′, children′, diseases′, agentImmunity ′, age′, sex′, initialSugar′) = (1) (population′, position′, vision′, agentSugar′, agentCulture′, applyMating(rndNewSweep(potentialMatingP airs), population, position, vision, agentSugar, agentCulture, metabolism, children, diseases, agentImmunity, age, maxAge, sex, initialSugar) (1) Call applyMating function on the agents in sequence. 4.11. Credit. MakeLoans ∆Agents ΞStep population′ = population sex′ = sex position′ = position vision′ = vision age′ = age maxAge′ = maxAge agentCulture′ = agentCulture agentImmunity ′ = agentImmunity diseases′ = diseases children′ = children metabolism′ = metabolism initialSugar′ = initialSugar (loanBook′, agentSugar′) = (1) applyLoans(rndNewSweep(position), population, position, agentSugar, age, sex, loanBook, step (1) Call applyLoans on each agent in turn. 68 JOSEPH KEHOE applyLoans : seq AGENT × P AGENT ×AGENT ֌ P OSIT ION ×AGENT 7→ N ×AGENT 7→ N ×AGENT 7→ SEX ×(AGENT ↔ (AGENT × (N × N))) ×N →(AGENT ↔ (AGENT × (N × N))) ×AGENT 7→ N ∀ population : P AGENT ; position : AGENT ֌ P OSIT ION; sex : AGENT 7→ SEX; age : AGENT 7→ N; agentSugar : AGENT 7→ N; head, ag : AGENT ; loanBook, loans : (AGENT ↔ (AGENT × (N × N))); tail : seq AGENT ; step : N; ∃ newAgentSugar : AGENT 7→ N; newLoans : (AGENT ↔ (AGENT × (N × N))); neighbours : P AGENT (1) neighbours = {b : AGENT vonNeumanNeighbour(lender, b, position)} (newLoans, newAgentSugar) = singleLenderLoans(lender, asSeq(neighbours), agentSugar, age, sex, loanBook, step) (2) (loanBook, agentSugar) applyLoans(hi, population, position, agentSugar, age, sex, loanBook, step) = applyLoans(hlenderi a tail, population, position, (3) applyLoans(tail, population, position, newAgentSugar, age, sex, newLoans, step) agentSugar, age, sex, loanBook, step) = (1) Construct the set of neighbours of an agent lender, the updated loan book and the updated sugar levels gotten by the lender giving loans to its neigh- bours; (2) Base Case: Nothing to do just return existing values; (3) Recursive case: Recursively call applyLoans on the remainder of the agents (excluding the first agent lender) and the new loan and sugar levels gotten by lender generating new loans. THE SPECIFICATION OF SUGARSCAPE 69 singleLenderLoans : AGENT × seq AGENT × P AGENT ×AGENT 7→ N ×AGENT 7→ N ×AGENT 7→ SEX ×(AGENT ↔ (AGENT × (N × N))) ×N →(AGENT ↔ (AGENT × (N × N))) ×AGENT 7→ N ∀ sex : AGENT 7→ SEX; age : AGENT 7→ N; agentSugar : AGENT 7→ N; head, lender : AGENT ; loanBook : (AGENT ↔ (AGENT × (N × N))); tail : seq AGENT ; step : N; singleLenderLoans(lender,hi, agentSugar, age, sex, loanBook, step) = singleLenderLoans(lender,hheadi a tail, agentSugar, age, sex, loans, step) = if canLend(age(lender), sex(lender), agentSugar(lender)) (loanBook, agentSugar) then ∧ willBorrow(age(head), sex(head), agentSugar(head),{head} ⊳ ran(loanBook))) ∃ newAgentSugar : AGENT 7→ N; newLoans : (AGENT × (AGENT × (N × N))); amt : N amt = min(amtAvail(age(lender), sex(lender), agentSugar(lender)), amtReq(agentSugar(head))) newAgentSugar = agentSugar ⊕ {lender 7→ agentSugar(lender) − amt, newLoans = loanBook ∪ {(lender, (head, (amt, step + DURAT ION)} singleLenderLoans(lender, tail, newAgentSugar, age, sex, newLoans, step) head 7→ agentSugar(head) + amt} else singleLenderLoans(lender, tail, agentSugar, age, sex, loanBook, step) singleLenderLoans calculates all loans that a particular agent can give to its neighbours. (1) If there are no loans in the sequence then just return the current loans and sugar levels as is; (2) If the loan sequence is not empty then apply the payment details to the first loan and make the payments on the rest: 70 JOSEPH KEHOE a) If the first loan is capable of being paid by the borrower we simply move the correct amount of sugar from the borrower to the lender; b) If the borrower cannot pay off the loan then they pay back half their sugar and the loan is renegotiated for the remainder. Using these functions we can now specify the P ayLoans part of the Credit rule. P ayLoans ∆Agents ΞStep population′ = population sex′ = sex position′ = position vision′ = vision age′ = age maxAge′ = maxAge agentCulture′ = agentCulture agentImmunity ′ = agentImmunity children′ = children diseases′ = diseases metabolism′ = metabolism initialSugar′ = initialSugar ∃ dueLoans, newLoans : P(AGENT × (AGENT × (N × N))) • dueLoans = loanBook ⊲ (ran(loanBook) ⊲ {a : (N × N) second(a) = step}) (1) (newLoans, agentSugar′) = makeP ayments(asSeq(dueLoans), ∅, agentSugar)(2) loanBook′ = (loanBook \ dueLoans) ∪ newLoans (3) (1) We create the set of due loans; (2) We now create the set of renegotiated loans and update the agentSugar levels using the makeP ayments function; (3) Finally we update the loan book by removing all loans that were due and adding any new renegotiated loans. 5. Added Spice 5.1. Introduction. We have defined all the rules so far under the assumption that there is only one resource (known as sugar). The final rule, T rade, is only defined for simulations with at least two resources. In fact the rules are meant to be general enough that they will work with any number of resources although we know of no sugarscape based simulation that used more than two resources. The second resource is known as spice. We will show how to extend the rules to deal with two resources. In order to avoid unnecessary clutter and make the differences as clear as possible we will THE SPECIFICATION OF SUGARSCAPE 71 show the differences between the one and two resource schemas with boldface. Any part of a schema that is not an exact copy of the previously defined version will be in boldface. 5.2. Basic Types. The basic types are copies of those already defined for sugar. MAXSP ICEMET ABOLISM : N (1) SP ICEGROW T H : N (2) MAXSP ICE : N (3) INIT IALSP ICEMIN, INIT IALSP ICEMAX : N (4) SP ICEP RODUCT ION, SP ICECONSUMP T ION : N(5) (6) SP ICECOMBAT LIMIT : N SP ICECHILDAMT : N (7) (1) Agents metabolise spice during each move at an individually set rate less than MAXSP ICEMET ABOLISM; (2) Spice grows back at a predefined rate; (3) Each location can hold a set maximum amount of spice; (4) Agents created after mating start with an initial spice endowment; (5) Pollution can be caused by production and consumption of spice; (6) SP ICECOMBAT LIMIT is required to help determine the reward from attacking an agent using the combat rule. (7) We posit that a minimum amount of spice is needed for agent mating to occur. Note that these constants are replicas of their sugar counterparts. 5.3. The SpiceScape. The spice grid contains everything in the Lattice scheme and just adds information on the extra spice resource. SpiceLattice Lattice spice : POSITION 7→ N maxSpice : POSITION 7→ N dom spice = dom maxSpice = POSITION ∀ x : POSITION • spice(x) ≤ maxSpice(x) ≤ MAXSPICE (1) (2) (1) Every location has an associated amount of spice and maximum carrying capacity; (2) Every position's spice levels are within the acceptable levels. 5.4. Agents. 72 JOSEPH KEHOE SpiceAgents Agents agentSpice : AGENT 7→ N initialSpice : AGENT 7→ N spiceMetabolism : AGENT 7→ N spiceLoanBook : AGENT ↔ (AGENT × (N × N)) dom spiceMetabolism = dom agentSpice == dom initialtSpicepopulation(1) (2) dom spiceLoanBook ⊆ population dom(ran spiceLoanBook) ⊆ population ∀ x : AGENT • x ∈ population ⇒ spiceMetabolism(x) ≤ MAXSPICEMETABOLISM (3) (1) Every agent has a spice metabolism and a spice store; (2) The spiceLoanBook has the same invariants as the original loanBook; (3) Every agents metabolism is less than or equal to the defined maximum. Finally we combine them into an overall schema as before: SpiceScape SpiceAgents SpiceLattice Step The initialisation scheme and tick schemas are also largely unchanged. THE SPECIFICATION OF SUGARSCAPE 73 InitialSpiceScape SpiceAgents ′ SpiceLattice ′ Step ′ step′ = 0 # population′ = INIT IALP OP ULAT IONSIZE loanBook′ = ∅ spiceLoanBook′ = ∅ ∀ a : AGENT • a ∈ population′ ⇒ age′(a) = 0 diseases′(a) = ∅ children′(a) = ∅ INIT IALSUGARMIN ≤ agentSugar′(a) ≤ INIT IALSUGARMAX INITIALSPICEMIN ≤ agentSpice′(a) ≤ INITIALSPICEMAX initialSugar′(a) = agentSugar′(a) initialSpice′(a) = agentSpice′(a) 74 JOSEPH KEHOE T ickspice ∆SpiceAgents ∆Step population′ = population position′ = position sex′ = sex vision′ = vision maxAge′ = maxAge metabolism′ = metabolism initialSugar′ = initialSugar initialSpice′ = initialSpice spiceMetabolism′ = spiceMetabolism agentCulture′ = agentCulture children′ = children loanBook′ = loanBook spiceLoanBook′ = spiceLoanBook agentImmunity ′ = agentImmunity diseases′ = diseases step′ = step + 1 ∀ x : AGENT • x ∈ population ⇒ (age′(x) = age(x) + 1 ∧ agentSugar′(x) = agentSugar(x) − metabolism(x) ∧ agentSpice′(x) = agentSpice(x) − spiceMetabolism(x)) 5.5. Rules. As well as defining the final rule, T rade, we will also expand the other rules to allow them to operate on a simulation with two resources. We define the new rule (T rade) first. 5.6. Agent Trade T . Agent Trade T : else continue; • Agent and neighbour compute their MRSs; if these are equal then end, • The direction of exchange is as follows: spice flows from the agent with the higher MRS to the agent with the lower MRS while sugar goes in the opposite direction; the bargaining price, p; • The geometric mean of the two MRSs is calculated-this will serve as • The quantities to be exchanged are as follows: if p>1 the p units of spice for 1 unit of sugar; if p < 1 the 1/p units of sugar for 1 unit of spice; THE SPECIFICATION OF SUGARSCAPE 75 • If this trade will (a) make both agents better off (increases the welfare of both agents), and (b) not cause the agents' MRSs to cross over one another, then the trade is made and return to start, else end. MRS is calculated simply for an agent as the fraction obtained by dividing its spice level times its sugar metabolism by its spice metabolism times its sugar level, as set out below. MRS : N × N × N × N ֌ A ∀ sugar, sugarMetabolism, spice, spiceMetabolism : N • MRS(sugar, sugarMetabolism, spice, spiceMetabolism) = (spice ∗ sugarMetabolism)/(spiceMetabolism ∗ sugar) tradeP airs constructs a sequence of all possible trading partners based on the proximity of the agents to each other. tradeP airs : seq AGENT × AGENT 7֌ P OSIT ION ֌ seq(AGENT × AGENT ) ∀ tail : seq AGENT ; positions : AGENT 7֌ P OSIT ION; a : AGENT • tradeP airs(hi, positions) = hi tradeP airs(hai a tail, positions) = trade(tail, positions a asSeq({b : agent adjacent(position(a), position(b)) • (a, b)}) 76 JOSEPH KEHOE T rade ∆SpiceAgents spiceLoanBook′ = spiceLoanBook loanBook′ = loanBook population′ = population initialSugar′ = initialSugar initialSpice′ = initialSpice sex′ = sex metabolism′ = metabolism spiceMetabolism′ = spiceMetabolism position′ = position vision′ = vision age′ = age maxAge′ = maxAge agentCulture′ = agentCulture agentImmunity ′ = agentImmunity children′ = children ∃ allP airs : seq(AGENT, AGENT ) allP airs = tradeP airs(asSeq(population), position) (1) (agentSugar′, agentSpice′) = executeT rades(allT rades(chooseExclusiveT rades(allP airs), (agentSugar, agentSpice), metabolism, spiceMetabolism)) T rade is similar to Mating in that trading must be done in exclusive pairs. An agent cannot carry out two simultaneous trades and the rule forces each agent to trade with all its neighbours in some sequence. As with Mating we construct conflict free sets of trading pairs that can proceed simultaneously and then order these sets. chooseExclusiveT rades : AGENT × AGENT ↔ seq P(AGENT, AGENT ) ∀ a, b : AGENT ; tradingP airs : AGENT ↔ AGENT • chooseExclusiveT rades(∅) = hi chooseExclusiveT rades(tradingP airs) = ∃ maxSet : AGENT ↔ AGENT maxSet ⊆ tradingP airs ∧ ((a, b) ∈ tradingP airs ∧ (a, b) 6∈ maxSet) ⇔ ∃ c : AGENT {(a, c), (c, a), (b, c), (c, b)} ∩ maxSet 6= ∅) hmaxSeti a chooseExclusiveT rades(tradingP airs \ maxSet) THE SPECIFICATION OF SUGARSCAPE 77 executeT rades : seq(AGENT × AGENT ) × AGENT 7→ N × AGENT 7→ N ↔(AGENT 7→ N, AGENT 7→ N) ∀ tail : seq(AGENT × AGENT ); head : (AGENT × AGENT ); agentSugar, agentSpice, metabolism, spiceMetabolism : AGENT 7→ N • executeT rades(hi, agentSugar, agentSpice, metabolism, spiceMetabolism) = executeT rades(hheadi a tail, agentSugar, sugar) = ∃ newAgentSugar, newAgentSpice : AGENT 7→ N (newAgentSugar, newAgentSpice) = • executeT rades(tail, newAgentSugar, newAgentSpice, metabolism, spiceMetabolism) allT rades(head, (agentSugar, agentSpice, metabolism, spiceMetabolism) (agentSugar, agentSpice) allT rades recursively goes through the sequence of trading partners and gets each trading pair to update the sugar and spice levels based on their trades. allT rades : seq(AGENT × AGENT ) × ((AGENT 7֌ N) × (AGENT 7֌ N))× (AGENT 7֌ N) × (AGENT 7֌ N) ↔((AGENT 7֌ N) × (AGENT 7֌ N)) ∀ head : AGENT × AGENT ; tail : seq(AGENT × AGENT ) sugar, spice, sugarMetabolism, spiceMetabolism : AGENT 7֌ N • allT rades(hi, (sugar, spice), sugarMetabolism, spiceMetabolism) = (sugar, spice) allT rades(hheadi a tail, (sugar, spice), sugarMetabolism, spiceMetabolism) = allT rades(tail, pairT rade(head, (sugar, spice), sugarMetabolism, spiceMetabolism) , sugarMetabolism, spiceMetabolism) Each trading partnership will execute a series of trades until their MRS scores cross over. pairT rade is complicated by the fact that there are multiple options: (1) If their MRS scores are equal then they perform no trades; (2) If their MRS scores are not equal then the direction of trade will depend on which MRS score is higher; a) Within a trade a value of p (based on MRS scores) determines the price of the resources. (3) Trades between the pair continue until their new and old MRS scores cross over. (mrsA > mrsB ∧ p ≤ 1) ⇒ (newSugar = sugar ⊕ {(a, sugar(a) + 1), (b, sugar(b) − 1)} ∧ newSpice = spice ⊕ {a 7→ spice(a) − p, b 7→ spice(b) + p}) (newSugar = sugar ⊕ {a 7→ sugar(a) + (1 div p), b 7→ sugar(b) − (1 div p)} newSpice = spice ⊕ {a 7→ spice(a) − 1, b 7→ spice(b) + 1}) (newSugar = sugar ⊕ {b 7→ sugar(b) + 1, a 7→ sugar(a) − 1} ∧ newSpice = spice ⊕ {b 7→ spice(b) − p, a 7→ spice(a) + p}) (newSugar = sugar ⊕ {b 7→ sugar(b) + (1 div p), a 7→ sugar(a) − (1 div p)} ∧ newSpice = spice ⊕ {b 7→ spice(b) − 1, a 7→ spice(a) + 1}) (mrsA ≤ mrsB ∧ p ≤ 1) ⇒ (mrsA ≤ mrsB ∧ p > 1) ⇒ 78 JOSEPH KEHOE pairT rade : (AGENT × AGENT ) × ((AGENT 7֌ N) × (AGENT 7֌ N))× (AGENT 7֌ N) × (AGENT 7֌ N) ↔((AGENT 7֌ N) × (AGENT 7֌ N)) ∀ a, b : AGENT ; sugar, spice : AGENT 7֌ N; metabolism, spiceMetabolism : AGENT 7֌ N ∃ p, mrsA, mrsB, newMrsA, newMrsB : A; newSugar, newSpice : AGENT 7֌ N mrsA = MRS(sugar(a), metabolism(a), spice(a), spiceMetabolism(a)) mrsB = MRS(sugar(b), metabolism(b), spice(b), spiceMetabolism(b)) p = √mrsA ∗ mrsB (mrsA > mrsB ∧ p > 1) ⇒ newMrsA = MRS(newSugar(a), metabolism(a), newSpice(a), spiceMetabolism(a)) newMrsB = MRS(newSugar(b), metabolism(b), newSpice(b), spiceMetabolism(b)) •pairT rade((a, b), (sugar, spice), metabolism, spiceMetabolism) = (sugar, spice) pairT rade((a, b), (sugar, spice), metabolism, spiceMetabolism) = (newSugar, newSpice) pairT rade((a, b), (sugar, spice), metabolism, spiceMetabolism) = ⇔ mrsA = mrsB ⇔ ((mrsA > mrsB ∧ newMrsA ≤ newMrsB) ∨ (mrsA < mrsB ∧ newMrsA ≥ newMrsB)) pairT rade((a, b), (newSugar, newSpice), metabolism, spiceMetabolism) ⇔ ((mrsA > mrsB ∧ newMrsA > newMrsB) ∨ (mrsA < mrsB ∧ newMrsA < newMrsB)) 5.7. Asynchronous Trade. THE SPECIFICATION OF SUGARSCAPE 79 T rade ∆SpiceAgents spiceLoanBook′ = spiceLoanBook loanBook′ = loanBook population′ = population sex′ = sex metabolism′ = metabolism spiceMetabolism′ = spiceMetabolism initialSugar′ = initialSugar initialSpice′ = initialSpice position′ = position vision′ = vision age′ = age maxAge′ = maxAge agentCulture′ = agentCulture agentImmunity ′ = agentImmunity children′ = children ∃ traders : P(AGENT × AGENT ) • (agentSugar′, agentSpice′) = (1) allT rades(tradeP airs(rndNewSweep(position), positions), (agentSugar, agentSpice), metabolism, spiceMetabolism) (1) The new sugar and spice allocations are derived by conducting all possible trades using the recursive helper function allT rades. 5.7.1. Growback. Growbackspice ∆SpiceLattice pollution′ = pollution maxSugar′ = maxSugar maxSpice′ = maxSpice sugar′ = sugar ⊕ {x : P OSIT ION • spice′ = spice ⊕ {x : POSITION • x 7→ min({sugar(x) + SUGARGROW T H, maxSugar(x)})} x 7→ min({spice(x) + SPICEGROWTH, maxSpice(x)})} (1) (2) (3) (1) Lattice is replaced with SpiceLattice. In all subsequent schemas SpiceLattice replaces Lattice and SpiceAgent replaces Agent; (2) maxSpice remains unchanged; (3) The new spice levels are calculated using the same simple formula used for sugar growback. 80 JOSEPH KEHOE 5.7.2. Seasonal Growback. SeasonalGrowbackspice ∆SpiceLattice ΞStep pollution′ = pollution maxSugar′ = maxSugar maxSpice′ = maxSpice ∀ x : P OSIT ION • (step div SEASONLENGT H) mod 2 = 0 ⇒ sugar′ = {x : P OSIT ION f irst(x) < M div 2 • x 7→ min({sugar(x) + SUGARGROW T H, maxSugar(x)})} ∪ {x : P OSIT ION f irst(x) ≥ M div 2 • x 7→ min({sugar(x) + SUGARGROW T H div W INT ERRAT E, maxSugar(x)})} ∀ x : P OSIT ION • (step div SEASONLENGT H) mod 2 6= 0 ⇒ sugar′ = {x : P OSIT ION f irst(x) < M div 2 • x 7→ min({sugar(x) + SUGARGROW T H div W INT ERRAT E, maxSugar(x)})} ∪ {x : P OSIT ION; y : N f irst(x) ≥ M div 2 • x 7→ min({sugar(x) + SUGARGROW T H, maxSugar(x)})} (1) ∀ x : POSITION • (step div SEASONLENGTH) mod 2 = 0 ⇒ spice′ = {x : POSITION first(x) < M div 2 • x 7→ min({spice(x) + SPICEGROWTH, maxSpice(x)})} ∪ {x : POSITION first(x) ≥ M div 2 • x 7→ min({spice(x) + SPICEGROWTH div WINTERRATE, maxSpice(x)})} ∀ x : POSITION • (step div SEASONLENGTH) mod 2 6= 0 ⇒ spice′ = {x : POSITION first(x) < M div 2 • x 7→ min({spice(x) + SPICEGROWTH div WINTERRATE, maxSpice(x)})} ∪ {x : POSITION first(x) ≥ M div 2 • x 7→ min({spice(x) + SPICEGROWTH, maxSpice(x)})} (1) Seasonal growback adds a rule for spice grow back that is an exact replica of the sugar rule. We note that we only use the one W INT ERRAT E instead of a separate rate for sugar and spice. In the absence of any explicit direction on this point this solution seems to be the most obvious. THE SPECIFICATION OF SUGARSCAPE 81 5.8. Movement. In order to update Movement we will need to implement a wel- fare function that can be used to measure the desirability of a location. With two resources the desirability of any location becomes a subjective measure, what one agent may rate highly another may not. Welfare is dependent on the agents current levels of spice and sugar, so an agent with low spice levels may consider a location containing spice more desirable than one containing sugar. Overall the desirability of a location is determined by the agents current resource levels (wealth) and the relative metabolism rates for each resource. This is in contrast to the previous approach where welfare just equaled the amount of sugar in a location. This welfare measure is precisely defined in the book and we follow this definition. welf are : N × N × N × N × N × N ֌ A ∀ agentSugar, sugarMetabolism, agentSpice, spiceMetabolism, locationSugar, locationSpice : N • locationSugar, locationSpice) = (locationSugar + agentSugar) ∗ (sugarMetabolism div ∗(locationSpice + agentSpice) ∗ (spiceMetabolismdiv (sugarMetabolism + spiceMetabolism)) welf are(agentSugar, sugarMetabolism, agentSpice, spiceMetabolism, (sugarMetabolism + spiceMetabolism)) Movement can now be restated by replacing the previous measure of a locations desirability (sugar level) with this new measure and the updating of spice levels. In all other respects the schema remains unchanged. 82 JOSEPH KEHOE MovementbasicSpice ∆SpiceScape step′ = step population′ = population maxSugar′ = maxSugar pollution′ = pollution sex′ = sex vision′ = vision age′ = age initialSugar′ = initialSugar initialSpice′ = initialSpice metabolism′ = metabolism maxAge′ = maxAge agentCulture′ = agentCulture loanBook′ = loanBook diseases′ = diseases agentImmunity ′ = agentImmunity children′ = children maxSpice′ = maxSpice spiceLoanBook′ = spiceLoanBook spiceMetabolism′ = spiceMetabolism ∀ a : AGENT ; l : P OSIT ION • a ∈ population ⇒ distance(position′(a), position(a)) ≤ vision(a) distance(position(a), l) ≤ vision(a) ∧ (l 6∈ ran position′)) agentSpice(a), spiceMetabolism(a), sugar(l), spice(l)) ⇒ welfare(agentSugar(a), metabolism(a), < welfare(agentSugar(a), metabolism(a), agentSpice(a), spiceMetabolism(a), (1) (2) sugar(position′(a)), spice(position′(a))) agentSugar′ = {a : AGENT a ∈ population • a 7→ agentSugar(a) + sugar(position′(a))} sugar′ = sugar ⊕ {loc : P OSIT ION loc ∈ ran position′ • loc 7→ 0} agentSpice′ = {a : AGENT a ∈ population • a 7→ agentSpice(a) +spice(position′(a))} spice′ = spice ⊕ {loc : POSITION loc ∈ ran position′ • loc 7→ 0} (1) This is a copy of the original proposition with the welf are function now (3) replacing the previous sugar level check; (2) Agents consume spice at their new locations; (3) Locations with agents present now have no remaining spice. THE SPECIFICATION OF SUGARSCAPE 83 5.9. Pollution Formation PΠ,χ. The MovementspiceP ollution schema has the same alterations as the MovementbasicSpice 84 JOSEPH KEHOE MovementpollutionSpice ∆SpiceScape step′ = step population′ = population maxSugar′ = maxSugar sex′ = sex pollution′ = pollution vision′ = vision age′ = age maxAge′ = maxAge agentCulture′ = agentCulture loanBook′ = loanBook children′ = children agentImmunity ′ = agentImmunity initialSugar′ = initialSugar initialSpice′ = initialSpice diseases′ = diseases metabolism′ = metabolism spiceMetabolism′ = spiceMetabolism spiceLoanBook′ = spiceLoanBook maxSpice′ = maxSpice ∀ a : AGENT ; l : P OSIT ION a ∈ dom(position′) • a ∈ population ⇒ distance(position′(a), position(a)) ≤ vision(a) (distance(position(a), l) ≤ vision(a) ∧ (l 6∈ ran position′)) ⇒ welfare(agentSugar(a), metabolism(a), agentSpice(a), spiceMetabolism(a), sugar(l), spice(l)) div(1 + pollution(l)) < welfare(agentSugar(a), metabolism(a), agentSpice(a), spiceMetabolism(a), sugar(position′(a)), spice(position′(a))) div (1 + pollution(position′(a))) agentSugar′ = {a : AGENT a ∈ population • a 7→ agentSugar(a) + sugar(position′(a))} sugar′ = sugar ⊕ {l : P OSIT ION l ∈ ran position′ • l 7→ 0} agentSpice′ = {a : AGENT a ∈ population • a 7→ agentSpice(a) + spice(position′(a))} spice′ = spice ⊕ {l : POSITION l ∈ ran position′ • l 7→ 0} pollution′ = pollution⊕ {l : P OSIT ION; x : AGENT position′(x) = l • l 7→ (P RODUCT ION ∗ sugar(l) + CONSUMP T ION ∗ metabolism(x))} THE SPECIFICATION OF SUGARSCAPE 85 5.10. Pollution Diffusion. P ollutionDif f usionspice ∆SpiceLattice ΞStep maxSugar′ = maxSugar sugar′ = sugar maxSpice′ = maxSpice spice′ = spice (step mod P OLLUT IONRAT E 6= 0) ⇒ pollution′ = pollution (step mod P OLLUT IONRAT E = 0) ⇒ pollution′ = {l : P OSIT ION • l 7→ (pollution(north(l)) + pollution(south(l)) +pollution(east(l)) + pollution(west(l))) div 4} 5.11. Replacement. ⊳ loanBook⊲ ∀ a : AGENT • Deathspice ∆SpiceAgents population′ = population \ {a : AGENT age(a) = maxAge(a) loanBook′ = population′ ∨ agentSugar(a) = 0∨ agentSpice(a) = 0 • a} {x : AGENT × (N × N) f irst(x) ∈ population′} ⊳ spiceLoanBook⊲ {x : AGENT × (N × N) first(x) ∈ population′} a ∈ population′ ⇒ spiceLoanBook′ = population′ (sex(a) = sex′(a) ∧ vision(a) = vision′(a) ∧ maxAge(a) = maxAge′(a) ∧ agentCulture(a) = agentCulture′(a) ∧ position(a) = position′(a) ∧ age(a) = age′(a) ∧ agentSugar(a) = agentSugar′(a) ∧ metabolism′(a) = metabolism(a) ∧ diseases′(a) = diseases(a) ∧ agentImmunity ′(a) = agentImmunity(a) ∧ children′(a) = children(a) ∧ initialSugar′(a) = initialSugar(a) ∧ agentSpice′(a) = agentSpice(a) ∧ initialSpice′(a) = initialSpice(a) ∧ spiceMetabolism′(a) = spiceMetabolism(a)) 86 JOSEPH KEHOE Replacementspice ∆SpiceAgents # population′ = INIT IALP OP ULAT IONSIZE loanBook′ = loanBook spiceLoanBook′ = spiceLoanBook ∀ a : AGENT • a ∈ (population) ⇒ (a ∈ population′ ∧ sex(a) = sex′(a) ∧ vision(a) = vision′(a) ∧ maxAge(a) = maxAge′(a) ∧ agentCulture(a) = agentCulture′(a) ∧ position(a) = position′(a) ∧ age(a) = age′(a) ∧ agentSugar′(a) = agentSugar(a) ∧ metabolism′(a) = metabolism(a) ∧ initialSugar′(a) = initialSugar(a) ∧ initialSpice′(a) = initialSpice(a) ∧ agentSpice′(a) = agentSpice(a) ∧ spiceMetabolism′(a) = spiceMetabolism(a) ∧ diseases′(a) = diseases(a) ∧ agentImmunity ′(a) = agentImmunity(a) ∧ children′(a) = children(a)) a ∈ population′ \ population ⇒ (age′(a) = 0 ∧ INIT IALSUGARMIN ≤ agentSugar′(a) ≤ INIT IALSUGARMAX ∧ initialSugar′(a) = agentSugar′(a) ∧ INITIALSPICEMIN ≤ agentSpice′(a) ≤ INITIALSPICEMAX ∧ initialSpice′(a) = agentSpice′(a) ∧ diseases′(a) = ∅ ∧ children′(a) = ∅) ∀ a : AGENT • 5.12. Agent Mating. THE SPECIFICATION OF SUGARSCAPE 87 AgentMating ΞLattice ∆Agents loanBook′ = loanBook ∧ spiceLoanBook′ = spiceLoanBook ∃ potentialMatingP airs : P(AGENT × AGENT ) potentialMatingP airs = {(a : AGENT, b : AGENT ) sex(a) 6= sex(b) ∧ isF ertile(age(a), sex(a)) ∧ isF ertile(age(head), sex(head)) ∧ adjacent(position(a), position(head))} , children′, diseases′, agentImmunity ′, age′, sex′, initialSugar′, (population′, position′, vision′, agentSugar′, agentCulture′, metabolism′ (1) spiceMetabolism′, agentSpice′, initialSpice′) = (2) concurrentMating(getConf ictF reeP airs(potentialMatingP airs), population, position, vision, agentSugar, agentCulture, metabolism, children, diseases, agentImmunity, age, maxAge, sex, initialSugar, spiceMetabolism, agentSpice, initialSpice) (1) Generate the set of all possible mating pairs; (2) Recursively proceed with concurrent mating within the conflict free subsets. 88 JOSEPH KEHOE concurrentMating : seq P(AGENT × AGENT ) × P AGENT ×AGENT ֌ P OSIT ION ×AGENT 7→ N1 ×AGENT 7→ N ×AGENT 7→ seq BIT ×AGENT 7→ N ×AGENT 7→ P AGENT ×AGENT 7→ P seq BIT ×AGENT 7→ seq BIT ×AGENT 7→ N ×AGENT 7→ N1 ×AGENT 7→ SEX ×AGENT 7→ N ×AGENT 7→ N ×AGENT 7→ N ×AGENT 7→ N ↔P AGENT ×AGENT ֌ P OSIT ION ×AGENT 7→ N1 ×AGENT 7→ N ×AGENT 7→ seq BIT ×AGENT 7→ N ×AGENT 7→ P AGENT ×AGENT 7→ P seq BIT ×AGENT 7→ seq BIT ×AGENT 7→ N ×AGENT 7→ N1 ×AGENT 7→ SEX ×AGENT 7→ N ×AGENT 7→ N ×AGENT 7→ N ×AGENT 7→ N ∀ tail : seq P(AGENT × AGENT ); head : P(AGENT × AGENT ); population : P AGENT ; position : AGENT ֌ P OSIT ION; vision : AGENT 7→ N1; agentSugar, agentSpice : AGENT 7→ N; agentCulture : AGENT 7→ seq BIT ; metabolism, spiceMetabolism : AGENT 7→ N; children : AGENT 7→ P AGENT ; diseases : AGENT 7→ P seq BIT ; agentImmunity : AGENT 7→ seq BIT ; age : AGENT 7→ N; maxAge : AGENT 7→ N1; sex : AGENT 7→ SEX; initialSugar, initialSpice : AGENT 7→ N; THE SPECIFICATION OF SUGARSCAPE 89 ∃ newpopulation : P AGENT ; newposition : AGENT ֌ P OSIT ION; newvision : AGENT 7→ N1; newagentSugar, newagentSpice : AGENT 7→ N; newagentCulture : AGENT 7→ seq BIT ; newmetabolism, newspiceMetabolism : AGENT 7→ N; newchildren : AGENT 7→ P AGENT ; newdiseases : AGENT 7→ P seq BIT ; newagentImmunity : AGENT 7→ seq BIT ; newage : AGENT 7→ N; newmaxAge : AGENT 7→ N1; newsex : AGENT 7→ SEX; newinitialSugar, newinitialSpice : AGENT 7→ N; (newpopulation, newposition, newvision, newagentSugar, newagentCulture, newmetabolism, newchildren, newdiseases, newagentImmunity, newage, newmaxAge, newsex, newinitialSugar, newspiceMetabolism, newagentSpice, newinitialspice) = applyMating(asSeq(head), population, position, vision, agentSugar, agentCulture, metabolism, children, diseases, agentImmunity, age, maxAge, sex, initialSugar, spiceMetabolism, agentSpice, initialspice) • agentSugar, agentCulture, metabolism, children, diseases, agentImmunity, age, maxAge, sex, initialSugar) = concurrentMating(hi, population, position, vision, (population, position, vision, agentSugar, agentCulture, metabolism, children, diseases, agentImmunity, age, maxAge, sex, initialSugar, spiceMetabolism, agentSpice, initialspice) concurrentMating(hheadi a tail, population, position, vision, agentSugar, agentCulture, metabolism, children, diseases, agentImmunity, age, maxAge, sex, initialSugar) = concurrentMating(tail, newpopulation, newposition, newvision, newagentSugar, newagentCulture, newmetabolism, newchildren, newdiseases, newagentImmunity, newage, newmaxAge, newsex, newinitialSugar, newspiceMetabolism, newagentSpice, newinitialspice) 90 JOSEPH KEHOE applyMating : seq(AGENT × AGENT ) × P AGENT ×AGENT ֌ P OSIT ION ×AGENT 7→ N1 ×AGENT 7→ N ×AGENT 7→ seq BIT ×AGENT 7→ N ×AGENT 7→ P AGENT ×AGENT 7→ P seq BIT ×AGENT 7→ seq BIT ×AGENT 7→ N ×AGENT 7→ N1 ×AGENT 7→ SEX ×AGENT 7→ N ×AGENT 7→ N ×AGENT 7→ N ×AGENT 7→ N ↔P AGENT ×AGENT ֌ P OSIT ION ×AGENT 7→ N1 ×AGENT 7→ N ×AGENT 7→ seq BIT ×AGENT 7→ N ×AGENT 7→ P AGENT ×AGENT 7→ P seq BIT ×AGENT 7→ seq BIT ×AGENT 7→ N ×AGENT 7→ N1 ×AGENT 7→ SEX ×AGENT 7→ N ×AGENT 7→ N ×AGENT 7→ N ×AGENT 7→ N ∀ population : P AGENT ; position : AGENT ֌ P OSIT ION; sex : AGENT 7→ SEX; vision : AGENT 7→ N1; age : AGENT 7→ N; initialSugar, initialSpice : AGENT 7→ N; maxAge : AGENT 7→ N1; metabolism, spiceMetabolism : AGENT 7→ N; agentSugar, agentSpice : AGENT 7→ N; agentCulture : AGENT 7→ seq BIT ; children : AGENT 7→ P AGENT ; agentImmunity : AGENT 7→ seq BIT ; diseases : AGENT 7→ P seq BIT ; head : AGENT × AGENT ; tail : seq(AGENT × AGENT ); • THE SPECIFICATION OF SUGARSCAPE 91 ∃ of f spring, a, b : AGENT ; newsex : AGENT 7→ SEX; newvision : AGENT 7→ N1; newmetabolism, newagentSugar, newinitialSugar : AGENT 7→ N; newspiceMetabolism, newagentSpice, newinitialSpice : AGENT 7→ N; newmaxAge : AGENT 7→ N1; newagentCulture : AGENT 7→ seq BIT ; newchildren : AGENT 7→ P AGENT ; newagentImmunity : AGENT 7→ seq BIT ; inheritedImmunity : seq BIT ; inheritedCulture : seq BIT ; of f spring 6∈ population a = f irst(head) ∧ b = second(head) newchildren : children ∪ {of f spring 7→ ∅ a 7→ children(a) ∪ {of f spring}, b 7→ children(b) ∪ {of f spring}} newsex = sex ∪ {of f spring 7→ male} ∨ newsex = sex ∪ {of f spring 7→ f emale} newvision = vision ∪ {of f spring 7→ vision(a)} ∨ newvision = vision ∪ {of f spring 7→ vision(b)} newmaxAge = maxAge ∪ {of f spring 7→ maxAge(a)} ∨ newmaxAge = maxAge ∪ {of f spring 7→ maxAge(b)} newmetabolism = metabolism ∪ {of f spring 7→ metabolism(a)} ∨ newmetabolism = metabolism ∪ {of f spring 7→ metabolism(b)} newspiceMetabolism = spiceMetabolism ∪ {offspring 7→ spiceMetabolism(a)} ∨ newspiceMetabolism = spiceMetabolism ∪ {offspring 7→ spiceMetabolism(b)} newinitialSugar = initialSugar⊕ {of f spring 7→ initialSugar(a)/2 + initialSugar(b)/2} newagentSugar = agentSugar ⊕ {of f spring 7→ initialSugar(of f spring), a 7→ initialSugar(a)/2, b newinitialSpice = initialSpice⊕ newagentSpice = agentSpice ∪ {offspring 7→ initialSpice(offspring), a 7→ initialSpice(a)/2, b {offspring 7→ initialSpice(a)/2 + initialSpice(b)/2} ∧ ∀ n : 1 . . IMMUNIT Y LENGT H • (inheritedImmunity(n) = agentImmunity(a)(n) ∨ inheritedImmunity(n) = agentImmunity(b)(n)) newagentImmunity : agentImmunity ∪ {of f spring 7→ inheritedImmunity} ∧ ∀ n : 1 . . CULT URECOUNT • (inheritedCulture(n) = agentCulture(a)(n) ∨ inheritedCulture(n) = agentCulture(b)(n)) newagentCulture : agentCulture ∪ {of f spring 7→ inheritedCulture} 92 JOSEPH KEHOE applyMating(hi, population, position, vision, agentSugar, agentCulture, metabolism, children, diseases, agentImmunity, age, maxAge, sex, initialSugar, spiceMetabolism, agentSpice, initialSpice) = (population, position, vision, agentSugar, agentCulture, metabolism, children, diseases, agentImmunity, age, maxAge, sex, initialSugar, spiceMetabolism, agentSpice, initialSpice) applyMating(hheadi a tail, population, position, vision, agentSugar, agentCulture, metabolism, children, diseases, agentImmunity, age, maxAge, sex, initialSugar, spiceMetabolism, agentSpice, initialSpice) = if ((∃ loc : P OSIT ION (adjacent(loc, position(ag))) ∨ adjacent(loc, position(head)) ∧ loc 6∈ dom position) ∧ (agentSpice(head) > initialSpice(head)) ∧ (agentSpice(ag) > initialSpice( applyMating(tail, population ∪ {of f spring}, position ∪ {of f spring 7→ loc}, newvision, newagentSugar, newagentCulture, newmetabolism, newchildren, diseases ∪ {of f spring 7→ ∅}, newagentImmunity, age ∪ {of f spring 7→ 0}, newmaxAge, newsex, initialSugar, applyMating(tail, population, position, vision, agentSugar, agentCulture, else newspiceMetabolism metabolism, children, diseases, agentImmunity, age, maxAge, sex, initialSugar, spiceMetabolism, agentSpice, initialSpice) 5.13. Culture. THE SPECIFICATION OF SUGARSCAPE 93 Culturespice ∆SpiceAgents population′ = population sex′ = sex position′ = position vision′ = vision age′ = age maxAge′ = maxAge agentSugar′ = agentSugar children′ = children loanBook′ = loanBook diseases′ = diseases initialSugar′ = initialSugar metabolism′ = metabolism agentImmunity ′ = agentImmunity agentSpice′ = agentSpice spiceMetabolism′ = spiceMetabolism spiceLoanBook′ = spiceLoanBook initialSpice′ = initialSpice ∀ a : AGENT • a ∈ population ⇒ agentCulture′(a) = f lipT ags(agentCulture(a), asSeq({b : AGENT adjacent(position(a), position(b))}), agentCulture) 5.14. Disease. 94 JOSEPH KEHOE ImmuneResponsespice ΞSpiceLattice ∆SpiceAgents ΞStep loanBook′ = loanBook spiceLoanBook′ = spiceLoanBook population′ = population sex′ = sex position′ = position vision′ = vision age′ = age maxAge′ = maxAge agentCulture′ = agentCulture diseases′ = diseases children′ = children agentSugar′ = agentSugar initialSugar′ = initialSugar initialSpice′ = initialSpice metabolism′ = metabolism agentSpice′ = agentSpice spiceMetabolism′ = spiceMetabolism agentImmunity ′ = {a : AGENT a ∈ population • a 7→ applyDiseases(agentImmunity(a), asSeq(diseases(a)))} ∀ x : AGENT • x ∈ population ⇒ agentSugar′(x) = agentSugar(x)− #{d : seq BIT d ∈ diseases(a) ∧ ¬ subseq(d, agentImmunity(a)} ∀ x : AGENT • x ∈ population ⇒ agentSpice′(x) = agentSpice(x)− #{d : seq BIT d ∈ diseases(a) ∧ ¬ subseq(d, agentImmunity(a)} THE SPECIFICATION OF SUGARSCAPE 95 T ransmissionspice ∆SpiceAgents loanBook′ = loanBook initialSugar′ = initialSugar spiceLoanBook′ = spiceLoanBook population′ = population sex′ = sex position′ = position vision′ = vision age′ = age maxAge′ = maxAge agentCulture′ = agentCulture agentImmunity ′ = agentImmunity children′ = children agentSugar′ = agentSugar agentSpice′ = agentSpice spiceMetabolism′ = spiceMetabolism initialSpice′ = initialSpice metabolism′ = metabolism ∀ a : AGENT • a ∈ population ⇒ diseases′(a) = diseases(a)∪ newDiseases(asSeq(visibleAgents(a, position, 1)), diseases) 5.15. Inheritance. 96 JOSEPH KEHOE Inheritancespice ∆SpiceAgents population′ = population ∧ sex′ = sex position′ = position ∧ vision′ = vision age′ = age ∧ maxAge′ = maxAge agentCulture′ = agentCulture ∧ children′ = children metabolism′ = metabolism ∧ spiceMetabolism′ = spiceMetabolism diseases′ = diseases ∧ agentImmunity ′ = agentImmunity initialSugar′ = initialSugar initialSpice′ = initialSpice ∃ dying : P AGENT ; inheritF romF emale, inheritF romMale : AGENT 7→ (N × N) • dom inheritF romF emale = dom inheritF romMale = population \ dying dying = {x : AGENT x ∈ population ∧ (age(x) = maxAge(x) ∨ agentSugar(x) = 0∨ agentSpice(x) = 0)} ∀ x : AGENT ; n, m : N x ∈ population \ dying • inheritF romMale(x) = (0, 0) getMother(x, children, sex) 6∈ dying ⇒ inheritF romF emale(x) = (0, 0) getF ather(x, children, sex) 6∈ dying ⇒ ∃ m : AGENT m = getMother(x, children, sex) ∧ m ∈ dying ⇒ (agentSugar(m) div #(population ∩ children(m) \ dying)) (agentSpice(m) div #(population ∩ children(m) \ dying)) inheritF romF emale(x) = ∃ d : AGENT d = getF ather(x, children, sex) ∧ d ∈ dying ⇒ inheritF romMale(x) = (1) (2) (3) agentSugar(d) div #(population ∩ children(d) \ dying)) (agentSpice(m) div #(population ∩ children(m) \ dying)) x ∈ dying x 6∈ dying ⇒ (agentSugar′(x) = 0∧ agentSpice′(x) = 0) ⇒ (agentSugar′(x) = agentSugar(x) +f irst(inheritF romMale(x)) + f irst(inheritF romF emale(x)) ∧ agentSpice′(x) = agentSpice(x) +second(inheritFromFemale(x)) + second(inheritFromMale(x))) loanBook′ = disperseLoans(loanBook, asSeq(dying), children) spiceLoanBook′ = disperseLoans(spiceLoanBook, asSeq(dying), children) (4) (1) Agents can now inherit two amounts, a sugar inheritance and a spice inheritance; THE SPECIFICATION OF SUGARSCAPE 97 (2) Death now occurs if either resource reaches zero; (3) The individual spice inheritance is calculated in the same way as the sugar inheritance; (4) Spice loans are dispersed amongst children. 5.16. Combat. Combat is defined only in terms of sugar. We can either accept this and assume combat is based only on sugar levels or we can extend combat by defining new versions of wealth and reward. We note that no simulations combining combat with more than one resource are presented in the book. We can extend the combat rule with a few simple assumptions. First the wealth of an agent is used to determine if we can attack that agent or if an agent can retaliate against us. In the single resource scenario we simply used the sugar that an agent carried. With two resources we need to combine both sugar and spice. The simplest approach is to add these two amounts together and in the absence of any guidelines this seems the sensible option. availMoves requires only minor changes to return the set of all safe moves that an agent can make. availMovesspice : AGENT × (AGENT 7֌ P OSIT ION) × (P OSIT ION 7֌ N) × (AGENT 7֌ N) ×(P OSIT ION 7֌ N) × (AGENT 7֌ N) × (AGENT 7֌ seq BIT ) × N 7→ P P OSIT ION ∀ x, agent : AGENT ; positions : AGENT 7֌ P OSIT ION; vision : N; sugar, spice : P OSIT ION 7֌ N; agentSpice, agentSugar : AGENT 7֌ N; culture : AGENT 7֌ seq BIT • availMovesspice(agent, positions, sugar, agentSugar, spice, agentSpice, culture, vision) = {l : P OSIT ION; x : AGENT l ∈ distance(l, positions(agent)) ≤ vision ∧ positions(x) = l ⇒ (agentSugar(x)+agentSpice(x) < agentSugar(agent) + agentSpice(ag) ∧ ((distance(positions(x), l) ≤ vision) ∧ tribe(culture(x)) 6= tribe(culture(agent))) ∧ tribe(culture(x)) 6= tribe(culture(agent))) ⇒ agentSugar(x)+agentSpice(x) < agentSugar(agent)+agentSpice(agent) +reward(l, sugar, positions, agentSugar, COMBAT LIMIT )) +reward(l, spice, positions, agentSpice, SPICECOMBATLIMIT) • l} (3) 98 JOSEPH KEHOE Combatspice ∆SugarScape step′ = step maxSugar′ = maxSugar maxSpice′ = maxSpice pollution′ = pollution loanBook′ = population′ spiceLoanBook′ = ⊳ loanBook ⊲ (population′ ⊳ (ran loanBook)) population′ ⊳ spiceLoanBook ⊲ (population′ ⊳ (ran spiceLoanBook)) (3) ∀ ag : AGENT ; l : P OSIT ION • ag ∈ population′ ⇒ (sex′(ag) = sex(ag) ∧ vision′(ag) = vision(ag) ∧ age′(ag) = age(ag) ∧ maxAge′(ag) = maxAge(ag) ∧ children′(ag) = children(ag) ∧ agentCulture′(ag) = agentCulture(ag) ∧ agentImmunity ′(ag) = agentImmunity(ag) ∧ metabolism′(ag) = metabolism(ag) ∧ initialSugar′(ag) = initialSugar(ag) ∧ initialSpice′(ag) = initialSpice(ag) ∧ spiceMetabolism′(ag) = spiceMetabolism(ag) ∧ diseases′(ag) = diseases(ag)) (population′, position′, sugar′, agentSugar′, agentSpice′) = applyAllCombatspice(asSeq(population), population, position, sugar, agentSugar, agentSpice, vision, agentCulture) THE SPECIFICATION OF SUGARSCAPE 99 applyAllCombatspice : seq AGENT × P AGENT × (AGENT 7֌ P OSIT ION) ×(AGENT 7֌ N) × (AGENT 7֌ N)) ×(P OSIT ION 7֌ N) × (AGENT 7֌ N) × (AGENT 7֌ N) ×(AGENT 7→ N) × (AGENT 7→ seq BIT ) →(P AGENT × (AGENT 7֌ P OSIT ION) × (P OSIT ION 7֌ N) ∀ head : AGENT ; tail : seq AGENT ; pop : P AGENT ; positions : AGENT 7֌ P OSIT ION; sugar : P OSIT ION 7֌ N; agSugar, agSpice : AGENT 7֌ N; vision : AGENT 7→ N; culture : AGENT 7→ seq BIT • applyAllCombat(hi, pop, positions, sugar, agSugar, agSpice, vision, culture) = (pop, positions, sugar, agSugar, agSpice) applyAllCombat(hheadi a tail, pop, positions, sugar, agSugar, agSpice, vision, culture) = if(head ∈ pop)then applyAllCombatspice(tail, singleF ightspice(head, pop, positions, sugar, agSugar, agSpice, vision, culture)) else applyAllCombatspice(tail, pop, positions, sugar, agSugar, agSpice, vision, culture) 100 JOSEPH KEHOE ×(AGENT 7֌ N) × (AGENT 7֌ N)) ×(AGENT 7֌ N) × (AGENT 7֌ N) × (AGENT 7→ N) × (AGENT 7→ seq BIT ) singleF ightspice : AGENT × P AGENT × (AGENT 7֌ P OSIT ION) × (P OSIT ION 7֌ N →(P AGENT × (AGENT 7֌ P OSIT ION) × (P OSIT ION 7֌ N) ∀ ag : AGENT ; population : P AGENT ; positions : AGENT 7֌ P OSIT ION; sugar : P OSIT ION 7֌ N; agSugar, agSpice : AGENT 7֌ N; vision : AGENT 7→ N; culture : AGENT 7→ seq BIT • singleF ightspice(ag, population, positions, sugar, agSugar, agSpice, vision, culture) = if (availMoves(ag, positions, sugar, agSugar, agSpice, culture, vision(ag)) = ∅) then (population, positions, sugar, agSugar, agSpice) else ∃ loc : P OSIT ION; available : P P OSIT ION; ∀ otherLoc : P OSIT ION loc, otherLoc ∈ ∧ otherLocation 6= location reward(loc, sugar, position, agSugar, COMBAT LIMIT ) availMovesspice(ag, positions, sugar, agSugar, agSpice, culture, vision(ag)) +reward(loc, spice, position, agSpice, SPICECOMBATLIMIT) +reward(otherLoc, spice, position, agSpice, SPICECOMBATLIMIT) ≥ reward(otherLoc, sugar, position, agSugar, COMBAT LIMIT ) (distance(position(ag), loc) > distance(position(otherLoc), position′(ag)) ⇒ reward(loc, sugar, position, agSugar, COMBAT LIMIT ) +reward(loc, sugar, position, agSpice, SPICECOMBATLIMIT) > reward(otherLoc, sugar, position, agSugar, COMBAT LIMIT ) +reward(otherLoc, spice, position, agSpice, SPICECOMBATLIMIT)) •(population \ {positions ∼(loc)}, (positions −⊲ {loc}) ⊕ {ag 7→ loc}, sugar ⊕ {loc 7→ 0}, agSugar ⊕ {ag 7→ agSugar(ag)+ agSpice ⊕ {ag 7→ agSpice(ag)+ reward(position′(ag), sugar, position, agSugar, COMBAT LIMIT )}, reward(position′(ag), spice, position, agSpice, SPICECOMBATLIMIT)}) THE SPECIFICATION OF SUGARSCAPE 101 SynchronousCombatspice ∆SugarScape step′ = step maxSugar′ = maxSugar maxSpice′ = maxSpice pollution′ = pollution loanBook′ = population′ spiceLoanBook′ = ⊳ loanBook ⊲ (population′ ⊳ (ran loanBook)) population′ ⊳ spiceLoanBook ⊲ (population′ ⊳ (ran spiceLoanBook)) population′ ⊆ population sugar′ = sugar ⊕ {p : P OSIT ION p ∈ ran position′ • p 7→ 0} ∀ ag : AGENT ; l : P OSIT ION • ag ∈ population′ ⇒ (sex′(ag) = sex(ag) ∧ vision′(ag) = vision(ag) ∧ age′(ag) = age(ag) ∧ maxAge′(ag) = maxAge(ag) ∧ children′(ag) = children(ag) ∧ agentCulture′(ag) = agentCulture(ag) ∧ agentImmunity ′(ag) = agentImmunity(ag) ∧ metabolism′(ag) = metabolism(ag) ∧ initialSugar′(ag) = initialSugar(ag) ∧ initialSpice′(ag) = initialSpice(ag) ∧ diseases′(ag) = diseases(ag) ∧ agentSugar′(ag) = agentSugar(ag) ∧ agentSpice′(ag) = agentSpice(ag) ∧ spiceMetabolism(ag) = spiceMetabolism(ag) +reward(position′(ag), sugar, position, agentSugar, COMBAT LIMIT ) +reward(position′(ag), spice, position, agentSpice, SPICECOMBATLIMIT) ∧ position′(ag) ∈ availMovesspice(ag, position, sugar, agentSugar, agentSpice, agentCulture, vision(ag))) ) ag ∈ population \ population′ ⇒ ∃ x : AGENT • position′(x) = position(ag) ∧ tribe(culture(x)) 6= tribe(culture(ag)) (l ∈ availMovesspice(ag, position, sugar, agentSugar, agentSpice, agentCulture, vision(ag)) ∧ reward(l, sugar, position, agentSugar, COMBAT LIMIT ) +reward(l, spice, position, agentSpice, SPICECOMBATLIMIT) ≥ reward(position′(ag), sugar, position, agentSugar, COMBAT LIMIT ) +reward(position′(ag), spice, position, agentSpice, SPICECOMBATLIMIT) ∧ distance(position(ag), l) < distance(position(ag), position′(ag))) ⇒ ∃ x : AGENT • position′ ∼(l) = x ∧ position(x) 6= l 102 JOSEPH KEHOE 5.17. Credit. Credit is defined with one resource. It is incorporated into a dual resource simulation but no extra information is given as to what changes were made, if any. The most logical approach is to assume that spice loans are admin- istered the the exact same way as sugar. We create a separate system of loans for spice that is dealt with in the exact same manner as sugar loans. Our specification is now split into two parts, the parts dealing with sugar, already specified, and the parts dealing with spice, which are copies of their counterparts. The amount of a resource available to be borrowed now depends on which resource we are talking about. amtAvailnew : N × SEX × N × N → N ∀ age, resource, baseAmt : N • amtAvailnew(age, male, resource, baseAmt) = resource div 2 ⇔ age > MALEF ERT ILIT Y END amtAvailnew(age, male, resource, baseAmt) = resource − baseAmt ⇔ amtAvailnew(age, male, resource, baseAmt) = 0 ⇔ (age ≤ MALEF ERT ILIT Y END ∧ isF ertile(age, male) ∧ resource > baseAmt) (age ≤ MALEF ERT ILIT Y END ∧ ¬ (isF ertile(age, male) ∨ resource > baseAmt)) amtAvailnew(age, f emale, resource, baseAmt) = resource div 2 ⇔ age > F EMALEF ERT ILIT Y END amtAvailnew(age, f emale, resource, baseAmt) = resource − baseAmt ⇔ amtAvailnew(age, f emale, resource, baseAmt) = 0 ⇔ (age ≤ F EMALEF ERT ILIT Y END ∧ isF ertile(age, f emale) ∧ resource > baseAmt) (age ≤ F EMALEF ERT ILIT Y END ∧ ¬ (isF ertile(age, f emale) ∨ resource > baseAmt)) THE SPECIFICATION OF SUGARSCAPE 103 P aySugarLoans ∆Agents ΞStep population′ = population sex′ = sex position′ = position vision′ = vision age′ = age maxAge′ = maxAge agentCulture′ = agentCulture agentImmunity ′ = agentImmunity children′ = children diseases′ = diseases metabolism′ = metabolism initialSugar′ = initialSugar initialSpice′ = initialSpice spiceMetabolism′ = spiceMetabolism spiceLoanBook′ = spiceLoanBook agentSpice′ = agentSpice ∃ dueLoans, newLoans : (AGENT ↔ (AGENT × (N × N))) • dueLoans = loanBook ⊲ (ran(loanBook) ⊲ {a : (N × N) second(a) = step}) (loanBook′, agentSugar′) = payExclusiveLoans(chooseConf lictF reeSets(dueLoans), agentSugar, loanBook) 104 JOSEPH KEHOE P aySpiceLoans ∆Agents ΞStep population′ = population sex′ = sex position′ = position vision′ = vision age′ = age maxAge′ = maxAge agentCulture′ = agentCulture agentImmunity ′ = agentImmunity children′ = children diseases′ = diseases metabolism′ = metabolism initialSugar′ = initialSugar initialSpice′ = initialSpice spiceMetabolism′ = spiceMetabolism loanBook′ = loanBook agentSugar′ = agentSugar ∃ dueLoans, newLoans : (AGENT ↔ (AGENT × (N × N))) • dueLoans = loanBook ⊲ (ran(spiceLoanBook) ⊲ {a : (N × N) second(a) = step}) (spiceLoanBook′, agentSpice′) = payExclusiveLoans(chooseConf lictF reeSets(dueLoans), agentSpice, spiceLoanBook) THE SPECIFICATION OF SUGARSCAPE 105 MakeLoansSugar ∆Agents ΞStep population′ = population ∧ sex′ = sex position′ = position ∧ vision′ = vision age′ = age ∧ maxAge′ = maxAge agentCulture′ = agentCulture ∧ agentImmunity ′ = agentImmunity diseases′ = diseases ∧ children′ = children metabolism′ = metabolism initialSugar′ = initialSugar initialSpice′ = initialSpice spiceMetabolism′ = spiceMetabolism agentSpice′ = agentSpice spiceLoanBook′ = spiceLoanBook ∃ newLoans : P(AGENT × (AGENT × (N × N))); ∀ ag, lender, borrower : AGENT ; amt, due : N • loanBook′ = loanBook ∪ newLoans ag ∈ dom newLoans ⇒ ag ∈ dom(ran newLoans) ⇒ ag 6∈ dom(newLoans) ∪ dom(ran newLoans) ⇒ willBorrow(age(ag), sex(ag), agentSugar′(ag), agentSugar′(ag) = agentSugar(ag) − totalLoaned(ag, newLoans) agentSugar′(ag) = agentSugar(ag) + totalOwed(ag, newLoans) agentSugar′(ag) = agentSugar(ag) ran(loanBook′ ∩ {a : AGENT × (AGENT × (N × N)) ¬ ∃ ag2 : AGENT • canLend(age(ag2), sex(ag2), agentSugar′(ag2)) borrower(a) = borrower(loan)})) ⇒ ∧ adjacent(position(ag2), position(ag)) amtAvailnew(age(ag), sex(ag), agentSugar(ag), CHILDAMT ) totalLoaned(ag, newLoans) ≤ totalOwed(ag, newLoans) ≤ CHILDAMT − agentSugar(ag) (lender, (borrower, (amt, due))) ∈ newLoans ⇒ (canLend(age(lender), sex(lender), agentSugar(lender)) ∧ willBorrow(age(borrower), sex(borrower), agentSugar(borrower), ∧ amt ≤min({amtAvailnew(age(lender), sex(lender), agentSugar(lender), ∧ due = step + DURAT ION ∧ adjacent(position(lender), position(borrower))) {borrower} ⊳ (ran loanBook)) CHILDAMT ), CHILDAMT − agentSugar(borrower))}) 106 JOSEPH KEHOE THE SPECIFICATION OF SUGARSCAPE 107 MakeSpiceLoans ΞSpiceLattice ∆SpiceAgents ΞStep population′ = population ∧ sex′ = sex position′ = position ∧ vision′ = vision age′ = age ∧ maxAge′ = maxAge agentCulture′ = agentCulture ∧ agentImmunity ′ = agentImmunity diseases′ = diseases ∧ children′ = children metabolism′ = metabolism spiceMetabolism′ = spiceMetabolism agentSugar′ = agentSugar loanBook′ = loanBook initialSugar′ = initialSugar initialSpice′ = initialSpice ∃ newLoans : P(AGENT × (AGENT × (N × N))); ∀ ag, lender, borrower : AGENT ; amt, due : N • spiceLoanBook′ = spiceLoanBook ∪ newLoans ag ∈ dom newLoans ⇒ ag ∈ dom(ran newLoans) ⇒ ag 6∈ dom(newLoans) ∪ dom(ran newLoans) ⇒ willBorrow(age(ag), sex(ag), agentSpice′(ag), agentSpice′(ag) = agentSpice(ag) − totalLoaned(ag, newLoans) agentSpice′(ag) = agentSpice(ag) + totalOwed(ag, newLoans) agentSpice′(ag) = agentSpice(ag) ran(loanBook′ ∩ {a : AGENT × (AGENT × (N × N)) ¬ ∃ ag2 : AGENT • canLend(age(ag2), sex(ag2), agentSpice′(ag2)) borrower(a) = borrower(loan)})) ⇒ ∧ adjacent(position(ag2), position(ag)) amtAvailnew(age(ag), sex(ag), agentSpice(ag), SP ICECHILDAMT ) totalLoaned(ag, newLoans) ≤ totalOwed(ag, newLoans) ≤ SP ICECHILDAMT − agentSpice(ag) (lender, (borrower, (amt, due))) ∈ newLoans ⇒ (canLend(age(lender), sex(lender), agentSpice(lender)) ∧ willBorrow(age(borrower), sex(borrower), agentSpice(borrower), ∧ amt ≤ min({amtAvailnew(age(lender), sex(lender), {borrower} ⊳ (ran spiceLoanBook)) agentSpice(lender), SP ICECHILDAMT ) SP ICECHILDAMT − agentSpice(borrower))}) ∧ due = step + DURAT ION ∧ adjacent(position(lender), position(borrower))) 108 JOSEPH KEHOE 5.18. Rule Application Sequence. T ickspice [# Growbackspice # SeasonalGrowbackspice] [# MovementbasicSpice # (MovementpollutionSpice # P ollutionDif f usionspice) # Combatspice] {# Inheritancespice}{# Deathspice[# Replacementspice # AgentMatingspice]} {# Culture}{# P aySugarLoans # P aySpiceLoans # MakeSugarLoans # MakeSpiceLoans} {# T ransmissionspice # ImmuneResponsespice}{# T rade} 6. Conclusions We have shown that it is possible to apply formal methods fruitfully in the area of ABSS and produced a full formal specification of the Sugarscape family of simulations. It is, to the best of our knowledge, the first formal specification of the entire Sugarscape simulation family. The purpose of the specification is to provide a clear, unambiguous and precise definition of Sugarscape. The specification has identified many ambiguities and/or missing bits of information in the original rule definitions. Where there is an obvious way of removing these ambiguities we have done so. If there is more than one possible solution we have identified them and chosen the most likely one. The issues with the model definition that we have encountered can broadly be grouped into three main types: Lack of Clarity, Missing Information and Sequential biases. Lack of Clarity: The rules, although simply stated in the appendix, lack clarity in their definition. Only one version of each rule is presented even when many variations are referred to in the text. The variations presented cannot always be used together, for example the Movement rule defined in the appendix is not the variant required if the pollution rule is also used. Our specification brings them all together in one place for ease of reference. Missing Information: Missing or incomplete information is the biggest cause for concern. In many cases we can work out the most likely answer based on context but in some cases there is not one definitive correct an- swer. If there was more than one arguably correct solution we chose the simplest. How we fill in these blanks can have a big effect on how the simulation proceeds. These effects may be important if we are trying to compare different implementations of Sugarscape. If we take the disease transmission rule, for example, questions that are unanswered include: (1) Once an agent gains immunity from a particular disease, do we re- move that particular disease from the set of diseases that the agent is carrying, or is the agent still a carrier? (2) When we transmit a disease do we only transmit diseases that we carry and have no immunity for, or, can any disease we carry be transmitted? THE SPECIFICATION OF SUGARSCAPE 109 (3) The Mating rule omits important information about parents contribut- ing half of their resources to their offspring. This has a huge effect on how mating works in a simulation. By replacing each ambiguous interpretation with one simple and precise interpretation we allow different developers to replicate their results and benchmark them against each other. All hidden assumptions that could serve to advantage one implementation over another are excised. Sequential Biases: Sugarscape is based on the assumption that it will be implemented sequentially. While this may have been a good assumption at the time it was written it is not now necessarily the case. Improvements in processing speed have recently been attained mainly through the introduc- tion of concurrency. Simulations are now almost always run on multicore or even multiprocessor machines. The Z specification is free from assumptions about implementation. It achieves this without having to specify or constrain in any way what conflict resolution or avoidance strategies are employed. This leaves developers the freedom to try out different approaches as suits their implementation platform. Because the specification is high level and only defines the before and after state of each rule it makes few assumptions as to how any rule will be imple- mented. All inherent biases towards a sequential implementation are removed. Implementers have complete freedom as to what programming model they em- ploy (Object-oriented, imperative, functional, or any concurrent approach). Any simulations that adhere to the standard can be properly compared in terms of performance or patterns of behaviour. This will put on a firmer foundation any claims made by researchers about their implementations. 6.1. Further Work. Further work remains to be done in getting agreement from the ABM community on the decisions made in producing this interpretation of Sugarscape. Any incorrect assumptions made in removing ambiguities need to be identified and agreed upon. This provides a route to address the issues of replication of experimental results in ABSS. Sugarscape can now be used as a benchmark (or rather set of benchmarks) for ABM implementers. This is particularly useful for those proposing new approaches to concurrency that promise performance improvements. Current trends, for exam- ple, include the use of Graphics Processor Units (GPUs)[Deissenberg et al., 2008, Lysenko and D'Souza, 2008, Richmond et al., 2009], containing hundreds to thou- sands of individual processors. These approaches tend not to be tested on the more complex rules in Sugarscape (such as Combat, Inheritance and Trade) as they are not easily parallelized. By providing a precise and full set of these rules it is now possible for researchers to properly compare how different models cope with more complex and more realistic ABMs. 110 JOSEPH KEHOE Z itself is rather verbose and can be hard to parse when reading. The lack of modularity made function definition signatures overly long. The available tools such as CZT [Malik and Utting, 2005] make the process of writing the specification easier but I have altered the specifications to remove bracketing where I thought it made the specification easier to read even if this was flagged as an error in the type checker. These issues could be overcome through the use of a variant of Z such as Object-Z or Alloy. The issue of whether ABM modellers would be willing to use formal specifications remains unknown. There are differences between the outcomes of the synchronous and asynchro- nous approaches. Sugarscape assumes an asynchronous approach and this affects the style of specification that we use. We have shown in the case of combat the differences in a synchronous and asynchronous specification. While we regard the synchronous specification as somewhat simpler to produce but others may disagree. We tackle the question as to which approach is the more correct elsewhere. References [Axtell and Axtell, 2000] Axtell, R. and Axtell, R. L. (2000). Why agents? on the varied moti- vations for agent computing in the social sciences. In Working Paper 17, Center on Social and Economic Dynamics, Brookings Institution, page 17. [Berryman, 2008] Berryman, M. (2008). Review of software platforms for agent based models. Technical report, DTIC Document. [Campillo et al., 2012] Campillo, X. R., Cela, J. M., and Cardona, F. X. H. (2012). Simulat- ing archaeologists? using agent-based modelling to improve battlefield excavations. Journal of Archaeological Science, 39(2):347 – 356. [Deissenberg et al., 2008] Deissenberg, C., van der Hoog, S., and Dawid, H. (2008). Eurace: A massively parallel agent-based model of the european economy. Applied Mathematics and Com- putation, 204(2):541 – 552. ¡ce:title¿Special Issue on New Approaches in Dynamic Optimization to Assessment of Economic and Environmental Systems¡/ce:title¿. [Epstein and Axtell, 1996] Epstein, J. M. and Axtell, R. (1996). Growing Artificial Societies: Social Science from the Bottom Up. The Brookings Institution, Washington, DC, USA. [Gilbert, 2004] Gilbert, N. (2004). Agent-based social simulation: dealing with complexity. The Complex Systems Network of Excellence, 9(25):1–14. [Inchiosa and Parker, 2002] Inchiosa, M. E. and Parker, M. T. (2002). Overcoming design and development challenges in agent-based modeling using ascape. Proceedings of the National Acad- emy of Sciences, 99(suppl 3):7304–7308. [Lysenko and D'Souza, 2008] Lysenko, M. and D'Souza, R. M. (2008). A framework for megas- cale agent based model simulations on graphics processing units. Journal of Artificial Societies and Social Simulation, 11(4):10. [Macal and North, 2009] Macal, C. and North, M. (2009). Agent-based modeling and simulation. In Simulation Conference (WSC), Proceedings of the 2009 Winter, pages 86–98. [Malik and Utting, 2005] Malik, P. and Utting, M. (2005). Czt: A framework for z tools. In ZB. Lecture, pages 65–84. Springer. [Perumalla, 2006] Perumalla, K. S. (2006). Discrete-event execution alternatives on general pur- pose graphical processing units (gpgpus). In Proceedings of the 20th Workshop on Principles of Advanced and Distributed Simulation, PADS '06, pages 74–81, Washington, DC, USA. IEEE Computer Society. THE SPECIFICATION OF SUGARSCAPE 111 [Railsback et al., 2006] Railsback, S. F., Lytinen, S. L., and Jackson, S. K. (2006). Agent-based simulation platforms: Review and development recommendations. Simulation, 82(9):609–623. [Richmond et al., 2009] Richmond, P., Coakley, S., and Romano, D. M. (2009). A high perfor- mance agent based modelling framework on graphics card hardware with cuda. In Proceedings of The 8th International Conference on Autonomous Agents and Multiagent Systems - Volume 2, AAMAS '09, pages 1125–1126, Richland, SC. International Foundation for Autonomous Agents and Multiagent Systems. [Richmond et al., 2010] Richmond, P., Walker, D., Coakley, S., and Romano, D. (2010). High performance cellular level agent-based simulation with flame for the gpu. Briefings in Bioinfor- matics, 11(3):334–347. [Spivey, 1989] Spivey, J. M. (1989). The Z Notation: A Reference Manual. Prentice-Hall, Inc., Upper Saddle River, NJ, USA. [Troitzsch, 2009] Troitzsch, K. G. (2009). Perspectives and challenges of agent-based simulation as a tool for economics and other social sciences. In Proceedings of The 8th International Con- ference on Autonomous Agents and Multiagent Systems-Volume 1, pages 35–42. International Foundation for Autonomous Agents and Multiagent Systems.
1706.02209	1	1706	2017-06-07T14:29:23	Improving Max-Sum through Decimation to Solve Loopy Distributed Constraint Optimization Problems	[ "cs.MA", "cs.AI" ]	In the context of solving large distributed constraint optimization problems (DCOP), belief-propagation and approximate inference algorithms are candidates of choice. However, in general, when the factor graph is very loopy (i.e. cyclic), these solution methods suffer from bad performance, due to non-convergence and many exchanged messages. As to improve performances of the Max-Sum inference algorithm when solving loopy constraint optimization problems, we propose here to take inspiration from the belief-propagation-guided dec-imation used to solve sparse random graphs (k-satisfiability). We propose the novel DeciMaxSum method, which is parameterized in terms of policies to decide when to trigger decimation, which variables to decimate, and which values to assign to decimated variables. Based on an empirical evaluation on a classical BP benchmark (the Ising model), some of these combinations of policies exhibit better performance than state-of-the-art competitors.	cs.MA	cs	Improving Max-Sum through Decimation to Solve Loopy Distributed Constraint Optimization Problems J. Cerquides1, R. Emonet2, G. Picard3, and J.A. Rodriquez-Aguilar1 1 IIIA-CSIC, Campus UAB, 08193 Cerdanyola, Catalonia, Spain {cerquide,jar}@iiia.csic.es 2 Université de Lyon, Laboratoire Hubert Curien UMR CNRS 5516, France 3 MINES Saint-Etienne, Laboratoire Hubert Curien UMR CNRS 5516, France [email protected] [email protected] Abstract. In the context of solving large distributed constraint optimization problems (DCOP), belief-propagation and approximate inference algorithms are candidates of choice. However, in general, when the factor graph is very loopy (i.e. cyclic), these solution methods suffer from bad performance, due to non-convergence and many exchanged messages. As to improve performances of the Max-Sum inference algorithm when solving loopy constraint optimization problems, we propose here to take inspiration from the belief-propagation-guided dec- imation used to solve sparse random graphs (k-satisfiability). We propose the novel DeciMaxSum method, which is parameterized in terms of policies to decide when to trigger decimation, which variables to decimate, and which values to assign to decimated variables. Based on an empirical evaluation on a classical BP bench- mark (the Ising model), some of these combinations of policies exhibit better performance than state-of-the-art competitors. 1 Introduction In the context of multi-agent systems, distributed constraint optimization problems (DCOP) are a convenient to model coordination issues agents have to face, like resource allocation, distributed planning or distributed configuration. In a DCOP, agents manage one or more variables they have to assign a value (e.g. a goal, a decision), while taking into account constraints with other agents. Solving a DCOP consists in making agents communicate as to minimize the violation of these constraints. Several solution methods exist to solve such problems, from complete and optimal solutions, to approximate ones. When dealing with larger scales (thousands of variables), approximate methods are solutions of choice. Indeed, complete methods, like ADOPT or DPOP, suffer exponential computation and/or communication cost in general settings [10,15]. As a consequence, in some large settings, approximate methods are better candidates, as evidenced by the extensive literature on the subject (see [1] for a complete review). One major difficulty for approximate method to solve DCOP is the presence of cycles in the constraint graph (or factor graph). Among the aforementioned methods, inference-based ones, like Max-Sum [3] and its extensions like [16], have demonstrated good performance even on loopy settings. However, there exists some cases, with numerous loops or large induced width of the constraint graph, where they perform badly, which translates into a larger number of messages, a longer time to convergence and a final solution with bad quality. One original approach to cope with loopy graphs is to break loops by decimating variables during the solving process. Decimation is a method inspired by statistical physics, and applied in belief-propagation, which consists in fixing the value of a variable, using the marginal values as the decision criteria to select the variable to decimate [13]. The decimation is processed regularly after the convergence of a classical belief-propagation procedure. In [11], decimation has been used in the constraint satisfaction framework, for solving centralized k-satisfiability problems [11]. Inspired by this concept, we propose a general framework for applying decimation in the DCOP setting. Other works proposed Max-Sum_AD_VP as to improve Max-Sum performance on loopy graphs [20]. The idea is to perform the inference mechanism through an overlay directed 7 1 0 2 n u J 7 ] A M . s c [ 1 v 9 0 2 2 0 . 6 0 7 1 : v i X r a 2 Cerquides et al. acyclic graph, to remove loops, and to alternating the direction of edges at a fixed frequency as to improve the sub-optimal solution found with the previous direction. One mechanism within one of these extensions, namely value propagation, can be viewed as a temporary decimation. Against this background, the main goal of this paper is to propose a general framework for installing decimation in Max-Sum for solving DCOP. More precisely, we make the following contributions: 1. We propose a parametric solution method, namely DeciMaxSum, to implement decimation in Max-Sum. It takes three fundamental parameters for decimation: (i) a policy stating when to trigger decimation, (ii) a policy stating which variables to decimate, and (iii) a policy stating which value to assign to decimated variables. The flexibility of DeciMaxSum comes from the fact that any policy from (1i) can be combined with any policy from (1ii) and (1iii). 2. We propose a library of decimation policies; some inspired by the state-of-the-art and some original ones. Many combinations of policies are possible, depending on the problem to solve. 3. We implement and evaluate some of these combinations of decimation policies on classical DCOP benchmarks (meeting scheduling and Ising models), against state-of-the-art methods like standard Max-Sum and Max-Sum_AD_VP. The rest of the paper is organized as follows. Section 2 expounds some background on DCOP and expounds the decimation algorithm from which our algorithm DeciMaxSum is inspired. Section 3 defines the general framework of DeciMaxSum, and several examples of decimation policies. Section 4 presents results and analyses of experimenting DeciMaxSum, with different combinations of decimation policies, against Max-Sum and Max-Sum_AD_VP. Finally, Section 5 concludes this paper with some perspectives. 2 Background This section expounds the DCOP framework and some related belief-propagation algorithms from the literature are discussed concerning the mechanisms to handle cycles in constraint graphs. 2.1 Disributed Constraint Optimization Problems One way to model the coordination problem between smart objects is to formalize the problem as a distributed constraint optimization problem. Definition 1 (DCOP). A discrete Distributed Constraint Optimization Problem (or DCOP) is a tuple (cid:104)A,X ,D,C, µ(cid:105), where: A = {a1, . . . , aA} is a set of agents; X = {x1, . . . , xN} are variables owned by the agents; D = {Dx1, . . . ,DxN} is a set of finite domains, such that variable xi takes values in Dxi = {v1, . . . , vk}; C = {u1, . . . , uM} is a set of soft constraints, where each ui defines a utility ∈ R ∪ {−∞} for each combination of assignments to a subset of variables Xi ⊆ X (a constraint is initially known only to the agents involved); µ : X → A is a function mapping variables to their associated agent. A solution to the DCOP is an assignment X ∗ = {x∗ N} to all variables that maximizes the overall sum of costs4: 1, . . . , x∗ M(cid:88) 4 Note that the notion of cost can be replaced by the notion of cost ∈ R∪{+∞}. In this case, solving a DCOP is a minimization problem of the overall sum of costs. m=1 um(Xm) (1) DeciMaxSum: Decimation in MaxSum 3 As highlighted in [1], DCOPs have been widely studied and applied in many reference domains, and have many interesting features: (i) strong focus on decentralized approaches where agents negotiate a joint solution through local message exchange; (ii) solution techniques exploit the structure of the domain (by encoding this into constraints) to tackle hard computational prob- lems; (iii) there is a wide choice of solutions for DCOPs ranging from complete algorithms to suboptimal algorithms. A binary DCOP can be represented as a constraint graph, where vertices represent variables, and edge represent binary constraints. In the case of n-ary constraints, a DCOP can be repre- sented as a factor graph: an undirected bipartite graph in which vertices represent variables and constraints (called factors), and an edge exists between a variable and a constraint if the variable is in the scope of the constraint. Definition 2 (Factor Graph). A factor graph of a DCOP as in Def. 1, is a bipartite graph F G = (cid:104)X ,C, E(cid:105), where the set of variable vertices corresponds to the set of variables X , the set of factor vertices corresponds to the set constraints C, and the set of edges is E = {eij xi ∈ Xj}. When the graph representing the DCOP contains at least a cycle, we call it a cyclic DCOP; otherwise, it is acyclic. A large literature exists on algorithms for solving DCOPs which fall into two categories. On the one hand, complete algorithms like ADOPT and its extensions [9], or inference algorithms like DPOP [15] or ActionGDL [19], are optimal, but mainly suffer from expensive memory (e.g. exponential for DPOP) or communication (e.g. exponential for ADOPT) load –which we may not be able to afford in a constrained infrastructure, like in sensor networks. On the other hand, approximate algorithms like Max-Sum [3] or MGM [8] have the great advantage of being fast with a limited memory print and communication load, but losing optimality in some settings –e.g. Max-Sum is optimal on acyclic DCOPs, and may achieve good quality guarantee on some settings. The aforementioned algorithms mainly exploit the fact that an agent's utility (or constraint's cost) depends only on a subset of other agents' decision variables, and that the global utility function (or cost function) is a sum of each agent's utility (constraint's cost). In this paper, we are especially interested in belief-propagation-based algorithms, like Max-Sum, where the notion of marginal values describes the dependency of the global utility function on variables. 2.2 From Belief-Propagation to Max-Sum Belief propagation (BP), i.e. sum-product message passing method, is a potentially distributed algorithm for performing inference on graphical models, and can operate on factor graphs m=1 fm(Xm) . The sum-product algorithm provides an efficient local message passing procedure to compute the marginal functions of all variables simultaneously. The marginal function, zn(xn) describes the total dependency of the representing a product of M factors [7]: F (x) =(cid:81)M global function F (x) on variable xn: zn(xn) =(cid:80){x(cid:48)},n(cid:48)(cid:54)=n F (Xn(cid:48)). BP operates iteratively propagating messages mi→j (tables associating marginals to each value of variables) along the edges of the factor graph.When the factor graph is a tree, BP algorithm computes the exact marginals and converge in a finite number a steps depending on the diameter of the graph [7]. Max-product is an alternative version of sum-product which computes the maximum value instead of the sum. Built as a derivative of max-product, Max-Sum is an approximate algorithm to solve DCOP [3]. The main evolution is the way messages are assessed, to pass from product to sum operator through logarithmic translation. And as a consequence, Max-Sum computes an assignment X ∗ 4 Cerquides et al. Algorithm 1: The BP-guided decimation algorithm from [11] Data: A factor graph representing a k-satisfiability problem Result: A feasible assignment X ∗ or FAIL 1 initialize BP messages 2 U ← ∅ 3 for t = 1, . . . , n do 4 5 6 7 8 9 10 11 12 return X ∗ run BP until the stopping criterion is met choose xi ∈ X \ U uniformly at random compute the BP marginal zi(xi) choose x∗ fix xi = x∗ U ← U ∪ {xi} simplify the factor graph if a contradiction is found, return FAIL i distributed according to zi i that maximizes the DCOP objective in Equation 1. Depending on the DCOP to solve, Max- Sum may be used with two different termination rules: (i) continue until convergence (no more exchanged messages, because when a variables or a factor receives twice the same message from the same emitter it does not propagates); (ii) propagate message for a fixed number of iterations per agent. Max-Sum is optimal on tree-shaped factor graphs, and still perform well on cyclic settings. But there exist problems for which Max-Sum does not converge or converge to a sub-optimal state. In fact, on cyclic settings [3] identify the following behaviors: (i) agents converge to fixed states that represent either the optimal solution, or a solution close to the optimal, and the propagation of messages ceases; (ii) agents converge as above, but the messages continue to change slightly at each update, and thus continue to be propagated around the network; (iii) neither the agents' preferred states, nor the messages converge and both display cyclic behavior. As to improve Max-Sum performance on cyclic graphs, [20] proposed two extensions to Max-Sum: (i) Max-Sum_AD which operates Max-Sum on a directed acyclic graph built from the factor graph, and alternates direction at a fixed rate (a parameter of the algorithm); (ii) Max- Sum_AD_VP which operates Max-Sum_AD and propagates current values of variables when sending Max-Sum messages so that factors receiving the value only consider this value instead of the whole domain of the variable. These two extensions, especially the second one, greatly improves the quality of the solution: Max-Sum_AD_VP found solutions that approximate the optimal solution by a factor of roughly 1.1 on average. However, the study does not consider the number of exchanged messages, or the time required to converge and terminate Max- Sum_AD_VP. 2.3 BP-guided Decimation In this paper, we propose to take inspiration from work done in computational physics [13], as to cope with cyclicity in DCOP. Notably, [5] introduced the notion of decimation in constraint sat- isfaction, especially k-satisfiability, where variables are binary, xi ∈ {0, 1}, and each constraint requires k of the variables to be different from a specific k-uple. Authors proposed a class of algorithms, namely message passing-guided decimation procedure, which consists in iterating the following steps: (1) run a message passing algorithm, like BP ; (2) use the result to choose a variable index i , and a value x∗ for the corresponding variable; (3) replace the constraint satis- faction problem with the one obtained by fixing xi to x∗ . The BP-guided decimation procedure is shown in Algorithm 1, whose performances are analysed in [11,13]. i i DeciMaxSum: Decimation in MaxSum 5 BP-guided decimation operates on the factor graph representing the k-satisfiability problem to solve. At each step, the variable to decimate is randomly chosen among the remaining variables. The chosen variable xi is assigned a value determined by random sampling according to its marginal zi. After decimation, the factor graph is simplified: some edges are no more relevant, and factors can be sliced (columns corresponding to removed variables are deleted). In some settings, BP-guided decimation may fail, if random choices assign a value to a variable which is not consistent with other decimated variables. Some comments can be made on this approach. First, relying on marginal values is a key feature, and is the core of the "BP-guided" nature of this method. Marginal values are exploited to prune the factor graph. Second, while in the seminal work of [11], this procedure is used to solve satisfiability problems, the approach can easily be implemented to cope with optimization problems. For instance, the inference library libDAI proposes an implementation of decimation for discrete approximate inference in graphical models [12], which was amongst the three winners of the UAI 2010 Approximate Inference Challenge5. 2.4 State of a Factor Graph Representation The previous BP-based algorithm operates on factor graph representing the problem. "Operates" means that the algorithms create a data structure representing the factor graph which evolves with time : marginal values change, variables disappear, messages are sent/received, etc. Commonly, the logical representation of a factor graph is a set of nodes connected depending on the connectivity of the graph. Each such node has a state which stores some useful values. Definition 3. The current state F Gt at time t of a factor graph F G = (cid:104)X ,C, E(cid:105) is the compo- sition of all the current states of the data structures used by the BP-based algorithm to operate on the related factor graph, including the marginal values zi, the messages mi→j, the set of decimated variables U, and other algorithm-specific data. We can consider that for a given problem, many factor graph states may exist. We denote S the set of possible factor graph states, and S(F G) ⊂ S the set of possible states for the factor graph F G. 3 DeciMaxSum: Extending Max-Sum with Decimation While mainly designed as a centralized algorithm and studied on k-SAT problems, BP-guided decimation could be utilized for solving DCOP with a few modifications. To the best of our knowledge, this approach has never been proposed for improving Max-Sum algorithm. Here we expound the core contribution of this paper, namely the DeciMaxSum framework and its components. 3.1 Principles The main idea is to extend the BP-guided decimation algorithm from [11] in order to define a more general framework, in which other BP-based existing algorithms could fit. First, the main focus is decimation, which means assigning a value to a variable as to remove it from the problem. As the name suggests, there is no way back when a variable has been decimated –unlike search algorithms, where variable assignments can be revised following a backtrack, for instance. Therefore, triggering decimation is an impacting decision. This is why our framework 5 http://www.cs.huji.ac.il/project/UAI1/ 6 Cerquides et al. is mainly based on answering three questions: (i) when is decimation triggered, (ii) which variable(s) to decimate, (iii) which value to assign to the decimated variable(s)? Several criteria can be defined for answering each question, and the DeciMaxSum specifies such criteria as decimation policies, that are fundamental parameters of the decimation procedure. Definition 4 (Decimation Policy). A decimation policy is a tuple π = (cid:104)Θ, Φ, Υ, Λ(cid:105) where: – Θ : S → {0, 1} is the condition to trigger the decimation process, namely the trigger policy, – Φ : S → 2X is a filter policy which selects some candidate variables to decimate, – Υ : X × S → {0, 1} is the condition to perform decimation on a variable, namely perform – Λ : X × S → DX is the assignment policy, which assigns a value to a given variable. policy, A rich population of decimation-based algorithm can be modeled through this framework by combining decimation policies. For instance, one can consider a DeciMaxSum instance, which (i) triggers decimation once BP has converged, (ii) chooses randomly a variable to decimated within the whole set of non-decimated variables, and (iii) samples the value of the decimated variable depending on its marginal values (used as probability distribution). By doing so, we result in the classical BP-guided decimation algorithm from [11] . However, as many more decimation policies can be defined and combined, we fall into a more general framework generating a whole family of algorithms. 3.2 DeciMaxSum as an Algorithm We can summarize the DeciMaxSum framework using Algorithm 2. It is a reformulation of BP- guided decimation, parameterized with a decimation policy. Here decimation is not necessarily triggered at the convergence (or time limit) of BP. Criterion Θ may relies on other components of the state of the factor graph. Contrary to classical BP-guided decimation, there may be several variables to decimate at the same time (like in some variants of DSA or MGM) and that variables can be chosen in an informed manner (and not randomly), using criterion Υ . Values assigned to decimated variables, are not necessarily chosen stochastically, but are assigned using the function Λ that can be deterministic (still depending on the current state of the FG). Since, here we're not in the k-satisfiability case, but in an optimization case, there is no failure (only suboptimality), contrary to Algorithm 1. Finally, once all variables have been decimated, the output consists in decoding the state F Gt, i.e. getting the values assigned to decimated variables. This means that finally DeciMaxSum is performing decoding while solving the problem, which is not a common feature in other DCOP algorithms, like classical Max-Sum or DSA. Indeed, once these algorithms halt, a decoding phase must be performed to extract the solution from the variables' states. While presented as a classical algorithm, let us note that decimation is meant to be imple- mented in a distributed and concurrent manner, depending on the decimation policy components. The rest of the section details and illustrates each of these decimation policies component with some examples. 3.3 Triggering Decimation (Θ criterion) In the original approach proposed by [11], decimation is triggered once BP has converged. In a distributed settings and diffusing algorithms like BP, this can be implemented using termination detection techniques. (cid:26) 1, if s is quiescent 0, otherwise Θconverge(s) def= (2) DeciMaxSum: Decimation in MaxSum 7 Algorithm 2: The DeciMaxSum framework as an algorithm Data: A factor graph F G = (cid:104)X ,C, E(cid:105), a decimation policy π = (cid:104)Θ, Φ, Υ, Λ(cid:105) Result: A feasible assignment X ∗ run BP until decimation triggers, i.e. Θ(F Gt) = 1 choose variables to decimate, X (cid:48) = {xi ∈ Φ(F Gt) Υ (xi, F Gt)} for xi ∈ X (cid:48) do 1 initialize BP messages 2 U ← ∅ 3 while U (cid:54)= X do 4 5 6 xi ← Λ(xi, F Gt) 7 U ← U ∪ {xi} 8 simplify F Gt 9 10 return X ∗ by decoding U // Sect. 3.3 // Sect. 3.4 // Sect. 3.5 // remove variables, slice factors This trigger consists in detecting the quiescence of the current state of the factor graph. This means no process is enabled to perform any locally controlled action and there are no messages in the channels [6]. Algorithms like DijkstraScholten can detects such global state by implementing a send/receive network algorithm, based on the same graph than F G [6]. Note that such techniques generates extra communication load for termination detection-dedicated messages. Due to the Max-Sum behavior on loopy factor graphs, convergence may not be reached [20]. The common workaround is to run BP for a fixed number of iterations in case there is no convergence. Setting this time limit (namely LIMIT) might be really problem-dependent. (cid:26) 1, if time(s) = LIMIT 0, otherwise Θtime(s) def= (3) (4) In synchronous settings (all variables and factors are executed synchronously, step by step), getting the iteration number of the current state of the FG, time(s), can done in a distributed manner, as usually done in Max-Sum. In the asynchronous case, one can either (i) use a shared clock, or (ii) count locally outcoming messages within each variables, and once a variable has sent a limit number of messages, decimation is triggered. In some settings with strong time or computation constraints (e.g. sensor networks [3], internet-of-things [17]), waiting convergence is not affordable. Indeed, BP may generate a lot of messages. Therefore, we may consider decimating before convergence at a fixed rate (e.g. each 10 iterations), or by sharing a fixed iteration budget amongst the variables (e.g. each 1000 iterations divided by the number of variables). We can even consider a varying decimation speed (e.g. faster at the beginning, and lower at the end, as observed in neural circuits in the brain [14]). (cid:26) 1, if time(s) mod f (s) = 0 0, otherwise Θfrequency(s) def= where f is a function of the current state of the system, for instance : – f (s) = RATE, with a predefined decimation frequency, – f (s) = BUDGET/X, with a predefined computation budget, – f (s) = 2 × time(s), for an decreasing decimation frequency. Finally, another approach could be to trigger decimation once a loop in the FG is detected. Indeed, decimation is used here to cope with loops, so decimating variables, which could potentially break loops, seems a good approach. 8 Cerquides et al. (cid:26) 1, if ∃xi ∈ X ,loop(xi) > 1 0, otherwise Θloop(s) def= (5) where loop(xi) is the set of agents in the same first loop that xi just discovered. Detecting loops in the FG can be implemented during BP, by adding some metadata on the BP messages, like done in the DFS-tree construction phase of algorithms like DPOP or ADOPT. 3.4 Deciding the Subset of Variables to Decimate (Φ and Υ criteria) Now our system has detected decimation should be triggered, the following question is "which variables to decimate?" In [11], the variable is chosen randomly in a uniform manner, while in [12], the variable with a the maximum entropy over its marginal values (the most determined variable) is selected. Obviously, exploiting the marginal values, build throughout propagation is a good idea. From which subset choosing the candidate variables to decimate? Both [11] and [12] select the only variable to decimate amongst the whole set of non-decimated variables (cf. line 5 in Algorithm 1). Here, Φ criterion is specified as follows: def=X \ U (6) However, this selection on the whole set of variables can be discussed when using local decimation triggers, like loop detection. In such case, selecting the variables to decimate within the agents in the loop, or the one which detected the loop sounds better. Another approach is to consider selecting agents depending on the past state of the system. For instance, if a variable has been decimated, good future candidates for decimation could be its direct neighbors in the FG: Φall(s) Φneighbors(s) def={x ∈ X \ U neighbors(x) ∩ U (cid:54)= ∅} with neighbors(xi) = {xj ∈ X j (cid:54)= i,∃eik, ekj ∈ E}. (7) (8) (9) Which criteria to decide whether the variable decimate? Now, we have to specify the Υ criterion used to decide which candidates decimate. In [11], it is fully random: it does not depends on the current state of the variables. It corresponds to make each variable roll a dice and choosing the greatest draw: (cid:26) 1, if ∀xj (cid:54)= xi ∈ X , rand(xi) > rand(xj) Υmax_rand(xi, s) def= 0, otherwise where rand(x) stands for the output of a random number generator (namely sample) using a uniform distribution (e.g. U [0, 1]). In [12], the variable with the maximal entropy over its marginal values is selected. This means the variable for which marginal values seems to be the most informed, in the Shannon's Information Theory sense, is chosen: (cid:26) 1, if ∀xj (cid:54)= xi ∈ X , H(zi(xi)) > H(zj(xj)) 0, otherwise with H(zk(xk)) = −(cid:80) Υmax_entropy(xi, s) def= zk(xk)(d) log(zk(xk)(d)). d∈Dk consider the maximal normalized marginal value: From this, other criteria can be derived. For instance, instead of using entropy, one can DeciMaxSum: Decimation in MaxSum 9 (cid:40) 1, if ∀xj (cid:54)= xi ∈ X , max 0, otherwise d∈Di (zi(xi)(d)) > max d∈Di (zj(xj)(d)) (10) Υmax_marginal(xi, s) def= If several variables can be decimated at the same time, one may consider selecting the set of variable having an entropy or a normalized marginal value greater than a given threshold, to only decimate variable which are "sufficiently" determined. Hence, this approach requires setting another parameter (namely THRESHOLD): (cid:26) 1, if H(zi(xi)) > THRESHOLD 0, otherwise Υthreshold_entropy(xi, s) def= (11) Of course, many combination of the aforementioned criteria, and other criteria could be considered in our framework. We don't discuss here criteria like in DSA which does not rely on marginal values, but on stochastic decision. Which subset of variables the decision to decimate a variable depends on? Behind this question lies the question of coordinating the variable selection. Indeed, if computing criterion Υ does not depend on the decision of other variables, the procedure is fully distributable at low communication cost, as for policies like (11). At the contrary, if the decision requires to be aware of the state of other variables, as for policies like (8), (9) and (10), the procedure will require some system-scale coordination messages. In [11] and [12], decimation only concerns all the variables, from which only one will be chosen. This requires a global coordination, or a distributed leader election protocol which may require an underlying network (ring, spanning tree, etc.), like the one used for quiescence detection, to propagate election messages [6]. In some cases, the decimation decision might be at local scale, when variables will make their decision depending on the decision of their direct neighbors, or variables in the same loop. In this case, less coordination messages will be required. For instance, if considering decimating variables in a loop, only variables in the loop will implement a leader election protocol. All policies, from (8) to (10), could be extended in the same manner, by replacing X by loop(xi), neighours(xi), or any subset of X . For instance: (cid:26) 1, if ∀xj (cid:54)= xi ∈ loop(xi), rand(xi) > rand(xj) Υmax_rand_loop(xi, s) def= (12) 0, otherwise 3.5 Deciding the Values to Assign To Decimated Variables (Λ criterion) Now variables to decimate have been selected, the question is "which values to assign?" Usually, in BP-based algorithms, the simplest way to select values for variables, after propagation, is to assign values with maximal marginal value (or utility). [12] is using such a criterion for inference: Λmax_marginal(xi, s) def= argmax d∈Di zi(xi)(d) (13) While, the policy is deterministic, in [11] the choice of the value is a random choice using the marginal values as a probability distribution: (14) Once again, these are only some examples of policies exploiting BP, and one can easily Λsample_marginal(xi, s) def= sample(zi(xi)) specify many more. 10 Cerquides et al. 4 Experiments In this section we evaluate the performance of different combinations of decimation policies in DeciMaxSum, on a classical optimization model (Ising model), against classical Max-Sum [3] and its extension Max-Sum_AD_VP [20], we have implemented in our own framework. Ising Model 4.1 Since we are interested in evaluating our algorithms in the presence of strong dependencies among the values of variables, we evaluate them on Ising model which is a widely used bench- mark in statistical physics [4]. We use here the same settings than [18]. Here, constraint graphs are rectangular grids where each binary variable xi is connected to its four closer neighbors (with toroidal links which connect opposite sides of the grid), and is constrained by a unary cost ri. The weight of each binary constraint rij is determined by first sampling a value κij from a uniform distribution U [−β, β] and then assigning (cid:40) rij(xi, xj) = κij −κij if xi = xj otherwise The β parameter controls the average strength of interactions. In our experiments we set β to 1.6. The weight for each unary constraint ri is determined by sampling κi from a uniform distribution U [−0.05, 0.05] and then assigning ri(0) = κi and ri(1) = −κi. 4.2 Results and Analysis In this section we analyse results of different DeciMaxSum combinations to solve squared-shape Ising problems with side size varying from 10 to 20 (e.g. 100 to 400 variables). We implemented the following combinations: – 11 DeciMaxSum instances with different decimation policies using the following criteria: • trigger policies (Θ criterion): ∗ Θconverge (from equation 2, noted converge), ∗ rate-based Θfrequency (from equation 4, noted 2-periodic, 3-periodic, 5-periodic, 10- periodic, 20-periodic, and 100-periodic), ∗ budget-based Θfrequency (from equation 4, noted periodic), • filter policy (Φ criterion): ∗ the one that selects the whole set of variables as potential variables to decimate (i.e. Φall from equation 6), • perform policies (Υ criterion): ∗ Υmax_rand (from equation 8, noted random), ∗ Υmax_entropy (from equation 9, noted max_entropy), • assignment policies (Λ criterion): ∗ deterministic Λmax_marginal (from equation 13, noted deterministic), ∗ sampled Λsample_marginal (from equation 14, noted sampling), – MaxSum, as defined in [3], – MaxSum_AD, as defined in [20], – MaxSum_AD_VP, as defined in [20], – Montanari-Decimation, as defined in [11], – Mooij-Decimation, as defined in [12]. DeciMaxSum: Decimation in MaxSum 11 (a) Final total cost (b) Final number of messages Fig. 1: Performances of DeciMaxSum and other solution methods on Ising problems (average of 10 problems per generator's parameter set, and average of 3 runs per problem). Figure 1 presents two performance metrics (final total cost and total number of exchanged messages). Considering optimality of the final solutions obtained by the different solution methods and DeciMaxSum instances, what appears is that very fast decimation combined with a deterministic decimation of the most determined variable (max_entropy) presents the best cost. Besides, very fast decimation also imply that few messages are exchanged compared to other solution methods, since decimation cuts message propagations. However, all the solution methods (except Montanari-Decimation and Mooij-Decimation) tend to a comparable number of exchanged messages. 5 Conclusions In this paper we have investigated how to extend Max-Sum method for solving distributed constraint optimization problems, by taking inspiration from the decimation mechanisms used to solve k-satisfiability problems by belief-propagation. We propose a parametric method, namely DeciMaxSum, which can be set up with different decimation policies stating when to trigger decimation, which variables to decimate, and which value to assign to decimated variables. In this paper, we propose a library of such policies that can be combined to produce different versions of DeciMaxSum. Our empirical results on different benchmarks show that some combinations of decimation policies outperform classical Max-Sum and its extension Max-Sum_AD_VP, specifically design to handle loops. DeciMaxSum outputs better quality solutions in a reasonable number of message propagation. There are several paths to future research. First, we only explore a limited set of decimation policies. We wish to investigate more complex ones, especially policies trigger when loops are detected by agents. In fact, since our overarching goal is to cope with loops, detecting them at the agent level seems a reasonable approach to initiate decimation in a cyclic network. This approach will require agents to implement cycle-detection protocol, by sending message history, while propagating marginals. In such a setting, several decimation election may arise concurrently in the graph. Second, we would like to generalize DeciMaxSum framework to consider Max- Sum_AD_VP as a particular case of decimation: iterated decimation. Finally, we plan to applied 12 Cerquides et al. DeciMaxSum on real world applications, with strong loopy nature, like the coordination of smart objects in IoT [17] or decentralized energy markets in the smart grid [2]. References 1. Cerquides, J., Farinelli, A., Meseguer, P., Ramchurn, S.D.: A tutorial on optimization for multi-agent systems. The Computer Journal 57(6), 799–824 (2014), http://dx.doi.org/1.193/comjnl/bxt146 2. Cerquides, J., Picard, G., Rodríguez-Aguilar, J.: Designing a marketplace for the trading and distribution of energy in the smart grid. In: 14th International Conference on Autonomous Agents and Multiagent Systems (AAMAS). pp. 1285–1293. International Foundation for Autonomous Agents and Multiagent Systems (2015), http://www.aamas-conference. org/Proceedings/aamas215/forms/contents.htm#I4 3. Farinelli, A., Rogers, A., Petcu, A., Jennings, N.R.: Decentralised coordination of low-power embedded devices using the max-sum algorithm. In: International Conference on Autonomous Agents and Multiagent Systems (AAMAS'08). pp. 639–646 (2008), http://dl.acm.org/citation.cfm?id=142298.142313 4. Kolmogorov, V.: Convergent tree-reweighted message passing for energy minimization. IEEE Transactions on Pattern Analysis and Machine Intelligence 28(10), 1568–1583 (Oct 2006) 5. Krzakala, F., Montanari, A., Ricci-Tersenghi, F., Semerjian, G., Zdeborova, L.: Gibbs states and the set of solutions of random constraint satisfaction problems. Proceedings of the National Academy of Science 104, 10318–10323 (Jun 2007) 6. Lynch, N.: Disributed Algorithms. Morgan Kaufmann (1996) 7. Mackay, D.J.C.: Information Theory, Inference and Learning Algorithms. Cambridge University Press, first edition edn. (Jun 2003) 8. Maheswaran, R., Pearce, J., Tambe, M.: Distributed algorithms for dcop: A graphical-game-based approach. In: Proceedings of the 17th International Conference on Parallel and Distributed Computing Systems (PDCS), San Francisco, CA. pp. 432– 439 (2004) 9. Modi, P.J., Shen, W., Tambe, M., Yokoo, M.: ADOPT: Asynchronous Distributed Constraint Optimization with Quality Guarantees. Artificial Intelligence 161(2), 149–180 (2005) 10. Modi, P., Shen, W., Tambe, M., Yokoo, M.: ADOPT: Asynchronous distributed constraint optimization with quality guarantees. Artificial Intelligence Journal (2005) 11. Montanari, A., Ricci-Tersenghi, F., Semerjian, G.: Solving constraint satisfaction problems through belief propagation- guided decimation. CoRR abs/0709.1667 (2007), http://arxiv.org/abs/79.1667 12. Mooij, J.M.: libDAI: A free and open source C++ library for discrete approximate inference in graphical models. Journal of Machine Learning Research 11, 2169–2173 (Aug 2010), http://www.jmlr.org/papers/volume11/mooij1a/ mooij1a.pdf 13. Mézard, M., Montanari, A.: Information, Physics, and Computation. Oxford University Press (2009) 14. Navlakha, S., Barth, A.L., Bar-Joseph, Z.: Decreasing-rate pruning optimizes the construction of efficient and robust distributed networks. PLOS Computational Biology 11(7), 1–23 (07 2015), http://dx.doi.org/1.1371%2Fjournal. pcbi.14347 15. Petcu, A., Faltings, B.: A scalable method for multiagent constraint optimization. In: International Joint Conference on Artificial Intelligence (IJCAI'05). pp. 266–271 (2005) 16. Rogers, A., Farinelli, A., Stranders, R., Jennings, N.: Bounded approximate decentralised coordination via the max-sum algorithm. Artificial Intelligence 175(2), 730 – 759 (2011), http://www.sciencedirect.com/science/article/ pii/S43721183 17. Rust, P., Picard, G., Ramparany, F.: Using message-passing DCOP algorithms to solve energy-efficient smart environment configuration problems. In: International Joint Conference on Artificial Intelligence (IJCAI). AAAI Press (2016) 18. Vinyals, M., Pujol, M., Rodríguez-Aguilar, J., Cerquides, J.: Divide and coordinate: solving dcops by agreement. In: International Joint Conference on Autonomous Agents and Multi-Agent Systems (AAMAS'10). pp. 149–156. IFAAMAS, Canada (2010) 19. Vinyals, M., Rodriguez-Aguilar, J., Cerquides, J.: Constructing a unifying theory of dynamic programming dcop algorithms via the generalized distributive law. Autonomous Agents and Multi-Agent Systems 22(3), 439–464 (2010), http://dx. doi.org/1.17/s1458-1-9132-7 20. Zivan, R., Peled, H.: Max/min-sum distributed constraint optimization through value propagation on an alternating DAG. In: van der Hoek, W., Padgham, L., Conitzer, V., Winikoff, M. (eds.) International Conference on Autonomous Agents and Multiagent Systems, AAMAS 2012, Valencia, Spain, June 4-8, 2012 (3 Volumes). pp. 265–272. IFAAMAS (2012), http://dl.acm.org/citation.cfm?id=2343614

Subsets and Splits

No saved queries yet

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